Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
listeners. This is the eighty thousand hours podcast
0:02
where we haven't usually end up conversations about the world's
0:04
most pressing problems, what I can do to solve them.
0:06
And whether if I were at least as handsome as Ryan
0:08
goes I'd be exactly as handsome as Ryan Gosling.
0:10
I'm Rob Woodland, head of research at
0:12
eighty thousand hours. Many people have asked
0:14
me to interview Allen Hayek over the years and this year
0:16
I finally got to do it when I went home for effective
0:19
autism global Australia back in July.
0:21
Alan is one of the most prominent philosophers
0:23
focused on resolving puzzles in probability and expected
0:25
value issues that, it turns out crop up
0:27
shockingly often when you try doing research
0:29
to figure out how to do the most good. We
0:31
did the first hour on at the conference, then another
0:34
few hours of recording back in Allan's office.
0:36
The first hour is very accessible, but I won't lie.
0:38
The later part of the conversation is among the
0:40
more advanced content that we've had on the show, so far.
0:43
To help you out, I've recorded a number of cut
0:45
ins to quickly define terms like Epicycle's
0:47
or Dutch books, which are familiar to
0:49
philosophers but not so familiar to the rest of
0:51
us. Katie Moore has also gone
0:53
through the transcript and added a lot of links that will
0:55
take you to articles that elaborate on concepts
0:58
and arguments as they come up, if you don't mind reading
1:00
interviews rather than listening to them. People
1:02
familiar with these topics should find this interview
1:04
super informative, while I hope those who aren't so
1:06
familiar will learn a lot even if they don't follow
1:08
every single point that's being made. Some of
1:10
the things we cover are simple tricks for doing philosophy
1:13
well, why fruquintism is misguided, what
1:15
probability fundamentally is, problems
1:17
with using expected value and whether we should use it
1:19
anyway, a fundamental problem with Pascal's
1:22
wager and why the dominant theory of kind of factual's
1:24
philosophy is unworkable. One
1:26
quick notice is that eighty thousand hours is
1:28
currently hiring a recruiter to help
1:30
grow our team with applications for that role closing
1:32
soon on November the second. I'll
1:35
say little more about that in the outro, but if
1:37
you'd like to know more about that recruiter role, you
1:39
can get a full job description on our job board
1:41
at eighty thousand hours dot org slash jobs.
1:43
Alright. With that further ado, I bring you. Alan Hayek.
1:48
Today, I'm speaking with Alan Hayek.
1:50
Alan is a professor of philosophy at the Australian
1:52
National University. Years ago, he
1:54
did his PhD at Princeton where he won
1:56
a fellowship for displaying in the judgment of the university
1:59
the highest scholarly excellence of all
2:01
graduate students across all of their disciplines.
2:03
These days, he has a broad ranging interest across
2:05
epistemology, philosophy of language and
2:07
philosophical methodology. His work
2:10
is of great interest to me and the effective
2:12
autism because he is one of the world's
2:14
top experts on the philosophy of probability,
2:16
bayesianism, decision theory, expected value
2:18
and counterfactuals. Some of his
2:21
more memorably titled papers include
2:23
fifteen arguments against frequentism, waging
2:25
war on Pascal's wager, vexing
2:27
expectations, most counterfactors are
2:29
false, and a follow-up paper titled, most manufacturers
2:31
are still false. He's currently working on
2:34
a book on Canifactual's title would work.
2:36
This is our first ever live recording of the show.
2:38
So for once, I should also welcome all of you
2:40
in the audience as well. And before we button
2:43
down for the recording, let's have a random applause
2:45
for Alan.
2:50
Thank you. Great to be here, Rob. Thanks so much
2:52
for having me on the show. Okay.
2:55
I hope we're gonna get to talk about whether
2:57
Besianism is the final big evolution
2:59
and probability. and why you think objective
3:01
consequentialism is even wronger than many people
3:03
can even imagine. But first, you started
3:05
out studying maths and which
3:07
are kinda practical fields in which you could have
3:09
been gainfully employed doing something of obvious
3:11
now you. Yes. What made you switch into philosophy
3:13
after you finished your undergrad degree? Yeah. I
3:16
was studying, as you say, math and statistics,
3:18
and I went to lots of courses on
3:20
probability. And my professors
3:23
would write equations on the board, p of
3:25
this equals p of that.
3:27
And I was wondering, what is this p? What does
3:30
probability mean, and I asked one of them,
3:32
what does probability mean? And he
3:34
looked at me like I needed medication. And
3:37
he said, well, that's a philosophical question.
3:40
And I began a master's in
3:42
statistics, and I got as far as photocopying
3:45
the article that my supervisor wanted
3:47
me to read and I sat down,
3:50
started reading it and I went, and
3:53
I don't know if you've seen pulp fiction, but there's
3:55
a moment where Jules the hitman
3:58
has a moment of clarity. Right. I
4:00
had a moment of clarity. I realized I didn't
4:02
want to go on with statistics -- Yeah. -- but
4:04
what to do instead traveled the world
4:06
hoping to find myself on the road.
4:09
And I did it happened in Western Ontario,
4:12
and a friend of mine who'd
4:14
had his moment of clarity was studying
4:16
philosophy there. He showed me
4:18
the courses that he was studying he
4:20
was looking at. And I
4:22
thought, wow. how cool is that? And
4:24
you could hear that Penny dropped several provinces
4:27
away. That was my second moment of
4:29
clarity. Yeah. That's what I wanna do,
4:31
philosophy. What what was it that bug I mean,
4:33
obviously, most statisticians kinda don't care
4:35
about these, deeper questions about what probability
4:37
is given that they can just, like, do operations on it.
4:39
And what what like, why couldn't you let it go?
4:42
Yeah. In a way, I was
4:44
less interested in the the sort of practical
4:46
payoffs. I I was just wondering,
4:49
where does probability fit into my of
4:51
conceptual scheme, you know -- Yeah. --
4:53
how does it relate to other things? What
4:55
does it really mean? When I say something
4:57
like the probability of rain
4:59
tomorrow is point three. What
5:01
have I said? I understand the stuff about rain,
5:04
but the probability stuff, I didn't understand.
5:06
And here I am, I'm still asking that question. What
5:08
does p mean? Yeah. Okay. Well, we'll
5:10
get to what p means later on.
5:12
But, like, what what are your main focuses at the
5:14
moment? I guess, so you're working on kind of actuals
5:16
primarily. Right? Yeah. That's right. And the
5:18
book would work. Thanks for the advertising.
5:21
And as always, I'm working on
5:23
probability. I still ask what
5:25
is pee. And that's
5:27
related to decision theory, something that I
5:29
think quite a lot about. That got me into
5:31
things like Pascale's wager, still
5:34
thinking about that. We'll probably
5:36
talk about my heuristics and
5:38
some of that comes up as well.
5:40
Yeah. Makes sense. So we're gonna
5:42
get into the details of probability and can affect
5:44
us later on. But first, I wanted to dive
5:46
into this, yeah, other passion of yours, which is
5:48
philosophical methodology. Mhmm.
5:50
Yeah. How did you first get into that topic?
5:52
Yeah. It began when I was a graduate
5:54
student at Princeton. I was
5:56
surrounded by these really good philosophers.
5:59
I wanted one day to be a good philosopher
6:01
or as good as I could be. And I
6:03
noticed there were these recurring patterns
6:05
of thought, sometimes in their work,
6:07
sometimes in conversations, sometimes
6:10
in Q and A. For example,
6:12
I would hear a really good question at
6:14
Q and A. Three weeks later, the same
6:16
person would ask a similar question.
6:18
I thought it worked last time. It
6:20
worked again. There's a recurring
6:22
pattern here. Yeah. I'll internalize
6:24
this. So I started to make a list of these
6:26
recurring patterns of thought, philosophical techniques
6:29
that seem to be fertile. And now
6:31
my list is hundreds long. Why
6:33
isn't this just really an obvious thing. You would think you would
6:35
kind of start of philosophy PhD just before,
6:37
like, here's all the tools in the toolkit.
6:39
Just like go and do this. Yeah. Yeah. It's
6:41
funny. I think that these are strangely
6:43
underexplored. We
6:46
teach our students logic, for example,
6:48
that's certainly one tool in the toolbox. Yeah.
6:50
I think we have all these other techniques and
6:52
we don't nearly discuss them
6:54
enough, think about them. And then I went on
6:56
to teaching at Celtic And
6:59
I had these very smart students for ten
7:01
weeks. How do I convey to them? How
7:03
philosophy is done? Of course, I had
7:05
them read the classics. They
7:07
cart and Hume and so on. But
7:09
along the way, I'd occasionally drop in these
7:11
philosophical heuristics, these
7:13
techniques,
7:14
the partly
7:15
just to show them how philosophy is done, but
7:17
also just it helps You
7:19
do philosophy. And, yeah,
7:21
philosophy is hard. Yeah. I think we we
7:23
could use all the help that we can get,
7:26
especially when you've just written
7:28
some philosophical paper,
7:30
some view of your own. It's curiously
7:32
hard to be your own critic, and
7:34
other people are lining up to
7:36
point out the problems with your view. Mhmm.
7:38
But these heuristics, I think, help
7:40
guide your mind to finding the problems before
7:43
others -- Others -- -- gleefully. -- before you point
7:45
them out. Yeah. That's right. Yeah. What are
7:47
your papers on on this topic? Well,
7:49
you sound extremely defensive in
7:51
a sense that you're you're like, worry
7:53
that other philosophers are gonna judge you daring
7:55
to write down a bunch of heuristics. Yeah. Yeah. Do
7:57
you wanna explain what what of what's going on with
7:59
that? Yes. It's as if I'm
8:01
like the magician who gives away the
8:03
the tricks of the trade.
8:05
And I find this a very strange
8:08
attitude. I mean, think of some other
8:10
area where there are heuristics like
8:12
chess. Mhmm. there's
8:14
no problem with the chess book actually giving
8:16
you some advice, Castle early
8:19
and often avoid isolated pawns
8:21
or in that there are various
8:23
fertile heuristics. But I
8:25
don't know, somehow in philosophy, some
8:27
people view it with a bit of suspicion
8:29
that this isn't really depth
8:32
this isn't really getting to the
8:34
profundity of philosophy. Mhmm.
8:36
And I'm not saying you should just
8:38
follow these heuristics mindlessly
8:41
just like when you're playing chess, you shouldn't
8:43
just mindlessly play
8:45
the things. Yeah. But I
8:48
think they help I think they get you
8:50
closer to your targets. And
8:53
I think they actually help
8:55
creativity too, that somehow, as
8:57
I say, philosophy is hard. So
8:59
These are just techniques to breaking down
9:01
a hard problem into easier sub
9:03
problems, and then you can make progress
9:06
on those. Okay. So, yeah, let's hear one
9:08
of these heuristics. What's what's one that really stands out
9:10
as as a special use one? Yeah. I I
9:12
like the one I call check
9:14
extreme cases. Okay? You're in
9:16
some domain. Extreme cases
9:18
are things like the first case or
9:20
the last case or the biggest or
9:22
the smallest or the
9:24
best or the worst or the smelliest
9:26
to or what have you. Yeah. Okay?
9:29
Now you've got this huge
9:31
search space. Someone gives a big
9:33
philosophical thesis. Suppose
9:35
you want to test it. You
9:37
know, stress test it. Are they counter
9:39
examples? Hard
9:40
problem. Somewhere
9:41
in the search space find
9:44
trouble. Find counter examples. Easier
9:46
subproblem. go to
9:48
the corner cases, go to the
9:50
extreme cases. Mhmm. And often the trouble
9:52
lurks there, if it looks
9:54
anywhere, and it's a smaller
9:57
search space. So that's the technique. I could
9:59
give you some examples. That looks like Yeah. Good
10:01
example to you. All right. Brandi
10:03
O's philosophical thesis every
10:06
event has a cause. Okay? Yeah.
10:08
Okay. And at first, you
10:10
might think, gee, I don't know. Is that
10:12
true or false? Is it it's kind of hard to
10:14
to tell? Alright.
10:15
Hard problem come up with
10:17
a counter example to every event has
10:19
a cause. Easier subproblem
10:22
consider extreme cases of
10:24
events. For example, the first
10:27
event. Call it the big bang.
10:29
Yeah. The big bang didn't have
10:31
a cause counter example, or
10:33
philosophers sometimes say that you should only
10:35
believe in entities that have
10:37
causal efficacy. they have
10:39
some poof. Yeah.
10:41
Like, that's maybe a reason to be suspicious
10:43
of numbers. Maybe numbers don't exist,
10:45
because they don't cause
10:47
anything. And then Lewis has
10:49
us imagine, well, what about the
10:52
entity which is the whole of history
10:54
There's causation. We never. But that
10:56
doesn't cause anything. Right. So according to this
10:58
principle, you shouldn't believe in
11:00
the whole of history. Okay. So
11:02
there the heuristics doing, I suppose,
11:04
negative work. It's -- Yeah. -- destructive
11:07
shooting down some position. But
11:09
I think it could also be constructive.
11:12
Right. Maybe just worth explaining a little bit
11:14
of, I guess, one of your theories of what philosophy
11:16
is. Yeah. Yeah. You want it? Oh, well,
11:18
I think you're thinking of that a lot
11:20
of philosophies, the demolition of common
11:22
sense, followed by damage control.
11:25
Yeah. I love that I love that question. Yeah.
11:27
And philosophy often comes up with
11:29
some radical claim like,
11:31
you know, we don't know anything.
11:33
Mhmm. And but then
11:35
we try to soften the blow a bit and we we
11:37
find some way. Maybe we know a little
11:39
bit. We know a little bit or we have to
11:41
understand knowledge the right way.
11:44
Anyway, so far the heuristics
11:46
have been this extreme cases heuristic
11:48
was somewhat negative. It
11:50
was pointing out a counter example to
11:52
some big thesis. I think
11:54
it could also be constructive.
11:56
I guess maybe long
11:58
termism could be thought of in this
12:00
way. Maybe
12:02
the thing that comes naturally really to us is to
12:04
focus on the short term consequences
12:06
of what we do and we think that's what
12:08
matters. Then you push that out a
12:10
bit And then that extreme case
12:12
would be, well, gosh, our actions
12:14
have consequences until the
12:16
end of time, you know, for the rest of history. So
12:19
maybe we should be more focused on
12:21
that. Yeah. And that's now the beginning of a
12:23
more positive movement.
12:26
Yeah. I guess, so a
12:28
simple question there might be like, for how long
12:30
should we consider the consequences of, like, what should be the
12:32
scope of our moral consideration? And here, say, well,
12:34
let's consider the extreme possibility. We should consider
12:36
all space and all time forever. That's
12:38
right. And then a related heuristic, so
12:40
I started with check extreme
12:43
cases. And sometimes you might just check near
12:45
extreme cases. So you back off a bit
12:47
and they're a little bit more plausible.
12:49
So maybe we don't need to look until the
12:51
end of time, but still look far
12:54
ahead and that
12:56
is still at some odds with
12:58
initial common sense. Yes, I
13:00
guess, people might often come back and say, sure and
13:02
they stream situation doesn't work. Like, lots of things
13:04
don't work in extremes. Like, it's it's more sensible
13:06
to focus on the on the middle cases. Yeah.
13:08
And so this isn't, like, actually, such a powerful
13:11
objection. What do you what do you think of that? I think for
13:13
that very reason that this is a fertile
13:15
heuristic because we spend our lives
13:17
mostly living among the normal cases.
13:19
So extreme cases don't come
13:21
so naturally do is even though they may
13:23
well be troubled for
13:25
some philosophical position. In fact, maybe
13:27
especially because they're extreme.
13:29
They're more troubled than
13:31
the middle cases. Yeah. I guess it I guess it
13:33
depends on whether the claim is like a more
13:35
pragmatic one about like how you ought to do things every day
13:37
or whether you're trying to claim that I've discovered
13:39
some like transcended fundamental truth, and you'll
13:41
be like, well, it doesn't work in this, like, one case.
13:43
That's it. You you said it. You you were claiming that this
13:45
was like something that should cover everything, and now
13:47
it doesn't. and philosophers often do that. They have
13:49
these grandiose claims. Every
13:51
event has a cause or what have you. Yeah.
13:53
And this is a good way of stress testing
13:55
such claims. Okay. Yeah. What's what's
13:57
another heuristic? Yeah. I
13:59
like to focus on the word
14:03
and I say see the word
14:05
in neon lights because
14:07
it typically comes with a presupposition.
14:09
VX typically
14:11
presupposes there's exactly one x.
14:14
There are two ways that could go
14:16
wrong. There could be multiple x's
14:18
or there could be none at
14:20
all. Right? So an example
14:22
of that. do the right thing.
14:25
Just rolls
14:27
off the tongue. Yeah. Rolls off the tongue.
14:30
Alright. The right thing.
14:33
Now that like there's exactly one right
14:36
thing to do. Well,
14:38
two ways we could challenge that. There could
14:40
be multiple right things.
14:42
Yeah. And maybe it's okay if you do any
14:44
one of them, but still that we're challenging
14:46
the pre supposition. There's exactly one.
14:48
This comes up by the way. I don't know if we'll get to
14:50
talk about Pascal's wager at
14:52
some point, but it turns out that there are
14:54
many ways to follow Pascal's
14:57
advice. Going in the other
14:59
direction, maybe there's no
15:01
right thing. Think of a moral
15:03
dilemma like Sophie's choice
15:05
you know, there's no right thing to do
15:07
or Sartre's case of the
15:09
student who's torn between
15:11
fighting in the war or
15:13
staying home with his mother. Yeah.
15:15
What's the right thing to do in this moral
15:17
dilemma? It's not clear that there is one.
15:19
Or at least if you if you start saying that
15:21
there is one thing, then you're making a claim, maybe,
15:23
without even realizing that you're making a claim, you feel
15:25
like you've slipped in an assumption -- Yeah. -- by
15:27
using the That that's right. Yeah. Any
15:29
other examples of that one? Of a of
15:31
a v or, like, people just assuming that there's one
15:34
answer. Yeah. Yes. I
15:36
I think
15:37
Yes, maximizing expected utility. We'll
15:39
probably talk about expected
15:41
utility later on and there is
15:43
– here we're being told do
15:46
the thing that maximizes expected
15:48
utility. If you hear it that way, then there are
15:50
problems on each side. There could be ties.
15:52
for expected utility. Yeah. Or there could be
15:55
nothing that maximizes it. Things just
15:57
get better and better without
15:59
end. It's slightly I guess it's not the same,
16:01
but it reminds me of this saying that in
16:03
philosophy, you can have, like, two two problems.
16:05
One is have, like, no answer to a question. Yeah.
16:07
And the other is to have like so many explanations.
16:09
You can't tell which one's the right one. That's right.
16:11
And this corresponds to the two
16:13
ways. And again, it's breaking down, so
16:15
to speak, harder problem into easier sub
16:17
problems. So you look on each
16:19
side, are there too many of these
16:21
things or are there not enough?
16:23
Yeah. Okay. Yeah. What's what's another
16:25
one? Yep. philosophers love to
16:27
talk about what's possible. They
16:29
love to say x is possible.
16:31
Yeah. I guess a classic case of that like,
16:33
the the zombie thing If I'm
16:35
a zombie yeah. That's right. Our zombie's
16:38
possible. Okay. But that is like a beings that would act
16:40
like people but have no conscious experience. That's
16:42
exactly right. Yeah. And
16:44
philosophers love to distinguish different
16:46
senses of possibility, logical,
16:49
metaphysical, no make what's
16:51
consistent with the laws of nature,
16:53
dogmatic, what's consistent with
16:55
one's beliefs, epistemic and so
16:57
on dontic. And
16:59
there are various techniques for
17:02
generating possibilities or
17:04
for for arguing that something
17:06
called an x is possible.
17:08
Yeah. And when you think about
17:10
it, there are two moving parts to that.
17:12
There's x and is
17:14
possible. possible
17:15
So focus on is possible first.
17:18
One kind of technique says,
17:21
look at some other property of
17:23
x. and it follows from x having this other property
17:25
that x is possible. Let
17:27
let's just do a few of those. Yeah. x
17:31
is actual Okay. So
17:33
suppose x is actual. So for a lot of
17:35
the modalities I just talked about, the
17:37
possibilities, that's a good way
17:39
to establish that x
17:41
is possible Hi, listeners. Rob here
17:43
with a definition to help you out.
17:45
Epistemic refers to
17:47
epistemology, which is the sub
17:49
field of philosophy that's studies,
17:51
the nature of knowledge, how do we know things, what
17:53
things do we know? Then we talked
17:55
about actual. Actual here refers
17:57
to things actually existing. So
17:59
actual is kind of a a
17:59
definite philosophy is used to mean something that's real in
18:02
this world. Okay. Back
18:03
to the show. Well, you gave a
18:04
good one. x is conceivable. That's
18:07
David Dave Charmers appeals to that one in
18:09
the zombie argument -- Yeah. -- x is
18:12
conceivable so it's possible so that it's
18:14
possible to imagine it. And therefore -- Yeah. --
18:16
you, like, start thing from that. Yeah. right.
18:18
Yeah. Yeah. Yeah. What so what what
18:20
are some examples of cases where people where philosophers make
18:22
this argument, I guess, we've got the zombie one. But
18:24
-- Yes. -- it seems like
18:26
like, it's maybe a slightly dubious style
18:29
of argument where you're saying, well, something can be imagined.
18:31
Therefore, like, I'm gonna conclude things from that.
18:33
Yeah. Yeah. That's right. And of course,
18:35
philosophers love thought experiments. And
18:37
I guess soon we'll be talking about some pretty
18:40
ricochet thought experiments. And even
18:42
if these things aren't actual,
18:44
like the Saint Petersburg game or something, still
18:46
we might say, look, we can conceive
18:48
of this. In that sense, it's possible and
18:50
we should take it seriously and
18:53
test our theories against it. And
18:55
that was, by the way, one just one way
18:57
to show the next is possible. Okay. Yeah. And
18:59
remember, is the other moving
19:01
part x as How about we look
19:03
at some other object, let's
19:05
call it, y, some other
19:08
thing, And why is
19:10
possible? We could all agree on
19:12
that. And X is appropriately related
19:15
to why and we
19:17
conclude that X must be possible too. So you're
19:19
related in what in what way.
19:21
So here are some ways. If y
19:24
is possible and y
19:26
entails x. Okay.
19:28
Then it seems pretty safe that x is
19:30
possible too. Yeah. If they're
19:32
compatible, then it seems that x is
19:34
possible too. We just use the term entail
19:36
or entailment. Intailment in
19:38
philosophy is a kind of a fancy
19:40
way of saying that something necessarily
19:42
implies something else So
19:44
if a and tails b, then
19:46
if a is true, then b is also
19:48
true. Okay. Back
19:49
to the show. Here's one
19:50
of my favorite ones. almost
19:54
x is possible. So let
19:56
y be almost x.
19:58
It's very similar to x in
20:00
some appropriate sense. Yeah.
20:02
And then you say, well, the small
20:04
difference between y,
20:06
which is almost x and x
20:08
won't make a difference. to possibility. An
20:10
example of that, according
20:12
to behaviorism, it's not
20:14
possible to have a
20:16
mental life and no
20:18
behavioral
20:19
manifestation of it whatsoever. Reply?
20:22
Yes, it is
20:23
possible because it's possible to
20:26
have almost no
20:28
behavioral manifestation of your mental life.
20:30
Think of Stephen Hawking, sadly
20:32
no longer with us. And
20:35
towards the end, I guess he had just very
20:38
minimal behavior movement of his
20:40
finger, not much more. Obviously,
20:42
he had a rich mental
20:44
life. And now imagine just just
20:46
that finger movement stopping, so he loses that
20:48
last bit of behavior. Yeah. It's
20:50
not like the lights suddenly go out
20:52
for Steven Hawkins. So it is possible.
20:54
Yeah. This is reminding me of another line of argument that
20:56
some people might have heard, which is, you know, some
20:59
people say, oh, you couldn't create a mechanical
21:01
mind or you couldn't out of machines. And you
21:03
say, well, let's take a human brain and let's just
21:05
replace one neuron with a mechanism that that
21:07
does what just what that neuron does. It'll be like, are
21:09
you not a person now? Are you not conscious? And I'll
21:11
be like, No. No. No. No. Really? So they're like,
21:13
what if we place another one? And then the other
21:15
person is either just forced to say, like, you're
21:17
becoming gradually, gradually less conscious. order
21:19
can see or or that there's some sharp cut
21:21
off, which also seems plausible. Excellent.
21:23
So this is another of the heuristics
21:25
of this kind. Let's call it extrapolation.
21:27
Mhmm. So you have a sequence of things and
21:29
you think that each of them is possible.
21:31
Well, now let's just go to the
21:34
next step. That should be possible too. Or
21:36
interpolation. Mhmm. Start with two
21:38
things that you know are possible, maybe
21:40
because they're actual. And now
21:42
interpolate on some whatever gradeable
21:45
scale, Hume's missing shade
21:47
of blue is like this. Okay?
21:49
Take two shades of
21:51
blue that are actual, therefore
21:54
possible. And now imagine a
21:56
missing shade of blue. It's not
21:58
actual, but it's
22:00
somehow between those two on some natural
22:02
scale. Yeah. Well, that seems to be possible
22:04
too. Okay. Yeah. I mean,
22:06
couldn't someone respond that it's not possible,
22:08
but actually too close. There's no shade in between.
22:10
They could say that. And you got to be careful
22:12
with your scales too, and
22:15
sometimes extrapolating or
22:17
interpolating will give you wrong
22:19
answers. Like think
22:21
of the sequence one over n,
22:23
you know, one half
22:25
third quarter, etcetera. Is it
22:28
possible for that sequence
22:30
to reach zero Well,
22:32
no. It gets as close as you like. So
22:35
almost reaching zero is possible, but
22:37
it never actually hits
22:40
zero. That's not possible for that sequence.
22:42
Well, what's an example of the thing we kind of repeat
22:44
the thing with the the sequence again and again and again. I
22:46
guess that's an example we are kind of strapolating. Are
22:48
there examples where you take some principal and then you
22:50
just like keep operating it and then you get some absurd conclusion?
22:53
How about maybe the
22:55
lottery paradox is I guess, you should think
22:57
about -- Yeah. -- or even better maybe the Preface
23:00
Paradox. Oh, I don't know that one. Yeah. Let's do
23:02
the Preface Paradox. This
23:04
puts pressure on the idea that
23:07
Belief is closed under
23:10
conjunction. Okay? You've just
23:12
written
23:12
a long book
23:13
and then you write the following
23:16
preface. I
23:16
believe everything
23:17
I say in this book, but
23:20
I'm confident that there's a
23:22
mistake somewhere in the book.
23:24
Yeah. Okay. So
23:27
I I did my best
23:29
to get sentence number one
23:31
right. I believe sentence number one.
23:33
I believe sentence number two. I
23:36
believe the final sentence of the book,
23:39
but I don't believe the conjunction of
23:41
the book on the contrary. I know that
23:43
in long books like this there's
23:45
always a mistake somewhere. Okay. So
23:47
this puts pressure on the idea that belief, even
23:49
rational belief is closed under
23:52
conjunction. Yeah. A different approach that
23:54
sometimes I apply is, like, trying to subdivide things,
23:56
like, super closely. Yeah. So so
23:58
every so often people say to me, batch
24:01
not possible to put a probability on, like,
24:03
things that have never happened before. So it's, like, fine
24:05
to say that the probability of a coin
24:07
flipping heads is fifty percent.
24:09
But if you say something like the probability of a
24:11
nuclear war in twenty twenty seven is
24:13
one point five percent, then this is just a bunch
24:15
of silliness and like because this never happened before.
24:17
then, of course, you can say, well, if you define
24:19
all events like sufficiently closely, then none of them
24:21
have ever happened before. Like, all events are unprecedented
24:24
once you, like, specify exactly what they are. That
24:27
that's right. If you specify the events too
24:29
precisely, then they're all unique. If
24:31
you specify them too
24:33
loosely, then, of course,
24:35
anything can be described as
24:37
something happens. Yeah. Okay. We have many
24:39
instances of that. Alright? And
24:41
now you want to find the right level of
24:43
grain that describes this
24:46
event in an informative way and
24:48
such that susceptible, let's go to
24:50
some probability. Now that I think about it, I mean, that
24:52
that potentially creates trouble for the, like, coin
24:54
being fifty percent. because you say, well, like,
24:56
each coin flip is unique in its own way. I mean,
24:58
like, exactly like whether in any
25:01
specific case, like, it depends on how they flipped it. And
25:03
so if you define and and if you like specify that, then
25:05
it's a hundred percent and zero percent. Yeah.
25:07
Maybe we'll soon be talking about frequentism,
25:09
but this is related to a famous problem
25:12
for the frequentist account of probability, the
25:14
problem of the single case. Right.
25:16
At some level of specification, every event
25:18
only happens once And then in
25:20
some events, that is quite a natural thing to
25:23
do. Like, this year's
25:25
NFL grand final That's
25:27
Australian rules football for the non Australian.
25:30
Okay. That seems to be a unique
25:33
event. And there may
25:35
be an issue, but how we assign probabilities
25:37
to it. But it seems as
25:39
far as frequencies go, it's a
25:41
one off thing. It's either one or zero for
25:43
say calling would my team winning. Yeah.
25:46
Yeah. Alright. Yeah. Are there any other any other
25:48
heuristics that people should keep in mind if they wanna kinda out
25:50
with their friends in conversation before we
25:52
move on? Yeah. 0II
25:53
like to think about
25:56
putting
25:56
contrastive stress on different
25:59
words in a claim and that that
26:01
makes you realize there's a there's a
26:03
contrast class this as opposed to
26:05
that. Let's take for
26:07
example, smoking a pack
26:09
a day causes lung
26:11
cancer. That's that seems
26:13
reasonable when you just say it like
26:15
that. Well, smoking as
26:17
opposed to other things you could do
26:19
with a cigarette like stroking it. Okay?
26:22
Yes. It seems like the smoking is relevant.
26:25
Yeah.
26:25
Smoking a pack
26:26
a day. Smoking one pack
26:29
a day Now as opposed to
26:31
none, yes, that seems
26:33
to cause lung cancer.
26:35
But what
26:35
about smoking one pack
26:38
a day as
26:38
opposed to three packs a
26:41
day. Now it's not so clear that
26:43
it's the one pack a day
26:45
that causes the
26:47
lung cancer as opposed to
26:49
three. And this now makes you
26:51
think that explanation must
26:53
be a contrastive matter.
26:55
Yeah. Isn't that and yeah. This thing
26:57
is opposed to lip band counterfactuals or like
27:00
an alternative. They're saying one is opposed to zero,
27:02
but we don't we don't say that most of the time. And
27:04
so we could eventually end up having a kind of
27:06
factual that you haven't really, like, properly thought through?
27:08
Yeah. Yeah. And, yes, this
27:10
makes you be honest. What am I
27:12
contrasting this thing with
27:14
maybe that was implicit and not
27:16
obvious. And this this technique where
27:18
you, I mentally, italicize different
27:20
words in a claim, and
27:22
and say, okay. In each case, this as opposed
27:24
to what. And then I realized, oh, okay.
27:26
I was making a presupposition or
27:28
something like that. Yeah. Okay. So
27:31
I suppose, like, all of these heuristics can
27:33
potentially provide kinda stress cases
27:35
or, like, I guess, considerations against
27:37
conclusions that people are putting forward. Yeah. In
27:39
general, do you think that there's very often, like,
27:41
decisive arguments against the position? Or or is it more
27:43
often that all of these weaknesses kind of
27:45
accumulate? And then you end up with kind of view that,
27:47
like, this many like little pro there's too many problems
27:49
here for this to be quite the right theory.
27:51
Yeah. This thing, like, most most theories you can throw
27:53
something at them and be like, it doesn't feel quite right in this
27:55
case. Yeah. Lewis famously
27:57
said, philosophical positions are
27:59
rarely refuted. Get
28:01
here may have done it. That's the justified
28:03
true belief, analysis of knowledge, Girdle
28:05
might have done it, but it's it's about
28:07
it. It's rare. Hey, listeners. We just
28:09
mentioned getier problems, which We
28:11
don't really need to understand in order to follow
28:13
the rest of the conversation, but I thought I'd add
28:16
in a little bit about that just because it's
28:18
quite fun. since the beginning of philosophy, that
28:20
being this question, what is
28:22
knowledge? And one of the most
28:24
common answers or one of the most popular
28:26
answers into the twentieth century going back to the
28:28
ancient Greek was that knowledge
28:30
is justifying true belief. So you have
28:32
to believe something. It's gotta be true and you
28:34
have to have good reason to think that
28:36
it's true. In nineteen sixty three, Edmund Gettya
28:38
published this paper that basically
28:40
to almost everyone's satisfaction kind of
28:42
demolished that definition of knowledge with a
28:44
really strong counter example. And the
28:46
cataract example is this. Imagine that I
28:48
walk into my colleague's
28:50
office say, and I see
28:52
them sitting at their desk. and I
28:54
formed the belief that my colleague, Neil,
28:56
say, he's in his
28:58
office. It'd be very natural to say I have a
29:00
justified true belief there because I've seen
29:02
Neil sitting at his desk. But
29:04
imagine that unbeknownst to
29:06
me, Neil has had an
29:08
amazingly realistic wax sculpture of
29:10
himself that he has placed in
29:12
his desk chair in his place, and I wasn't
29:14
looking very carefully. So I hadn't
29:16
realized that it wasn't actually Neil sitting at
29:18
the desk. But
29:19
also unbeknownst
29:20
to me, Neil is in his office. He's
29:22
hiding under the desk as part of a prank
29:24
involving this wax sculpture that he's
29:26
produced of himself.
29:27
Now, I believe that Neil was in his
29:30
room. He is
29:30
in his room. He's hiding under the desk. And I
29:32
had good reason to think that Neil was in his
29:34
office because I saw him there.
29:36
But
29:36
can I really be said to have known
29:39
that? Given that, what I saw
29:41
was the wax sculpture, so it's
29:43
only by pure coincidence that my belief
29:45
is true. There's not actually a real
29:47
connection between the justification that
29:49
I had for thinking that Neil was in the office and
29:51
the fact that he is. It's just luck that
29:53
he also happens to be hiding under his desk.
29:55
this is the case where it seems like one can have
29:57
a justified true belief, but it doesn't
29:59
really
29:59
feel like knowledge because the
30:02
truth of the claim is not connected to the
30:04
justification that I have in my mind.
30:06
That's a
30:06
getier case or a getier problem. It shows
30:08
up all over the place where one can think
30:10
the right thing, but for the wrong reason, Another
30:12
amazing thing is that Edmund Gettya published this paper,
30:15
his knowledge justified true belief in
30:17
nineteen sixty three, but
30:19
that's basically the only thing that he's known for doing. I I don't know
30:21
that he made any other notable contributions
30:23
to philosophy other than one of the most
30:25
famous arguments made in philosophy across the
30:27
entire twentieth century. So I
30:29
guess he's kind of the one hit one, the the the macarena of his fields,
30:32
so to speak. Alright. That's get your
30:34
problems. Let's go back to the
30:36
interview. I think there are some
30:38
killer problems for various positions
30:40
we might talk about some later.
30:42
Yeah. But yes, often it's more
30:44
a matter of accumulation. There are some
30:47
smaller problems, but they start adding up.
30:49
And overall, the
30:51
position doesn't
30:51
seem tenable. It doesn't seem to
30:54
best capture the data points. Yeah.
30:56
The people sometimes give
30:56
up on them because it feels like, okay. You you could
30:58
like patch the first three things kind of, but then just like
31:00
more and more cracks start appearing and you're, like, trying
31:02
to patch it up everyone. You're, like, it starts to feel
31:05
like more complicated. Just patch it. Yes. Epi
31:07
cycles. Right? Yeah. It feels like more complicated to patch the
31:09
theory than just to come up with a new one.
31:11
That's right. Some might say that that
31:13
happened to the justified true
31:15
belief account of knowledge that
31:17
that can't be quite right. It seems because of
31:19
get here and so on. And then we add an
31:21
Epicycle and another one and another one.
31:23
And if it goes too far, it starts to
31:25
feel like a degenerating research
31:28
program it's like, yeah, the explanation so long that it's lost.
31:30
It's original intuitive appeal. It's no longer
31:32
simple either. That's right. And even if
31:34
it turned out, alright, I
31:36
can't come up with any counter examples to
31:38
this tremendously long complicated
31:41
analysis, but do I really feel
31:43
it? Illuminate. Suppose that really
31:45
is knowledge is this
31:47
thing with many Epi cycles. Yeah.
31:49
Do I feel like I understand it better? Yeah.
31:51
Yeah. Not really. hey, listen.
31:53
Rob here with with another quick aside.
31:55
This one's
31:56
a fun one. We just use the term EpiCycles,
31:58
which
31:59
philosophy around quite a bit, but I think I
32:02
hadn't heard until I started talking to people
32:04
who'd done philosophy PhDs. That year
32:05
of EpiCycles, is basically you're adding
32:08
lots more complexity to a
32:10
theory in order to salvage it
32:12
from cases in which it doesn't match in reality or it
32:14
doesn't match your intuition. But really
32:16
what you want to do is throw out the theory
32:18
altogether and and then start again because it's the
32:20
wrong approach. It goes back to
32:22
how the ancients used to
32:24
predict the the the motions of planets
32:26
when they thought that everything was
32:28
going around the Earth one way or another.
32:31
Obviously, the planets weren't growing around the
32:33
Earth fundamentally. So how
32:35
did they manage to explain the apparent
32:37
motions of of the planets and then them coming
32:39
closer and then and then going further
32:41
away. Basically, they modeled the planets
32:43
traveling around the earth, but they were
32:45
also traveling in little circles
32:47
around the circle that they were traveling
32:49
around the Earth on. They were traveling at the same time on
32:51
two different circles. Then this allowed them
32:53
to match up the predictions of that model with their
32:56
observations of where the planets were. So the main
32:58
big circle that they were traveling on around the
33:00
Earth was called the deferrant or deferrant,
33:02
I'm not sure how it's pronounced. And
33:04
the little circles that they were traveling on around
33:06
that circle were called EpiCycles. I
33:08
think it I think it comes from Greek meaning,
33:10
yeah, circle moving on another circle.
33:13
anyway, so you can see how this has come to mean adding
33:15
complexity to a model in order to
33:17
salvage it when really it's it's the wrong model
33:20
to start. What they should done is that the
33:22
planets were moving in ellipses around the
33:24
sun rather than adding more and more circles
33:26
to try to match things up. Interestingly
33:28
though, by adding in enough Epicycle's like
33:30
this, they were able to make their predictions of
33:32
the planet's motions match almost exactly
33:34
that their observations because they
33:36
just had enough degrees of freedom in the model to allow
33:38
them to place the planets in in any particular place
33:41
at any particular point in
33:43
time. what can happen when you're you add a
33:45
lot of complexity to a model even if it's mistaken.
33:47
Okay. That's Epicycle's back to the show.
33:49
Okay.
33:49
Well, what what are some common moves that
33:52
philosophers make trying to debunk at a other
33:54
philosophers of ideas that you think are kind of overrated.
33:56
It's like not as strong an objection as as people
33:58
make out. Well, a whole
34:00
movement in philosophy see that I now distance
34:02
myself from. I think this speaks to your
34:04
question. Flossers love to
34:06
appeal to other worlds, and
34:08
in particular, similar
34:10
worlds. This may come up when we
34:12
talk about cataractials. The
34:14
standard analysis of cataractials
34:16
appeals to
34:18
most similar possible worlds, but not just
34:20
counterfactuals. Knowledge, we often
34:22
talk about so called safety and
34:26
sensitivity, And that's a a matter
34:28
of what's going on in nearby worlds where you believe something or where the
34:30
thing you believe is is true or not true.
34:34
And I I used to
34:36
just talk this way too. I used to love all this
34:38
talk of similar worlds
34:40
and and other worlds in general. And
34:42
now I've now i've jumped off
34:44
the bus and I think one should not be
34:46
casting everything in terms of
34:48
other worlds and similarity relations.
34:52
I prefer to do it in terms
34:54
of probability and stay in the actual world, probabilities anchored
34:56
here and it will give
35:00
my preferred accounts of things. It doesn't quite speak to your question, but I think
35:02
it's close. Okay? Yeah. Any
35:04
yeah. Any other advice or burning young philosophers in the
35:06
in the audience, but before you push on to concrete
35:09
concrete questions. But high value for
35:11
me in philosophy is clarity. I
35:14
really care about clarity. And by
35:16
the way, these heuristics, I think,
35:18
help in that mission as
35:20
well, trying to just really get clear on what a
35:22
position is and what possible
35:24
defects are.
35:26
And as I think SIRL you understand
35:28
what you're saying. Yeah. And I
35:30
would begin with that. Yes. I'm
35:32
not sure a hundred percent agree.
35:36
So philosophers do tend to just be absolutely obsessed with, like,
35:38
making sure that the thing applies to every case.
35:40
And, like, when you read philosophy papers, you really do
35:42
get this impression that it's like a long series
35:45
of like paragraphs added to, like, like, every objection that some
35:47
philosopher might raise, so there's, like, nothing that can be
35:49
said in response. So it's in, like, one sense that's
35:51
extremely clear. In another sense, it feels like maybe you're,
35:53
like, losing the phrase for freeze
35:55
sometimes. Yeah. Like, in considering off of, like, small object
35:58
like, clever, small object rather than, like, is
35:59
the is the core issue because of the
36:02
thesis sound. Yeah. Maybe being too
36:04
pedantic. Yeah. Yeah. yeah, come on. You know
36:06
roughly what I meant. But Yeah. Since some
36:08
degree, like, a philosophy is a kind of professional pedents.
36:10
That's kind of a The whole thing is to find the education.
36:12
That's right. And I'm sure
36:14
you know the way that philosophy is often right.
36:16
It's often in this very defensive
36:18
style because you're preparing yourself
36:20
for the most uncharitable objection because -- Yeah. -- you didn't
36:22
quite nail the point, you know,
36:24
given your wording, there's some counter
36:28
example. And, yeah, this should
36:30
be tempered with the principle of charity that you should try to give a charitable
36:32
interpretation of what's being said
36:34
and and look for more
36:37
interesting
36:37
more profound deeper objections.
36:40
Yeah. Yeah. It is the most challenging part
36:42
of, I think, interviewing philosophers is is that they
36:44
really wanna be right. And just like,
36:46
can you the love of God, say something wrong. Say something wrong, but less
36:48
words, please. Yeah. Right.
36:50
And, yeah, you you ask a question and, well,
36:52
let's not quite right the way you said it.
36:54
Yeah. But Yeah.
36:56
If if at least the the spirit of it's
36:58
conveyed that that is often good
37:00
enough. Yeah. I I suppose yeah. If you if
37:02
you get I guess if you get too sloppy in that direction, then you can't like do
37:04
philosophy so much anymore. But then you just become like
37:06
ordinary people who accept common sense and like
37:08
I'm not looking for like the little ways that
37:11
like end up you, like, investigate more deeply, it does actually demolish common
37:13
sense when you think about it properly. Yeah.
37:16
Okay. Let's push on to a topic that is
37:18
dear to all of our heart's probability. Yeah. But
37:20
kind of the core question I
37:22
really wanna ask here is whether
37:24
basicism as kind of as it's often practiced by people in
37:26
the audience, whether it's true in
37:28
some sense. Yeah. But before I
37:30
ask that, we should probably wind back a step. At
37:32
a high level, kinda what are the different
37:34
theories of what probability what
37:36
probability is? Like, what yeah. What what ideas do people
37:38
have about it? Yeah. This is what my
37:40
professor should have said when I asked
37:42
the question, what is
37:44
p? And he thought I
37:46
needed medication. this would be the
37:48
beginning of a reply. Let's
37:50
distinguish three main kinds of
37:52
probability. The first, let let's
37:54
call it objective chance. Out
37:57
there in the world, there's probability and it's mind dependent.
37:59
It doesn't care what
38:00
we think. So I think maybe of
38:04
radioactive decay. The
38:06
radium atom has a decay law. It's got
38:08
a half life of sixteen hundred years,
38:10
which is a probabilistic fact. It
38:13
doesn't care what anyone thinks. It's just
38:16
out there in the world. It's like mass and
38:18
charge. It's some property
38:20
out there. That's the first.
38:22
Yeah. Can I make sense? Yep. Yeah. Second one,
38:24
let's call it Oh, I guess, under that one, do we have
38:26
kind of quantum mechanics issues? It's like the
38:28
probability of an electron being here rather than
38:30
there. Is that that in the same
38:32
category? Yeah. And that's a big question in the interpretation
38:34
of quantum mechanics whether there really
38:37
are objective probabilities or
38:40
is in fact everything deterministic and really the
38:42
chances are all ones and zeros -- Okay.
38:44
-- on one view. Wow. But
38:47
Let's get that. Okay. Second
38:50
interpretation, let's call it
38:52
subjective
38:53
probability and That's
38:55
now more about rational agents, people like
38:57
us. Call it Credence,
39:00
if you
39:02
like. and it's degree of confidence. We
39:04
sometimes we outright believe
39:07
things, but it's often
39:10
our degrees of confidence are more nuanced than that.
39:12
We believe some things more than others as we
39:14
might say. We have degrees of belief, and that
39:16
that would be the second interpretation. But
39:20
he had not just anything goes, it seems these are rational degrees
39:22
of belief, and it'll
39:23
turn out according to Beijingism,
39:25
they're constrained by
39:28
probability
39:28
theory. So would there be other, like, the correct credence?
39:30
Well -- Yeah. -- excellent. This this this
39:32
will get us to a big issue in
39:36
Beijingism. at
39:37
one extreme, you've got, I guess, a
39:39
very permissive radical subjectivism,
39:42
where the only
39:44
constraint is a
39:46
bay the probability calculus. They'll shelter
39:48
bay, columnar of sections, something like
39:50
this. Like, everything's got a sum to one
39:53
or the probability something happening as one. That
39:55
that's right. probabilities are non negative. Yeah. They have a top
39:57
value of one they add in a certain
39:59
way. Yeah. Okay. Now that
40:02
seems very seems very permissive.
40:04
You know, I can assign
40:07
probability even one to,
40:10
you
40:10
know, a meteor quantum
40:14
tunneling through this lecture theater
40:16
before we're finished and going out the other side --
40:18
Yeah. -- as long as I get probability zero to it,
40:20
not doing so and so
40:22
on. And probability, very high probabilities to the
40:24
sun not rising tomorrow and
40:26
so on. And it
40:28
doesn't violate
40:30
any fundamental axioms are probably
40:32
That's right. Yeah. But it seems somehow
40:34
too unconstrained. Right. Now at
40:36
the other end of the spectrum, This
40:39
is the so called uniqueness thesis
40:41
that your
40:41
evidence uniquely
40:42
constrains you and there's
40:44
exactly one creams you should have and
40:48
in fact There was let's call it an Ere prior. No.
40:50
There was the correct
40:52
thing. Correct. Before you saw anything about it. probability
40:54
before any evidence, you should
40:58
start there. and just update it on your evidence as it comes
41:00
in. Now that seems too extreme
41:02
perhaps in the other
41:04
direction that given my
41:06
evidence, there's a sharp probability
41:08
I should give to it
41:10
raining tomorrow. And -- Yeah. --
41:12
if I said seventy one point two rather than
41:14
seventy one point one, I wrong to think of --
41:16
Yes. Yeah. -- that's right.
41:18
And you'd think that different people could
41:20
disagree to some extent,
41:22
but not according to this thesis, not if
41:24
they're rational. And then you've got all the
41:26
positions in between in this
41:28
spectrum, and we could then
41:30
have discussions
41:32
about just how much grievances are constrained.
41:34
Actually, this nicely brings us to the third
41:36
main interpretation. Let's call
41:40
it evidential probability. And
41:41
the thought is that
41:42
evidence does put
41:44
some constraints
41:46
on
41:47
on hypotheses or propositions
41:50
degrees of support. Let's
41:51
call them. Okay? And probabilities
41:53
measure these degrees
41:56
of support. So for example, maybe given your
41:58
evidence, it's very
41:58
probable that the
41:59
sun will rise tomorrow. Yeah.
42:02
And then we might say the evidential
42:04
probability of
42:06
Sunrise given your evidence is
42:07
high. So I haven't quite tracked that. Yeah. How is this
42:10
different than the previous
42:12
ones? Yeah. So the previous
42:14
one, we started with just subjective
42:16
probability. Yeah. And notice
42:18
how we
42:18
sort of morphed into the third one,
42:20
But first, we we said we had the
42:21
permissive radical subjectivism that
42:24
said anything It's it's up to
42:26
you. Yeah. Your choice. Leave whatever you
42:28
like. Yeah.
42:30
Your credence just run with
42:32
it and then update it appropriately. And that definitely
42:35
was not evidential probability.
42:38
And then probably when we get
42:40
to the other extreme, the
42:42
uniqueness thesis, then it does
42:43
seem that that would be some something very
42:46
much like evidential
42:47
probability that would would
42:49
constrain you. And then there are the positions
42:51
in between. So I I don't wanna make
42:53
a sharp division between
42:55
subjective
42:55
and evidential probability. Really,
42:58
it's a spectrum and this
43:00
corresponds to what you
43:01
might call subjective bayziness
43:03
or more objective basingism
43:05
and degrees of that as we move
43:07
down the spectrum? Yeah. Okay. So
43:09
so the spectrum between, like, the
43:11
subjective basicism and these
43:13
kind of evidential a reasonism. So, like, where you can
43:15
believe anything and where have to believe, like, one specific thing. I like both of them feel but I feel
43:18
pretty terrible in their own own
43:20
own way. And you're saying, like,
43:22
maybe the most plausible thing or the most intuitive thing.
43:24
It's gonna be somewhere in the middle where it's like, you could there's
43:26
a range of stuff that you can believe that's not stupid.
43:30
Yeah. Yes. but you can, like, go too far. And it's also not just like one specific thing, you know,
43:32
like an idiot. If you don't believe the exact exact
43:34
truth. Yeah. I guess, whenever you have a spectrum like
43:36
that, it's gonna feel a little
43:38
bit arbitrary. a point you draw on there.
43:40
So, like, how would we even begin to know, like, what what is the correct point along that spectrum? Yeah. Well,
43:42
we start adding some constraints
43:44
to to the radical
43:47
subjectivism. Maybe for
43:48
example, the principle of indifference. You might
43:51
think that's a kind of
43:54
evidential probability explain that? Yes. Okay. When your
43:56
evidence bears symmetrically
43:58
on a range of
43:59
possibilities, not favoring
44:02
any other then
44:04
you should give probability these cases. It's most
44:07
plausible in gambling cases
44:09
like Etorso coin head's
44:12
tails. My evidence doesn't favor heads over tails.
44:15
It seems I should give half
44:17
half to those possibilities. And
44:20
then we can complicate that and find
44:23
brain
44:23
some problem and then say, well,
44:25
either you have no
44:26
evidence So you should
44:29
give equal probability to the cases or you do have
44:31
evidence that
44:31
bears symmetrically on them and you
44:33
should give equal
44:34
probabilities. Now a
44:36
big discussion could be had. There there
44:39
meant to be various serious problems. There are
44:41
serious problems with the principle of indifference
44:43
as I just stated it. and
44:45
various people think it's bankrupt
44:48
because of these problems.
44:49
Yeah. But it's funny how
44:52
people often reach
44:53
for it intuitively, even
44:56
when they've disavowed it a minute
44:58
ago, for example, in the Monty Hall
45:00
problem, may maybe you know that. Yeah.
45:02
You know? behind one of
45:04
three doors, there's some prize, what probability
45:06
should you give to
45:09
it being behind a particular
45:11
door? And then the problem continues. Now
45:13
no one
45:14
says, oh, I give probability
45:17
one seventeenth to behind
45:20
door number one because I'm just unconstrained,
45:22
and that's what I feel like. And
45:24
one one seventeenth, you feel the
45:27
pull of the principal in the difference.
45:29
I should give third, if you've just a
45:31
moment ago, disavowed the
45:32
principle of indifference. So it seems like
45:34
in some restricted cases, at least,
45:38
Yeah. It
45:38
has some pull. But I suppose one we've been talking here
45:40
about, like, what you should believe or just this is, like, smuggling
45:42
in this idea that there is, like, anything that you
45:45
should believe. We're, like, would be, like, practically, brought
45:47
in, like, ethics before, like, some ethical considerations almost into, like, into what
45:49
you should believe. And and and maybe that'd be too much
45:51
to get into here. This we
45:54
may be getting into ethics later
45:56
too, which is not my area
45:58
of specialization, but I'm always happy
46:00
to try. Okay. With that kind of
46:02
scene setting about what probability might be out of
46:04
the way. Real quick, I wanna ask, you know, me and
46:06
my friends, when we're like hanging out, when
46:08
we're chatting, watching the news or talking about events in the world. You know, I might say something
46:10
like I think that's like an eighty percent chance that Boris
46:12
Johnson is gonna resign in the next week or
46:14
two. And someone else would say no,
46:16
I think seventy percent, and then we, like, go
46:18
and check the news and, like, moment as to resign, and we're, like, okay, I think it's, like, eighty five percent now.
46:20
Yeah. How are you doing? How
46:22
are you doing right now? Yeah. Okay.
46:26
But are we doing the right thing? Is is this like the correct thing to do?
46:28
Or are we just playing some game that's like fun
46:30
for us and is like not really any better
46:32
than any other approach? or or at
46:35
least like not like uniquely privileged as the correct approach. Yeah. So I am
46:37
a big fan of Beijingism. I guess
46:39
I
46:39
say on Mondays,
46:42
Wednesdays, Fridays, I
46:43
call myself a baysian, Tuesdays,
46:46
Thursdays, Saturdays, I'm not so
46:48
sure. Today's Saturday. Today's
46:50
Saturday. Okay. Maybe I'm not
46:52
so sure. But look, I think there's something right about
46:54
this. We have these
46:56
degrees of confidence. I think
46:58
we have to allow this. It's not just
46:59
all or nothing belief
47:02
It does seem there are better and worse ways to
47:04
manage these degrees of
47:06
confidence. There are various good reasons
47:09
to think that they should obey probability theory. We may talk
47:12
about that Dutch books, for
47:14
example. There are ways
47:14
that they should update, so called
47:18
conditionalization. And we can then have disputes about how much
47:20
constraints there should be on
47:22
the priors that you have we
47:25
we had a bit of that discussion before. Yeah. But it it
47:28
seems like a very good model and it's
47:30
very fertile too. Lots of
47:32
good things are closely
47:34
related to this. Maybe soon we'll talk about decision
47:36
theory, another thing that I'm fond
47:38
of, and probabilities figure
47:40
centrally in that. And so I
47:41
think it's a good good place to
47:44
start. Hey,
47:44
listeners. One more definition. The
47:46
term
47:46
Dutch book is from
47:49
gambling. It's a set of odds or bets that when created
47:51
by a bookmaker, ensure that the bookmaker is
47:53
going to make a profit. Now, how is
47:55
that relevant to decision procedures and
47:58
beliefs and and all of that kind of
48:00
thing. So
48:00
one weakness that set
48:02
of ideas about grievances or
48:05
decision making could have is
48:07
that someone could be presented with a series of bets that they could be
48:09
offered to make or a series of trades that they
48:11
could be offered to make. that
48:14
would cause them to just constantly get worse off and worse
48:16
off. Basically, they have less and less money until
48:18
they bankrupt. How could that happen?
48:20
how could that happen So
48:21
normally, in philosophy, a standard thing is
48:24
that with preferences say, is that you would
48:26
assume transitivity. So if you think a
48:28
is better than b, and you think b is
48:30
better than c, and you should also
48:32
think that a is better than c.
48:34
Now you might
48:34
want to say, I don't want to buy this
48:36
idea of transitivity. I want to think that
48:38
I could believe that a is better
48:40
than b that b is better than c,
48:42
but also that c is better than a. So
48:45
you have kind of a circle here where you'd be
48:47
willing to give up c in exchange for b.
48:49
you'd give up b in exchange for a, but you'd also
48:51
give up a in exchange for c. But one
48:53
of the reasons why giving up this
48:55
principle of transitivity isn't very
48:57
desirable. is that if you have circular preferences like
48:59
that, then it's possible for someone to just
49:01
keep trading with you. So if you think b
49:03
is better than c, then you'd be willing to
49:06
give up c and potentially pay
49:08
a bit of extra money in order to get
49:10
B. But you can see that someone could just
49:12
keep trading you, b for c, a
49:14
for b. c for a
49:16
and so on and so on and so on in a
49:18
circle until you were left with no money. So
49:20
you said you might end up with c again, but then you've
49:22
lost all the other things that you traded as we were
49:24
going through this circle. where each time you thought things were getting better, but then you just end
49:26
up back where you started. So when
49:28
philosophers say that a set of ideas would allow
49:30
you to be
49:32
Dutch booked, They're saying that they're inconsistent some someone
49:34
to in principle offer you a series of exchanges
49:36
that you would all accept because you
49:38
think that they are good
49:41
because individually, they look good, but then
49:43
collectively, they would clearly just leave you
49:45
better off, potentially just just bankrupt. Okay.
49:47
Back
49:47
to the show. Yeah. Go
49:49
go go go go go go go go go go go go go go go
49:51
go go go go go go go go go go go
49:53
go go go go go
49:55
go go go go go go But for some of
49:57
the week, you're not into vaziness -- Yeah. -- or not into like this approach. Yeah. Okay. Let let's
49:59
do that one first. Like, what what are the reservations
50:01
that you have? Yeah. We
50:04
start with some
50:06
excimatization of probability, and
50:08
Colmogorff is is the
50:10
standard. Okay? And
50:14
he excimatized unconditional probability first and
50:16
then conditional probabilities defined in
50:18
terms of unconditional probabilities. The
50:20
probability of a given b
50:22
is a
50:24
ratio probability
50:25
of A probability of B,
50:26
assuming that the probability of
50:28
B is positive. I actually
50:31
be other way around. I think the conditional
50:34
probability is the basic notion,
50:36
the fundamental notion. I think
50:38
there are problems
50:40
with Colmogorov's exmeritization,
50:42
as I've just put it, he did have
50:44
further subtleties, which I can't get into
50:48
here. But two kinds of problem. Remember I had this
50:50
proviso provided that the
50:52
bottom line of that ratio is positive. You
50:54
can't divide by zero. But here's
50:56
now another problematic
50:58
feature of the standard theory with
51:01
real valued probabilities. Yeah. It
51:03
seems that probabilities zero
51:05
events can happen
51:08
Now
51:08
that's not intuitive. We normally think if something
51:10
can happen, it
51:11
must have positive probability. Yeah. Or
51:13
what what's an
51:16
example? Yeah. Okay? Think, for example, of the
51:18
radio atom
51:19
decaying exactly
51:21
at noon tomorrow,
51:24
say, what's the probability
51:26
of that? On the standard theory
51:28
where it's real valued probabilities,
51:30
in this case, an exponential
51:34
decay law, The
51:34
probability of it decaying exactly at
51:36
noon is zero. Now
51:38
it decaying in
51:39
some infinitesimal region
51:42
around noon, that could be something more. But
51:45
noon itself, that
51:47
point gets probability
51:48
probability zero
51:49
or or throw a dart at a dartboard.
51:52
What's the probability that it hits
51:54
randomly? It hits the
51:56
exact center
51:58
Yeah. Okay. So this really didn't do much math at university. This
52:00
is this is this is this classic thing. So imagine
52:02
that we're gonna randomly choose a real number
52:04
or, let's say, any number between zero and
52:06
one. the probability of any
52:08
being any specific one of them -- Yep.
52:10
-- any specific number is zero because there's,
52:12
like, uncountably infinite any numbers. Indeed,
52:14
I think the probability of you picking like
52:16
randomly picking a number that could ever even be written down is zero.
52:19
Right? Yeah. Well, you think it's computable? Yeah. And
52:21
it's not computable. Right? Yeah. And there are only
52:23
cannibalizing many of those and Okay.
52:25
Okay. So you're saying each of these, like, specific
52:27
numbers has a probability of being chosen of zero or
52:29
any any specific example that you pick a zero. And yet, we know
52:31
that one of them will chosen. Yes. Okay.
52:34
So yeah. Okay. So it's a bit like the lottery
52:36
paradigm. It is. It's like an uncountable
52:38
lottery paradigm. And this is a
52:40
puzzling feature. And
52:42
there is this this rather strong intuition.
52:44
Come on. If if it can happen,
52:46
it should get some positive probability.
52:48
We want to distinguish it from
52:50
an impossible event and
52:52
it seems the way to do that is, well, it's got greater probability. And
52:54
actually, this is another way in
52:57
which we might go
52:58
away from this
53:00
numerical call megal effect immunization. You
53:03
might say more fundamental is
53:06
comparative probability. And now the
53:08
primitive is something like x is
53:10
more
53:10
probable than y or maybe
53:12
better still
53:13
w given x
53:15
is more probable
53:16
than y given z or
53:18
at least is probable and now
53:20
make that the fundamental notion
53:23
and then
53:23
we can maybe recover the numerical
53:26
theory out of certain axioms and so
53:28
on on those fundamental
53:30
comparative probabilities. Yeah. hey, here
53:32
with another definition. We just use the term
53:34
primitive, a philosophical primitive.
53:36
In philosophy,
53:37
a primitive notion is
53:39
a concept that is kind of fundamental
53:42
and can't be broken down and defined in terms
53:44
of other previously defined
53:46
concepts.
53:46
In geometry, for example, a
53:48
point or a line or contains or
53:50
potentially space might be primitive notions that
53:53
one can't really break down and define in terms
53:55
of other notions. They're kind of the
53:58
starting point. Okay. Back
53:58
to the show. I'm just thinking about the the
53:59
the, like, choosing a specific number case. Okay. So so it's
54:02
it's actually quite a bad situation because
54:04
you're saying, So normally,
54:06
with basicism, we're we're thinking we got a prior, we
54:08
got like a prior probability that any number be chosen,
54:10
and then we update a lot based on the evidence.
54:12
Yeah. But because our prior and any any number is gonna
54:14
be zero. And when you, like, multiply anything by
54:17
zero, it's still zero. So so
54:19
random number gets chosen, you're like, you have
54:21
your prior you observe the and you still think
54:23
it wasn't chosen because it would be it was impossible for any number to be chosen,
54:25
whatever it was. Well, it's like, quite bad. Impossible
54:27
in this probabilistic sense, and that's
54:29
the trouble that and it
54:31
seems that probability, so to speak,
54:34
has blurry vision. It does not
54:36
distinguish between the genuine impossibilities
54:38
-- Yeah. -- and these very
54:40
unlikely, but in some good
54:42
sense possible cases. And this
54:44
is related to my
54:46
worry about the ratio formula, again, you divide
54:48
by zero. Now what happens
54:50
if you learn in this
54:53
experiment where you randomly
54:55
choose a real number from say zero
54:57
one. You learn its value, but it's something
54:59
you antecedently gave
55:02
probability zero. If you just
55:04
looked at the ratio
55:06
as standardly understood as
55:09
involving
55:09
a ratio of real
55:11
numbers, It was zero divided
55:12
by zero, it is zero that you'd get that particular
55:14
number. Now how do you update on
55:16
that? How do you, as we say, conditionalize
55:18
on something you originally gave?
55:21
probability zero. Now I have to
55:23
say that this quickly gets us into
55:25
very sophisticated mathematical refinements.
55:28
Colmogorff himself was well aware of this problem
55:30
and he had sophistication. People bring on
55:32
hyperreal numbers. They bring on
55:33
richer number systems that are meant
55:35
to preserve this
55:38
intuition that if something's possible, it'd it'd better have positive
55:40
probability -- Yeah. -- and we can
55:42
let that debate unfold. But it's
55:44
it's
55:44
just to say, now this is Saturday,
55:47
you know, I'm saying, I'm just
55:49
trying to bring out there are various problems
55:51
with just this simple statement. straightforward.
55:54
It seems like. Just
55:56
follow Colmogorff probabilism says
55:57
your probabilities should
55:59
obey these axioms well. It's
56:02
more complicated. should
56:03
we take conditional probability as primitive?
56:06
Do we have to worry about the probability
56:08
zeros? Do we enrich the
56:10
number system beyond
56:11
the reels, which will take comparative
56:14
probabilities as basic lots of
56:16
debates to be had. Yeah. That this might be
56:18
a stupid frivolous question. But is it is it
56:20
maybe that I guess it not
56:22
possible. Is it possible to have a
56:24
probability that's like a negative
56:26
probability or like a probability above one
56:28
or like Well, I'm just thinking, you know, obviously, we have imaginary
56:30
numbers now, like, off the real number line. Yes.
56:32
Like, is is any of this coherent?
56:34
Yeah. Well, some people
56:36
think so Feynman thought they could be
56:38
negative probabilities.
56:40
And, of course, you're not obeying
56:43
on the Gaurav. the chromogor effect
56:45
amortization or thou shalt have non
56:47
negative probabilities. There
56:49
are some thoughts
56:52
along these lines that there's something that behaves probability
56:54
like some function that seems
56:56
to have probabilistic properties. And
57:00
then you see that it sometimes goes negative. There are meant to be some
57:02
physical examples of this. And then you
57:04
say, well, those are the negative
57:07
probabilities. There are some problems with
57:09
this,
57:09
just to take one,
57:12
consider the
57:13
usual formula for independence. Two
57:15
events A and
57:15
B are independent, just in
57:18
case the probability of A and
57:20
B equals the probability of A
57:22
times the probability
57:24
of B. now
57:24
suppose you have independent negative probability
57:27
events, their
57:29
product, something negative times,
57:31
something negative. Yeah. is
57:34
positive. Alright. Yeah. That seems to be
57:36
a problem if if you think the event
57:38
itself and maybe the conjunction of
57:40
it with something else has negative
57:42
probability. How does independents pan out?
57:44
Yeah. Huge, you know,
57:46
issue we could get into there too. Yeah. Are
57:48
there any other downsides to
57:50
bayzunism that important or or or way on you at all? Here's
57:52
one thing that just
57:53
bothers me a bit, and I'll throw
57:55
it out there. As a
57:57
slogan, I'll say, subjective bay
57:59
zionism is anchoring
57:59
and adjustment, and I
58:01
need to explain what I mean
58:03
by that. Anchoring and
58:05
adjustments are heuristic that
58:07
often use when estimating some
58:10
quantity, they're given the so called
58:12
anchor some starting point for
58:14
thinking about the value of
58:16
that quantity and then they adjust until they reach an estimate
58:18
that they find plausible. The trouble
58:20
is that sometimes the
58:21
anchor is in highly
58:24
irrelevant to the quantity, and it just
58:26
should be ignored, yet it
58:28
still influences the
58:30
final estimate. the adjustment
58:32
is insufficient. And there are a couple of
58:34
classic examples, which I can give you. To
58:36
our skin condiment, had a famous study.
58:39
They asked people to watch the
58:41
spin of a roulette wheel, which was rigged to
58:43
land
58:43
on either ten or
58:46
sixty five. And
58:48
then they were asked whether the percentage of African countries
58:50
in the United Nations was
58:53
higher or lower
58:56
than
58:56
the number that they saw. And then they were asked
58:58
to estimate the percentage. Okay.
59:00
Those who saw a low number
59:03
tended
59:03
to give substantially lower estimates
59:06
for the percentage than those who
59:08
saw a high number.
59:10
And look, of course, they knew
59:12
that the RULET number, the anchor
59:14
provided no information whatsoever
59:16
about the percentage, yet it still
59:18
influenced their estimate. And
59:20
that
59:20
just
59:22
seems absurd. That's just seems crazy.
59:24
There's another famous
59:25
study. It was Arieli and Co.
59:27
They asked NBA students
59:30
at MIT to write
59:32
down the last two digits of
59:34
their Social Security number. And
59:36
then they were
59:36
asked whether they would pay this
59:38
number of dollars for some product
59:40
say, a bottle of wine box of fancy
59:43
chocolates and so
59:43
on. And then they were asked, what
59:46
was the maximum amount they were willing
59:48
to pay for
59:50
the product. Those who wrote down higher
59:53
two digit numbers were
59:54
willing to pay substantially more
59:58
And of course, they knew that their Social Security number
59:59
was completely uninformative
1:00:02
about the value of the product, but still
1:00:04
they anchored on it and
1:00:07
it influenced their final valuation. Okay. So the
1:00:09
idea is that the residue of the
1:00:11
anchor remained even after the
1:00:13
adjustment of thinking well,
1:00:15
how valuable is this product really? Okay.
1:00:18
Now these seem to be paradigm cases
1:00:20
of irrationality. Okay. But
1:00:22
now consider consider a putative
1:00:24
paradigm of rationality,
1:00:27
subjective
1:00:27
bayziness. Okay? Here
1:00:28
you start with a prayer That's
1:00:32
your initial probability distribution
1:00:34
before you get any information. And
1:00:36
the only constraint on this
1:00:38
is that it abaze the probability calculus.
1:00:41
That that's the version
1:00:43
I'm thinking of.
1:00:44
Okay. That's your anchor. Your
1:00:45
prayer is your anchor. And
1:00:47
then you get some information and you
1:00:50
update by conditionalizing on
1:00:52
it, as we say. So your
1:00:54
new probabilities are your old
1:00:56
probabilities conditional on that
1:00:58
information? That's your
1:00:59
adjustment. But the trouble
1:01:01
is that your prior
1:01:03
has no evidential value. It's it's
1:01:06
not based on any
1:01:08
information and you know this. That's what
1:01:10
makes it a prior. and its
1:01:12
residue remains even
1:01:14
after the adjustment often.
1:01:16
Now we can imagine that your
1:01:19
prior was even determined by the spin
1:01:21
of a roulette wheel or by your Social Security
1:01:23
number. As long as it obeys
1:01:25
the probability calculus, and
1:01:28
still it influences your final probabilities, your posterior
1:01:31
probabilities as we say. And now the
1:01:33
worry is, why isn't that
1:01:36
just as absurd as before. We were laughing
1:01:38
at the people in the
1:01:40
African United Nations experiment or
1:01:42
the wine and chocolate experiment.
1:01:46
what's the relevant difference? And look, there are
1:01:48
things that one can say, but I
1:01:50
just put that out there as something
1:01:52
that that needs some attention. I
1:01:55
think
1:01:55
in theory, if you get a lot of empirical
1:01:58
information over time, then the
1:01:59
influence of your
1:02:02
starting point. becomes more and more relatively irrelevant over time as this
1:02:04
kind of washed out by all these updates you're
1:02:06
making, you know, conditionalizing on things that you've
1:02:08
observed. So there's there's one thing, at least
1:02:10
if you you're around long enough and
1:02:12
collecting enough evidence to to move away
1:02:14
from the very first priority you started
1:02:16
with. I guess another difference that that I
1:02:18
imagine is that I think in theory you're meant to
1:02:20
start with something like an uninformed prior or I'm sorry. Which was
1:02:22
just to say a prior that is extremely
1:02:24
agnostic. I mean, like, is really pretending to
1:02:26
know all that much.
1:02:28
Now it's think it's a bit
1:02:30
hard to define exactly what is the appropriate, uninformed prior, but the hope is that
1:02:32
your views are gonna be very flexible
1:02:34
initially because it would be
1:02:38
presumptions, it would be foolish to think that before you look to any evidence, you
1:02:40
should have a really strong view about things. So maybe that's
1:02:42
one way in which the prior case
1:02:45
is well, the priors should hopefully be doing not that
1:02:47
much work. Even if we do concede that it
1:02:49
is largely arbitrary. Yep.
1:02:50
Excellent. Both good replies. The
1:02:53
first one was the washing
1:02:55
out of the prize in the long run. And
1:02:57
there are these famous convergence
1:03:00
theorems for Beijingism that
1:03:02
in the limit, under certain
1:03:04
appropriate conditions, effectively the
1:03:07
prior is completely washed out.
1:03:09
And that's good. But
1:03:10
that then
1:03:11
as Kain said in the long run, we'll all
1:03:13
be dead. And at
1:03:16
any point in our
1:03:18
finite lives, we will
1:03:20
not have reached that limit.
1:03:22
And the worry is that
1:03:24
the residue of the prior will remain
1:03:27
at all of these finite stages, which
1:03:29
is all we ever have.
1:03:31
Okay? So it's nice to know that there are
1:03:33
these theorems that we'll get there in the end,
1:03:35
but the problem is we're never at the
1:03:37
end. And regarding the second reply, and that that's certainly a good one.
1:03:40
Yes. I think actually that's the way to go that you
1:03:42
want some sort
1:03:44
of constraints beyond the
1:03:46
probability calculus on the
1:03:48
priors. And as you said,
1:03:51
maybe uninformative priors as we might
1:03:54
say, maybe there are some
1:03:56
constraints on what the price should be,
1:03:58
not just anything goes. And in fact,
1:04:00
there's the so called uniqueness thesis that
1:04:02
says, no, there's even exactly one.
1:04:04
There's only one rational starting
1:04:06
point, which is somehow, you know,
1:04:09
maximally uninformative, you might say,
1:04:11
For example, the principle of indifference might
1:04:13
kick in at that point. And now we
1:04:15
get into a good debate about
1:04:17
just how constrained prize should
1:04:19
be. So anyway, I raised anchoring and adjustment worry
1:04:22
as a concern for that completely
1:04:26
unconstrained version of
1:04:28
Beijingism that just said, thou shalt obey
1:04:30
the probability calculus. But beyond
1:04:32
that, it's okay, whatever you do.
1:04:35
it seems that you need to be a
1:04:37
bit more constrained than that. And then as
1:04:39
you say, the analogy to
1:04:41
anchoring an adjustment starts to to go away.
1:04:43
For example, we can't just spin a roulette
1:04:46
wheel or we we can't just
1:04:48
get your Social Security number and
1:04:50
make that
1:04:52
your prior. That's not the right starting
1:04:54
point. We haven't got five minutes left
1:04:56
on stage. So I wanted to give you a chance to
1:04:59
for the audience a few criticisms of fruquentism, which is
1:05:01
always. So firstly, like, what is fruquentism? And
1:05:03
secondly, like, what are some of the problems that we haven't yet mentioned
1:05:05
with it? Okay. Frequentism.
1:05:08
Remember my taxonomy of
1:05:10
objective probability, subjective probability,
1:05:13
evidential probability, frequentism would be
1:05:15
a version of objective. probability
1:05:17
out there in the world, mind
1:05:20
independence, there are these
1:05:22
probabilities. What are they? They're
1:05:24
relative
1:05:24
frequencies. Mhmm. To take a simple case
1:05:27
what's the
1:05:28
objective probability
1:05:29
that the coin lands heads,
1:05:31
let's suppose it's a fair
1:05:33
coin? Well, you toss
1:05:35
the coin a number see it lands and count the
1:05:37
number of heads over
1:05:39
the grand total. if
1:05:42
that number turns out to be a half,
1:05:44
then Bingo, you've got a chance
1:05:46
of a half for heads. Yeah. And
1:05:48
and now we generalize that. So
1:05:51
in general, the
1:05:52
probability of some, let's
1:05:54
call
1:05:55
it, an attribute relative
1:05:57
to some reference class So
1:05:59
attribute
1:05:59
a relative to reference class
1:06:01
r is the relative
1:06:04
frequency of a's
1:06:06
in that reference class. makes
1:06:09
sense, I think. And I think it may be still the most popular
1:06:11
account is among scientists.
1:06:14
Yeah. It's always scientist,
1:06:16
isn't it? They're the ones that They need to
1:06:18
talk to philosophers. Yeah. Absolutely. on so
1:06:20
many things. Yeah. Sorry. Okay. So this is kind of
1:06:22
a this has an intuitive appeal, and it's a bit like how
1:06:24
you're taught about probability. high school or something.
1:06:26
Yes, that's right. Yes. And we all know
1:06:28
that there's got to be some close
1:06:30
connection between probability
1:06:31
and frequency
1:06:34
and
1:06:34
frequentism posits the tightest connection of all
1:06:36
identity. And now we could distinguish
1:06:38
two kinds of frequentism.
1:06:40
You might call it actual frequentism,
1:06:43
you just look at what actually happens. You just,
1:06:45
in fact, toss the coin some
1:06:46
number of times -- Yeah. -- count the
1:06:49
number of heads divide
1:06:51
by the total of trials done. And then we
1:06:53
could
1:06:53
go hypothetical and
1:06:54
say, well, no, I really
1:06:56
meant some long run of heads
1:07:00
maybe I didn't get a long run,
1:07:01
so I go hypothetical, I imagine
1:07:04
counterfactually a long run
1:07:06
of trials and relative
1:07:08
frequency I would get maybe in the
1:07:10
limit. If I have infinitely many trials,
1:07:12
we have two versions. And
1:07:13
you probably want to ask why I don't
1:07:15
like -- Yeah. --
1:07:17
frequentism. I think version
1:07:18
of frequentism has the following kind of
1:07:20
problem. I think it's just built
1:07:22
into the very notion of probability
1:07:25
that fixing
1:07:26
the probability of something is compatible with
1:07:29
just any pattern and
1:07:31
any frequency of the corresponding
1:07:34
outcomes. Let's let's do it. Yeah. Yeah. Let's
1:07:36
do it for the coin toss. Let's suppose we
1:07:38
have a fair coin by which
1:07:40
I
1:07:40
mean the chance of heads as a half.
1:07:42
I
1:07:43
say that's compatible with any distribution of
1:07:45
heads and tails, including heads on
1:07:47
every toss, tails on
1:07:50
every toss, and everything in
1:07:52
between. And I can even tell you what the
1:07:54
probabilities of those outcomes are.
1:07:56
You know, it's one on two to the end
1:07:58
for end trials for each
1:08:00
exact sequence. But
1:08:02
it seems you can't say that
1:08:04
if you're a frequentist because
1:08:06
if the coin lands heads every
1:08:08
time, then that's probability one according to
1:08:10
frequentism. So there are various problems. We we
1:08:12
talked about the the problem of the single
1:08:15
case -- Yeah. -- earlier. there's
1:08:17
the problem of the double case. If I
1:08:19
toss a coin twice, then I can only have probabilities
1:08:22
of zero half
1:08:25
And one, off hand, I would have thought there could be other biases. And
1:08:27
as I like to
1:08:30
point out, it turns
1:08:32
out that
1:08:34
you cannot toss a fair coin
1:08:36
an odd number of times
1:08:39
according to frequentism, at
1:08:41
least this actual frequentism. because just
1:08:43
by definition according to them, if it's
1:08:45
an odd number of times, you can't
1:08:47
have exactly a
1:08:50
ratio of half of heads. Yeah. That
1:08:52
just doesn't seem – it seems pretty
1:08:54
weird. That seems pretty, pretty weird.
1:08:57
You'll also have I guess, a version of the the
1:09:00
gambler's fallacy. Suppose you
1:09:02
somehow know that
1:09:03
the
1:09:04
chance of heads is
1:09:06
a half of this coin and you
1:09:09
start tossing the coin and you
1:09:11
see a surprisingly long run
1:09:13
of heads as could happen. Yeah.
1:09:15
But if you know that somehow God tells you that
1:09:17
the chance of heads is a half,
1:09:20
you know that tails must be coming.
1:09:22
Oh, right. because they've got to make
1:09:24
up the run of
1:09:26
heads that's happened so far. Again, that was -- Yeah. -- actual frequentism. Now let's go
1:09:28
to the other kind,
1:09:31
the hypothetical frequentism. Yeah. But
1:09:35
now we have some counterfactual. This is not what actually happens. If
1:09:37
you were to toss the coin some large
1:09:39
number of times, maybe
1:09:42
infinitely many times, this
1:09:43
is what would happen. I think
1:09:45
for that, this is a
1:09:47
very weird counterfactual.
1:09:49
Imagine tossing an actual
1:09:51
coin like the
1:09:51
twenty cent coin in my pocket infinitely
1:09:54
many times. What would that even mean,
1:09:57
you know, that this coin would would
1:09:59
disintegrate before
1:09:59
then. So so we'd have to violate the
1:10:02
laws of nature so that
1:10:03
it would survive infinitely
1:10:06
long. I think it's very strange. I think there's a fact of
1:10:08
the matter of exactly how the the
1:10:10
queen would land in this infinite
1:10:14
sequence and even what it's frequencies
1:10:15
would be? Yeah. So you're trying to make this
1:10:17
like
1:10:17
very practical theory or that's just kind of a it's
1:10:20
virtue, but then it
1:10:22
seems to involve bizarre scenario that could never happen
1:10:24
to to back it up. Yeah.
1:10:26
Exactly. And I'll give actual frequentism
1:10:28
credit. At least, it was I
1:10:30
suppose practical anchored in the
1:10:33
world, gave you numbers that you
1:10:35
could ascertain. Yeah. Now when when the
1:10:36
problems start piling up and I gave
1:10:39
you a few of them, You start retreating
1:10:40
to this other version, the hypothetical frequencies, but
1:10:43
this seems to have other problems. It's
1:10:45
not practical. It's
1:10:48
not ascertainable. I
1:10:48
can't make sense of these counterfactuals. They seem
1:10:51
to be about the wrong thing about some idealized coin that would
1:10:53
never disintegrate and
1:10:55
what have you
1:10:56
what have you and
1:10:57
so put it all together and I
1:10:59
I frequentism is pretty bad. know there are lots
1:11:02
of frequentists out there read
1:11:05
my papers, the fifteen arguments against finite frequentism, finite fifteen arguments against
1:11:08
hypothetical -- Yeah. -- frequentism. I
1:11:10
I think it's in a bit of
1:11:12
trouble. Alright.
1:11:14
Well, we have to go upstairs because other people need the stage here. But,
1:11:17
yeah, we're gonna go chat about ways of
1:11:19
breaking expected value, counterfactuals, and why and
1:11:21
your your your many objections to
1:11:24
objective utilitarianism. But for now,
1:11:26
can everyone give random applause to to Alan? Thanks so much.
1:11:34
Alright. We're we're we're back in your
1:11:36
in your office, Alan. So I I was just
1:11:38
saying before we started recording that. In general,
1:11:40
I have a prejudice against doing interviews
1:11:43
in front of live audiences. because my my experience
1:11:45
listening to other other shows is that guests tend to kind of panned it to the crowd
1:11:47
a little bit, especially if they're talking about, like, charge issues. They
1:11:49
don't wanna have the courage to
1:11:51
say things of things wouldn't like. I
1:11:53
hope I didn't pander too much. No. No. I think Fortunately, we're we're mostly safe on this topic, except
1:11:55
perhaps for quizzes, although -- Right. -- you
1:11:58
probably have, like, sympathetic and unusually sympathetic
1:12:00
crowd. criticisms
1:12:02
or fruquintism. Hopefully, yes. Cool. But now
1:12:04
we can do the really hard hitting critical challenging
1:12:07
stuff. That was very good trouble. I'm
1:12:09
ready. Oh, okay. So before we came in
1:12:11
here, we were chatting about criticisms of fruquintism.
1:12:13
I guess one that you didn't bring
1:12:15
up, that jumps to mind for me. is
1:12:17
it it it seems like very odd that for Quenchism, it's saying like, okay, we've got like
1:12:19
one particular coin and we're saying
1:12:22
like, what is the
1:12:24
chance? that this coin is gonna land
1:12:26
heads or tails. It seems like it depends then on like all these other coins or like the same
1:12:28
coin and like
1:12:31
other points in time why
1:12:33
should the properties of this coin or, like, our knowledge of them,
1:12:35
like, depend on things far away in space and time.
1:12:37
It's very odd
1:12:39
in that respect. Yep. This radium
1:12:42
atom point to a particular atom -- Yeah. -- the case with probability a half
1:12:45
in sixteen
1:12:48
hundred years that seems to be an
1:12:50
intrinsic property of this atom. It seems
1:12:51
a little odd that its probability depends
1:12:53
on how all
1:12:56
these other atoms maybe very far away in space and
1:12:58
time happen to go decaying or not. Yeah. But I guess
1:13:00
the atom case sharpens it because with a
1:13:02
coin, you can flip it many times, but
1:13:04
each radium atom
1:13:06
can only decay once. That's it. And so you can imagine a scenario where what if there was there was only one radium or
1:13:08
like what if there was lots of
1:13:10
radium atoms and then you've got some frequency
1:13:14
then you, like, shrink it down such that now there's only one left. Yep. I
1:13:16
guess the fruquintis has to say now there's no
1:13:19
probability left anymore because there's just
1:13:21
one or zero. Yeah. Exactly. At that point, Yeah. It
1:13:23
seems very strange that probability depends on these
1:13:25
very extraneous facts. You'd think
1:13:28
it's just the the protagonist
1:13:30
is right here and now it's this Adam,
1:13:32
we're talking about -- Yeah. --
1:13:34
reminds me a bit of Hume's theory of causation about constant conjunction. And
1:13:38
take a paradigm case of causation like
1:13:40
I I put my hand
1:13:42
in a flame by accident and
1:13:46
I feel pain it seems somewhat odd to me to say,
1:13:48
well, what makes that causal claim
1:13:50
true is a fact about,
1:13:53
the you know,
1:13:54
these very disparate events across space and time, whether
1:13:56
putting hands in flames
1:13:58
were followed by
1:13:59
pains across space
1:14:02
and time? No. It seems like the protagonist Protagonist are
1:14:05
right here and now. Hey, listeners. Alan just
1:14:07
used the term constant conjunction. I
1:14:09
didn't know what that meant. But since that constant conjunction
1:14:11
is a relationship between two events where
1:14:14
one event is always invariably
1:14:16
followed by the other. So if
1:14:17
the occurrence of a is always
1:14:19
followed by b, and a and b are said
1:14:21
to be constantly conjoined. So in this case, I guess, you got the putting your hand in the
1:14:23
fire is always associated with
1:14:26
it with it being burned. And
1:14:28
I think cubes suggested that causation is just a
1:14:30
matter of constant conjunction. Okay. Back to the show. I see. So
1:14:32
so this is
1:14:34
an account of causality where
1:14:36
saying, the flame
1:14:38
causes the pain -- Yeah. -- if insufficiently, like, many cases or, like, hypothetical cases, the flame
1:14:41
and the
1:14:44
pain occur related really strongly. Yeah. Well So
1:14:46
saying, like, what does that got to do with it? Yeah. That's yeah. So who had a version of causation like that,
1:14:48
this account involving constant
1:14:50
conjunction? I think that frequentism
1:14:54
actual
1:14:54
frequentism, especially, is quite a lot like that.
1:14:56
Mhmm. And
1:14:57
I have the same reaction just
1:14:59
as you did that, look,
1:15:01
we should just be
1:15:03
staring at my poor hand and
1:15:05
the flame or the the radium atom, it seems odd
1:15:08
that its its chance
1:15:10
depends on
1:15:10
these maybe very distant
1:15:13
maybe very distant
1:15:14
facts and health, they they pan out. Yeah. Okay.
1:15:16
I guess causality is its own
1:15:19
With buddulamax. So that's a
1:15:21
big topic. Let's go back to probability.
1:15:23
Yeah. I think that's that's probably enough of fruquintas and people can go and check out
1:15:25
your your papers demolishing it. Yep. demolishing it, I
1:15:27
guess, thirty different ways and
1:15:29
maybe some extra my original paper was called thirty
1:15:32
arguments against fricotism. Mhmm. And I
1:15:34
I sent it in and I was
1:15:36
told it was it was a good
1:15:38
paper, but it was much too long. In fact, it twice as long as what could publish. So they
1:15:40
said that you need to
1:15:42
cut it. It was easy. fifteen
1:15:46
arguments against finite frequentism, fifteen arguments
1:15:48
against hypothetical frequentism -- Right. -- going to
1:15:51
okay. So if people people will often
1:15:53
cite both, you get double the citation. Absolutely.
1:15:55
Yeah. I should've got a mobile that's some more
1:15:58
I actually well, apparently, this was a big
1:16:00
problem in genetics. where you would have people
1:16:02
who were doing like whole genome studies and they decided to they they realized that they could that the studies of,
1:16:04
like, the effects
1:16:07
of different genes have break it,
1:16:09
like, in one paper for every chromosome. So they potentially get, like, thirty
1:16:11
different papers out of basically, like, exactly the same study. Yeah. That
1:16:15
that's how you you pump up your citation, your
1:16:17
various markers of
1:16:20
productivity. Yeah. Yeah. Okay. So what
1:16:22
you got from fruquintisum? I'm wondering
1:16:24
like Are there
1:16:26
other ways that, like, I could in practice start, like
1:16:28
reasoning differently? You know, start, like, in my daily life, thinking about the
1:16:30
probability of Boris Johnson being deposed in in different ways. I guess,
1:16:34
The adjustment that I'm, like, most familiar with with sync, we will make, is going from these points, estimates, probabilities to,
1:16:36
like, ranges and so on.
1:16:38
Yeah. Yeah. I'll I'll be glad
1:16:42
talk about that. That's right. Then you might say
1:16:45
this sharp
1:16:46
probabilism where you're
1:16:47
supposed to assign a
1:16:50
sharp real number is just psychologically
1:16:52
implausible. It's not something that we could do. Take
1:16:54
again the case of my probability of rain tomorrow. It's
1:16:57
not sharp to infinitely
1:16:59
many decimal places. what
1:17:02
do I say? What's the probability of
1:17:04
rain? Well,
1:17:05
it's zero point six
1:17:07
ish. It's in the region
1:17:09
of zero point six Now the
1:17:11
first move you might make is to say, well,
1:17:13
the probabilities should be intervals,
1:17:15
maybe point five to point seven
1:17:17
in my case. Yeah. But in
1:17:20
a way, that just makes the problem of
1:17:22
precision worse. So now you've got two precise numbers, the endpoints of the interval -- Right. -- point
1:17:25
50000
1:17:28
and point 70000 Yeah. That doesn't seem
1:17:30
to quite And that's the problem. Yeah. I suppose I suppose if
1:17:32
you wanted to defend it, you
1:17:35
would say, well, you've got two two numbers now
1:17:37
that you've chosen, but, like, they're not as important somehow. Yeah. Is that no.
1:17:39
Yeah. Not not look at convinced. Well, you you could say
1:17:41
that or or you could say
1:17:43
these probabilities get determined
1:17:46
by something else. Like, maybe there
1:17:48
are
1:17:48
the judgments that I make that I've
1:17:50
been frozen would say. Yes. And now
1:17:52
you just look at all of
1:17:54
the probability functions that
1:17:56
respect my judgments, and they
1:17:59
will form
1:17:59
a set, probably not just a
1:18:02
singleton set. but
1:18:03
the set itself will have some sharp boundaries and
1:18:05
it could well work out, but that
1:18:07
set has a boundary
1:18:09
point five and point seven and we we
1:18:11
represent my credence with in principle. Yeah.
1:18:14
And, yeah, I I may
1:18:16
not so easily access these
1:18:18
probabilities too. It may be unbadtered
1:18:20
in prospecting as Williamson would say, maybe my
1:18:22
credences are not luminous -- Right. -- to myself.
1:18:24
But I might
1:18:27
still have them and
1:18:28
they they might have these forms like intervals
1:18:30
or or sets. Yeah. So it gets something that's
1:18:35
appealing about saying, you know, the probability of raining camera tomorrow is between
1:18:37
point five and point seven on a in
1:18:39
a practical sense is that it
1:18:41
helps to indicate your level
1:18:44
of uncertainty to other people, whereas if you just say point
1:18:46
it's point six. Mhmm. Well, I mean, first of all, someone laughed at you because it sounds so precise. It
1:18:48
sounds like ridiculous to be to be so
1:18:50
sure. I mean, they're less like they're laughed
1:18:53
you if if you give a range. Yeah. I guess,
1:18:55
actually, it it does potentially indicate something technical, which is like how quickly you
1:18:58
would update your beliefs as
1:19:00
you get new information. If you have like a very wide range and
1:19:02
you're saying, well, a wide range of point estimates would be plausible and reasonable.
1:19:04
And so, you know, as I see the weather
1:19:06
report, I'm gonna, like, shift a lot. really
1:19:10
feel like it's also it's interesting. If you say it's like point six, in some sense, you're making a claim that you're like a hundred percent sure that it's point
1:19:12
six. And then you should say, well, I and you'll just
1:19:14
never change your mind about that even after it rains or
1:19:16
doesn't. I'm
1:19:19
not sure committed to that. I think it's okay.
1:19:22
You've got the point six
1:19:25
credence initially and then you conditionalize, as we
1:19:27
say, you update that, as the evidence comes in, like,
1:19:30
like, maybe you see the
1:19:31
rain for example. So, yeah,
1:19:33
that becomes a new
1:19:35
sharp value of one,
1:19:37
that's that's okay. But I guess it would be a
1:19:39
misunderstanding to
1:19:39
interpret someone who's saying the probability
1:19:42
of it raining in camera tomorrow
1:19:44
is sixty
1:19:46
percent with one hundred percent probability. That's like not
1:19:49
what's being claimed. Yeah. Well, that brings
1:19:51
us to another
1:19:52
issue with another
1:19:54
choice point, by the way, for varieties of
1:19:57
bayesianism, whether you allow
1:19:59
higher order
1:19:59
probabilities, do you have, for example,
1:20:02
credences of credence probabilities or probabilities in
1:20:04
general. And you could say no, that
1:20:05
doesn't make sense or that maybe they
1:20:08
collapse down to
1:20:10
just the first order probabilities
1:20:12
But you could say, no, I
1:20:14
have probabilities about various ways the world could be, including
1:20:18
what
1:20:18
my credence is r. Yeah.
1:20:20
Because that's part of the world. Why not have
1:20:22
intermediate credences for that too? Alright.
1:20:23
So maybe the thing that it feels
1:20:25
more like you're doing when
1:20:27
you introspect and try to, like, give
1:20:29
these ranges to actual questions. Is your inspect you're like you you pull like in
1:20:31
different numbers in your mind and then
1:20:34
see how much the mind
1:20:36
revolves at that number. So if you say,
1:20:38
like, the probability of rain tomorrow is one percent, then you're, like, no, that's, like, that's too crazy. And then you,
1:20:40
like, kind of, have a,
1:20:42
like, level of, like, craziness all
1:20:45
of these different numbers. And then that gives you some
1:20:47
sort of distribution like, numbers plausible. of what you think you ought to
1:20:49
believe, that's kind of maybe what you're measuring.
1:20:51
Yeah. And maybe we
1:20:55
could
1:20:55
compare your credences to say Lotteries. What do
1:20:57
you think is more probable? Rain
1:20:59
tomorrow or I, you
1:21:01
know, say, a hundred
1:21:04
ticket lottery tickets
1:21:04
one to sixty. One of
1:21:06
those is the winner. Yeah, it feels
1:21:09
maybe slightly higher
1:21:09
than that, but now I
1:21:12
make it one
1:21:14
to seventy or maybe slightly lower than that
1:21:16
and maybe I can somehow hone in
1:21:19
on what my probability is. That's a
1:21:21
way of a listing -- Yeah. --
1:21:23
means from myself. Yeah. Okay. So, yeah, a
1:21:25
philosopher's working on any adjustments to how we do these things that could
1:21:27
possibly affect how I reason about uncertain events,
1:21:29
like on a day to day basis,
1:21:31
that might like actually
1:21:34
help me live better. Starting I suppose with the imprecise credences, okay, maybe it's too much
1:21:36
to ask of you
1:21:38
to have the sharp probabilities
1:21:43
and update by conditionalizing all the time. But this
1:21:45
was meant to be a
1:21:47
friendly amendment. Jeffrey calls
1:21:49
it bayesianism with a
1:21:52
human face. And
1:21:53
it seems more psychologically plausible that you've got these
1:21:55
ranges of probabilities, you know, intervals
1:21:56
or sets. So that's
1:21:59
one kind of
1:22:00
humanizing. guess
1:22:02
so so you were saying but but you objected to that or you
1:22:05
offered the objection that well, now you've just chosen
1:22:07
two numbers. You've made it, like, you now
1:22:09
you've got double double the problem. Yeah. Is is that a
1:22:11
good objection interview? Well, I I think I think it has to be taken
1:22:13
seriously, but but now maybe things
1:22:15
start getting worse. 0II
1:22:17
don't want to have this
1:22:19
exact sharp interval from
1:22:21
point five to point seven in
1:22:23
my example. Maybe what I should have is
1:22:26
a
1:22:26
probability distribution over numbers that's not smooth.
1:22:28
It's smooth and
1:22:30
maybe it hits a peak somewhere in the middle near point six and it tapers off towards the edges and
1:22:32
maybe doesn't just stop it sharply
1:22:34
at point five and point seven.
1:22:39
but now we just raise the problem again. So, really,
1:22:41
I've got this sharp exactly
1:22:43
that sharp probability function over the
1:22:45
-- Yeah. -- range of values.
1:22:47
Usually, you're on you're unsure about the
1:22:49
value of the probability that you should give to each probability. And so -- Yeah. -- and and we have
1:22:51
us -- Okay. -- uncertainty all all the
1:22:54
way up. That's right or all the way
1:22:56
down. And
1:22:58
then one reply might be, well, look,
1:23:01
we're
1:23:01
just representing things. We're
1:23:03
providing models. So don't reify
1:23:05
this so much. Don't take this all literally that
1:23:07
this has to be in your head. But
1:23:09
these models, you know, with
1:23:11
various levels of sophistication,
1:23:13
may better or
1:23:15
worse represent what things
1:23:16
in what Going any My brain
1:23:18
is doing some Right. Right. Right. Hey, listeners. Rob here with another definition. We just
1:23:20
use
1:23:23
the term Raytheying something or ratification is
1:23:25
the mistake of imagining or representing that something
1:23:28
that is
1:23:30
really just an abstraction. making the mistake of thinking that it's material
1:23:32
or concrete thing that's that's actually real,
1:23:34
where that isn't. I mean, so
1:23:38
inflow recursion actually not always a problem. Right? You could just say, well, there is just uncertainty
1:23:40
distributions all the way up. Yeah.
1:23:42
Right. And maybe that's actually fine.
1:23:46
And it like caches out. And I suppose you could, in fact,
1:23:48
represent those uncertainties, like, as far as
1:23:50
you like. And so they they could all
1:23:52
cash out to this point estimate. if you like, you
1:23:54
can go one level up and represent uncertainty there, and you can go another one. Yep. At some point, it doesn't feel like it's adding value to you to
1:23:56
do it any further. Yeah. If it
1:23:59
if it does point first, maybe
1:24:02
things stop at a fixed point or
1:24:05
maybe you get a bit more
1:24:07
information at each level you
1:24:09
go back. This is another of the heuristics, by
1:24:11
the way, you asked for
1:24:13
some more infinite regresses, a
1:24:16
technique often used. And often
1:24:18
infinite
1:24:18
regress is thought to be a bad thing and it's a fatal problem for a view if faces such
1:24:21
a regress. But
1:24:24
that's not always
1:24:26
clear, some some notions
1:24:28
seem to be well understood in
1:24:30
terms of infinite hierarchies like that.
1:24:32
Lewis' notion of convention is
1:24:34
like that in in terms
1:24:36
of common knowledge or common belief, which is
1:24:38
a recursive thing
1:24:39
about knowing of each other. what
1:24:43
they believe and so on for higher orders. And we don't
1:24:45
just say, well, that can't be right because we
1:24:47
have an infinite regress. I
1:24:50
mean, especially if the
1:24:51
if the regress converges on something where it's just like it just
1:24:53
becomes the same every time you're just like forever. Oh,
1:24:55
the fixed point. Okay. That's right. And then you
1:24:57
argued it's like, well, that's just fine. Yeah.
1:24:59
That's right. Yeah. Yeah. Okay. Interesting. Yeah. Okay. So
1:25:01
we've got these, like what are these calling the the range of probabilities? What's the term for that?
1:25:03
Yeah. Well, sometimes
1:25:07
more distribution on the probability? So
1:25:09
sometimes we call it the the presenter. Okay. Representative is the set of
1:25:12
probability functions. that
1:25:15
represent your grievances, and they're meant
1:25:17
to be
1:25:18
all of the
1:25:19
precipifications of
1:25:23
your imprecise. credence, but that are faithful to it, but then
1:25:25
fill in in all of the precise ways.
1:25:27
Yeah. So this really does feel
1:25:29
like a friendly amendment to me. I feel
1:25:31
like -- Yeah. -- this is kinda it's the same spirit, and we're just gonna, like, get a little bit
1:25:33
better or, like, represent high levels of uncertainty. All
1:25:35
good. That that's right. And and
1:25:37
again, this takes me back to
1:25:39
earlier when I was saying that I'm
1:25:41
Amazon on Mondays, Wednesdays, Friday and so on. Look, so many things deserve the
1:25:43
name, Amazon,
1:25:44
and
1:25:48
I I shouldn't really say that I've
1:25:50
jumped ship on on Tuesdays and today, Saturday just because I
1:25:52
I make certain choices in
1:25:54
that tree of choice points. Yeah.
1:25:57
Yeah. Yeah. It makes sense. Are
1:25:59
there any more radical departures? Oh, I suppose, are there people who are
1:25:59
probability nihilists? yes,
1:26:05
I suppose you could say that all there is is
1:26:07
just these all or nothing states for start when
1:26:10
it comes to credences, beliefs, as
1:26:12
we might say. Look,
1:26:13
there's I mean, some of the kind of
1:26:15
people
1:26:15
who who are just
1:26:17
like, look, there is what is actual
1:26:19
probability else probability zero. Yep. And that's kind of all that there is
1:26:22
to say about probability. And and you should believe the thing
1:26:24
that will happen, another thing that is true.
1:26:26
Yeah. You shouldn't believe the thing that is
1:26:28
false. and that's kind
1:26:30
of the end of the story. And that
1:26:32
just shows you that objective norms are sometimes tough to
1:26:34
live up to. Yeah. Right. Not not easy.
1:26:37
your your ought to be amnescient. So -- Yeah. -- it's
1:26:39
it's tough to be amnescient, but so that that's not the fault of the
1:26:41
norm. It just says the the norm
1:26:43
is is hard to to
1:26:47
to live up to. Well, that
1:26:48
does seem a little extreme, I
1:26:50
guess. It's certainly not
1:26:52
it's it's it's unhelpful even
1:26:54
if it's it's not giving you good advice, for example. Okay.
1:26:56
So that is quite a lot on
1:26:58
probability. I guess, let's move down the
1:27:00
track now to, I guess, an application
1:27:03
of probability estimates, which is expected
1:27:05
value. Yeah. We've got a whole cluster of questions around that. I guess to start
1:27:07
though, what is expected value? It's it's it's I turned that we throw around a
1:27:09
lot, but I think it's actually like
1:27:11
basically not used among
1:27:14
the general how I was like, it it is actually
1:27:16
quite technical in a way. Yeah. That's right.
1:27:18
And they'll quickly turn into a discussion
1:27:21
of expected utility. But Mhmm. Alright. It comes
1:27:23
up to especially in decision theory,
1:27:26
expected utility theory. And
1:27:28
you
1:27:28
have a choice among certain
1:27:30
options. What should you do?
1:27:33
And
1:27:33
let's assume that the world could be in various ways. They're very states of
1:27:35
the world, and you don't have control of of
1:27:37
what they are, but you
1:27:39
assign probabilities to the
1:27:42
various states. Yeah. And the
1:27:44
combination of a particular
1:27:46
action of yours and the
1:27:48
state of the world together
1:27:50
that
1:27:51
determines an outcome and you can value
1:27:53
the outcomes more or less. Yeah. And we
1:27:55
could put numbers, which measure
1:27:57
how much you
1:28:00
value Now
1:28:00
the expected value is a
1:28:01
weighted average of these values where the
1:28:04
weights are
1:28:06
the probabilities. So that turns out to be a
1:28:09
sum of products. You
1:28:11
take probability multiplied by
1:28:14
the value and add across all
1:28:16
of the states. And that weighted
1:28:18
average
1:28:18
is the expected value. And
1:28:22
now think of that as like a figure of merit.
1:28:24
That's how choice worthy
1:28:25
each of your actions. Hey,
1:28:28
listen. One quick definition.
1:28:30
Choice
1:28:30
worthy knows. is this philosophy jargon term where choiceworthiness
1:28:32
corresponds to the strength of the
1:28:34
reasons for choosing a given option.
1:28:37
How good choices is
1:28:39
this for us? Okay. back to the show. She
1:28:41
was Trust where he was another one of these. So he said no
1:28:43
no no no no more non philosophy. But would that That's
1:28:46
exactly where I've been. Yeah.
1:28:49
And and now you should
1:28:51
maximize that quantity, do whatever action or actions because
1:28:53
maybe
1:28:54
it's not one. We'll
1:28:56
they will yield
1:28:58
the highest value along that score. Cool. So I I guess, like yeah. To to make that concrete
1:29:00
really
1:29:01
simple, if you
1:29:03
had a bat where
1:29:06
you got like one dollar if coin heads, comes it's fair
1:29:09
coin, then the
1:29:12
expected value I
1:29:14
guess, in dollars in this case is one dollar and
1:29:16
a half or 111 dollar fifty. That's right.
1:29:18
And of course, so normally, we'd we'd rather talk
1:29:20
about utility or well-being or something because we don't value
1:29:23
dollars the same as as as especially as the amounts become larger or maybe you
1:29:25
don't value dollars at all. So utility is the
1:29:27
thing where it ultimately caches out. That
1:29:29
that's right. And very soon, we'll
1:29:32
we'll talk about some motivations, I think, for for
1:29:34
why we should
1:29:34
really shift to, you know, payoffs just measured in terms
1:29:36
of dollars -- Yeah. -- to
1:29:38
to this other thing, utility. Yeah.
1:29:42
So,
1:29:42
yeah, expected value is just
1:29:44
super important to the project of effective altruism
1:29:46
and doing good because like any project you
1:29:48
take on is gonna have a very wide
1:29:50
range of possible outcomes. At least if you're doing anything remotely interesting -- Yes. -- that, you know, it could include, like, bad outcomes, potentially
1:29:53
lots of neutral outcomes,
1:29:55
possibly, like, things go wells.
1:29:58
You have a really good outcome. And you're just gonna
1:30:00
be absolutely screwed, absolutely paralyzed. If you
1:30:03
don't have a way of making
1:30:05
decisions under that uncertainty where you're able to
1:30:07
weigh up like, what should I do given that
1:30:09
things might go badly or might go well or
1:30:11
might might be neutral? And expected value provides this
1:30:13
extremely natural way of weighing those things up where
1:30:16
it says, well, something is twice as likely. It's twice
1:30:18
as important. Something that's good. If it's twice as likely, it's twice as good and that should get, like, twice as much weight in
1:30:22
your in your decision. So in every in every day situations like where they go,
1:30:25
you know, what to buy at the shops, like, what
1:30:27
to watch on Netflix, the kind of expected
1:30:29
value approach produces like extremely natural
1:30:31
answers that's in sensible basically
1:30:33
everyone, at least in principle, even though of course, we never
1:30:35
actually well, we never consciously almost done consciously calculating
1:30:39
expected value, but actually heard from
1:30:41
from neuroscientists that they've likely doing research on like how we make decisions. And apparently, there is like a process in
1:30:43
the brain that
1:30:48
effectively represents expected value, basically. Well, like, particular neurons
1:30:50
will fire based on, like, their expected reward. And then, like, the one that fires most frequently
1:30:52
or, like, or harder in some way,
1:30:54
like, basically, ends up winning the decision.
1:30:58
some like choice point in the brain. Anyway, I probably
1:31:00
just said something that's completely wrong. Well, no.
1:31:02
In fact, I think it goes even further. I
1:31:04
think it's not just people, I think, nay. Even
1:31:06
bees are Absolutely. thought to do this too. I mean, it kind of maybe it maybe makes sense that
1:31:08
it has to be this way
1:31:10
because evolution's gonna push us towards
1:31:14
doing things that, like, get us the most food or whatever. And, yeah,
1:31:17
obviously, when you have, like, different options on the table,
1:31:19
yeah, the brain's gotta choose the
1:31:21
bigger one. Yep. And you will be selected again if
1:31:23
you don't if you don't make the right choices that roughly -- Yeah. --
1:31:25
that maximized
1:31:25
expected value or
1:31:26
close to it. Yeah. So maybe
1:31:29
we we can say, like, expected value is, like,
1:31:31
probably deeply embedded in how we're kind of wired in in
1:31:33
some sense, even though we're not actually doing the math most of
1:31:35
the time, it's embedded in the like, in our
1:31:37
instincts about what what risks to take and which
1:31:39
one's not to. But, yeah, philosophers have looked at a
1:31:41
whole bunch of much stranger situations to see if they can make expected value break, whether they
1:31:44
can, like, make it stop working
1:31:46
or at least, like, arguably, stop
1:31:48
working. I guess one of
1:31:50
these that you're really into is the Saint Petersburg Paradox,
1:31:52
which some people have heard of, but other people won't have. Yeah. Can
1:31:54
you can you layout what the Saint Petersburg Paradox is? Absolutely.
1:31:58
A fair coin will be
1:31:59
tossed repeatedly until
1:32:00
it lands heads for the first time,
1:32:03
and you will get
1:32:05
escalating rewards the longer it
1:32:07
takes, the better for you, as
1:32:08
follows. If the
1:32:10
coin lands heads immediately,
1:32:13
you'll get two dollars. If
1:32:15
it takes two two trials for
1:32:17
the first head, so it's tail and
1:32:19
then heads, you'll get four dollars.
1:32:21
If it takes three trials,
1:32:23
you'll get eight dollars In general, if it
1:32:25
takes N trials for the first heads, you'll get two to
1:32:27
the N dollars. Okay?
1:32:29
And
1:32:30
so on. All right. How good is
1:32:32
that at least in
1:32:34
terms of expected dollar amount?
1:32:36
Yes. Well, let's
1:32:39
do the calculation. with probability half, you
1:32:42
get two dollars and multiply those. Half times two is one.
1:32:44
With
1:32:45
probability quarter,
1:32:46
you get four dollars
1:32:50
quarter
1:32:50
times four is
1:32:52
one. And now we keep adding terms
1:32:54
like that. It's one plus one plus
1:32:56
one, one over two to the
1:32:58
n times two to the is one
1:33:00
forever.
1:33:01
One plus one plus one added up
1:33:03
forever is infinity. So the
1:33:06
expected value of this
1:33:08
game in terms of dollar amount
1:33:10
is infinity. So prima facie, it looks like you should be
1:33:12
prepared to
1:33:14
pay any finite amount to
1:33:17
play the game just once -- Mhmm. -- and you should think you're
1:33:19
getting a fantastic deal. You're just just a
1:33:24
finite amount, but that's very
1:33:26
unintuitive because very probably you'll
1:33:27
make a rather
1:33:28
very probably you make a rather
1:33:30
small amount small amount for example,
1:33:32
most people wouldn't pay a hundred dollars
1:33:34
to to play this game. It seems because very probably they
1:33:39
will make a lot less than Yeah. And
1:33:40
therein lies a paradox. They're various
1:33:42
paradoxical aspects of this game. And
1:33:44
it is paradoxical that every
1:33:47
possible outcome is finite. Okay?
1:33:50
You know you'll get a finite outcome, finite payoff. And yes, you value
1:33:52
it in instantly.
1:33:55
How did that happen? Yeah.
1:33:58
Yeah. Something's gone wrong
1:33:59
here, it seems. Yeah.
1:34:02
It reminds me, Garrison
1:34:04
Kaeluses, is it
1:34:05
like Woebe Gornham, all of the
1:34:07
children are above average. Yeah. And yeah.
1:34:09
So in in this case, all
1:34:11
of the payoffs are below average, so
1:34:13
to speak. I see. Right. Right. Okay.
1:34:15
a variety of the paradox that, like, most numbers are small or, like actually, almost all
1:34:18
numbers are small. Right? Right. Because, like, no
1:34:20
matter where the no matter what cutoff
1:34:22
you choose for, like, numbers being small below
1:34:24
that, most numbers aren't larger
1:34:26
than that. Okay. So sympathetic paradox. Yeah. Seems seems weird. Seems like we should evaluate this thing internally,
1:34:28
but we do not. What what like,
1:34:30
how how can we fix it? Absolutely.
1:34:35
just look at the the moving parts in expected
1:34:37
value theory,
1:34:38
and we could tweak
1:34:40
each of them. So I think
1:34:42
there are five different ways of replying
1:34:44
to the Saint Petersburg Paradox. We
1:34:46
could do something to the probabilities. So don't
1:34:48
just stick in
1:34:49
the probabilities as they
1:34:52
were before. maybe some function of
1:34:54
them. And Lara Bushak has a very nice theory, a risk a risk weighted utility
1:34:56
theory, which will
1:34:59
tweak the probabilities. Now
1:35:02
look at the payoffs, look at these rewards, which were dollar Well,
1:35:04
maybe
1:35:04
we should
1:35:06
tweak that, and this
1:35:08
was Bernoulli
1:35:11
solution, actually, that how much
1:35:13
you value things is not just
1:35:15
a linear function
1:35:15
in the in this
1:35:18
case, dollar amount. there's what we call
1:35:20
diminishing marginal utility. The richer you are,
1:35:22
the less you know, we'll value some
1:35:24
incremental amount,
1:35:26
like an extra dollar Okay? And in fact, he thought that
1:35:28
how much you really value
1:35:29
the money goes more like by
1:35:31
the logarithm of
1:35:34
the dollar amount rather than by the face value
1:35:36
-- Mhmm. -- the dollars themselves. And
1:35:38
it turns out if you replace the
1:35:40
dollar amounts by the
1:35:43
logarithm of their amounts, then
1:35:44
you get a convergent series, you get a
1:35:46
sum that is finite and Yeah. Okay. Then we can
1:35:48
tweak
1:35:49
the the formula
1:35:50
the combination rule So
1:35:54
previously, was this weighted average? Or
1:35:56
maybe we could do something to
1:35:58
that. Well, there
1:35:59
would be
1:36:00
more radical departures like maximin.
1:36:02
It's got nothing to
1:36:04
do with expected value, but
1:36:06
maybe that's
1:36:07
the decision rule we
1:36:10
should follow. An alternative that I like
1:36:12
was introduced by an
1:36:14
economist too and studied and
1:36:17
developed further by Chris
1:36:19
Bottomley. former AIU student. Yeah. And a former student
1:36:21
of mine, Tim Williamson, not
1:36:23
not the famous
1:36:27
epistemologist at Oxford, but funnily the younger
1:36:28
Tim Williams, and he's also with Oxford
1:36:30
down and I'm sure he'll be famous
1:36:34
soon too. And they have been working on
1:36:37
a theory called weighted linear
1:36:39
utility theory. Oh, yeah. So what
1:36:41
are
1:36:41
what are the other options on
1:36:43
the Now hold fixed
1:36:45
the combination rule, which was this
1:36:47
sum of products and
1:36:50
have some different operation and what you do to previously,
1:36:52
it was maximize that
1:36:54
thing,
1:36:55
that expectation.
1:36:56
Well, maybe we
1:36:58
don't have to maximize it.
1:37:00
it Well, an
1:37:01
alternative would be to minimize it. That would be that would be a
1:37:03
pretty crazy theory. Yeah. But more plausible would we
1:37:07
we satisfy So it's
1:37:09
it's
1:37:09
just good enough to
1:37:10
to get a sufficiently high expected
1:37:13
value. You don't have
1:37:15
to literally maximize it.
1:37:17
So, I mean, I can
1:37:19
see the appeal of stratospising. So, stratospising
1:37:21
would be like maximizing the probability
1:37:23
of being above some particular
1:37:25
value. Is that right? Well, may not necessarily maximize
1:37:27
it, but just provided your maybe
1:37:29
sufficiently high up in the
1:37:32
ordering of your
1:37:34
act actions. Oh, that's good enough. I
1:37:36
said then you stop valuing any improvement beyond
1:37:38
that. Yeah. Something something like that. So is
1:37:40
this the same as kind of having
1:37:42
bounded utility in a way? Well, So so bounded
1:37:44
utility would be, say, well, my well-being can't go above some particular
1:37:47
level. So any any money beyond that would
1:37:50
be worthless. So I was thinking it was different that maybe
1:37:52
you were allowing unbounded utility,
1:37:54
but
1:37:55
you're not insisting
1:37:56
on maximizing this
1:37:58
overall quantity. You
1:37:59
just say, you know, near enough
1:38:02
is is good enough. Benson and Hedges cigarettes used to have this slogan
1:38:04
when
1:38:07
only the best will do.
1:38:09
And Satisfising says, oh, no, you don't need the best.
1:38:11
And in fact, voltaire
1:38:15
said, that the best is the enemy of the good, and maybe you don't
1:38:17
always have to strive for the best. Yeah. So this
1:38:19
will get us out of the
1:38:21
Saint Petersburg Paradox because at
1:38:24
some point, rather than just adding plus one,
1:38:26
plus one, plus one, you'd say, well, beyond this point, I would be above my satisfying level. And now I don't
1:38:28
value any any additional wins. Yeah.
1:38:30
And I, you know, if if I
1:38:34
instead
1:38:34
just received thirty,
1:38:36
dollars maybe that would be good
1:38:39
enough. And I
1:38:40
don't have to play the
1:38:42
game. Okay. So changing the decision rule. What what other options are there? And I've left for last, the one that I've
1:38:45
actually argued for
1:38:48
in print So I'd
1:38:50
better say something for it, which is biting the bullet. Yeah. And how do you do that? Well,
1:38:52
maybe it's not so
1:38:54
crazy after all to value
1:38:58
for
1:38:58
some petersburg game infinitely. And here's an
1:39:01
argument
1:39:01
for that. By the way, I've made
1:39:03
a switch to expected utility
1:39:05
theory where I'm now replacing the dollar
1:39:07
amounts, the values with how much you value
1:39:09
them in the utility sense. And that's
1:39:11
the quantity
1:39:12
and expected utility
1:39:15
that you're maximizing. Well, Let's
1:39:16
agree that expected utility theory is plausible
1:39:19
at least for the finite cases and somehow it
1:39:21
went wrong, it seems in
1:39:23
the infinite case. Well,
1:39:26
did it. Imagine
1:39:27
various truncations of the
1:39:30
Saint Petersburg game. For example,
1:39:32
if the first heads does
1:39:34
not happen by the tenth toss. We call the game off and we go home. That's the
1:39:38
end
1:39:39
of the game. Well, it
1:39:41
seems that the value of that game is ten.
1:39:43
Alright? Yeah. Now truncated the
1:39:44
eleventh toss. It
1:39:45
seems the value
1:39:47
of that's eleven. truncated
1:39:50
the twelfth toss, the values twelve
1:39:52
and so on. Now the Saint Petersburg
1:39:54
game is strictly better all of these.
1:39:56
This is so called dominance reasoning. Yeah.
1:39:59
Come what
1:39:59
may the Saint Petersburg game
1:40:01
is at least as good as each
1:40:03
of these
1:40:04
and with
1:40:04
some probability
1:40:07
it's genuinely better. Okay. So the Saint Petersburg
1:40:09
game, it seems it's better than ten. It's better than eleven. It's better than
1:40:11
twelve. Don't don't don't keep going. Yeah.
1:40:14
It's
1:40:14
better than all. Better than any inter
1:40:17
better than any any interject. So that seems what's left.
1:40:19
A reason to value it infinitely. And that's
1:40:21
that's a way of biting
1:40:24
the bullet. but
1:40:26
there's a
1:40:27
revenge problem. Okay. Maybe
1:40:29
we've solved the original Saint
1:40:32
Petersburg problem. But
1:40:34
there's a Saint Petersburg like game that we can easily introduce, which
1:40:37
will thwart even
1:40:40
this proposal Now
1:40:42
let the payoffs go up not exponentially as they did in Saint Petersburg.
1:40:45
Let them
1:40:48
go up super exponentially. The payoffs
1:40:50
are not two to the end. They're two to the two to the end. Now
1:40:52
take logs as Bernoulli
1:40:55
would have us do. take
1:40:58
the
1:40:58
log of two to the two to the end and you get to the end,
1:41:01
you get you're right back where you
1:41:03
started. Yeah. And so now it's
1:41:05
a sequence again of, like, adding up yeah, one plus one
1:41:07
plus one plus one utility this time. Yes. Yeah.
1:41:09
Yeah. That that's right. So it's
1:41:11
not enough just to have diminishing
1:41:14
marginal utility. It seems really what
1:41:16
this solution is asking
1:41:17
for is bounded utility. So I can't just keep
1:41:19
on ramping up the payoff
1:41:22
sufficiently to get arbitrarily high
1:41:26
utility. So that's the solution
1:41:28
now to to be discussed. Should
1:41:30
utility be bounded and arrow
1:41:33
and hard and and almond and
1:41:35
and various luminaries
1:41:36
have advocated this solution. And
1:41:38
it's
1:41:39
actually implicit in various
1:41:41
theories too that really utilities bounded.
1:41:43
Yeah. So this seems like a super compelling response.
1:41:45
Right? So that we don't value each dollar
1:41:47
equally as much. Yeah. This, like,
1:41:50
captures, I think, the intuition for
1:41:52
why people don't wanna pay so much because they're
1:41:54
like, well, even if there's some, like, infinitesimal probability of winning, like, an enormous amount of money,
1:41:56
I I just don't
1:41:59
value the
1:41:59
money enough to make it to make it worth it because
1:42:02
but but then I'm so unfathomably rich, more money is, like, not worth very much to me. And that
1:42:06
may be true of humans and we probably have saturation point and
1:42:08
there's only so much we
1:42:11
can value things given
1:42:13
our finite heads
1:42:15
and so on. but there's still
1:42:17
I think an in principle problem. You might ask why should utility
1:42:19
be bounded in various
1:42:24
other quantities are not bounded. Like,
1:42:26
length is not bounded. Volume is not bounded time.
1:42:28
Space time curvature.
1:42:30
Mhmm.
1:42:30
Various things are not
1:42:33
bounded.
1:42:33
They're unbounded. And why should utility
1:42:35
be? And normally, when you do
1:42:37
have a bounded
1:42:39
quantity, you can say why
1:42:41
it is and you can say what the
1:42:44
bound is, like think of say angle. And
1:42:46
if you think of it one way angle
1:42:48
is bounded
1:42:49
by zero to three hundred and sixty degrees, and
1:42:51
it's easy to explain that.
1:42:52
greens
1:42:53
it's easy to explain that probabilities
1:42:55
bounded. Yeah. to
1:42:57
one. So value of one, bottom value of zero. Not so easy to say it
1:42:59
in the case of utility. And the
1:43:03
problem gets worse if
1:43:06
we make utility depend
1:43:08
on one of these other unbounded quantities
1:43:10
in a what seems to be
1:43:12
an unbounded way, his appointed
1:43:14
example, I I had a student, Celtic, who hated
1:43:16
i had
1:43:17
his hometown
1:43:20
so much that
1:43:23
he
1:43:23
said that for him, utility was
1:43:25
distance from his hometown.
1:43:27
The further away, the
1:43:29
better. And then length being unbounded
1:43:31
than gave him unbounded utility. He was he
1:43:33
was joking, but you see the point
1:43:36
that -- Yeah. -- in principle,
1:43:38
you you there could be these
1:43:40
relationships between utility and some unbounded
1:43:42
quantity that would yield unbounded utility. Also, in
1:43:45
this case,
1:43:45
we're talking
1:43:46
about the rationality of individual
1:43:50
action, you know, what should
1:43:52
you do? But fairly
1:43:53
soon expected utility
1:43:54
like reasoning applies to say population
1:43:59
ethics. And you can imagine a Saint
1:43:59
Petersburg like Paradox where
1:44:02
the
1:44:02
Lokaiya
1:44:03
value different people in a population
1:44:06
-- Mhmm. -- and we we can
1:44:08
run a
1:44:09
paradox for that. And
1:44:11
now the analogical
1:44:11
replies don't seem so
1:44:14
good diminishing marginal value of
1:44:18
people doesn't sound so good. You
1:44:20
want each new person to
1:44:22
count equally.
1:44:23
Yeah. Yeah. And you certainly don't
1:44:25
wanna bound. It seems Yeah. The the total
1:44:27
value across people as as the population
1:44:30
grows.
1:44:30
Yeah. I mean, to start with
1:44:32
that, has the counterintuitive
1:44:34
conclusion that, like, how valuable it might be to create an extra person
1:44:36
on Earth might depend on, like, how many aliens there
1:44:38
are or something like that? Or is it like,
1:44:41
you have to know how many beings there
1:44:43
are in the entire universe in order to tell how
1:44:45
much it's it's how good it is to add an extra allowance. Yeah.
1:44:47
Seems odd. That's that's it. So for all
1:44:50
these reasons, I I don't
1:44:51
like the bounding utility solution. That's why I
1:44:53
I look elsewhere -- Okay. --
1:44:55
and even bite
1:44:58
the ultimate bullet at least in
1:44:59
print that's maybe that verdict from expected
1:45:01
utility theory is
1:45:02
not so bad. Yeah.
1:45:05
Okay. So tend towards just like buying the bullet
1:45:07
and saying, well, we should maybe just accept
1:45:09
this. But it doesn't seem like in an
1:45:12
actual practical situation. Like,
1:45:14
would you pay an infinite amount? Would you
1:45:16
pay any any finite amount if someone actually came up
1:45:18
and offered this to you? Like, it doesn't seem very
1:45:20
action guiding to bite the bullet in a sense.
1:45:22
Yeah. And yet, I found something attractive about that dominance --
1:45:25
Yeah. -- reasoning that was iterated. It makes
1:45:27
sense. Yeah. I see the argument.
1:45:30
Yeah. Yeah. So So I guess where did it leave us? So if if if you're willing
1:45:32
to bite the bullet on something like that. Uh-huh. Yeah.
1:45:34
So there's one thing is you might say,
1:45:36
well, in any practical situation
1:45:39
where someone came up and offered you this game,
1:45:41
you wouldn't believe them. No. No. It's a pascara slugging, I think. Yeah. So
1:45:43
there's such to approach this, like, other thing, which are some some
1:45:45
people in the audience who'll be familiar with this
1:45:47
pascara slugging situation. But
1:45:50
this might be an escape or like a way of reconciling the
1:45:52
bullet with, like, not actually playing the game in practice.
1:45:54
Is that who would say, well, sure. If I
1:45:56
was, like, a percent sure, that I was playing
1:45:58
the Saint Peter's Boat Paradox game -- Yeah. -- with a particular kind of setup. And, yes, I would. Valley were infinite dollars,
1:46:04
but because I don't believe that I ever
1:46:06
am because it's like not even possible. I don't even think it's possible in the universe to someone to deliver these
1:46:08
gains, then I'm not actually compelled
1:46:10
in a real situation to take it.
1:46:14
Yeah.
1:46:14
Let me make the Saint Petersburg
1:46:16
game
1:46:16
more paradoxically. So in
1:46:19
the
1:46:19
first telling of
1:46:20
the story, you just
1:46:23
genuinely believe this offer and you you think
1:46:25
you
1:46:25
should take
1:46:26
it at any price. Richard Jeffrey
1:46:28
famously says anyone
1:46:29
who offers you the
1:46:31
Saint Petersburg game is
1:46:34
a liar -- Mhmm. -- because they're
1:46:36
pretending to have an indefinitely large bank account. And that
1:46:38
I suppose that's true. That's where
1:46:39
we we would think someone
1:46:42
who offers you this game as a liar. But
1:46:44
I don't think that gets us
1:46:46
out of the paradox so so
1:46:49
easily. Okay? because
1:46:49
the paradox hits you with its full force as long as you assign just some
1:46:52
positive
1:46:53
probability
1:46:56
to
1:46:56
the offer being genuine. Okay. Suppose
1:46:59
someone comes up to you and
1:47:00
offers you
1:47:02
the Saint Petersburg game and
1:47:05
you give probability point 00000000001
1:47:07
to
1:47:10
them telling the truth.
1:47:12
Okay? Yeah. As we would ordinarily
1:47:14
say, you think they're a liar. You give overwhelming probability that they're
1:47:18
not telling
1:47:19
the truth. But
1:47:21
now think about that point 00001 The paradox
1:47:23
hits you with its full force because you now
1:47:25
you multiply
1:47:27
that by infinity And
1:47:29
even that extremely low probability of the Saint Petersburg
1:47:31
game has expected infinite
1:47:34
expected value for you
1:47:37
given
1:47:37
your assigning positive credence however small. One in the Googleplex -- Yeah. -- will
1:47:40
still will still keep
1:47:42
the game alive enough that
1:47:46
you should value that prospect infinitely.
1:47:48
So it seems the only way
1:47:50
to escape the paradox is
1:47:53
to genuinely give probabilities
1:47:55
zero to it. This brings us back to
1:47:58
our earlier discussion that probability zero should be reserved for impossibility and
1:47:59
that anything that's
1:48:03
possible should get positive
1:48:05
probability, perhaps one in the Googleplex, but but but something
1:48:07
positive. I guess, some people might be inclined to
1:48:09
say, well,
1:48:12
an infinite pay off or infinite value isn't possible in
1:48:14
the universe as it is, but we don't know that. It it could it may maybe it's impossible,
1:48:18
but that that's right. are you so sure that you
1:48:20
really can zero it out? And now the way
1:48:22
I like to put this, it's like
1:48:25
a dilemma, you
1:48:27
can either be so
1:48:30
to
1:48:30
speak practically irrational or
1:48:33
theoretically irrational. Mhmm. Now
1:48:35
if you assign any
1:48:37
positive probability whatsoever to the Saint Petersburg gamble to the game,
1:48:39
then the infinity
1:48:42
then
1:48:42
clubbers you and you
1:48:45
think
1:48:45
you're already enjoying infinite expected utility right
1:48:48
now. And
1:48:52
that seems practically,
1:48:53
irrational and you'd be prepared to you think
1:48:55
that actually, you'll think
1:48:55
pretty much anything you do
1:48:58
is infinitely good because There's
1:49:01
some prospect that you'll be playing the
1:49:03
Saint Petersburg game at the end. In terms of expected value, it's infinitely
1:49:04
good. That's a
1:49:06
worry on
1:49:07
the practical side. But
1:49:09
now solution, I'm imagining is you you give probabilities zero. But now
1:49:11
the worry is that that's theoretically
1:49:13
irrational because
1:49:14
your evidence is not
1:49:18
so decisive. I mean, can you really with
1:49:21
such
1:49:21
confidence -- Yeah. -- rule
1:49:23
it out. Namely, it it
1:49:25
has the same probability as
1:49:27
a genuine impossibility Yeah. It
1:49:28
seems well. It's a contingent matter. Yeah. You've you've
1:49:31
got all this evidence Inceivable. Yeah. It's conceivable. And it's
1:49:33
you've got all this
1:49:36
evidence against the
1:49:37
Saint Petersburg game, but
1:49:39
not so decisive it seems that it just absolutely rules it out. Yeah.
1:49:42
It feels like structurally
1:49:44
improve like structurally this
1:49:46
whole, like, this path down to
1:49:48
solving some Petersburg Paradox has become extremely similar
1:49:50
to Pascal's wager. Absolutely. Oh, okay. Right.
1:49:52
Right. Yeah. So I guess, Many people will
1:49:54
have heard of Pascal's wager at the same stage, but maybe
1:49:56
do you wanna, yeah, remember, repeat it, give everyone a refresher?
1:49:59
Delighted too, because I've thought
1:49:59
a lot
1:50:02
about Pascal's wager. This is Pascale's argument for
1:50:03
why you should believe in God or
1:50:06
cultivate belief in God. And just
1:50:08
to locate it
1:50:11
historically, we should contrast Pascale's
1:50:13
wager to predecessors, which
1:50:14
purported to establish the existence
1:50:17
of god, prove the
1:50:19
existence of god, And
1:50:21
I'm thinking of things like
1:50:24
the ontological argument, St. Dan Selm and
1:50:26
Descartes had one.
1:50:26
Thomas Aquinas had five ways
1:50:28
aquinas had five ways Daycart
1:50:30
had a cosmological argument and
1:50:33
there the conclusion was,
1:50:35
yeah, God exists. Pascal won't
1:50:37
have a bar of this. He
1:50:39
says, reason can decide nothing here.
1:50:42
Okay? You can't just buy some clever
1:50:44
proof established the existence
1:50:46
of God. But he turned his attention to the attitude we should have to, the
1:50:48
existence of God. Should you
1:50:50
believe in God or not?
1:50:54
in particular. And
1:50:56
that's now a decision problem. And that that's
1:50:58
why it's relevant to our discussion about
1:51:01
decision theory. And
1:51:02
he argued that you should believe
1:51:04
in God or at least wager for God
1:51:06
as he said. Think of that as
1:51:09
cultivate belief in God. short
1:51:10
version of the argument is because it's
1:51:12
the best bet. And
1:51:14
in fact, hacking writes that
1:51:16
this was the first ever exercise
1:51:19
of decision theory -- Alright. -- which is
1:51:21
interesting -- Uh-huh. -- because of
1:51:23
all the cases, this is such a
1:51:25
problematic -- Right. -- case for this
1:51:27
in theory. I see. Which is ironic. Yeah. opening with, like, the with
1:51:29
the paradox. Basically, we're opening with the
1:51:31
paradox. Anyway, Lifetosh
1:51:34
said that every research program
1:51:36
was born, refuted, and maybe you you could
1:51:38
say that of this very case, decision theory, was born, refuted with
1:51:41
a problematic case. Here's
1:51:43
how the argument goes. There
1:51:46
are
1:51:46
two ways the world could be.
1:51:48
God exists. God does not exist. Two things
1:51:50
you could choose to do. Believe in God,
1:51:54
or not believe or as Pascal says,
1:51:56
wager for God
1:51:57
and wager against God.
1:51:59
And here are the payoffs. If
1:52:02
God exists and you infinite
1:52:05
reward,
1:52:08
infinite utility, and
1:52:10
infinity of infinitely happy lives as
1:52:12
Pascal says. And now
1:52:13
in every other case, God
1:52:15
does not exist or you don't
1:52:17
Believe in God, you get
1:52:20
some finite payoff. And
1:52:21
there's some controversy about the
1:52:24
case where God
1:52:25
does exist and you don't believe maybe
1:52:28
you get negative infinity, maybe you have infinite damnation.
1:52:30
But I think actually detection What's zero, I
1:52:32
guess, So
1:52:34
is it it's sufficient to put zero there?
1:52:36
Well, isn't it? Oh, no. III
1:52:38
think Pascal himself in the text
1:52:41
is telling us that really that's a
1:52:43
finite term that wager against God or don't
1:52:44
believe in God. God exists
1:52:47
as
1:52:47
only finite in in
1:52:49
utility, not not infinitely
1:52:52
bad.
1:52:52
Okay. So
1:52:52
that's the first premise. That's
1:52:55
the decision matrix as we say. Those are the utilities, Infinity,
1:52:59
the war for believe
1:53:00
in God, God exists or wager
1:53:02
for God, God exists, finite everywhere else. Then
1:53:06
then the premise about the probability. The probability
1:53:08
that God exists should be positive.
1:53:10
So your credence, as we would
1:53:12
say, should be positive.
1:53:14
Non zero. Non zero?
1:53:17
as
1:53:17
we might say, it's possible that God exists.
1:53:19
So you should respect that by giving a positive probability. Okay? This theme
1:53:22
keeps coming
1:53:23
up. And now
1:53:25
PASCAL does what we
1:53:27
recognize as an expected
1:53:29
utility calculation and just
1:53:31
do the sum,
1:53:33
you've got infinity
1:53:34
times some positive probability, plus some finite stuff,
1:53:36
add
1:53:37
it
1:53:38
up, you get infinity.
1:53:40
up you get infinity So
1:53:42
it looks like wagering for
1:53:44
god, believing in god has infinite
1:53:46
expected utility. Mhmm. And wagering against god
1:53:49
not believing in God. The expected value was some finite stuff plus some
1:53:51
finite stuff which is finite. Infinity
1:53:56
beats finite. Therefore,
1:53:58
you should believe in
1:53:59
God. That's
1:54:00
Pascal's wager.
1:54:02
I see. So is this
1:54:05
actually analogous to the St. Petersburg paradox
1:54:07
as as you're biting the bullet on it? Or are are are there any differences here there? It's
1:54:12
it's structurally similar
1:54:14
in that infinite utility is is what you get in in
1:54:15
the punch line. Yeah. Notice we got to it
1:54:18
in a different way in Pascal's
1:54:20
wager from
1:54:23
Saint Petersburg. In Saint Petersburg, we
1:54:24
were adding finite terms. Every
1:54:27
possible payoff was finite, but
1:54:29
just because of
1:54:30
the the way they are summed,
1:54:33
you get infinity in Pascal's wager.
1:54:34
You you you just to get a this single hit of infinity, it's
1:54:39
this this one possible outcome that
1:54:42
just gets you the infinite utility in in one shot. Okay? So that's, I guess, structural
1:54:44
difference, but but
1:54:45
I think there are
1:54:47
other parallels. Yeah. Okay.
1:54:50
And I guess in both cases, the infinity is like
1:54:52
really messing us up here because even like a
1:54:54
tiny possibility of an infinity just like
1:54:57
swamps everything else. and -- Absolutely. -- unless
1:54:59
you unless you literally assign zero probability to it then
1:55:01
-- Exactly. -- it's basically the only thing that matters in
1:55:03
the entire process. Yeah. Yeah. Yeah. Now,
1:55:06
let's let's do it. A strict
1:55:09
atheist who gives probability zero
1:55:11
to God's existence, I
1:55:13
guess, is not going to be moved
1:55:15
by Pascal's wager, but that may seem overly
1:55:18
dogmatic, you know, surely you should give at
1:55:20
least
1:55:20
some probability to
1:55:22
God's existence may be extremely small. As as you might say, maybe one in the Googleplex, but not
1:55:28
that
1:55:28
small. And as soon as you
1:55:30
give it positive probability, you toast. You toast. Or you you have to or you have to -- Yeah.
1:55:33
-- wager for
1:55:36
God. Yeah. But
1:55:37
I think there is now a
1:55:39
a revenge problem for Pascal. Okay. Yeah. As
1:55:39
I like
1:55:43
to point out, Pascal's
1:55:44
wager is invalid. Let's grant him
1:55:46
his premises. So we grant him the
1:55:49
matrix of utilities, as
1:55:51
I said before. grant
1:55:54
him you should get positive
1:55:56
probability to God's existence. Grant him
1:55:58
that you should maximize expected
1:55:59
utility. Fine.
1:56:01
fine does
1:56:02
not follow that you should wager
1:56:04
for God by those lights.
1:56:06
Why not? What's that?
1:56:08
Yeah. He's an
1:56:10
alternative strategy. and Pascal's definitely not recommending this.
1:56:12
Toss a coin,
1:56:13
heads you believe
1:56:15
in God tells
1:56:17
you don't. k?
1:56:18
Yeah. What's the expected utility
1:56:20
by Pascal's
1:56:21
lights for that? But with
1:56:23
probability half the Queensland
1:56:25
heads, and then you get
1:56:27
the infinite expectation that Pascal was talking
1:56:30
about. With probability half,
1:56:30
you get some finite expectation. the
1:56:35
expectation of
1:56:35
this mixed strategy, as we say, is
1:56:38
still Infinity. So
1:56:40
by the lights of expected utility theory,
1:56:42
this is equally good. This is another
1:56:44
way get infinite expected value.
1:56:46
Question on his own baton. Well, that's it. That's right. And and I've just started. Okay.
1:56:48
Yeah. Well, I was gonna say, won't you
1:56:50
have a problem? It's like, okay. So
1:56:54
flip the coin and you get whatever result says that you shouldn't believe in
1:56:56
God. Yeah. Now it seems like you wanna flip the
1:56:58
coin again. Yeah. Okay. And that's often a
1:57:01
problem with mixed strategies. Maybe you don't like. the way
1:57:03
the point lands, you want to have
1:57:05
another shot. But it is a
1:57:07
strategy that is alternative to
1:57:09
Pascal. He was certainly not recognizing the
1:57:11
same. This has the same expected
1:57:13
utility. And
1:57:14
now we just
1:57:16
run down the slippery
1:57:19
slope okay, now suppose you wager for god
1:57:21
if and only if your lottery
1:57:23
ticket wins
1:57:23
in in the next
1:57:25
lottery and let's suppose there's
1:57:28
a billion tickets in the lottery, one
1:57:30
in a billion times infinity, still infinity. And so do the calculation, infinite expectation
1:57:32
for that
1:57:36
strategy -- Yeah. -- the
1:57:38
lottery ticket. I
1:57:39
wait to see whether a meteor quantum
1:57:44
tunnels through this room before the
1:57:46
end of our interview, some tiny probability of this happening, I don't know, one in the Googleplex,
1:57:48
call it, multiply
1:57:51
that by infinity, that by infinity
1:57:54
And
1:57:54
I
1:57:55
have infinite expected utility for this
1:57:57
strategy, wager for God, if and
1:57:59
only
1:57:59
if the material happens. And now it
1:58:02
starts to look like whatever I do -- Yeah. -- there's some
1:58:03
positive probability that
1:58:06
I will
1:58:07
get
1:58:08
the -- Yeah.
1:58:10
-- info payoffs. Into the
1:58:12
payoff. And and that has infinite expected utility. Even
1:58:14
if I try to avoid belief in God, there's some probability that
1:58:16
I'll fail. Yeah. You'll accidentally
1:58:18
believe in God. And then I
1:58:22
get the infinite expected utility at
1:58:24
that point. And so
1:58:25
my conclusion was not only as
1:58:28
Pascal's wager invalid,
1:58:30
namely the conclusion that you should wait for God does not follow from the premises.
1:58:32
It's invalid in, so
1:58:34
to speak, the worst possible
1:58:36
way
1:58:36
in so to speak the worst possible
1:58:38
way namely, recommends everything equally. Everything
1:58:40
by those lights should be
1:58:42
equally good. Whatever you do
1:58:45
has infinite expected
1:58:46
utility? Yeah. It is it's it's interesting
1:58:48
that the the power of the infinity to create
1:58:51
the paradox here is so powerful that
1:58:53
it also destroys the paradox or Just, like,
1:58:55
ends up producing some garbage result. Like, everything is
1:58:57
permissible. Yeah. That's it. So the the
1:58:59
great strength of Pescale's wager was
1:59:01
you didn't need to worry about
1:59:03
the probability that you were assigned to God's existence. Infinity
1:59:06
just swamps it -- Yeah. -- as long as it's
1:59:08
positive. And
1:59:10
now the revenge problem is
1:59:13
well, that's swamping effect of infinity.
1:59:15
Now it affects just anything
1:59:17
you choose to. There'll be
1:59:19
some
1:59:19
probability that you wind up wagering for God and away we
1:59:21
go. Yeah. I guess you don't even have to
1:59:23
say
1:59:23
that you necessarily get the
1:59:26
infinite payoff via the original, like,
1:59:28
envisaged believing a God. It
1:59:30
could be like just anything you do might pay off in infinity. Right? That's that's right. There's all kinds of different ways that happen. Yeah.
1:59:32
That's it. Yeah. No. Okay. Anderson Psberg
1:59:34
gay maybe waiting down the road. Exactly.
1:59:39
Yeah. You you you go past the church, go to
1:59:41
the casino, and yeah. Yeah. Yeah.
1:59:43
Okay. It seems like there's a
1:59:45
very natural response here, which is
1:59:48
to say, well, under a in
1:59:50
in situations where your you have, like, different possible paths to infinite positive
1:59:52
values. Yes. Wouldn't it
1:59:54
be very natural to choose
1:59:56
as a tie breaker, the one that has the highest probability of Yeah. So if
1:59:58
you're, like, now I'm, like, weighing up I'm weighing up the casino with this a bit of a game in the
2:00:01
church, and then I kind of
2:00:03
have to make a decision one
2:00:05
has a higher probability of getting me getting me
2:00:07
infinite value. That's right. Yeah. Yep. Very natural reply. And a few people
2:00:09
have made it Schlessinger, for example, has has
2:00:11
a version of that. and
2:00:15
it's certainly very plausible. Well
2:00:17
first point, notice you've
2:00:18
added an extra premise, maybe that
2:00:20
you could put it this way.
2:00:22
It's like a lexical ordering. When some
2:00:25
action uniquely maximizes expected
2:00:28
utility, do that.
2:00:30
When you have ties,
2:00:32
you
2:00:33
need a tie breaker and you
2:00:35
offered a tie breaker that now look at the probabilities and maximize the
2:00:36
probability of, in
2:00:38
this case, the insured payoff.
2:00:42
That's
2:00:42
okay. That's fine. But
2:00:44
I think my
2:00:45
point still stood from before that
2:00:47
the argument as stated
2:00:49
was
2:00:50
invalid and it needed help from
2:00:52
this extra premise, this tie breaker -- Yeah.
2:00:54
-- when you have the ties. A lexical decision making
2:00:56
procedure
2:00:56
is one that
2:00:59
has multiple steps where you only proceed
2:01:01
to the later steps if there's a draw on the first process. It's a generalization actually, I think,
2:01:03
from lexical ordering
2:01:07
being alphabetical ordering, where when you're ordering
2:01:09
things alphabetically, you look at the first letter and you only bother to look at the second letter and then
2:01:11
order it based on the second letter
2:01:14
if the first letter is the
2:01:16
same. So with alphabetical
2:01:18
ordering, of course, you look at the first letter and you bring all of the a's together and then all of the b's together, only then do
2:01:20
you look at the second letter and rank them
2:01:22
by the rank them by the second letter?
2:01:27
So so decision making process where you the first step and then a second
2:01:29
step. Okay. Back to the show. This takes
2:01:31
us back to
2:01:33
our discussion of
2:01:35
heuristics before and you've got multiple ways of
2:01:37
realizing, in
2:01:38
this case, the maximal expected value.
2:01:40
And just to say
2:01:42
a bit more about this reply.
2:01:44
It's it's very intuitive what you said.
2:01:46
It'd be nice to to generalize what really
2:01:49
is the rule. So
2:01:51
in this case, we had something of
2:01:53
infinite utility, but what what's the general rule that we're applying?
2:01:56
Because it seems like so,
2:01:58
yeah,
2:01:58
you said lexical, which is kinda
2:02:00
like indicating there's two
2:02:02
steps here. It seems like the rule is if you if infinite utility is possible, maximize the probability
2:02:04
of infinite -- Yeah.
2:02:06
-- payoff. If it's not,
2:02:09
then do normal expected value. Yeah. I guess a trouble here is gonna be that infinite value is on the
2:02:11
table. We said we assign it, like, non
2:02:14
zero value possibility in, like, any
2:02:16
action and
2:02:19
say that actually we never get to the second step. Normally expected value never
2:02:21
arises because we're always stuck on the
2:02:23
first step. Just trying to
2:02:26
maximize the probability of infinite
2:02:28
payoff. Well, and then there are so
2:02:30
many routes to getting this infinite expected utility. And it
2:02:32
seems like you're going to be reaching for
2:02:34
the tie breaker all the time now. I'll
2:02:36
be there
2:02:38
were just too many ties. And
2:02:40
it'd be
2:02:40
nice to to really clarify what
2:02:43
the rule is, this lexical
2:02:45
priority rule, it
2:02:46
made sense in Pascal's wager as you said
2:02:48
it, but that seems to be a special case of some
2:02:50
more general thing and it would be nice to
2:02:54
give that some foundations. And utility theory
2:02:57
has it seems
2:03:00
some
2:03:01
firm
2:03:01
foundations in
2:03:04
terms of preference axioms and so called representation
2:03:06
theorem, which a lot of people appeal
2:03:08
to. It'd be nice to
2:03:10
have something parallel to that for
2:03:14
this enhanced decision theory where you have this lexical rule. And
2:03:16
I'll just put
2:03:19
in a plug for Paul
2:03:21
Barthor who has offered something like that, I I think he calls it relative
2:03:23
utility theory -- Okay. -- which
2:03:26
is kind of a generalization
2:03:30
of
2:03:30
the
2:03:31
familiar expected utility theory. It
2:03:33
involves ratios -- Uh-huh. -- utility ratios.
2:03:35
I see. It it's a
2:03:37
generalization into the infinite cases. That's right.
2:03:39
It's robust to them. Yes. So he he he,
2:03:41
I think, can deliver your intuition
2:03:43
in Schlessinger's rule Yeah.
2:03:46
And it will fall out, which is
2:03:49
good because now it's giving a
2:03:51
more foundational support
2:03:52
for this rule. Okay.
2:03:53
So zooming out for a minute. We
2:03:56
got the Saint Petersburg game. We got the Pascal's wager, which, like,
2:03:58
they're both introducing infinities by different passages, and then they seem
2:03:59
to, like, just
2:04:03
really create an awful lot
2:04:03
of trouble for expected value. I guess, so
2:04:06
I wanna go out tonight and, like,
2:04:09
choose what movie to watch and, like, make
2:04:11
decisions based on expected value. And so,
2:04:13
like, waiting things by their probability,
2:04:15
clearly. Yeah. I don't wanna feel like I'm
2:04:17
doing the right thing here, but those people are coming up and
2:04:20
saying, hey, I've got these paradoxes for expected
2:04:22
value that produce, like, garbage results or at
2:04:24
least, like, require
2:04:26
totally rethinking it. Yeah. How comfortable should I
2:04:28
feel when I use expected value to
2:04:30
make decisions in in life? Yeah.
2:04:32
Are these are these, like, wacky
2:04:34
cases with convergences to infinities and putting infinities and things. Are
2:04:37
they, like, fundamentally a problem? Or are
2:04:39
they just more curiosities? Yeah.
2:04:42
Well, one one solution is I guess I mentioned it earlier, you
2:04:44
really do just zero out these crazy
2:04:46
cases. Yeah. You don't even give them
2:04:48
one in the Google plex, I
2:04:50
see. Credence. Yeah. And that that would
2:04:53
certainly quarantine them. Yes, I have made versions of
2:04:55
this worry in
2:04:55
a few places
2:05:00
how even
2:05:01
just everyday decisions seem to be contaminated
2:05:03
by, in this case, Infinity, I've also talked about it
2:05:05
in relation to a
2:05:07
game that has no
2:05:10
expectation at all, the so
2:05:12
called Pasadena game.
2:05:13
And the game
2:05:14
itself may seem pathological than
2:05:17
But if you give it any credence, then even
2:05:19
a simple choice, like, where should I go out for dinner tonight? Will be Chinese
2:05:24
or pizza? if you give
2:05:26
those prospects, if you give some probability to the crazy stuff happening
2:05:28
at the end, that
2:05:31
easy decision gets infected Yeah.
2:05:34
So so I guess you
2:05:36
have to do the dogmatic
2:05:38
thing and just say, look, I'm
2:05:40
just zeroing out. or even more even more
2:05:42
because you can go dogmatics. Well, you can choose your dogmatism. You can either be when things become sufficiently weird, I
2:05:44
give them zero probability -- Yeah. -- it
2:05:46
seems dogmatic. Or I guess you can be
2:05:50
I refuse to consider infinities. Just give them some finite
2:05:53
positive value and and lead at that. Yeah. Or
2:05:55
you just have to become a
2:05:57
fanatic who pursues in for the values -- Well, the
2:05:59
time. -- well,
2:05:59
and you heard me before was putting
2:06:02
in an argument for for the crazy
2:06:04
thing. That's right.
2:06:06
And So for practical purposes, I I think you
2:06:08
you have
2:06:08
to be dogmatic. And maybe
2:06:10
even in some cases, not
2:06:13
just being dogmatic and giving probability
2:06:15
zero to scenarios. In some you just don't even consider them.
2:06:17
They're just not even in your space of
2:06:19
possibilities to begin with.
2:06:22
It's it's not that you you recognize it and give it probability
2:06:24
zero. This is one statisticians
2:06:26
reply that I've heard. You
2:06:30
just don't even put it in your model of the world.
2:06:32
Yeah. Yeah. I mean okay. So
2:06:34
to speak
2:06:34
up for being crazy for
2:06:37
a minute, like imagine that we we really did think
2:06:39
that infinite utility
2:06:40
was a a live possibility that we
2:06:42
had some, like, theory of the universe
2:06:45
in which, you know, we could
2:06:46
possibly let us say that we we didn't, for example, think that
2:06:49
the universe was gonna, like, peter out either become very
2:06:51
big or very small -- Yeah. -- and
2:06:53
such that, like, we're in a steady state universe and maybe
2:06:55
set up a system where you like do live forever. There's like nothing that interferes
2:06:57
with your life. And so maybe you could get
2:06:59
an infinite utility that
2:07:02
way. So we have some, like, theory that makes
2:07:04
it, like, feel not infinestimally likely, but maybe,
2:07:06
like, one in a thousand likely. Mhmm. Then
2:07:09
it feels less crazy to say, well, you
2:07:11
should orient your life around trying to do that, trying to get the infinig utility forever because
2:07:14
the universe permits that.
2:07:17
Yeah. So maybe maybe we can buy
2:07:19
the bullet. I guess yeah. Another way to go is to give infinity a
2:07:24
more nuanced treatment. Mhmm. So so far, I
2:07:26
was imagining it'll be hard to convey this just over the podcast,
2:07:28
but I'm I'm sort of drawing
2:07:30
the figure eight of infinity. It's that
2:07:33
figure eight on its side, infinity. And that's the un nuanced infinity that
2:07:35
seems to
2:07:36
have these problems.
2:07:39
If you have it, or
2:07:42
you multiply it by one in the Google plex,
2:07:44
you still get the
2:07:46
same sideways, figure eight infinity
2:07:48
back. But if you had a
2:07:50
more mathematically
2:07:51
nuanced treatment of infinity
2:07:53
where halving something or multiplying
2:07:55
it by when in the
2:07:57
Googleplex made a difference then we
2:07:59
might
2:07:59
get the ordering
2:07:59
that we
2:08:00
want again. This is another way
2:08:02
of handling the problem by
2:08:04
the way of which led
2:08:07
to your lexical rule Maybe
2:08:09
if we just
2:08:10
distinguish among different infinities. Mhmm. God, I'm scared of it.
2:08:13
This
2:08:13
this just seems
2:08:16
like it's create more problems. And
2:08:18
that and it it's also scary just the sheer mathematics
2:08:20
of it -- Yeah.
2:08:23
-- is formidable. But it
2:08:25
turns out that there
2:08:28
are
2:08:28
these systems, for example,
2:08:30
the surreal numbers,
2:08:31
hyperreal numbers where you
2:08:33
have infinities and multiplying them makes a difference, multiplying by
2:08:36
half
2:08:36
or what
2:08:39
have you will change
2:08:41
the value will make it smaller in
2:08:43
this case. And so maybe now you've get the ordering
2:08:44
that you're hoping
2:08:46
for and you can choose
2:08:50
Chinese over pizza after all if
2:08:52
you keep track of the sizes of
2:08:54
all of these infinities. Yeah. And that's that's
2:08:57
been a bit of a cottage industry too
2:08:59
of -- Okay. -- doing these highly
2:09:01
technical, highly sophisticated essay. Refinements of decision theory.
2:09:03
Okay. Yeah. Okay. Let let me
2:09:05
make another line of
2:09:08
argument here. Infinity's
2:09:09
mess shit up. Yeah. So
2:09:10
okay. Some sublisted might be familiar with
2:09:12
the Bannik Tarsky
2:09:15
Paradox. Yeah. So basically, the
2:09:17
idea is, like, take take a sphere, a solid sphere. Yeah. If you divide it into an infinite number of
2:09:19
points, the mathematicians in the
2:09:22
audience might be annoyed by
2:09:25
divide it into an infinite number of points and
2:09:27
then move them around in some special way. And it seems like you can get two full spheres out of
2:09:29
the matter of the
2:09:32
other, like, the volume of the
2:09:34
original sphere. Yeah. So it's like you've doubled the amount of volume that you have just by like splitting something into infinite points and then putting
2:09:39
it together again. I don't think that
2:09:41
that could happen in the universe. Probably, it doesn't seem like that happens. And there's, like, maybe just whenever we
2:09:43
put infinities into
2:09:47
into these decisions, we're just gonna
2:09:49
find lots of problems and lots of things that won't that will never happen in the real world and we and
2:09:51
so we should be okay to dismiss infinities
2:09:56
and say, and, like, throw them out because
2:09:58
of, like, some just just other ways that they make a life unlivable. I know. Great. Yeah. Feynman was
2:09:59
told about the Baraktarsky Paradox,
2:10:02
and it was presented to him
2:10:05
involving an orange. Yeah. You've got an orange of a certain size and by suitably
2:10:07
cutting it up, you can
2:10:11
create two oranges of
2:10:14
that size, in fact, you can
2:10:16
keep multiplying them. And Feynman
2:10:18
bet that that was just
2:10:20
nonsense. That wasn't true.
2:10:22
And then someone explained
2:10:23
to him how you do it. There's this infinitely precise
2:10:25
surgery and it involves non measurable
2:10:27
sets and so on.
2:10:29
And Feynman said come on, thought
2:10:31
you meant a real orange. Now, of
2:10:34
course, we understand that
2:10:36
reaction. But
2:10:38
I feel like saying, yeah,
2:10:40
but that doesn't really solve the paradox,
2:10:42
you know. Well, thank God, we can't
2:10:45
do infinitely precise surgery on
2:10:47
oranges. hence our theory of measure, okay, is is safe. You
2:10:49
feel like saying no that, of course,
2:10:51
this is highly implausible that
2:10:54
you can't actually do this.
2:10:56
But aren't you
2:10:57
worried that there's something wrong with our theory of measure
2:10:59
that it seems to allow this this result?
2:11:01
And I feel like saying
2:11:03
something similar about Decision
2:11:05
theory noticed that Richard Jeffrey's reply was rather like Feynman's
2:11:07
regarding
2:11:11
Bannaktarsky Jeffrey
2:11:13
said with regard to the Saint Petersburg Paradox, anyone who offers
2:11:15
you the Saint Petersburg game as a liar. And
2:11:18
of
2:11:19
course, that's true no
2:11:22
one
2:11:22
in the real world is gonna offer you the
2:11:24
Saint Petersburg game genuinely. But
2:11:26
I still have that niggling
2:11:28
feeling too look, there's still something
2:11:31
wrong with with just our theory of measure in the banach Tarsky case
2:11:33
of expected utility
2:11:35
and rational decision.
2:11:36
expected utility and rational decision
2:11:39
in the case of Saint Petersburg. And it'd be
2:11:41
nice
2:11:41
to solve that that
2:11:43
problem. Yeah.
2:11:44
Maybe that's now
2:11:46
the philosopher in me rather than the physicist or the engineer in me. Well,
2:11:48
it's a it's a very common theme, I guess, in
2:11:50
philosophy. That one flips between like the
2:11:53
sublime realm of ideas. Yeah. You're,
2:11:56
like, highly idealized situations, and then you, like, bring it
2:11:58
back into the world. And you have to say, like,
2:12:00
is this still relevant? It's like you
2:12:02
do a bunch of math and you're like, does right. guess judgments
2:12:04
on whether it's still relevant. Yeah. As
2:12:06
you've made it like stranger and stranger.
2:12:11
Yep. That's right. Flossers often have highly
2:12:13
fanciful thought experiments
2:12:16
to make some philosophical
2:12:19
points like Frank Jackson imagined Mary
2:12:21
in a room and she she knows all the physical facts,
2:12:23
but she's never seen red
2:12:27
and then she sees red for the first It seems that she's
2:12:29
learned something. The
2:12:31
Chinese room from
2:12:32
Sirl is a
2:12:34
famous thought experiment, Putnam, had
2:12:36
Twin Earth and so on.
2:12:38
Now it seems to me,
2:12:42
philosophically unsatisfying to reply,
2:12:43
well, there's no such room. There's there's no there's
2:12:45
no room with Mary in it.
2:12:47
There's no Chinese
2:12:50
room. Twin Earth comes as all rubbish. This is all rubbish. There's
2:12:52
no Twin Earth. Yeah. Of course, we know
2:12:54
that. We never said there was. Yeah.
2:12:57
But you still feel these these thought experiments
2:12:59
in Saint Petersburg are put in the same
2:13:02
category -- Yeah. -- like Tarske, they
2:13:04
are
2:13:05
putting pressure on some
2:13:07
entrenched notion of ours. Yeah. I mean, is it because you
2:13:09
want your theories of things to be,
2:13:11
like, to apply in all cases, to
2:13:14
not be contingent on, like, specific empirical
2:13:16
facts about how how
2:13:18
finally you can cut things. Yeah. Exactly. Yeah. Yeah. In in case of theory, you
2:13:23
know, fingers crossed that there
2:13:25
are no sufficiently large bank accounts and then you look at the world. Okay. We're good. In the case
2:13:28
of Balaktarsky,
2:13:32
Fingers crossed that you can't do
2:13:34
that surgery on oranges. Few. Yeah. Turns
2:13:36
out you can't and
2:13:39
all is good. No.
2:13:40
You you still think the
2:13:42
mere possibility, the conceivably as we were
2:13:43
talking about before or is
2:13:45
enough to
2:13:46
break it? Is enough to to
2:13:48
make one worry -- Yeah. -- the thought experiments
2:13:51
are a problem enough. Okay. Well, we should we should move on
2:13:53
for our expected
2:13:55
value. I mean, for people in the audience who have
2:13:57
reservations about expected value in, like, ordinary so many people, they hear about these cases
2:13:59
and they're like,
2:14:02
I'm going to now has serious reservations about using expected value
2:14:04
to make decisions in my life on what to
2:14:06
do. Yeah. I guess I probably know
2:14:09
a very outsized frac of the number of people in the world
2:14:11
who who actually do this. But do do you have
2:14:14
any advice for them? It it feels like this
2:14:16
field hasn't wrapped
2:14:18
up yet. We haven't answered this one. Yeah. the some
2:14:20
revisions of expected utility theory
2:14:22
that
2:14:22
might provide some therapy, like
2:14:27
the risk weighted theory or the weighted
2:14:29
linear utility theory, that might provide some relief. I'm
2:14:32
just guessing that
2:14:32
all of these are
2:14:33
gonna have their own paradoxes. But I don't
2:14:35
I You actually
2:14:38
in fact, let let's just cut to the chase. They
2:14:40
do. And in fact, they'll
2:14:43
be either
2:14:43
reformulations or the, you know, the negative
2:14:46
version like a negative Saint Petersburg game that
2:14:48
trouble for some of these theories.
2:14:50
So problems are still working. But I in a way,
2:14:52
I
2:14:53
think we we can
2:14:55
quarantine the problem. It's
2:14:57
a little bit like what Hume
2:14:59
said. He said something along these lines that when he leaves, know, is philosophizing
2:15:01
and
2:15:01
he goes out
2:15:04
into the world,
2:15:06
he plays billions or whatever.
2:15:09
You know, he he leaves his
2:15:11
philosophical problems
2:15:11
behind. And and of
2:15:13
course, we we shouldn't be
2:15:16
paralyzed by the
2:15:16
Saint Petersburg game or what have you -- Yeah. -- about
2:15:18
I mean, I think I think they're important to think about
2:15:20
conceptually while
2:15:23
we're being philosophers in the real
2:15:26
world, we will
2:15:26
not write out a a decision matrix that'll have these problematic features. Pascal's
2:15:28
wagerie actually is interesting --
2:15:30
Yeah. -- in that case, I
2:15:34
think there is a possibility of creating, like, new
2:15:36
universes or, like, infinite like, infinities
2:15:38
do seem possible maybe in in the universe
2:15:40
is sorry. Maybe that's not the point
2:15:42
you want. But Yeah. And also, I mean, people really do
2:15:45
take seriously what what Pascal
2:15:47
said. Mhmm. Okay. Yeah. Of course,
2:15:49
people do that. Yeah. There are
2:15:51
lots of lots of Christians
2:15:52
who say that Pascal said it
2:15:55
exactly right and not just Christians, there there'll be other religions with which
2:15:57
will acknowledge some source
2:15:59
of
2:15:59
infinite utility. And
2:16:02
so now this becomes a
2:16:04
practical problem. This is
2:16:05
not like banaktarshi in cutting up
2:16:08
oranges, you know, given certain
2:16:09
beliefs that people people really have in the real world, it seems like a
2:16:11
life issue. Yeah.
2:16:13
like alive issue
2:16:15
Yeah. I guess, It seems like
2:16:17
we at least don't have an impossibility proof. We don't have an impossibility
2:16:19
theorem yet that shows that we can't have a satisfying
2:16:22
decision theory. So so So
2:16:24
the dream so the dream
2:16:26
remains. Keep keep working. Keep philosophers employed and keep GPI more
2:16:31
research -- funded. -- we got we got to pay highest salaries to get better people into reservoirs. Absolutely.
2:16:37
Look at how the problems are. Alright. Let's push
2:16:39
on and talk now about kinda factuals, which have
2:16:42
been one of your your big recent passions.
2:16:44
Yeah. To start with yeah. Can you
2:16:46
explain what are kind of factors? I guess, we use the term loosely quite a lot, but, like,
2:16:51
yeah, precisely, what are they? Yeah. They're
2:16:54
conditionals. They're if then, statements and typically with it were
2:16:56
the case, the p, it
2:16:58
would be the case the
2:17:00
q, or if it had
2:17:02
been the case, the p, it
2:17:05
would have been the case that
2:17:07
queue. And typically, they presuppose that the first
2:17:09
bit what we call the antecedent, the p bit
2:17:12
is false. and
2:17:15
then they have us say
2:17:17
something about a hypothetical scenario. Yeah.
2:17:19
I mean, sometimes we might allow the
2:17:22
antecedent to be true, but the typical
2:17:24
typical cases where p in if p then
2:17:26
q is false. So we've got to say so
2:17:29
If I were able to fly, I could
2:17:31
travel to New York. The antecedent is I were
2:17:33
able to fly. Yep. But -- Yeah. -- there's a proposition that's
2:17:35
false in the actual world I
2:17:38
I can fly and then
2:17:39
we we imagine
2:17:40
a a situation in which that is realized. Yes, I see. Okay. So
2:17:42
that that's the antecedent. And what's what's the the
2:17:48
second bit is the consequence. Yes.
2:17:50
So if p then q, p is the antecedent q is the
2:17:52
consequent. Is
2:17:55
there, like, a a classic, what do you think,
2:17:57
the paradigm example of a counter factual that is used in philosophy when when you're
2:17:59
teaching this to students? Yes. Well, I'll give you a couple. Actually,
2:18:03
this is a good way of
2:18:05
bringing out the difference between two kinds of conditionals, the counterfactual as
2:18:07
opposed to what's called indicative conditional. If
2:18:13
Oswald didn't kill Kennedy, then
2:18:15
someone else did. Now
2:18:17
that seems true, and this
2:18:20
is an indicative conditional,
2:18:22
So
2:18:22
we're assuming that Kennedy really was killed. We know that,
2:18:25
but we're not sure who did
2:18:27
it. If Oswald didn't kill Kennedy,
2:18:29
then someone else did because one way or another he was killed, I see. Okay. Now compare
2:18:31
that to
2:18:35
if Oswald hadn't killed Kennedy,
2:18:37
someone else would have Now that's
2:18:38
quite different, quite different and maybe far less
2:18:43
plausible. I mean, that suggests
2:18:45
there was a backup assassin
2:18:47
maybe some conspiracy, and we could easily deny that. Yes.
2:18:51
Anyway, or or maybe even
2:18:53
easy how about We're not completely sure who
2:18:55
wrote Hamlet. If Shakespeare didn't write
2:18:57
Hamlet
2:18:57
than someone else
2:18:59
did. That seems
2:19:01
true because we know
2:19:03
Hamlet was written. Yeah.
2:19:04
Okay. If Shakespeare hadn't written, Hamlet
2:19:06
-- Someone else would. -- sounds crazy.
2:19:08
That sounds crazy as if Hamlet was just
2:19:10
fated to be written and it Shakespeare
2:19:14
happened to be the vehicle for for it,
2:19:16
but someone else would have stepped in if need be able
2:19:18
to put put it and sit it in their mind.
2:19:21
Yeah. So what what's the name
2:19:23
for these two different kinds? Okay. Indicative.
2:19:26
Yeah. Okay. And counterfactual. Subs you we use the subjective
2:19:28
conditional that
2:19:30
mood typically to express the counterfactual.
2:19:32
Yeah. We'll call
2:19:33
those counterfactuals where typically there's the presupposition that the antecedent
2:19:35
-- Yes. -- false. Yeah.
2:19:39
What what's
2:19:39
what's the history of the study of counterfactuals?
2:19:41
I imagine, surely, I mean, people have been making statements of this kind -- Yeah. -- like, since humanity began speaking, you imagine.
2:19:44
Yes. But It
2:19:48
seems like discussion of this kind of wood
2:19:51
could if then it doesn't seem like that's
2:19:53
kind of in the classic pantheon of old
2:19:55
school philosophers. They they didn't seem to think
2:19:57
about very much. I think you're right. They they did talk about conditionals,
2:19:59
for example, the
2:19:59
stoics, talked
2:20:05
about conditionals, theodaurus, creecepus, But often they
2:20:07
were talking more about what we'd now call the indicative conditional, for
2:20:09
example, they had what
2:20:12
we would now call
2:20:14
the material conditional. That has
2:20:16
a certain truth table. Basically, it's
2:20:19
true. However, P and Q turn
2:20:21
out except where P is true
2:20:23
and Q is false. then if
2:20:25
p then q is false, otherwise true
2:20:27
in every combination. Why did they care about that?
2:20:29
They were doing logic. They care about logic more
2:20:32
generally. And certainly
2:20:34
the material conditional is the standard part -- Yeah.
2:20:36
-- of of logic and they they got onto it early. So
2:20:38
yeah. So when did when did the study of this kind of
2:20:41
a first flourish. Counterfactual? Yeah. I I
2:20:43
would say, more recently, it I'd
2:20:45
say it started to hit its hay day. Well, in
2:20:48
the forties, Chism and
2:20:50
Goodman began to write about
2:20:52
it.
2:20:52
Goodman wrote some
2:20:55
classic stuff, especially in fact fiction
2:20:57
and forecast that book, there's
2:20:59
a classic treatment of counterfactuals. And
2:21:01
then I think
2:21:02
the real heyday for counterfactuals came a
2:21:06
bit later in the 60s,
2:21:09
perhaps towards the end of
2:21:11
the 60s, especially Storlacher and Lewis and their classic other
2:21:15
possible world's accounts, you know. Roughly the
2:21:17
idea is that if it were the case,
2:21:18
the p, it would be the q is true just
2:21:22
in case, at the closest p world
2:21:25
q is the case -- I see. -- by p world, I mean,
2:21:27
a world where p is true. Hold Yeah.
2:21:31
And then they had this famous
2:21:33
debate about, well, is there a unique closest -- Mhmm. -- p
2:21:35
world that was thought not
2:21:38
the was thought not And
2:21:39
by the way, notice this
2:21:41
as a use of a heuristic from earlier
2:21:43
when Stormnacko talks about the closest p world there
2:21:48
are
2:21:48
two ways you could challenge with that. There could be
2:21:50
more than one, p, well, there could be ties
2:21:53
-- Yes. -- and I think. If if Bizzais
2:21:55
and Verity were compatriots Would they
2:21:56
have been French? Would they have been
2:21:58
Italian? Seems maybe that that they're
2:22:00
equally close possibilities. And
2:22:02
going in the other direction, maybe
2:22:04
there's no Yeah.
2:22:05
No. Closest world, just ever, closer world. So
2:22:07
I'll
2:22:07
give you an
2:22:09
example of that if I want
2:22:11
to. It could be relevant later.
2:22:13
Lewis, imagines
2:22:14
the following case. If I were taller
2:22:16
if i were than
2:22:17
seven feet, how tall
2:22:19
would I be?
2:22:20
And let's imagine that
2:22:23
the closer I am hypothetically to my actual height, the
2:22:26
better
2:22:27
for close in in a sense
2:22:29
that's relevant here. would I be
2:22:31
seven foot one? Mhmm. Well, Lewis has thought is that that's a
2:22:33
gratuitous departure from my
2:22:35
actual height. Seven foot
2:22:37
half an inch would
2:22:40
be closer seven foot quarter of an inch
2:22:42
closer still -- Yeah. -- an infinite
2:22:44
sequence, a bit like Zeno's paradox --
2:22:46
Yeah. -- ever closer world's none closest
2:22:48
and that's that's been to be trouble for stormnaker.
2:22:51
Perhaps later on, I'll argue it's it'll turn out
2:22:53
to be trouble for Lewis. But and it was so that was a classic
2:22:55
period in the study of counterfactuals. Is
2:22:59
this under the banner of like a modal
2:23:01
logic? Is this basically what this is? Yeah.
2:23:03
It's yeah. You could certainly say that certainly artifacts will seem to have a modal element What
2:23:07
does Moto mean here? It means
2:23:09
something like it's not just a matter of how things
2:23:12
are actually
2:23:14
actually somehow
2:23:15
possibility is involved or
2:23:17
perhaps necessity. Yeah. I would call probability a modality
2:23:19
too, but not just things
2:23:24
as they actually turn out. Okay. So why
2:23:26
should we care about counterfactuals and and conditionals
2:23:28
and stuff? Like, from one point of view,
2:23:30
it seems like all pretty straightforward. We like
2:23:32
to use terms all the time. You wanna get confused. Everything seems
2:23:34
fine. But, like, yeah, how does it relate to,
2:23:39
like, actually important questions? Very good. For
2:23:41
start, it's philosophically important to study counterfactuals
2:23:43
because so many philosophers
2:23:44
reach for counterfactuals in
2:23:47
the study of other things.
2:23:49
things
2:23:49
like causation,
2:23:51
dispositions, explanation, laws
2:23:54
of
2:23:55
nature, free
2:23:56
will
2:23:58
will, perception, confirmation.
2:23:59
Flosses are often
2:23:59
analyzing
2:24:01
things or at least
2:24:03
referring to counterfactuals in the
2:24:05
study of these other important philosophical concepts. So starting with
2:24:07
philosophy, all over the
2:24:12
place. Yes. Then science,
2:24:13
I think traffics in counterfactuals in various ways. You could
2:24:16
ask a question like
2:24:18
if I were to
2:24:20
drill a hole through
2:24:22
the world, through the earth, and
2:24:24
drop a ball, what would happen?
2:24:27
Oh, it would be a harmonic oscillator. The
2:24:29
physics textbook tells you. Okay? Yeah. That would
2:24:31
be a
2:24:31
kind of factual. I
2:24:34
think the social sciences
2:24:37
traffic in counterfactual's history, for example, if
2:24:39
the archduke hadn't been
2:24:41
if the archduke hadn't been assassinated,
2:24:43
they wouldn't have been
2:24:45
World War one. Yeah. You
2:24:47
might say, economics, worries about the incremental benefit
2:24:50
of some commodity or the change
2:24:53
to the economy. Psychology -- Yeah. -- counterfactual is very
2:24:55
important. Think of say
2:24:56
regret regret
2:24:59
is is often informed by a counterfactual.
2:25:01
Actually, I I'm somewhat prone to regret. And when I psychoanalyze myself, I think that's why I
2:25:03
got so interested in counterfactors.
2:25:09
And that's why I especially want them to come
2:25:11
out false because then the
2:25:13
the counterfactuals that underpin my regret I
2:25:15
could I could banish. And we think
2:25:18
of relief you know, thank god that this happened because if it hadn't happened,
2:25:22
this bad thing would have happened.
2:25:24
So I think psychology is permeated
2:25:27
with cat effectuals?
2:25:27
Yeah. Okay. So it's, like, important in answering lots of other relevant philosophical questions. And
2:25:29
I guess also it's just, like, it's
2:25:31
such a core part of human reasoning
2:25:34
all the time. We constantly have to
2:25:36
think about how things would have been
2:25:38
in order to establish causation and think about what
2:25:40
is good and bad relative to other stuff. Exactly. And that
2:25:42
was gonna be my next point to just in daily life.
2:25:45
seems we need to be thinking
2:25:47
about counterfactuals, about decision making, and I
2:25:49
could add that to the philosophical list too, that that's an important
2:25:51
account of rational decision traffics
2:25:55
in counterfactuals. But never mind the philosophy,
2:25:57
just common sense, is permeated with thinking, well, if I were to do this, this would happen if I were to
2:26:00
do
2:26:00
that. something
2:26:05
else would happen? What should I do?
2:26:07
Okay. Daily life is
2:26:10
riddled with
2:26:11
counterfactuals. So when philosophers first, like, look at kind
2:26:14
of factors, like, what is the problem that they identify?
2:26:16
Why are
2:26:18
they interesting or
2:26:18
potentially challenging? For start, they can't just
2:26:21
be given the straightforward truth conditions of, for example,
2:26:23
the material conditional. That that might be your first
2:26:25
stab at analyzing if if it were
2:26:27
the case the p, it would be
2:26:29
the q. What what would that be?
2:26:31
It would be that's true in every case
2:26:33
except where p is true and
2:26:36
q is false. Okay. But
2:26:37
that would be a disaster. Okay. Because
2:26:39
now all counterfactuals where they're
2:26:41
genuinely counterfactual -- Mhmm. --
2:26:43
antecedents falls. Yeah. That would just come out
2:26:45
true. I see. And you wouldn't be able to
2:26:48
make distinctions
2:26:50
consider
2:26:50
a counterfactual of the
2:26:51
form if I were to let
2:26:53
go of the cup it would fall and I don't actually let go
2:26:55
of the cup. Now
2:26:59
on
2:26:59
on this material conditional analysis,
2:27:01
that comes out true because it has a false antecedent.
2:27:03
So far so good, I guess.
2:27:05
But now
2:27:07
if I were to let go of
2:27:09
the cup, I would finish up on the moon.
2:27:10
That would come out true as well because it still has a false antecedent
2:27:13
Okay.
2:27:16
So obviously, we want to make
2:27:18
distinctions among
2:27:19
these with false antecedents. Yeah.
2:27:21
I see. Some of them true,
2:27:23
some of them false. we're
2:27:25
gonna need some more sophisticated machinery. Hey,
2:27:27
listen. Just a
2:27:27
reminder that in
2:27:30
these kind of factual statements, we're
2:27:32
saying if
2:27:33
he, then q, then
2:27:34
he is the antecedent and q is
2:27:37
the consequent. So if I went
2:27:39
to Spain, I would have a
2:27:41
great time. Then me
2:27:42
going to Spain is the antecedent. It's the counterfactual thing that didn't happen.
2:27:46
And then me having a great time is
2:27:48
the consequent. The thing that would happen if
2:27:50
that happened. Okay. Back to the show. Got it. Okay. So the with the statement like, if
2:27:53
I
2:27:55
let go of the comp, then this other
2:27:57
thing will follow. If you never let go overcome, then possibly anything could come
2:27:59
after without the statement
2:28:02
being false because the if condition is not
2:28:04
met because you did not do it. Well,
2:28:06
and And intuitively, that's the wrong answer. We want to be able to say, well, these are the
2:28:10
true ones, and these are the false ones.
2:28:12
Not just anything goes, if some false antecedent were
2:28:15
the case. Right. Right. Right. And
2:28:17
now we need to have some
2:28:19
subtle way of distinguishing between the
2:28:21
the true ones and the false ones. Okay. Can't just be the material conditional. So how did
2:28:23
you try to do
2:28:27
that? Okay. Well, now philosophers like to
2:28:29
reach the possible world's And possible
2:28:30
worlds have been very influential, successful in
2:28:34
the study of modal logic,
2:28:37
like necessity, possibility. For example, we say that something's necessary if it's true
2:28:39
in all possible
2:28:43
worlds. Something's possible. If
2:28:44
it's true in some possible
2:28:47
world. we may have to restrict
2:28:47
the world's suitably, but that that's the first step. And now
2:28:50
the thought is
2:28:51
let's do something
2:28:54
similar to that for counterfactuals.
2:28:56
and
2:28:56
the thought is, but we don't
2:28:58
just want to look at all the
2:29:00
world's. Let's let's look at
2:29:02
certain privileged world's, the one that's ones that
2:29:04
matter And
2:29:05
the way we say that
2:29:07
is, well, the most similar worlds
2:29:10
where the
2:29:10
antecedent
2:29:11
is true. And roughly
2:29:12
this this style of analysis says,
2:29:14
P would queue, if P, where
2:29:16
the case queue would be the
2:29:18
case, is true just
2:29:19
in case. the most
2:29:21
similar p world's ARQ, and
2:29:23
maybe all of them, all of
2:29:25
the most similar p world's ARQ
2:29:27
and then let the debate So
2:29:28
hopefully, this
2:29:29
might make things clearer. So
2:29:31
you've got an issue where you
2:29:33
wanna say, if I let go
2:29:35
of the cup, then it would
2:29:37
fall down. But you're like, this leaves a
2:29:39
scenarios in which you, like, over it's like
2:29:42
a very wide range of possible scenarios. Lots of
2:29:44
things a bit different as well. That's right.
2:29:46
And, like, what for what for example, if you let
2:29:48
go of the cup and also suddenly a table appear
2:29:50
underneath the the cup to to catch it. that's, like,
2:29:52
then it then it wouldn't fall down and and hit
2:29:54
the floor. I make a lot of that very point
2:29:56
actually. That's right. So then you have to be, like, so
2:29:58
which out of the, like, vast space of possible,
2:30:01
like, kind of factual wells in which you dropped the
2:30:03
cup. Yes. Are we actually talking about when we
2:30:05
make some kind of factual claim like this? Okay.
2:30:07
And you're saying, The standard the standard accountant philosophy is
2:30:09
to say, well, it's similar and it's similar to the
2:30:11
actual world in every respect except for this one change
2:30:13
where you let go of the cup. which sounds very intuitive. Yes. You you have to make
2:30:15
some You can't
2:30:19
just tweak this one fact and keep
2:30:22
everything else as it was. because you might end up with some
2:30:23
inconsistency. That's right. Okay. Yep. Yep. can't just insert, you know,
2:30:26
the one change, like,
2:30:28
I release
2:30:29
a cup where,
2:30:31
in fact, I didn't because
2:30:33
there'll be all sorts of
2:30:35
ramifications
2:30:35
of that, ripple effects as you
2:30:37
have to take into account. But now the most
2:30:39
similar world where
2:30:41
the most similar world where all
2:30:42
of that stuff is taken care
2:30:45
of or such worlds. The thought is those are the worlds
2:30:46
that matter to evaluation of the kind of factual many
2:30:51
many people listening will be like, yeah, obviously,
2:30:53
you wanna make the minimal change when you're producing counterfactual. If if you're if the operation you're doing
2:30:56
is say, kind
2:30:59
of factual that changes that I like all the cup.
2:31:01
You know, you shouldn't be adding a table underneath it
2:31:04
has problems. Well,
2:31:08
there are three things I don't like
2:31:10
about the most similar world's accounts,
2:31:14
namely the most similar and
2:31:16
worlds. Yeah. Okay. And otherwise, I'm
2:31:18
right on the way. And and I
2:31:21
should say for years, I've I've
2:31:23
just assumed this philosophical,
2:31:24
let's call it, orthodoxy. But
2:31:26
more recently, I've come
2:31:27
to have my misgivings, which
2:31:29
I will now gladly share with
2:31:31
you. So let's go through them.
2:31:33
the
2:31:34
most. That means that we're supposed
2:31:35
to just attend to the
2:31:38
closest worlds, the ones that are
2:31:40
first in this ordering and we don't
2:31:41
look further back. We only look at the front row so to
2:31:43
speak of
2:31:45
the antecedent worlds. And I say,
2:31:48
well, no, sometimes you've got to look
2:31:50
further back.
2:31:50
I'll give you an example. Consider the
2:31:52
consider the last
2:31:55
US election and consider
2:31:57
this counter factual If Trump
2:31:59
or
2:31:59
Biden had won the
2:31:59
election,
2:32:00
or biden
2:32:02
the president would
2:32:03
be a Democrat. Does
2:32:05
that sound true? So that sounds intuitively wrong because it seems like in the cases where Trump won, it
2:32:07
wouldn't be. Exactly.
2:32:12
Spot on. Okay. But now let's run
2:32:14
it through the similarity semantics. What's the most
2:32:17
similar world where
2:32:19
the antecedent is true Because
2:32:21
one more Biden one. Right? It's the one more simple one. It's right here. It's the actual world.
2:32:24
We're we're
2:32:27
we're standing in it.
2:32:28
And it is true in the actual
2:32:30
world that the president as a democrat. Yeah. So this should come out true
2:32:33
as
2:32:36
according to this similar account
2:32:38
where as as
2:32:38
does seem intuitive, the the most similar world to
2:32:41
the actual world is
2:32:43
itself. So it seems
2:32:45
like
2:32:45
in that case when
2:32:47
you're saying if Trump or Biden one, what you're trying to
2:32:49
you're making you're trying to make a claim about all
2:32:52
the world in which, like, either of those two facts hold -- Yeah. --
2:32:54
not just about the one world that's most similar. Yeah. So I I
2:32:56
say you
2:32:58
have to yeah. As you say, you also
2:33:00
have got to consider the Trump world and that will push you
2:33:02
further back from the front row. Oh, and then I guess you've got a question of how
2:33:04
deep because
2:33:07
-- Okay. -- like, including all of them,
2:33:09
including the ones where the world exploded for no reason. Now
2:33:12
conditional account, you
2:33:14
have to look at all of the world.
2:33:17
And then maybe that's gonna be contextually circumscribed in some way, but that
2:33:19
-- Yeah. -- that does give you some impetus to look beyond Is
2:33:23
it just the front row? It's not as
2:33:25
gonna be b? Yeah. Not just the closest. Yeah. Okay. And let me give you another
2:33:28
case. And I'm
2:33:30
glad we talked about Lewis' seven foot
2:33:33
example -- Mhmm. -- earlier because now
2:33:35
I think it's gonna backfire. Yeah. Again, remember according to him in in that example, what
2:33:40
matters to similarity is just closeness
2:33:42
to my actual height. And he
2:33:45
had this sequence of ever
2:33:47
closer worlds getting
2:33:48
closer and closer to my actual
2:33:50
height, none closest. Alright. Now let me make the smallest tweak to apparent counter example.
2:33:54
the parent counter example and I
2:33:57
think it backfires on him. Yeah. If
2:33:59
I were
2:33:59
at least seven feet tall, how
2:34:02
tall would I be? So if
2:34:04
I were greater
2:34:04
than or equal to seven feet. Mhmm. Tool.
2:34:07
How tall would I be? Well, then it's
2:34:09
a hard bound, so it's easier. It's seven
2:34:11
seven feet. Yeah. Well, there you go. according
2:34:13
to that ordering, which he used, so I think it's fair
2:34:15
for me to now
2:34:18
use it
2:34:19
against him. The unique – well,
2:34:21
in this ordering, the closest worlds are going
2:34:23
to be the exactly seven foot seven point 0000
2:34:27
to infinitely many decimal place. Right.
2:34:29
Right. He told us that's that leaves so much in the ordering. So by
2:34:31
his lights, if I were at least
2:34:33
seven feet tall, I would
2:34:36
be seven point zero zero
2:34:38
to infinitely many decimal places. Exactly.
2:34:41
Exactly. And I say, whoa. Really,
2:34:43
that that comes as a surprise
2:34:45
to me. If anything, I I
2:34:47
would say, Well,
2:34:48
I suppose I might be, but it's
2:34:50
highly unlikely I'd be exactly so precisely
2:34:52
seven foot. I might be, you
2:34:54
know, a little bit more than
2:34:56
seven foot I
2:34:57
think in these cases,
2:34:59
the similarity account is
2:35:01
giving implausibly
2:35:04
specific verdicts Yeah. It's committed to the truth
2:35:06
of implausibly specific counterfactuals. Mhmm. In the
2:35:08
seven foot case, it was if I
2:35:10
were at least
2:35:10
seven feet tall, I'd be exact
2:35:13
actually seven feet tall really. So
2:35:15
specific in the
2:35:16
Trump or Biden case, the specific
2:35:18
verdict that the president would be a Democrat
2:35:20
when
2:35:22
that doesn't take into account the
2:35:24
Trump possibility. And again, I think this is
2:35:26
just symptomatic of only looking at the front row
2:35:30
of the world's, the closest world's, sometimes
2:35:32
you need to look a bit further
2:35:34
back where, for example, I'm a bit taller than seven feet or where the alternative in
2:35:39
the disjunction, Trump winning the election has
2:35:41
to be taken seriously. Right. Yes. It seems like when we're describing lots of these anti sedans,
2:35:44
the if thumb
2:35:47
or Biden one. If I were at least seven
2:35:50
feet tall, we're actually trying to indicate a range of possible worlds. the most just picks out of them like Albert
2:35:52
Fraterially. sense.
2:35:57
Exactly. And one thing that might push
2:35:59
us towards
2:35:59
is the strict
2:36:02
conditional account where you look
2:36:04
at all of the antecedent worlds,
2:36:06
and they're perhaps contextually restricted in
2:36:08
some
2:36:08
way. And by the way,
2:36:10
earlier, I should have mentioned on
2:36:12
Finto and Gileys' components of that kind of view. Or you
2:36:14
could go another way. This would be my way and this
2:36:19
might be getting to my positive view later.
2:36:21
where
2:36:21
I'd say you have to look at the worlds that
2:36:23
have positive chance at
2:36:25
the relevant time and that will take
2:36:28
us sometimes further back than just the
2:36:30
most similar worlds. Yeah. Okay. So that's you've attracted to the you've attracted to most.
2:36:35
Let's now object to to similar. Similar.
2:36:37
Lots of problems there, I think. I think in
2:36:39
the early days of this similarity approach,
2:36:41
it was
2:36:43
assumed that
2:36:44
similarity was a common sensical thing.
2:36:46
It's what the folk would regard as resemblance. Mhmm. And Gitmeid up with
2:36:48
what seems to be
2:36:51
a devastating counter example
2:36:54
to that understanding of similarity
2:36:57
his example was cast your
2:36:59
mind back to sixties, the
2:37:01
Cold War and consider the counterfactual
2:37:03
if Nixon had pressed the
2:37:06
button
2:37:06
on the nuclear bomb, there would
2:37:08
have been a holocaust. And that seems intuitively true.
2:37:10
Yeah. Let's say we want that to come out true
2:37:14
But holocaust make a big difference.
2:37:16
You know, if holocaust worlds are not
2:37:18
similar to our world where the holocaust didn't happen. Right. Okay?
2:37:22
more
2:37:23
similar would be a world
2:37:25
where Nixon presses
2:37:26
the button and then the mechanism just fizzles. No bomb is triggered and it's business pretty
2:37:28
much
2:37:29
as
2:37:32
usual. Okay. So
2:37:33
whatever disturbance is created by the
2:37:35
button not working is much much
2:37:37
smaller than the disturbance. to the
2:37:40
actual world created by the holocaust
2:37:42
versus not. That's right. So if you're assuming commonsenseical similarity, it seems that
2:37:47
we're gonna get the wrong verdict. We'll judge
2:37:49
that counterfactual to be false -- Yeah. -- by the likes of common sense equivalence. Yeah. In fact, if you could confidently make the
2:37:52
statement that if
2:37:56
Nixon had pressed the nuclear button, it wouldn't
2:37:58
have worked. Yeah. Exactly. It would have
2:38:01
fizzled. Yeah. because that would be the
2:38:03
most similar way of realizing Verizon pressing
2:38:05
the button wrong answer. Yeah. and Lewis took
2:38:07
this very seriously and then
2:38:10
fashioned a set
2:38:11
of priorities of what
2:38:14
matters to similarity well, really to handle this fine case, the And
2:38:15
fact, you might worry
2:38:18
that when he comes
2:38:20
up with
2:38:21
the zed hock,
2:38:23
you know, just reverse engineer to this particular
2:38:25
case. And you wonder how much it'll
2:38:27
generalize. And in fact, as we will
2:38:29
say in a moment, it doesn't seem
2:38:32
to generalize So well, anyway, here's
2:38:34
what Lewis said. First priority
2:38:36
is to avoid big miracles,
2:38:38
as we might say, widespread diverse
2:38:40
violations of law.
2:38:42
Okay? First priority. Second
2:38:45
priority.
2:38:45
Maximize perfect match
2:38:48
of history. Third
2:38:48
priority, avoid small miracles,
2:38:50
small violations of law. And and there's
2:38:52
a bit more. But that that's the
2:38:54
main idea. This was supposed to handle
2:38:56
supposed to handle The
2:38:58
deterministic case. Interestingly, Lewis assumes determinism
2:39:00
in a way I don't want him
2:39:02
to because it seems like the
2:39:03
universe isn't Yes. In fact,
2:39:06
he himself in in other work
2:39:08
thinks that the actual world we live in
2:39:10
is indeterministing. So one doesn't want to assume that
2:39:12
it's deterministic
2:39:15
to handle this case. And to
2:39:17
be fair to
2:39:18
him, he did then later also consider the priorities
2:39:20
for indeterministic
2:39:22
worlds -- Mhmm. -- and he
2:39:24
introduces the notion of a quasi
2:39:26
miracle. I should tell you a bit about that. Well, quasi miracle is
2:39:30
something
2:39:31
that while it's consistent
2:39:32
with the laws of nature,
2:39:34
it's not a genuine miracle. It's somehow remarkable
2:39:36
It's like a
2:39:38
pattern of outcomes seem to be conspiring
2:39:41
in a surprising way. Now, it's a little
2:39:42
bit hard to pin that down exactly. And in
2:39:43
fact, that's perhaps
2:39:47
a problem, but this very notion of quasi
2:39:49
miracles is a little bit shaky. But anyway, to to give you the sense of it, go go
2:39:51
back to the Nixon example, and
2:39:56
imagine a a world where all
2:39:58
traces of Nixon pressing the button are just erased. There's no trace that he pressed the
2:39:59
button. Now,
2:40:04
in a deterministic world, it seems that would
2:40:06
take a big miracle because you'd
2:40:07
have to
2:40:10
erase this trace and that trace and
2:40:12
another one over there. you'd need widespread
2:40:14
miracles to remove all of those. But
2:40:17
in a
2:40:18
chancy world, you
2:40:19
don't need any miracles quite
2:40:21
lawfully a
2:40:22
sequence of chancy events could collectively erase all
2:40:25
traces of the button pressing
2:40:28
just by chance Yeah.
2:40:29
They all look at thermodynamic miracles. That sort of thing quantum mechanics
2:40:31
or whatever.
2:40:34
Yes. And then Lewis wants
2:40:36
to say that quasi miracles detract
2:40:39
from similarity. we don't want to say, for example, if Nixon had pressed the then all
2:40:44
traces would have been erased. And we
2:40:46
want to somehow vanish the quasi miracles too. The quasi miracles, things that are consistent with the laws of physics, but I guess are
2:40:48
a tiny fraction
2:40:51
of worlds that
2:40:53
are consistent. And
2:40:55
they're somehow remarkable And so they detract from similarity and
2:40:58
virtue of that. Alright? And so now
2:41:00
we have the more complicated package. We've got,
2:41:02
sorry, big miracles. Big miracles. Big then similarity of
2:41:04
history. Then no small miracles. Then
2:41:06
no big match of history, then no match
2:41:08
of history. Okay. Then then no small miracles. And no priority
2:41:10
is no quasi miracle. Yeah. And actually, you might now wonder how these
2:41:14
priorities interact with each other,
2:41:16
the indeterministic and the deterministic
2:41:18
ones. So avoid the big miracles and then dispersed between
2:41:22
that and the avoid small miracles we've
2:41:25
got this other priority, maximize perfect match. Already, you might think that's
2:41:27
a bit strange. Like, where do I insert avoid
2:41:31
medium sized miracles. Yeah. It's it's
2:41:33
feeling definitely a bit arbitrary. Yeah. You'd think there's there's a continuum among
2:41:35
the miracles and it's little
2:41:39
odd that they're interrupted by this
2:41:41
different kind of priority. But now how do we square that stuff with the quasi miracle stuff?
2:41:44
So what square that stuff with
2:41:46
the quasi miracle stuff so what how
2:41:48
do we prioritize, avoid big
2:41:51
quasi miracles as opposed to
2:41:53
avoid medium sized and genuine
2:41:55
miracles? And so gets more more complicated.
2:41:58
And and one point I want to make is notice we've come
2:42:00
some
2:42:02
distance from the original common sensible resemblance. Now
2:42:04
-- Yeah. -- we've got this quite complicated
2:42:06
theory of how similarity works for counterfactuals. So taking a step back and looking at
2:42:10
the progression here, I guess, so we started
2:42:13
out with this kind of appealing, very simple, very clean statement, the
2:42:15
most similar world. If, like, if that worked in
2:42:18
all the cases, it'd be like beautiful. So
2:42:20
elegant, it's very parsimonious. Let's just keep that.
2:42:22
Now we've got this far more cumbersome theory. We've got, like, four different stages and where, like, you're you're getting
2:42:25
to suspect that if we looked at
2:42:27
more cases, we'd find more problems. And
2:42:29
now it's now we're gonna have
2:42:31
seven different steps in order to try to
2:42:33
make it work. And then you have a question.
2:42:36
As a philosopher's suppose, you can say, on the one hand, maybe
2:42:38
just the way humans use language with kind of factuals is super messy and
2:42:40
complicated. in
2:42:42
fact, like, even seven wouldn't be enough because there'll be
2:42:45
some weird cases with that. And maybe it just in order
2:42:47
to actually know what people are referring to when they use catapactuals you
2:42:50
just have to have the full intuition of a
2:42:52
human mind to, like, understand what is being referred
2:42:54
to. On the other hand,
2:42:54
you might think, actually, counterfactual should be cleaner than this. There should be a more simple statement.
2:42:59
And now we're in Epi cycles where we're trying to
2:43:01
fix this theory that was broken fundamentally at its core and we need a different approach. Is is
2:43:03
this kind of right that you could see
2:43:07
see you could you'd see what's going on
2:43:09
two different ways, Megan. I think that's a good way
2:43:11
to put it. And I think even when we add the epic cycles, as
2:43:14
we've been doing. We're still not
2:43:16
done. Okay. In fact, let's let's
2:43:18
Yeah. Let's do something. He's he's more trouble.
2:43:21
So
2:43:21
let's take Lewis' priorities
2:43:23
and consider the following cateffectual.
2:43:26
The story is that I
2:43:28
did not scratch my finger
2:43:30
yesterday. Let yesterday be the twenty
2:43:32
four hour period that ended at midnight last night,
2:43:34
and I'll include midnight as part of yesterday.
2:43:39
okay,
2:43:39
I did not scratch my
2:43:41
finger yesterday. If I had scratched
2:43:43
my finger, I would have done so at midnight precisely at midnight no earlier.
2:43:44
the midnight precisely at
2:43:46
midnight and know earlier Yeah.
2:43:49
It seems comes out true
2:43:51
on the Louisian priorities. Why is
2:43:53
that? Well,
2:43:54
we don't need any big miracles
2:43:56
to get the antecedent to come
2:43:58
out true. So we moved to the second priority. We now want to maximize perfect match of
2:43:59
history, but
2:44:03
we get to do that by delaying
2:44:06
my hypothetical scratch as late as possible because we get more and more match of what actually happened.
2:44:10
That means that the maximal match
2:44:12
will happen if I delay scratch until midnight. Exactly. That's
2:44:14
where I would have scratched exactly. That already seems
2:44:18
very Say that again.
2:44:19
If I had scratched my finger yesterday,
2:44:21
I would have done it at midnight exactly.
2:44:23
That's crazy. Yeah. That seems crazy. Notice it's another example of what call implausible specificity
2:44:29
up. Yeah. Why then exactly? Now to make things
2:44:31
even worse, let's suppose that it's a somewhat probabilistic matter
2:44:36
when I scratch I start the day a
2:44:38
little bit itchy, I'm more likely to scratch earlier
2:44:42
in the day. And as the day
2:44:44
progresses, it becomes less and less
2:44:46
likely that I'll scratch and in fact midnight is the least likely time for me
2:44:52
to scratch. Nevertheless, by those
2:44:54
priorities, seems to come out true that if I'd scratched yesterday, I would have done so at midnight,
2:44:57
namely
2:45:00
the least probable time, and I think that
2:45:02
makes the problem even worse. So we've tried patching it, and now we've got this other problem that's appeared another
2:45:04
case of excessive specificity. Are
2:45:06
there any other issues with
2:45:09
the Lewis attempted patch of
2:45:11
his theory? Yeah. look good on
2:45:13
him for going out on a limb and
2:45:15
telling us what matters to similarity. The trouble is
2:45:18
I think he'll get some camera examples and they'll have this form of
2:45:20
implausible specificity.
2:45:22
Now you
2:45:23
might retreat again and say,
2:45:26
well, similarity, it's
2:45:26
context dependent, and it's vague, and
2:45:30
it's complicated and don't
2:45:32
really say what matters
2:45:34
to similarity. And Goodman, by the way, said similarity is
2:45:36
insidious. It's
2:45:40
sort of impostor. It's it's
2:45:42
a quack. And he thinks that similarity is not suitable to
2:45:44
he thinks that similarities not
2:45:46
suitable to ground
2:45:48
any philosophical project. Oh, wow. Yeah. Okay. Okay.
2:45:50
Well, let let's come back to this in a second. Yeah. What's what what are some other kind of examples? Yeah. Well, so now I'm
2:45:51
imagining that we retreat
2:45:57
and
2:45:57
we we just say similarity. I won't tell you
2:45:59
much about it. It's
2:46:03
this context dependent,
2:46:04
vague, complicated thing. Well, now
2:46:05
I guess it's harder to come up with counter examples. But now it's not clear
2:46:08
what the
2:46:10
theory is saying. It's not making predictions. it's
2:46:12
not providing any explanation. And I like to put
2:46:14
this in the form of a dilemma.
2:46:17
And to explain this, I should just
2:46:19
say something about Harry Potter. I I went
2:46:21
to one of the Harry Potter movies. I
2:46:24
don't remember which one. This was a long
2:46:26
time ago. And to be honest, I didn't
2:46:28
like it as much as everyone else. seem
2:46:30
to him. For the following reason, well, early on in the movie,
2:46:32
early on in the movie Harry
2:46:34
has got his magic wand and
2:46:36
you see him doing all these cool
2:46:38
magic tricks Then later on
2:46:39
in the same movie, Harry's in danger and we in
2:46:41
the
2:46:44
audience are supposed to feel scared
2:46:46
for him. But I
2:46:46
felt like yelling at the screen just usually bloody wand. Yeah. Right.
2:46:49
And my complaint was I felt
2:46:51
that we weren't properly told the
2:46:53
rules of the game. You know?
2:46:56
Right. What could the one do and not do
2:46:58
-- Mhmm. -- I didn't know that suddenly
2:47:00
the one couldn't get him out of this difficult situation.
2:47:02
Yeah. Okay. So now I'm in problem of fantasy and
2:47:05
fiction, I think. Yeah. Yeah. That's it.
2:47:07
Now I think there's a worry about the
2:47:09
similarity account if you do this retreat. If you
2:47:12
just say, Oh, similarity. Well, I won't
2:47:14
tell you much about it. I won't tell you
2:47:16
the rules of the game. It's just whatever it needs to be to
2:47:18
make the counterfactual come out true, where we want them to come out true.
2:47:22
Well,
2:47:23
that's like this Harry
2:47:24
Potter as a wand theory. It's a
2:47:26
magic wand similarity. If ever you find yourself in a corner, you just like
2:47:31
change the rules and say, I know I meant something
2:47:33
else. Yeah. That's right. Or or you you just don't
2:47:35
specify the rules in the first place and so there are there are no counter examples. Mhmm. Okay. that's
2:47:39
the Harry Porto Horn of a dilemma
2:47:42
-- Yeah. -- or if you do the the more, I think, philosophically honest thing like
2:47:44
Lewis did,
2:47:44
and
2:47:47
just say, look, I'm gonna try hard to
2:47:49
tell you what matters to similarity, then I worry they're gonna be counter examples like the stretching, the
2:47:52
case. Yeah. Okay.
2:47:55
So some people in the audience, I think,
2:47:57
might understand what we think. Like, we're we're being awfully fussy here about, you know, exactly what the
2:47:59
what kind of factual are we referring to in these cases where in
2:48:05
actual reality, if two people are having a conversation and made statements
2:48:07
like this, there would be
2:48:09
no confusion about what they're referring to --
2:48:11
Yeah. -- to defend the fussiness for a minute.
2:48:13
The
2:48:14
challenge here is that in these, like, everyday cases where someone says, like, nicks and pressed the button, then why, you know,
2:48:16
if I'd scratch my finger yesterday,
2:48:18
then said, intuitively, we know what
2:48:21
what we're what we're communicating through
2:48:23
all of the context. but
2:48:25
we're going to try to develop rules
2:48:27
about the nature of counterfactuals and, like,
2:48:29
what reasonable, like, logic you can
2:48:31
apply to that. from these cases. And then we're gonna
2:48:33
start applying it to, like, very unusual cases, to,
2:48:36
like, strange things like, you know, if we can
2:48:38
conceive of pees on bees, then like other thing -- Yeah. -- where we're not gonna have the same intuition about what the situation really
2:48:40
is or, like, what
2:48:42
logic can reasonably apply.
2:48:44
Yes. And if we
2:48:46
have if we develop rules
2:48:48
that, as it turns out, don't actually work in
2:48:50
all of the, like, cases where we do have an
2:48:53
intuition about what works and what doesn't. Yeah. Then we could,
2:48:55
like, extend this, like, incorrect logic all
2:48:57
these other cases where we're not gonna be able to see
2:48:59
the error. Yeah. I think that is like one reason why
2:49:01
we really would ideally like to pin down what isn't isn't legitimate when you're
2:49:03
doing kind of natural reasoning. In
2:49:05
the cases where we might be able to see mistakes.
2:49:07
Mhmm. So then in case of what we can't see
2:49:09
the mistakes, we feel more more solid ground. That's right. By the way, this is a very general problem
2:49:12
in philosophy I
2:49:14
think that we often fashion
2:49:16
our conceptual analyses to familiar cases,
2:49:19
and then we hope that they still
2:49:22
apply to these maybe more ricochet
2:49:24
cases. And sometimes, philosophers say, well, it spoils to
2:49:26
the vector. You know, in some far fetched case,
2:49:30
we haven't earned the right to
2:49:32
some firm intuition about these cases
2:49:34
because they're strange and just let the theory dictate what
2:49:38
we should say about those cases. On
2:49:40
the other hand, we talked about this earlier too. Sometimes we
2:49:42
really do look to the strange cases like Saint Petersburg to stress test
2:49:47
an analysis that was formed to handle more
2:49:49
familiar cases. Okay. But let let me take the other angle for a minute. We'll say language is
2:49:52
always mess sees.
2:49:55
Like, people are constantly saying things that, like,
2:49:57
technically, aren't true, but, like, they they communicate the broad thrust of the argument, and people
2:49:59
know what they're saying. Yes. here
2:50:03
you're applying like a real fine to combed everyone's
2:50:05
statements and saying, oh, this isn't precisely right. But that's because you're taking the kinds of things that people
2:50:07
might say in ordinary life and then treating them as if they're like, factual
2:50:13
to a level that they were never designed to be. That's
2:50:15
an important part of my overall
2:50:18
view about counterfactuals. I think that the
2:50:20
things we say, the counterfactuals that we
2:50:22
use in daily life are mostly loose talk and let me say something
2:50:24
about
2:50:27
that. I think that
2:50:29
they're false but approximately
2:50:30
true and close enough to the truth that they convey
2:50:34
useful information. And I think this
2:50:36
is perfectly familiar perhaps soon, I'll say something about
2:50:38
why I think most artifacts are false.
2:50:41
and And
2:50:42
people think this is some
2:50:45
crazy, radical view. And I say, no, come on. Most
2:50:47
of what we say could be false -- Yeah. -- for completely
2:50:48
intelligible
2:50:53
reasons. Sometimes we're joking. Sometimes we're
2:50:55
being ironic. Sometimes, yeah, we're exaggerating a bit loose talk. I a common explanation.
2:50:57
Just
2:50:58
think about the
2:51:00
cases not involving counterfactuals.
2:51:03
We say things like six
2:51:05
and a half million people
2:51:07
have died of COVID,
2:51:08
sadly. Well, we don't say really,
2:51:11
so six million specific time. five
2:51:13
hundred thousand. No. And of course,
2:51:15
we didn't intend to convey that
2:51:18
what we meant was something like
2:51:20
roughly six and a half million people have
2:51:22
died. And that's approximately true even though the
2:51:24
exact truth is something slightly different or
2:51:26
someone says, I'll I'll be there at eleven.
2:51:28
Exactly. Yeah. Exactly. The probability of
2:51:30
arriving at any specific instant is zero.
2:51:33
That's it. See how a familiar phenomenon
2:51:35
this is. Mhmm. No one truly
2:51:37
arrived at eleven when they said
2:51:39
they'll arrive at eleven. Of course,
2:51:41
we we charitably understand them roughly eleven. We
2:51:44
think of the claims
2:51:46
say tennis balls are spherical.
2:51:48
Mhmm. Well, they're not
2:51:49
in the
2:51:51
mathematical sense, spheres because that has a
2:51:54
very specific meaning. They're approximately
2:51:56
spherical.
2:51:56
And close enough to
2:51:59
spherical that
2:51:59
we can treat them as if there's fears
2:52:02
for for most practical purposes. For
2:52:04
example, I'm packing a box with
2:52:05
tennis balls and I wanna calculate how
2:52:07
many tennis balls Can
2:52:09
I fit in the box? I
2:52:12
won't go far wrong if I
2:52:14
treat the balls as perfect spheres because they approximately
2:52:16
aren't. then I
2:52:18
do the calculation using perfect spheres and
2:52:20
it'll work well enough for the approximate
2:52:22
spheres, the tennis balls really are. Yeah.
2:52:24
Okay. So if
2:52:25
we kind of relax our attempts to
2:52:27
come up with a super precise theory of kind
2:52:29
of factuals like the most similar world and now and
2:52:32
we accept that they're fuzzy
2:52:34
and a bit messy and contextual and so
2:52:36
on. Where does that leave us as philosophers
2:52:38
or as people using counterfactual reasoning? First, I've distinguished the truth values of the counterfactual's
2:52:42
from the way we
2:52:44
use them, the assertability, the acceptability of them.
2:52:46
Should I go into each of those aspects? Yeah. Yeah. I'm not sure what this
2:52:47
means. Yeah. Alright. Good.
2:52:48
Maybe
2:52:52
first I'll hit you with a shocker -- Yeah.
2:52:54
-- that most counterfactuals are false. Mhmm.
2:52:57
And people will think I've lost my
2:52:59
philosophical marbles. And then I'll soften the blow
2:53:02
by saying something about assertability and -- Yeah.
2:53:05
-- acceptability. Go for
2:53:06
it. Great. For start, most
2:53:08
counterfactuals are false. consider the
2:53:10
coin in my pocket. Let's assume it's a fair coin.
2:53:13
I'll never
2:53:14
toss it. If I were to
2:53:15
toss it, it would land heads.
2:53:17
hotels, it would land heads. Doesn't seem right? That doesn't seem right. Thank you. I don't
2:53:19
think that's
2:53:22
right. I think that's false. And why? Well, it's
2:53:24
a chancy coin on imagining -- Mhmm. -- and
2:53:26
it if I were to toss it, might land heads and might land tails. Alright. Now let's
2:53:32
make
2:53:32
the coin heavily biased to Heads, let's
2:53:34
say ninety nine percent chance of heads, one percent chance of tails. If I
2:53:39
were to toss the coin at wood,
2:53:41
land, heads, not tails, Still bad. Yeah. Let's say it still
2:53:43
might land tails. Consider a
2:53:47
huge lottery. Let's say it has
2:53:49
a million tickets that's never
2:53:51
played. if the lottery were
2:53:51
played, ticket number one
2:53:52
would lose. Mhmm.
2:53:54
I say no.
2:53:55
And and notice
2:53:57
by the way the
2:53:59
problem there If you say that of
2:54:01
ticket number one, it seems you better say about
2:54:04
ticket two, ticket three would lose, blah, blah, blah,
2:54:06
blah, ticket number million would lose. seems you to ticket lose. that has win. One's there's
2:54:08
got to be a winning
2:54:10
ticket. So that in
2:54:12
fact, you contradict yourself if
2:54:14
you said all of that. Yeah. Okay.
2:54:17
And now consider
2:54:18
your favorite intuitive,
2:54:20
common sensical counterfactual.
2:54:23
I'm holding a cup if
2:54:24
I were to release the cup,
2:54:26
it would fall. Now, I know it's
2:54:28
very tempting to say that's true.
2:54:31
I still say it's false. because
2:54:32
it's a lottery. It's a
2:54:34
chance process, I say. If the
2:54:36
cup were released,
2:54:37
it might not fall because someone might quickly lace a table
2:54:39
under it, a very surprising
2:54:43
updraft of air might suddenly lift
2:54:45
it rather than letting it fall
2:54:47
positive things might happen. I know
2:54:50
some of them extremely
2:54:51
improbable. I don't I don't
2:54:54
mind. Just as in the lottery
2:54:56
case, it was
2:54:58
extremely improbable that ticket number one
2:55:00
would be the
2:55:01
winner. Yeah. Okay? So these things
2:55:03
aren't like absolutely certain,
2:55:05
it's not true in, like, every possible kind
2:55:07
of factual world that the cut does fall. I guess, some
2:55:09
people might wonder, it's like, does it really matter that in some, like,
2:55:12
infinite testimony action,
2:55:14
not like possible counterfactual worlds. The consequent doesn't
2:55:17
actually occur. Or or is this like or are you being a bit of juice here? Yeah.
2:55:19
III get that a lot. Yeah. Maybe
2:55:24
I am. I choose well,
2:55:27
I'm pedantic, but I I
2:55:29
do think that our use
2:55:31
of counterfactuals commits us to
2:55:33
this
2:55:33
kind of patentry in in various ways. Mhmm. For example,
2:55:36
look at
2:55:38
the logic of counterfactuals. Lotus Poonan
2:55:40
seems plausible, and that's that's
2:55:43
the rule if p then q, p therefore q. So modus ponents will
2:55:47
fail. It seems if you lower
2:55:50
the bar for
2:55:50
the chanceiness below one. One. So this is, like,
2:55:53
if p then probably
2:55:55
q Yes. p then q then,
2:55:57
that doesn't go through. Yeah. That that's it. Right? If you thought all you needed
2:55:59
for the truth of the
2:55:59
counterfactual was
2:56:03
the probability, high probability of q given p,
2:56:06
something like that, then you could
2:56:07
easily have a case where p is true, the probability is high, and
2:56:11
then it didn't happen. Like, it was
2:56:13
very probable that ticket number one would
2:56:15
lose should say better, the
2:56:17
ticket that, in fact, wins
2:56:20
in the
2:56:20
lottery, it was very improbable
2:56:23
that it would win but you
2:56:25
don't want to say if the lottery
2:56:27
were
2:56:27
played, it would lose just because its probability
2:56:29
of losing was high. Maybe this is a misplaced rant. But in general,
2:56:31
I I wonder whether logic
2:56:35
as it's like like formal logic as it's
2:56:37
taught, exactly things like f p, then q p, so q. Whether that stuff is as
2:56:39
useful as people like to think, because the world just doesn't
2:56:41
so
2:56:45
rarely affords us the opportunity to use these, like,
2:56:47
certain logical rules. Like, if you're
2:56:49
in computer science or something, maybe fine.
2:56:51
Yeah. But, like, in actual everyday life,
2:56:53
all we ever almost all we ever have is, like, if p,
2:56:55
then probably q. Yes. p, so probably
2:56:59
q. Yeah. That's right. And so that
2:57:01
kind of more informal reasoning is basically
2:57:03
Yeah. All all the the luxury that we're ever afforded. And then it means that, for example, like, strict
2:57:05
logical fallacies are not
2:57:07
so interesting. Instead, you
2:57:09
wanna be looking at,
2:57:11
like, probabilistic fall like, way in which, like, arguments
2:57:14
are not as powerful as they seem
2:57:16
like they might be and
2:57:17
so on. Yeah. That's right. Let's consider one logical rule, the
2:57:19
way that conjunction behaves.
2:57:22
If p is true and
2:57:24
q is true, then the conjunction
2:57:26
p an q is true and vice versa. Earlier, we talked about
2:57:28
some problems
2:57:31
to do with conjunction -- Mhmm. --
2:57:33
that was the lottery paradox and
2:57:35
the preference paradox, and where we have the interaction of, say, rational belief and conjunction.
2:57:40
So you might say in that
2:57:42
case,
2:57:42
look, rational belief is not bound
2:57:45
by a strict conjunction rule
2:57:48
because we saw a case where you could
2:57:50
rationally believe each of the sentences in your
2:57:52
book, but not rationally believe the conjunction
2:57:54
of them. actually something very parallel to
2:57:57
this is relevant to counterfactuals. The
2:57:59
way there seems to be a sort of conjunction introduction rule for counterfactuals. if
2:58:04
I were to release the
2:58:05
cup, it would fall. If I were to release the cup, it would break. Therefore, if I were
2:58:07
to release the cup, it would
2:58:08
fall
2:58:14
and break. Mhmm. So that's a conjunction introduction in the consequent.
2:58:17
I think that's
2:58:18
valid, and I want to --
2:58:20
Okay. -- perspective. And actually notice
2:58:22
how if you set the bar for probability less than one for the truth of the cat effectuals, you'll
2:58:28
violate this conjunction rule in the
2:58:30
consequent. I did it. In fact, for the lottery. If the lottery were played, ticket number
2:58:33
one
2:58:33
would lose, says
2:58:35
someone who thinks that
2:58:37
high probability is good
2:58:39
enough for truth. background.
2:58:40
So ticket number two would lose. And
2:58:42
now I can join all of the tickets
2:58:44
would lose. And notice I I did that
2:58:46
and you didn't blink. You didn't stop
2:58:50
me and say, hey, Al, you can't
2:58:52
conjoin those consequence like that. I think
2:58:54
it's very intuitive that you can. I did. And
2:58:57
then you get the absurdity that if the
2:58:59
lottery were played, every ticket would lose --
2:59:01
Mhmm. -- which we all agree is false. Yeah. So most kind of factors are false
2:59:03
in this respect.
2:59:05
this respect What
2:59:07
does that imply for listeners? Or what does that
2:59:09
imply for people who are using kind of factuals in in their
2:59:11
reasoning? Great. And this brings me to the next bit I wanted to talk about the assertability and
2:59:16
the acceptability of counterfactuals. When I
2:59:18
go around saying that most counterfactuals are false, a lot of floss is
2:59:22
think I've lost my philosophical
2:59:24
marbles. This is just crazy. Actually, just
2:59:26
a little sociological observation. Some of them think I'm crazy and some of them think, well, I've just given
2:59:28
some of them think i'm crazy and some of them think well
2:59:31
i've just given good arguments and that's
2:59:33
it's I'm exactly right. The former,
2:59:35
the ones who think I'm crazy. They
2:59:37
tend to be more philosophers of language
2:59:39
and philosophers of mind who think that
2:59:41
principles of charity and humanity are operative. And I'm
2:59:44
attributing to competence speakers
2:59:46
-- Mhmm. -- some deep confusion or
2:59:48
something. Yeah. And I must be just
2:59:50
getting how semantics works. wrong. The latter, who thinks that I've given conclusion's right.
2:59:53
They they tend to
2:59:55
be more philosophers of
2:59:57
science or or maybe
3:00:00
meta physicians who just take seriously,
3:00:02
well, this is what chance. What it
3:00:04
means is what implies for counterfactuals. We live in
3:00:06
a chancy world. That's just what you get. Mhmm. Now
3:00:09
how do I soften the blow? I don't
3:00:12
think falsehood is so scary. You know? I
3:00:14
don't think it's so troubling if if stuff that we say turns out
3:00:16
false Lots
3:00:18
of things we say. Lots of
3:00:20
things. And in fact, Gil Harmon
3:00:22
and Donald Davidson apparently did a bit of a early x five bit of experimental philosophy
3:00:28
I don't know if they ever published
3:00:30
this. They eavesdropped on conversations in
3:00:32
bars and cafes, just listening to
3:00:35
how much truth and false would do
3:00:37
people say in normal conversation. And they concluded that, remember, almost Much
3:00:39
or most of what we say is false for
3:00:44
completely understandable reasons. I
3:00:47
mentioned some before, sometimes
3:00:49
we're joking. We're exaggerating
3:00:51
loose talk is very important here. Sometimes
3:00:53
we we just have false beliefs. Sometimes we have true beliefs, but we just
3:00:56
take short cuts.
3:00:58
We just don't wanna say the full thing.
3:01:01
It would be wordy. And so we just cut to the chase. So
3:01:03
I think it's like that with counterfactuals. I I think they're in fact false. but
3:01:08
they're researchable. The ones that we think
3:01:10
are good. And I have a story
3:01:12
about that. Roughly well, first, I
3:01:14
should tell you my truth conditions, four
3:01:17
counterfactual themselves, and then then you'll see the rest.
3:01:19
I think that if p
3:01:21
where the case q would be the
3:01:23
case is true just in case the
3:01:25
chance of q given p is one at a suitable
3:01:27
time. And that
3:01:30
suitable time is when p was not
3:01:32
yet settled. things could have gone either way with respect to
3:01:34
p. Mhmm. It had positive chance and not p
3:01:38
had positive chance. Okay? Now that's very
3:01:41
demanding truth condition. And that's why I get results like if I were
3:01:43
to let go of the cup, it would fall is
3:01:46
false because the chance
3:01:48
isn't one, the falling
3:01:51
given, release it It's very close to one.
3:01:54
It's approximately one. That's the
3:01:56
key to my understanding acceptability
3:01:58
of counterfactuals. They're acceptable if they're approximately true, if the conditional is
3:02:00
close enough to one.
3:02:03
And that's very much
3:02:06
like stuff I said about
3:02:08
the tennis balls and about the
3:02:10
COVID cases. You know, these claims were
3:02:13
false pretty obviously, but they were useful. They
3:02:15
were useful. They're approximately true good
3:02:18
to reason with acceptable. Acceptability is
3:02:20
a story about what's what's useful
3:02:22
to reason with. And they're a certable
3:02:25
And now let let me
3:02:27
give the story about assertability
3:02:29
for counterfactuals. We take some possibility
3:02:31
seriously in our conversations and
3:02:34
that's a context sensitive matter.
3:02:36
And I say that if p with the
3:02:38
case, q would be the case, is certable
3:02:42
just in case the conditional
3:02:44
chance is one and it's the
3:02:47
following conditional chance, q given p and the serious possibilities, the stuff that we think
3:02:53
our live possibilities in our
3:02:55
conversation. That's context dependent. Mhmm. I
3:02:57
don't think that the truth
3:02:59
of counterfactuals is context dependent -- Okay. --
3:03:01
that everyone says that nearly, but me
3:03:04
but I think that the assertability is
3:03:06
and I think it depends on, like,
3:03:08
adding another conjunction additional condition. That's an extra
3:03:10
condition about what possibilities we're taking
3:03:13
seriously. In a normal context, we
3:03:16
don't take seriously the released
3:03:18
cup suddenly being saved by a table or an updraft.
3:03:22
But if I then draw
3:03:24
your attention to these possibilities,
3:03:26
then the context shifts. And I think now it would start adding a bunch of conjunctions saying,
3:03:30
and no thermo dynamic verticals. And like no one
3:03:32
catches it. And -- Yeah. -- that's it. And
3:03:34
then it becomes even unassertable because in that context, we're taking seriously these rather
3:03:37
odd possibilities. Yeah.
3:03:39
But I I
3:03:42
now locate the context
3:03:44
sensitivity in the assertability
3:03:46
conditions and not the truth.
3:03:48
And I think we can explain away
3:03:50
the intuitions that these counterfactuals are true
3:03:54
My error theory is they're
3:03:56
false, but they're researchable. They're acceptable.
3:03:58
They're good to reason with. And that's because they're approximately
3:04:00
true and
3:04:03
that's because the conditional chance is
3:04:05
close to one. It's like the number of COVID deaths is close
3:04:07
to six point five million. So a
3:04:12
lot of this is currently kind of resting
3:04:14
on you that we live in an indeterministic world, which I guess we have good reason to think that we do, but
3:04:16
but hypothetically
3:04:20
if we that's a study doing kind
3:04:22
of factors. Yeah. If we lived in a
3:04:25
deterministic world, would a bunch of this problem
3:04:27
go away? I think it doesn't. And
3:04:29
so I think even under determinism, our counterfactuals will come up
3:04:31
mostly false. And it
3:04:35
is actually a live debate and
3:04:37
some physicists actually think the world is deterministic. Even quantum mechanics
3:04:39
can be given a deterministic Yeah.
3:04:43
It's just some like interpretation. It turns out
3:04:45
that things collapse for reasons that we don't
3:04:47
yet know exactly what it is and there's only one path or certain interpretations bones and deterministic.
3:04:52
That's a live debate, and and
3:04:54
some physicists might say I've taken
3:04:56
the wrong wrong term. Turn here.
3:04:58
But now let let's take
3:05:00
the deterministic case Now I think there's a different
3:05:03
problem. It's not so much choanciness. It's
3:05:06
what I would call unspecificity. Let's take
3:05:08
a case that floss was a
3:05:10
fond of talking about, Sophie sadly missed the big parade of,
3:05:13
I think, baseball
3:05:16
players. At a
3:05:18
certain point, Pedro danced she
3:05:20
would have loved to see that. Sadly,
3:05:23
she missed that she didn't go to
3:05:25
the parade. If Sophie had gone to the parade, she would
3:05:27
have seen Pedro Dads. Alright? Now
3:05:30
let this be a deterministic case. There's
3:05:32
no chanceiness. And we don't know that might have gone to the
3:05:34
bathroom. She she might have gone to the bathroom exactly
3:05:39
just the wrong time or the case that
3:05:41
that is usually considered, she might have got
3:05:43
stuck behind a tall person. Mhmm. If so, if you'd gone to the parade and got stuck behind a tall person, she
3:05:49
would have seen Pedro dance. That seems
3:05:51
false. She wouldn't have seen. Okay.
3:05:53
Now I I wanna make a lot
3:05:55
of that. If she'd gone to
3:05:58
the parade, somehow or rather. That was the antecedent. She would have
3:06:00
seen specifically Pedro dance.
3:06:02
I say, well, no.
3:06:04
It depends on how
3:06:07
the antecedent is realized. If she'd gone
3:06:09
to the parade and got stuck behind a tall
3:06:11
person -- Yeah. -- she'd gone to the parade and
3:06:13
gone to the bathroom at the wrong time, it's not true
3:06:15
that she would have seen Pedro Dantz
3:06:17
and that might have happened. These are ways
3:06:19
of realizing the antecedent. For the same reason you
3:06:22
want to add in, I'm guessing a whole bunch
3:06:24
of additional conditions to the antecedent
3:06:26
lag. Yeah. And she wasn't stuck behind a tall
3:06:28
person and she didn't go to the bathroom. All of the stuff that is being assumed
3:06:30
in normal language -- Yes. -- because we engage in lose talk.
3:06:35
but actually should be there if we
3:06:37
wanted to formalize it. Exactly right. And
3:06:39
now some people will say to me well, context will determine what should
3:06:42
be added or not. Mhmm. And I
3:06:45
say that's a matter of assertability. But as far as the truth
3:06:47
goes, I take the counterfactuals that face value, you know, you said if
3:06:52
she'd gone to the parade, she
3:06:54
would have seen Pedro dance. Now
3:06:56
you're telling me that defeat her
3:06:59
of her seeing Pedro Danes is getting
3:07:01
stuck behind a tall person, and she might
3:07:03
get stuck behind a tall person. So I
3:07:05
say it's not true that she would have
3:07:07
seen Pedro Danes She might not have. She
3:07:10
might have got stuck. Yeah. Are there any
3:07:12
any other approaches to kind of factors
3:07:14
that it's worth people having in mind or
3:07:16
alternative ideas that people are for for it to make
3:07:18
sense of all of this? Sure. Yeah.
3:07:22
An important alternative approach to counterfactual involves
3:07:24
conditional probabilities. And I like
3:07:27
this approach. I'm thinking
3:07:29
of Adams, Edgington, Skirms.
3:07:31
Now, first aid, think that
3:07:33
counterfactuals don't have truth values, but they can
3:07:35
be assigned probabilities.
3:07:38
According to atoms, for example,
3:07:40
your probability for if p
3:07:42
would queue is your conditional probability of queue given
3:07:46
p before learning that p
3:07:48
was false. And for
3:07:50
Edgington, the correct probabilities, the conditional chance of q given
3:07:55
p just before it turned
3:07:57
out. The p and Skirm's identifies
3:07:59
the counterfactual's assertability for
3:08:02
you with your
3:08:05
expectation of the
3:08:08
conditional chance. According to these accounts, counterfactuals don't
3:08:10
have truth conditions. Lightgab does give
3:08:12
them truth conditions. He says if p
3:08:14
would queue is true, just in case,
3:08:18
the conditional chance is high. And
3:08:20
and my own accounts influenced by all
3:08:22
of these. I like conditional probabilities. I really like
3:08:26
conditional chances for counterfactuals. And
3:08:28
I really, really like truth conditions given in
3:08:30
terms of those conditional probabilities, those conditional chances. So
3:08:36
that's an important alternative. The basic idea
3:08:38
here is taking kind of away from
3:08:40
kind of the realm of strict logic
3:08:42
where you're like, if p then q,
3:08:44
and bring it into the realm of war
3:08:46
of probability or bayzianism or making claims
3:08:49
about correlations or associations between things. So
3:08:51
it's cause of relationships maybe we're
3:08:53
saying, if p then q is more likely or something like that. And
3:08:55
it's also a
3:08:59
a good approach to reasoning and
3:09:01
arguing with counterfactuals. And this goes hand in hand
3:09:03
with this Adam's alternative
3:09:06
approach to thinking about reasoning. So
3:09:08
don't think in terms of
3:09:11
classical validity, which is truth preservation. Because again, the thought
3:09:14
is that these conditionals, in
3:09:17
this case, counterfactual, don't have
3:09:19
truth values. Mhmm. But we want a reason in such a way that
3:09:23
we go from high probability premises
3:09:26
to high probability conclusions. We don't want to have our premises being probable
3:09:31
and our conclusion being improbable. And
3:09:33
Adams has this nice theory
3:09:35
of so to speak high probability preservation. And fits in with
3:09:37
that whole alternative program.
3:09:40
So it's not classical
3:09:42
validity. It's not truth values.
3:09:45
It's high probabilities. Okay. Let's push
3:09:47
on and think about maybe an
3:09:49
application of this set of objections.
3:09:52
I think you you you
3:09:54
reckon that this this sort of reasoning
3:09:56
about counterfactuals or recognizing the trouble that comes
3:09:58
with counterfactuals that can potentially present a problem for
3:10:01
a flavor of utilitarianism called objective utilitarianism
3:10:03
or, I guess, objective consequentialism of M and A
3:10:05
kind. Yes. I guess. Most of what I consequentialism is when
3:10:07
when you judge the
3:10:10
value or, like, the goodness of actions or
3:10:12
what would be right to do based on the consequences that
3:10:14
they have. Most people have heard of consequentialism in some form, but what is objective consequentialism? Roughly
3:10:20
this, Action one is objectively better
3:10:23
than Action two, if not least,
3:10:25
the consequences of Action one, better
3:10:27
than those of action two.
3:10:29
And I think here we're imagining really the long term consequences, not just the immediate consequences,
3:10:32
but really perhaps
3:10:37
to the end of history. Yeah. And now we get into
3:10:39
a big discussion, which
3:10:42
I know is is close to
3:10:44
the hearts of many listeners many
3:10:46
listeners. about the long term consequences of what we do. But anyway, what I'm
3:10:49
about to say, I
3:10:51
think, will generalize beyond
3:10:53
just objective consequentialism, but
3:10:55
that's a good place to start.
3:10:58
Alright. So let's take a case.
3:11:00
You have a choice. You could help
3:11:02
the old lady across the street or
3:11:05
something else. Go to the pub. Yeah.
3:11:07
What should you do? What's the right thing to
3:11:09
do? Now, let's suppose, in fact, you take the old lady across the street. You
3:11:12
help her. I
3:11:15
don't have any problem with taking a
3:11:18
total of the all of the goodness, whatever the happiness or the
3:11:20
welfare after
3:11:23
that. I'm happy to rule out there's
3:11:25
a fact of the matter of the
3:11:27
total value, the total goodness, the consequences of that. but
3:11:30
what about the thing you didn't do? You did
3:11:32
not go to the pub. That's where my worry is
3:11:34
gonna kick in. Okay. First thing we we should make clear that this is a counterfactual the
3:11:39
way I I just stated it before,
3:11:41
notice the the carelessness of it. Action one is objectively better
3:11:43
than action two, if not only if the consequences of action one, are
3:11:48
better than those of Action two.
3:11:50
Well, in this case, Action two didn't happen. It was nonactual, and it didn't have any consequences.
3:11:56
So we must be talking about counterfactual
3:11:58
consequences. And now my worries of about counterfactuals
3:12:00
are gonna start to kick in. Yeah. Alright.
3:12:02
So let's take the thing you didn't do.
3:12:05
You didn't go to the pub. Well, case one,
3:12:07
the world is chancy. Well,
3:12:10
let's consider the very first
3:12:12
chancy coin toss that never
3:12:15
happened. how would it have landed if that coin had been tossed, it would landed heads,
3:12:21
not tails, heads, No. No. I say that I find
3:12:23
implausible it might have landed tails. Consider the
3:12:26
first lottery that never
3:12:28
happened if the lottery
3:12:31
had taken place But ticket number seventeen would have
3:12:33
won. No, I say, can't save any
3:12:36
ticket that it would have won it.
3:12:38
Some other ticket might have won instead. Alright.
3:12:40
I've hardly started. Now I
3:12:43
I know in the cluelessness
3:12:45
industry that this worry
3:12:48
about consequentialism, there's a lot of
3:12:50
discussion of how our actions
3:12:52
have these far reaching consequences.
3:12:54
There are these ripple effects It's
3:12:56
not like ripples in a pond that
3:12:58
tend to dampen down as you go
3:13:00
further out. No. These these just keep on
3:13:02
rippling for the rest of history. you
3:13:05
know, unborn children, which children would have
3:13:07
been born or not. You know, depends acutely
3:13:09
sensitively on the size of the population. I think I think the
3:13:12
things that that
3:13:15
we do, you know, very minor changes
3:13:17
in what we do. Alright. So now let's go back to the pub,
3:13:19
the hypothetical visit to the pub. The first child to
3:13:24
have been born thereafter. Well, the
3:13:26
child to be conceived, hypothetically,
3:13:30
that depends on which
3:13:32
sperm fertilizes an egg and it's a
3:13:34
lottery which sperm wins the race to the
3:13:36
egg. So there would have to be
3:13:38
a factor of the matter of which
3:13:40
sperm wins the lottery to fertilize the egg
3:13:42
to make it this child that would have
3:13:45
been conceived and not some other one that
3:13:47
would have been the winner of another. a
3:13:49
different sperm winning the race. And I've still barely started. That's
3:13:51
the first child. But
3:13:54
now consider that child's children and grandchildren
3:13:56
and great grandchildren and now the
3:13:58
rest of history. And all of the people who they interact with, and all the people even slightly speed up or slightly
3:14:03
delayed. All of that. All of that. That's right.
3:14:05
Mhmm. Now, I find that wildly implausible that there is a fact of
3:14:07
the matter. We're still considering the Chancy
3:14:10
case where all of
3:14:12
these Chancy Chancy processes
3:14:15
would be resolved in one particular way
3:14:17
and no other way, but it
3:14:20
makes a huge difference to
3:14:22
how we evaluate these artifactual histories
3:14:24
which way things go. So on
3:14:26
one hypothetical scenario, the children that
3:14:28
happen to be conceived, you know,
3:14:31
latter day, gandis, and Einstein's
3:14:33
and and a wonderful world follows.
3:14:35
And with just a small tweak, we
3:14:37
now get a different kind of actual
3:14:40
history with a
3:14:42
latter day Hitler followed by a latter
3:14:44
day style and in a horrible world, and
3:14:46
everything in between. And all of this, again, is acutely sensitive to how
3:14:51
things are initiated and also how the
3:14:53
chance processes go. Yeah. So there's a lot here. Let's let's say
3:14:55
step by step. So we're considering that ethical
3:14:59
question, like, should you help this all
3:15:01
laid across the road or should you
3:15:03
go to the And it turns out that from a, like, total,
3:15:07
like, forever consequentialist point of view, the question
3:15:09
of how would the kind of factual would be of you going to the pub is
3:15:11
going to depend on, like, the main timing
3:15:15
of exactly when you go to the pub, like,
3:15:17
what cars you happen to work in front of and and slightly delay. When you go in, like, order a beer versus or not,
3:15:19
because that could, like, that
3:15:24
ultimately actually will, we think given the setup
3:15:26
of the world, end up affecting the identities of old people potentially in future times it will
3:15:31
change the exact moments of fertilization and so
3:15:33
on. Yep. Now, given that the action of going to the pub
3:15:35
could be extremely good or extremely
3:15:38
bad based on like a one second difference of
3:15:40
when you do it. You're gonna say that we need
3:15:42
to specify much more precisely what the counterfactual is. We need to say
3:15:46
than this, like, extremely precisely specified set of
3:15:48
actions that you're going to go and engage
3:15:50
in because otherwise it's just not not defined. Yeah. Okay.
3:15:53
Now so far I've been assuming
3:15:55
indeterminism, but now let's assume determinism,
3:15:57
which I think is the best -- Yeah. --
3:15:59
for objective consequentialism. And
3:16:02
now we're imagining if given a
3:16:04
precise specification of the initial conditions
3:16:07
and the laws of nature, we
3:16:09
get an entire history that's determined thereafter.
3:16:11
That's the best case. And now, maybe it's more
3:16:13
plausible that there's a fact of the matter of what would
3:16:16
have happened had
3:16:19
I gone to the pub,
3:16:21
but now the problem of unspecificity kicks in. This is not choanciness
3:16:23
at all. This is under
3:16:28
determinism if I'd gone to the
3:16:30
pub, somehow or how exactly would
3:16:33
I have gone to the pub, and
3:16:35
it makes a difference, you know, would
3:16:37
I have entered the pub at 603
3:16:40
and seventeen
3:16:43
milliseconds rather than eighteen
3:16:45
milliseconds and given the acute sensitivity of what follows
3:16:47
thereafter, it matters
3:16:50
which of these initial conditions
3:16:52
is realized. So I say,
3:16:54
This parallels what I said about Sophie a moment ago. Even under
3:16:58
determinism, I just don't think there's a
3:17:00
fact of the matter given this loosely specified
3:17:02
and antecedent. I go to the pub somehow or other counterfactually,
3:17:08
that you get this very
3:17:10
specific consequence, it'd be this realization, you know, the seventeen
3:17:14
millisecond one perhaps, not some
3:17:17
other one. And this, by the way, is another example of
3:17:19
implausible specificity. I think even
3:17:22
under determinism, I think it's implausibly
3:17:24
specific. lose antecedent, all too tight,
3:17:27
consequence. Okay. So I'm gonna try
3:17:29
to rescue it here. So
3:17:31
the person says, well, when I meant
3:17:33
go to the pub, what I meant was this, like,
3:17:35
whole portfolio of outcome of, like, possible
3:17:37
ways that you could do that. Yes. And you'll
3:17:39
say, yeah, but that's like a massive mixture
3:17:41
of, like, extremely good and extremely bad outcomes
3:17:44
that, like, depending on exactly how you
3:17:46
specify those actions. Then we say, okay. Alright.
3:17:48
Well, Let's say that I could exactly specify
3:17:50
the exact path. I'm gonna, like, hypothetically, we could
3:17:52
write out in this book like perfect instructions of
3:17:54
the exact moments I should make each of
3:17:56
these movements. like, on a space, that might rescue
3:17:58
this issue of, like, indeterminacy, although it is
3:18:01
now a bit weird. And I should say
3:18:03
more about the weirdness. Yeah. Yeah. Yeah. Okay.
3:18:05
So now let's very precisely specify
3:18:08
the way I go to the pub.
3:18:10
And now and we're resuming determinism still,
3:18:12
which I'd rather not, but let's Let's
3:18:14
spot that. Okay. So now given the initial condition precisely
3:18:17
specified, determinism, I
3:18:19
am happy enough
3:18:22
that there is history
3:18:24
that follows than the one that's determined.
3:18:26
But now we have a different problem.
3:18:28
I'll call it clumsiness, and I'm trying
3:18:30
to rip off some cluelessness. way, it's
3:18:36
hard for us to know or
3:18:38
to have confidence reasonable belief in which of these histories would be realized. I
3:18:40
think the problem is much worse
3:18:42
than that. I think it's a
3:18:45
metaphysical problem that there's just no
3:18:47
fact of the matter of
3:18:49
what would be Now we fix
3:18:51
that one. And we fix that one
3:18:54
with the determinism and the exact initial condition. And I say, the problem is, I'll call it,
3:18:56
clumsiness. it's
3:18:59
just not under our control
3:19:02
to realize the initial conditions in one very precise way rather than another,
3:19:04
to enter
3:19:08
the pub exactly at
3:19:10
seventeen milliseconds after 603
3:19:13
rather than eighteen Yeah. It's
3:19:15
just not somebody is likely to do
3:19:17
the good one as the bad one because they're so adjacent one another. That's basically
3:19:20
if that it's
3:19:22
not under my volition to so
3:19:24
finely tune exactly what happens, you know, in this case
3:19:26
down to the millisecond, let alone all of extra
3:19:31
stuff, you know, exactly when I ordered the
3:19:33
beer and so on. Even if I knew what I ordered What I ordered do,
3:19:35
it would be this this is the trajectory I want to get on let's
3:19:41
say, the seventeen millisecond one, not the eighteen one.
3:19:43
It's just not something that's
3:19:45
an option for me. I can't choose
3:19:47
seventeen rather than eighteen milliseconds after 603
3:19:50
So to to step back, put this in the form of a dilemma. Now what are the things that we're evaluating
3:19:57
what are the objects of moral evaluation? On the one
3:19:59
hand, let them be these rather
3:20:03
coarse grained statements like, I help
3:20:05
an old lady, as opposed to I go to the pub
3:20:07
somehow or rather. So now it's it's
3:20:11
pretty plausible. These really are my options,
3:20:13
but now I say it's not plausible
3:20:15
that there's an entire counterfactual history would be
3:20:18
realized given that
3:20:22
unspecificity of the option.
3:20:24
That was the first horn
3:20:26
of the dilemma. Second horn
3:20:28
now let the options in inverted
3:20:30
commas be. It's very precisely specified things,
3:20:33
you know, the exact millisecond I enter
3:20:35
the pub and so on. Now I think it's more
3:20:38
plausible that there is a fact of the matter of the entire
3:20:40
counterfactual history but
3:20:43
it's not plausible that that's an
3:20:45
option that anyone could actually act on it. That's right. Indeterminacy or
3:20:47
a relevance. Yeah. Yeah. It's not not the appropriate evaluation of
3:20:53
a moral theory. It's not something that I can realize
3:20:55
as an act of my
3:20:59
evolution. Okay. So one way
3:21:01
of dealing with this problem would be, I guess,
3:21:03
to specify extremely precisely what
3:21:05
you're going to do. So with, let's
3:21:07
say, you know, you're gonna go to
3:21:09
the pub at this exact time with these exact movements or help the
3:21:11
person in these very
3:21:16
exact ways. And that can
3:21:18
potentially Maybe that's overly precise, but it might at least allow you then to say in principle
3:21:20
that there is a
3:21:23
specific consequence. I guess,
3:21:26
again, always assuming determinism here.
3:21:28
for the sake of for the
3:21:30
sake of the conversation. That's right.
3:21:32
Well, for start, what does determinism mean? It
3:21:34
means given an entire time slice of the world,
3:21:38
and the laws of nature,
3:21:40
the rest of history follows. But I
3:21:42
don't think that it's enough just to be told the
3:21:47
details of your going to the
3:21:50
pub, for example, even an exact specification of your pubbing, I say, doesn't determine a
3:21:52
unique history
3:21:56
thereafter, even under determinism. Because you're just
3:21:58
a tiny part of the world, you
3:22:01
know, in the sweep of world history,
3:22:03
you're just a speck I'm sorry
3:22:05
if this comes as news to
3:22:07
you. But even describing precisely the
3:22:09
details of your going to
3:22:11
the pub, you know, the exact
3:22:14
millisecond of your arrival and exactly
3:22:17
your movements thereafter. Duchess
3:22:19
Falls way short of determining the
3:22:21
entire world's initial conditions at that time and under determinism, it's the initial
3:22:23
conditions of the entire world. and
3:22:29
the laws that entail the rest
3:22:31
of history. So so
3:22:33
so put it powerfully, according
3:22:35
to determinism, a snapshot of the
3:22:37
world at the time will give you the
3:22:39
rest of history.
3:22:41
But I'm imagining we've only just
3:22:43
got a selfie. We've only just
3:22:46
got this little part of the world, you know, you entering
3:22:50
the pub. And that's not
3:22:53
enough to entail the rest of history.
3:22:55
You're just a tiny
3:22:57
part of a portion of
3:23:00
a small bit of a
3:23:02
fragment of a time slice of history. So so in order to specify sufficiently closely,
3:23:06
we would have to not only talk about
3:23:08
your actions, but I guess specify the entire
3:23:10
initial conditions of the universe or all of the different atoms that could affect all of the different ones, which guess it's
3:23:17
becoming an absurdly bloated set of instructions. we need
3:23:19
not just information about you, we need to know about the hypothetical, you
3:23:24
know, other people and the hypothetical
3:23:26
dogs and frogs and bees and trees and photons and electrons and
3:23:31
everything else, you know, hypothetical
3:23:33
pandemics and natural disasters And I claim you don't
3:23:35
get all of that for free out of
3:23:40
the specification of your going to
3:23:42
the pump. even precisely specified. Okay. So this is all very clever,
3:23:44
but it it does it does
3:23:46
do make it easier. Well,
3:23:49
let's just start. Let's just start
3:23:51
here. I'll take it. It it does feel
3:23:53
like some kind of tricks being played here. Like, surely
3:23:55
we can patch this theory. So let's just, like, try
3:23:57
to do it in a kind of common sense way.
3:23:59
So what did someone really mean when they said, if I got into would be better. What they were
3:24:02
talking about was like some portfolio
3:24:04
of different scenarios in which you
3:24:06
went to the pub and we might
3:24:09
sample from them. And then, like, hypothetically, we
3:24:11
could see what the four long term consequences
3:24:13
of them would be. And then, you're gonna
3:24:15
choose one of these different pathways somewhat at
3:24:17
random from this portfolio of things that
3:24:19
plausibly match going to the pub.
3:24:21
And then we're gonna average, I
3:24:24
expected value of the of the different ones that
3:24:26
you might sample from. And then you should do it
3:24:28
if, like, the expected value of the, you know, outcomes of these
3:24:30
different options that you might end up taking. In fact, in
3:24:32
practice, would would
3:24:34
would be positive versus negative. That seems
3:24:36
to fix it. It's I think a big improvement. I had
3:24:38
noticed you really had to change the view. Remember, the original view said
3:24:43
one should perform the action that has
3:24:45
the best consequences. It wasn't any probabilistic thing. It wasn't an unexpected value thing. Mhmm.
3:24:47
It was just the best consequences.
3:24:52
Yeah. And I've been challenging that
3:24:54
thought. Yeah. And in particular, the
3:24:56
thought that there's a fact or
3:24:58
the matter of the consequences for
3:25:00
something you didn't actually do. Now,
3:25:02
I like this much better where you somehow probability of this thing, you take the expected consequences
3:25:08
where that's a probabilistic notion. And by the
3:25:10
way, we're gonna have to talk about
3:25:13
what the nature of probability is. I
3:25:15
like how various topics that we've been talking
3:25:17
about, all of a sudden, but now the
3:25:19
interpretation of probability matters, I think here
3:25:21
the right way to go is it's
3:25:23
some sort of reasonable subjective probability
3:25:25
that that we're putting in at this point. But notice we have gone a long
3:25:27
way from the original objective
3:25:32
consequentialism. Maybe, yeah, for practical
3:25:34
purposes, maybe it feels quite
3:25:36
similar, but I suppose we
3:25:38
have now this quite foundational level of
3:25:41
the theory, we started sticking in
3:25:43
subjective probability judgments and expected value calculations. These things that originally distinct like philosophical
3:25:48
issues. Now our moral theory is this,
3:25:51
like, combination of consequentialism plus expected value, plus,
3:25:53
like, as to some view on, like,
3:25:55
how people ought to form views about
3:25:57
things will happen. Yes. So an unpleasant consequence of
3:25:59
this is that if
3:26:01
if we identify problems with, like, subjective
3:26:04
opinion forming or, like, how we aggregate
3:26:06
them to expected value or so on. Yes. That's this is, like, all really deep in the moral theory, and so it's gonna
3:26:08
create a problem for,
3:26:10
like, the the the
3:26:12
subjective utilitarianism as well.
3:26:15
Yeah. Spot objection to objective
3:26:17
consequentialism was that it's presupposes
3:26:19
this bizarre matter physics, I find
3:26:21
impossible. To be fair, I should
3:26:23
say just very quickly, some,
3:26:26
you know, really serious good philosophers do,
3:26:28
I think, believe that there are these, which we might call,
3:26:30
count effects, or he's definitely gonna call some count effects.
3:26:34
That's the matter of what would
3:26:36
happen given any antecedent for the
3:26:38
entire world. This goes back to the molyneux who thought that God
3:26:43
has these these knowledge of these
3:26:45
kind of kind of actuals. And it seems to
3:26:47
to go pretty well with the early storenacko view that there's is
3:26:52
a unique closest possible world
3:26:54
where an antecedent is realized. John Hawthorne runs with that few, Sarah Moss,
3:27:00
I've mentioned stories Stephanson, Richard Bradley, there there are a
3:27:02
number of people who have views, I think, in
3:27:05
this neighborhood. So III have to
3:27:07
take this view very seriously. even though I've said
3:27:09
I don't buy it. I don't buy it in the end. So
3:27:11
so far, I've
3:27:15
I've been making the point that
3:27:17
objective consequentialism presupposes this very questionable metaphysics. Mhmm. And and
3:27:19
even if it turns out
3:27:22
the metaphysics is correct
3:27:24
and it's defended by
3:27:26
all of these really good philosophers,
3:27:28
we should be upfront
3:27:30
that this ethical theory,
3:27:32
this theory about morality
3:27:35
has this deep metaphysical commitment. Now
3:27:37
I think you're making the next
3:27:39
point that if we fix the theory as
3:27:41
I think we should and we go probabilistic and
3:27:43
we involve, we get expected value
3:27:47
into the calculations. Now
3:27:49
the foundations are probability theory
3:27:51
and expected value theory And
3:27:54
we've talked earlier about some of
3:27:56
the problems there. And again, the moral theory has
3:27:58
got these foundational presuppositions. We want to make sure they're in good order. Yeah.
3:28:04
I see. So it makes I'd say,
3:28:06
like, solving the problems with that, like,
3:28:08
even more important. That's that's without that
3:28:10
consequentialism, like, just It doesn't doesn't even, like,
3:28:13
really make make sense. Yeah. Yeah. And quite
3:28:15
a while back, you asked me, why should we care about kind of effectuals and why should we care about probability
3:28:20
in the expected value theory. And now
3:28:22
we see yet another implication that
3:28:25
some moral theories are gonna depend
3:28:28
on how cat effectuals pan out. Some
3:28:30
of them will now depend on how probability
3:28:33
and expect the value turns out. And
3:28:35
by the way, just to finish an
3:28:37
earlier thought, I began by saying, I'll put my
3:28:39
objection as an objection to Mhmm. But I
3:28:42
actually think it's it
3:28:45
will generalize beyond that.
3:28:48
Namely any moral theory that
3:28:50
takes consequences at least
3:28:52
somewhat seriously. So even
3:28:55
a deontologist should take some account
3:28:57
of consequences, especially when mistakes are sufficiently
3:28:59
high. Yeah. I think virtually all
3:29:01
do think that likely consequences matter
3:29:03
as well. Yeah. Yeah. And especially given this
3:29:06
more recent movement about how our actions
3:29:08
have these huge consequences for for the rest
3:29:10
of history, I've got right. It's like,
3:29:13
someone's like, 0IIII shouldn't, like, steal
3:29:15
this thing. Of course, like, that ends up changing everyone's identity
3:29:17
for the same reason -- Yeah. -- going to the picture actually does. Yeah. You know, so it's like
3:29:20
massive consequence. That's
3:29:22
it. The deontologist who says, I ought
3:29:24
to keep a promise, but now imagine some history we're in so doing
3:29:26
it. Right? So doing it, you know, it was in Stalin's later
3:29:32
are created. So even the the
3:29:34
ontologist has to take some account of consequences. Similarly, a virtual ethicist, virtual ethics is partly a matter I hope,
3:29:41
promoting good consequences for people, but now we have to
3:29:43
take seriously these
3:29:46
worries about counterfactual consequences and much
3:29:48
of the stuff I said before
3:29:50
kicks in again. Yeah. Okay. Yeah. So it's really yeah. It's gonna impact any theories that give some way to consequences.
3:29:55
I suppose it might also if you totally
3:29:57
constrain your moral picture to dust being, like, it's wrong to steal. You shouldn't steal here and now. sufficiently
3:29:59
constrained it probably like, it it's not gonna
3:30:02
create a problem -- Yeah. -- because you can
3:30:04
enter you can say if I took this
3:30:06
course of action, I would be much less
3:30:09
likely to steal. Although it says now we're in the
3:30:12
likelihood, but like you could say, yeah. Well, I I'm
3:30:14
gonna specify a path of action where that doesn't include stealing. And
3:30:17
I can do that set of action sufficiently closely that
3:30:19
I'll be confident that I won't steal -- Yes. -- to
3:30:21
and so that's safe. But as soon as you start considering broader issues and play them,
3:30:23
this is gonna come back That's
3:30:26
right. And now this takes us back
3:30:28
to long termism. So, yeah, maybe if we're myopic and
3:30:30
just look at the very short term consequences of what we do the
3:30:36
considerations that I've mentioned don't
3:30:38
kick in. But if horizon, then
3:30:43
I think it's a very live issue.
3:30:45
Slightly reminding me of the paralysis argument that I talked about with Warren McCaskill in my second
3:30:47
interview with him a couple of years ago. yeah,
3:30:52
we guess we don't have to have to go into that now, but
3:30:54
it turns out if you're a deontologist and you
3:30:58
place different values on, like, harms that you actively
3:31:00
cause versus harms that you prevent through your actions. And you
3:31:02
can end up in a very sticky situation where a bunch
3:31:05
of lots and lots of actions are prohibited to
3:31:07
you because of their foreseeable unintended consequences, basically.
3:31:09
It has a similar flavor. Yeah. Yeah.
3:31:11
That's right. Although even
3:31:14
being paralyzed and just locking yourself
3:31:17
up at home so that you don't
3:31:19
have any effects on all these other
3:31:21
people that you can't control and you
3:31:24
can't foresee. Well, there'll still be consequences of your
3:31:26
locking yourself up at home. Both well,
3:31:30
unforeseen, there'll be cluelessness there.
3:31:32
clumsiness too by the way. That's right. The
3:31:34
exact way you you quarantine yourself.
3:31:37
And I think you you just
3:31:39
can't avoid it. Even what are you supposed
3:31:42
to do? Just stay at home and not even move. And
3:31:44
then you die and then people are gonna come
3:31:47
and hit. Oh, yeah. If you worry that. Get
3:31:49
your body. Get your body in that gonna have consequences as such to
3:31:51
rule and so it becomes gonna be unclear what it's stating. Yeah. Yeah. I think
3:31:53
that this is this is a
3:31:55
this is a this is a
3:31:57
huge cannibalization. So yeah. People
3:32:00
we wanna what we were just talking about, probably have
3:32:02
to listen to my my second interview with
3:32:04
Will. Yeah. But anyway, just to summarize, I guess, what
3:32:06
we've said here, I I think it's underappreciated how these methodical
3:32:09
positions, whatever, objective consequentialism
3:32:12
or some other kind of
3:32:14
consequentialism, maybe subjective with inspected value,
3:32:18
and even at the end, the
3:32:20
ontology and virtue ethics, they may
3:32:22
involve some very questionable metaphysics, I say, about
3:32:26
how the counterfactuals pan out.
3:32:28
Or if we now probabilify
3:32:30
them in terms of expected value,
3:32:33
then the foundations of probability
3:32:35
and expected utility theory, very relevant. Yeah.
3:32:37
Just as a reminder, people who might have forgotten, deontology
3:32:40
is the class
3:32:43
of ethical theories where it's it's like rule
3:32:45
based ethics, like, you know, you shouldn't lie, you shouldn't steal, I guess, classically, that that kind of thing,
3:32:48
especially prohibitions although
3:32:51
usually provisions that can sometimes be swamped by
3:32:53
other considerations that might be really important. Yep. That's right. Thank you. And I guess, virtual ethics is kind of the theory that puts front
3:32:55
and center cultivation of,
3:33:01
like, good qualities of the actor. I think, you know, you should try to be a courageous
3:33:03
I don't know. You should
3:33:05
be I'd try to be a kind person, things like
3:33:07
that. That's right. And then when you spell out these things,
3:33:09
so that there's somewhat sensible consequences will have to be part of
3:33:11
the picture. And
3:33:14
my worry is about the the
3:33:16
non actual consequences as before. Okay. So this has been
3:33:18
a very nice segment to, I guess, like, bring together almost all of the different threads that
3:33:21
we've that we've talked about through
3:33:23
the interview. It's been a lot
3:33:25
to build. And it's been lots
3:33:28
of would not be surprised if some people, including me,
3:33:30
might have to listen to this request to fully
3:33:32
to fully grasp their speech has gone ahead. We should wrap
3:33:34
up though. We've got other sessions at the conference to get
3:33:36
to. I suppose yeah. Final
3:33:38
question. We we we managed to get through our Lexing
3:33:40
on stage earlier. We're without without too much too much. We didn't do anything too embarrassing. Or
3:33:42
if we did, I guess, listeners won't know because we'll have cut it out. Yeah.
3:33:47
you worse have done, like, tons of presentations. There's tons
3:33:49
of stuff on stage over the years. And III
3:33:51
think has gone, like, I either incredibly well incredibly poorly for you. I
3:33:55
hope some of them have gone incredibly
3:33:57
well at best for others to judge. I can tell you about
3:33:59
one that went very surprisingly, shall we say, This
3:34:04
happened in Argentina as at
3:34:06
a conference and I gave a talk and as is
3:34:10
often the case, there was an empty
3:34:12
glass in front of me and and a bottle, filled
3:34:14
bottle. Towards the end of the talk, I was getting thirsty. Mhmm. q
3:34:19
and a began, and I I
3:34:21
poured poured the bottle into the glass. And then I had a nice big drink from the
3:34:24
glass. Yeah. and
3:34:28
my head nearly exploded. And my
3:34:30
first reaction was just to
3:34:34
spill over the audience.
3:34:36
And I thought I can't do that.
3:34:39
So I just swallowed. What
3:34:41
was it? Well, my head was
3:34:43
hit by a sledgehammer. Okay. What
3:34:45
the hell do they put in their water here
3:34:47
in Argentina? Alright. Well,
3:34:49
I have to explain behind me was
3:34:52
a whiteboard and you were supposed to
3:34:54
use marker pens and then you erase
3:34:56
what you've written on the board with
3:34:58
a bit of eraser fluid which is
3:35:00
one hundred percent Pure ethanol. Jesus. So I just
3:35:02
just had a big drink of of pure
3:35:08
ethanol. So I was
3:35:10
smashed. Do people notice? Sure. Tell us, listen. I mean, how would you avoid making a scene? And the first
3:35:12
question is still coming at
3:35:14
me and now, you know,
3:35:16
the the room is just
3:35:18
spinning. Okay. For me. And I
3:35:21
said, hey folks, just a a
3:35:23
word of caution. Next time, you
3:35:25
put an empty glass in a bottle
3:35:27
next to the glass. make sure
3:35:29
it's not cleaner fluid. Make sure
3:35:31
it's not cleaner fluid. Water would be
3:35:33
better. And in a way, it's sort of worked out.
3:35:36
Now normally, I
3:35:38
maybe get a little nervous at talks and
3:35:40
maybe in Q and A, perhaps jump in a
3:35:42
little too fast or something. But I was so
3:35:46
mellow. Right? I was so relaxed, you
3:35:48
know, because I was smashed. I'm just gonna answer the
3:35:50
question when I damn well feel like it. Yeah. So now I'm thinking
3:35:52
maybe a
3:35:55
future. When I give a talk at
3:35:58
a conference, I should bring along a little hip flask of pure ethanol just to relax me. serious.
3:36:00
Anyway, that
3:36:05
was one of the more surprising things to happen when I gave it
3:36:07
all. Yeah. Well, I'm
3:36:09
glad you didn't hurt yourself. It's I was
3:36:12
is it isn't that dangerous? It I think
3:36:14
it is dangerous. I was okay. Yeah. The story does end pretty well. Okay. Yeah. I think
3:36:19
I don't think there was permanent damage,
3:36:21
but it was a little scary. Yeah.
3:36:23
Yeah. I think we can say if there's any brain damage done, it has not been apparent to me. Thanks so
3:36:26
much. And fortunately,
3:36:29
we didn't have
3:36:32
such mishap doing this interview. podcast today has been
3:36:34
Alan Hayek. Thanks so much for coming on
3:36:36
the eight eight thousand hours podcast, Alan. Thanks
3:36:38
so much for having me wrong with this. If
3:36:40
you like that episode, you might
3:36:42
also like some of our other
3:36:44
more challenging episodes such as number
3:36:47
ninety eight, Christian Tasani on future bias and a possible solution to moral fanaticism or perhaps
3:36:49
episode eighty six, Hillary
3:36:51
Graves on Pascal's mugging,
3:36:54
strong long termism and whether
3:36:56
existing can be good for us.
3:36:58
Sticking with the practical philosophy theme. If you've
3:37:00
made it this far, there's a decent chance
3:37:02
you've heard of the Oxford philosopher Nick Bostrom.
3:37:04
who is pioneer in thinking about existential
3:37:06
risk and humanity's potentially grand futures. A
3:37:08
friend of mine has started a
3:37:11
podcast feed called Radio Bostrom, features professional audio
3:37:13
readings of Bostrom's papers and articles
3:37:15
on those topics, as well as
3:37:17
Wise philanthropy, the ethics of AI,
3:37:19
transhumanism, and the case for speeding up some investments in technology while slowing down others. If that sounds interesting
3:37:21
to you, just search for radio
3:37:23
bostrom in any podcasting app or
3:37:25
visit radio bostrom dot com to
3:37:28
learn more. Finally, as
3:37:30
I mentioned in the intro, if you're interested
3:37:32
in working with us at eighty thousand hours, we are
3:37:34
currently hiring a recruiter in order to help grow our team.
3:37:37
Not being able to hire fast enough is actually
3:37:40
one of our biggest bottlenecks as an organization. And
3:37:42
the person in this role will help grind hiring rounds, source candidates, and help improve our recruiting
3:37:44
processes. and
3:37:46
I'll probably be key to increasing our impact
3:37:49
in the coming years. This might be a particularly good fit
3:37:51
for someone excited about building a career in operations or HR. and
3:37:55
you don't need any previous experience with recruiting
3:37:57
to apply. If that sounds interesting to you, then you need to get your application in pretty quickly because applications close on November
3:38:00
the second. and
3:38:04
you can find out more about that position at eighty
3:38:06
thousand hours dot org slash jobs or
3:38:09
by checking out our blog eighty thousand hours
3:38:11
dot org slash latest. Alright. The eighty thousand
3:38:13
hours podcast is produced and edited by Karen Harris,
3:38:15
audio mastering and technical editing by Ben
3:38:17
Cordell and Ryan Kessler. Full transcripts and an extensive
3:38:19
collection of links to learn more are available on
3:38:21
our site and put together by Katie Moore. Thanks for joining. Talk to you
3:38:24
again soon.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More