Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
hello. Good evening, good afternoon, good morning,
0:02
wherever you are. This is Off
0:04
the Fence, a transatlantic podcast where
0:07
three guys in Madrid and
0:09
in the New York metropolitan area get
0:12
together and try to put the world to this week we're gonna
0:14
talk about developments in banking,
0:16
in US banking and specifically Apple
0:19
offering a savings account. It's
0:21
gonna actually pay a competitive interest
0:23
rate which should be interesting given the way
0:26
deposits have been moving in the US and
0:28
and other and other developments.
0:31
But Luis, this was your topic.
0:33
Why didn't you, why don't you take it, tell us what you
0:35
think what the issues are. Thank
0:38
you Peter. I think, we discussed this topic
0:41
in previous podcasts and.
0:44
My impression is that when you're presented
0:47
with such a fantastic spread
0:49
between deposits that are
0:52
site deposit that pay nothing and federal
0:54
concentrated 5% or
0:56
500 basis points spread between what
0:59
banks pay for deposits and
1:01
treasuries short data treasuries,
1:04
there will be somebody somewhere that will take
1:06
advantage of that to take market share
1:08
in deposits and do something with it. And
1:11
we have, we, we talked about it and we couldn't,
1:15
I couldn't figure out who would be the disruptor.
1:17
Apple yesterday broke
1:20
the ice. It is, in my opinion, a much
1:22
more significant attacker than
1:24
any of the attackers that we saw in the early 2000,
1:27
such as I n g Direct or
1:29
the online brokers. Various reasons.
1:31
First of all, apple as an, as the.
1:35
Owner of the operating system
1:37
in which a lot of people's lives work
1:40
has an enormous amount of subscribers
1:43
already many of whom
1:45
use different functions that the
1:48
Apple Pay app allows them
1:51
to use. It is also
1:53
a big contender in the buy now pay
1:55
later layaway system of buying
1:57
for retail purchases. And
2:00
therefore the fact that they can get funding
2:02
at a competitive rate for them and
2:05
maybe not for them, maybe somebody else
2:08
who does the actual financial management
2:10
of both the asset and the liabilities of
2:12
these in, of this
2:14
book. I think it's quite a threat
2:16
to some lenders to
2:19
consu to some consumer lenders
2:22
in particular. Unsecured
2:24
consumer loans. One has to
2:27
believe that it won't be
2:29
too long before Amazon thinks
2:31
that maybe this is also their domain,
2:35
that they should probably get into this. And
2:38
perhaps other of the platforms
2:40
that are involved in either payments or
2:43
consumer Retail might,
2:46
will also develop their own special,
2:49
the interesting savings products
2:52
with this. I think the
2:54
one belonged before this trend moves
2:56
from west to east as these trends generally
2:58
do. And discuss with
3:00
you guys in previous podcast. I
3:03
believe that the current excitement with cyclical
3:07
Companies, specifically banks
3:09
in, in the Eurozone might be
3:11
a little bit ahead of itself because,
3:14
this would be one of the threats
3:16
that we could think of for that
3:19
markets. What do you think? I,
3:24
go ahead Alex, please. No, please,
3:26
Peter. I'm having
3:28
a look on online right now. What big
3:30
banks. Our paying is an average
3:32
rate, and it's about 0.1%.
3:36
And most if you look across
3:38
smaller banks, they're, they get to around
3:40
3%. I see you have some
3:42
special offers. PNC has an account
3:44
at 4%. But this
3:47
is a competitive move. Banks
3:50
are slow to raise deposit rates.
3:52
We know that they're a lot quicker to raise their lending
3:55
rates. It's which is natural.
3:58
Goldman Sachs, who will be servicing
4:00
this, took a stab at retail
4:02
banking in a, with a unit called Marcus that
4:05
hasn't gone so well. It hasn't
4:07
been unwound, but it's been it's been reduced
4:10
that back. Apple and Goldman
4:13
have been working together for a while. And
4:16
especially on the card side. I,
4:18
and, but this is a new and interesting move.
4:21
It's interesting also cause Apple has none
4:23
of the capital or I
4:25
believe liquidity concerns that
4:27
other banks might have
4:30
in, in backing up a savings account.
4:32
And this has become an issue in the last few, we, few
4:35
months as because Yeah. What sovereign
4:38
Silicon Valley Bank was was liquidity.
4:41
The liquidity can move incredibly
4:43
rapidly. And this is a huge
4:45
concern for regulators and supervisors.
4:47
Supervisors right now, I don't think,
4:50
I don't think Apple offers a deposit cause Apple
4:52
doesn't own a bank. I think the deposit
4:55
have to look into the small print is probably deposit
4:57
at Goldman Sachs or some third
4:59
party bank. Apple just
5:01
is the facilitator. I think. The marketer.
5:04
Yeah. But that's the whole point, right? So few things.
5:06
That's right. So the re what
5:08
I'm trying to say, Alex, is that the regulatory constraints
5:10
on the deposits will be with the department taking
5:13
bank will be with Goldman. That,
5:16
that's the whole question. The funny part is that Marcus offers
5:19
today 3.9% versus four
5:21
and change percent. You're not gonna your
5:24
own product are you? Yeah, of course. Exactly.
5:26
You're gonna love that. And it shows you the power of the thing.
5:29
So a few things you mentioned at the beginning,
5:31
the impetus for this and it makes sense is,
5:33
the fact that the spreads have widen significantly
5:36
and that the banks being slow to raise rates,
5:38
as Peter mentioned. Given that we've had, a
5:40
good 10 years of customers being used
5:42
to not earning anything in their deposit
5:45
accounts. This is, marketing
5:47
wise kind of an important move for a.
5:50
National player as opposed to a regional player. Regional
5:53
recently certainly have had to fight for
5:55
the deposits by paying higher rates. It's
5:57
a move that makes a whole lot of sense. And then actually like similar
6:00
to SVB in some ways, what you worry about
6:02
is with an inflow of significant
6:04
amount of savings in all one swoop
6:06
in a fairly big
6:08
marketing push by a very important
6:11
player that has a lot of customers, you
6:13
really wonder where, where those deposits
6:15
are going to go and who's taking the gap
6:17
book decision. You'd hope that it's
6:19
not Apple, as you mentioned, since they
6:22
don't have expertise, credibility,
6:24
or even a structure to do it. So as you say,
6:26
most likely somebody else is doing it. And if it's Goldman,
6:29
fantastic. It's Goldman Sachs, I mean
6:31
at the Apple. There you go. Starting
6:34
today, apple Card users can choose to grow their
6:36
daily cash rewards with savings account from Goldman
6:38
Sachs, which offers high yield annual
6:40
percent in yield, annual percentage
6:43
yield of four point 15% rate.
6:45
That's one than 10 times the national average. No
6:47
fees, no minimum deposits. Yeah. It's
6:49
a perfect marketing thing, but the argument that I'd
6:51
like to make, just to move the conversation elsewhere
6:54
is that for sure we have an opportunity
6:56
right now because we've had this, the rates back
6:58
up and expectations have been very low and it takes
7:00
a while for customers. Customers will look at this as being a good
7:02
deal. That's great, but my argument is a different
7:04
one, which is that Apple should.
7:07
Itself, long term
7:09
B, in the transaction
7:12
processing business, because one
7:14
of the main factors that keeps the wholesale
7:16
rate of transaction processing in the US at 2%
7:19
right is fraud. And
7:22
if there's one thing that the marriage of software
7:24
and hardware in your hand, has
7:27
been able to handle is biometric verification
7:29
and really getting an end-to-end between
7:31
the receiver and the sender of
7:34
a if not trusted, but. Verified
7:38
identity ecosystem, which is quite
7:40
strong. The apple has both the
7:43
physical hardware and the software
7:45
to run transactions.
7:48
Again, forget crypto, forget, how it gets handled.
7:50
But the idea is for them to be a
7:52
payment processor, given that they can ensure
7:54
what is probably the largest piece.
7:57
Of the puzzle in terms of the cost of wholesaling,
7:59
transaction processing I think
8:01
long term ease and attractive is
8:03
an attractive business. It could very well be that they wanna grow
8:06
their marketing shops at
8:08
this point, which makes sense, and turn that into
8:11
over time, building a real financial business
8:13
where they do transaction process. Not take, necessarily
8:16
back and balance sheet risk, but
8:18
just be there to make sure that you have the. Funds
8:21
in your account and that you're making an accurate payment.
8:23
That to me, I think is a long term play That would be a
8:25
significant source of growth for Apple. And
8:31
what does it mean for banks? I
8:34
don't know if you can hear me. Can you hear me? We
8:36
can. Yes. And just because of that, I'm gonna
8:38
answer it. The the problem with banks,
8:40
I think is that the physical existence
8:43
of a bank branch may have
8:45
outlived its relevance. And as
8:47
such, you have a lot of legacy
8:49
costs associated in a cost structure
8:51
that makes it fairly difficult to, to compete. Those
8:54
kind of banks are gonna be, this is not a new theme.
8:56
People have been talking about it for 20 years and it hasn't happened,
8:59
of course. But I think more and more my
9:01
sense is particularly as you get
9:03
to alternative ways of processing
9:06
transaction, which are almost
9:08
costless, it's gonna be very difficult for people to
9:10
accept the fees associated with traditional
9:13
banking. That's, and that I agree
9:15
completely. And the, and place where fees
9:17
are most obvious is
9:19
in payments. Yep. And for
9:22
years, banks basically had a monopoly
9:24
on payments. They no longer do. Yet
9:26
there's still competitive, there's still
9:29
room for new competition in payments and
9:31
and Apple is part of it. Others
9:35
are in the field and and banks are feeling
9:37
it and bank, especially in Europe. Agreed,
9:41
yep. I think if
9:44
I may say something very
9:46
obvious the big difference between. A
9:49
bank like i n g direct that
9:51
in the early nineties started collecting
9:53
deposits by paying a competitive rate
9:56
and then didn't know what to do with those deposits.
9:58
And the current situation is
10:00
that Apple also
10:02
originates an asset product, which is
10:04
the, these layaway loans. And
10:08
they are the platform where you can do both. Whoever
10:11
they partner with that runs the
10:13
both portfolios. We'll be
10:15
very keen to see that, this
10:17
works as, as well as it does into
10:20
the advantage of Apple and
10:22
the financial partner they have. And I think that
10:25
if they continue on this vein
10:27
and their opportunity for other products,
10:31
it's probably different in various parts of
10:33
the world, but, I am
10:35
going to go to meet some people
10:37
from Apple because, there is an enormous
10:39
opportunity for a brand like Apple to develop
10:42
a mass affluent independent financial advisor
10:45
business in Europe, which is
10:47
probably one of the areas of the market where there's the most
10:49
fat in the world. Yes. And
10:51
and this would be terrific for the
10:53
vast majority of the. Population
10:56
of the European Union or and other
10:58
Europeans. But what
11:01
I would also remind you of is that
11:03
a few years ago, Michael Milken, in
11:06
one of his appearances at one of
11:08
the conferences with, asked
11:10
to provide advice to some of
11:12
these platforms. And what he
11:15
told them is, don't get into finance because
11:17
the regulation will bug you down. I
11:20
talked this morning I should correct myself.
11:22
I exchanged WhatsApps
11:25
this morning with one of the most senior
11:27
people that I have
11:30
access to WhatsApp at eight o'clock in the morning
11:32
in finance. He was the chairman of a bank
11:34
in Spain. He was the vice chairman of one of the
11:36
top four investment banks for Europe. And
11:39
I asked him what, where he thoughts about this,
11:41
and one of the things he said is, this
11:44
company was doing very
11:46
well, UN, until or
11:48
has done very well. Let's hope they don't want to get
11:50
into the MOAs of being regulated
11:53
as a financial entity. And
11:55
I, and my, what I understand
11:57
is great about this deal with Goldman
11:59
Sachs is that Goldman Sachs already pays all the fixed
12:01
costs of being regulated as a financial entity
12:04
and if they can get the additional
12:06
business, Of being
12:09
Apple's joint venture partner in this,
12:12
everybody's better off and
12:15
including the public probably. So I think
12:17
it's a, it's an incredible, incredibly
12:20
I think it is as
12:23
material to the banking to the
12:25
retail banking world as the cash
12:28
management account in 1973 was
12:30
to the retail banking world
12:32
when Marin h introduced that. I
12:35
think you're right. And there's something else about it, which
12:37
is in recent years
12:39
I have heard an argument in banking,
12:41
which goes something like this. A bank does three
12:43
things. It takes deposits, it
12:46
lends money, and
12:48
it makes payments. It affects
12:51
payments. The payments business
12:53
it's going it's, if it's not gone, it's
12:55
going. Others have entered. You
12:57
don't really need a banking license to do it. They're
13:00
companies that can do it without the legacy. The
13:03
lending business is also is
13:05
also going private equity
13:07
and others are now lend money. Amazon
13:11
lends money to small and medium enterprises
13:13
and securitize to to fund it.
13:16
There's a lot going on in the lending space. Buy
13:18
now, pay later is done by a
13:20
non-bank, bank financial institution
13:23
or can be done. So there's a lot going on
13:26
in that space. But deposits,
13:28
taking deposits is where the regulation
13:31
kicks in when you're a deposit taking institution,
13:33
and thus others weren't gonna go into it.
13:36
The, they may try to create Anna,
13:38
the instruments that look like deposits, feel
13:40
like deposits, but they're not gonna be able to offer
13:43
customers guaranteed deposits because they don't
13:45
wanna be regulated. As Luis has just pointed out.
13:48
This seems to be a way in.
13:52
You, this partnership with Goldman Sachs
13:54
to to get a big tech into
13:57
that space, into the deposit taking space.
13:59
And these will be deposits because I'm,
14:01
they'll be guaranteed by the F D I C. So
14:05
I, I agree. It's, it
14:07
this is potentially a big move. I
14:10
agree. I think what's interesting about here is
14:12
you have two best of breeds, honestly, in, in
14:15
my mind, between Apple being knowledgeable
14:18
on the software hardware side, consumer
14:20
loyalty, privacy brand
14:23
and obviously Goldman Sachs on probably
14:26
regulatory management and all different financial
14:28
capabilities. So you have two best of breeds,
14:31
which for the moment early are,
14:34
have compatible. And corresponding
14:37
strengths in trying to address this. My,
14:39
my question would be long term, I
14:41
think to both your points, how
14:45
forward integrated into this, does Zapp want
14:47
to be long term? Right now it's easy.
14:49
It's a marketing deal. That's no problem. Longer
14:52
term, it's gonna be a business for them. But if it is, it's
14:55
a completely different business than what they do and
14:57
would be, enormously risky. And I think for
15:00
the investment base, for Warren Buffet, for everybody
15:03
else around would be viewed as something that
15:05
to due to gingerly. So it's gonna be very interesting
15:07
to see. And agreed. I think
15:09
Warren Buffet will probably not
15:11
be around long enough to see that, but
15:14
cuz just statistically speaking
15:16
but. He does like
15:18
banks and he does like insurance products.
15:20
And imagine Apple being hub
15:22
for banking and insurance products and
15:25
Warren Buffett sitting at the helm. I cannot
15:27
think of a more powerful platform to steward
15:30
retail products ever in
15:32
the history of mankind. Yes, agreed.
15:35
Completely agree. Again and
15:37
probably fairer, I think even for everybody
15:39
around, interestingly enough. That's
15:42
one would hope. But I can, let me tell you
15:44
what's happened in banking in a small
15:46
backward countries such as Spain, 20
15:49
years ago, there were, Peter
15:51
would probably know the numbers better, but
15:54
probably about 80, 90 banks
15:56
of national, of them
15:58
with branches in Madrid. Yeah.
16:02
Barcelona and some of the largest cities. And
16:05
banking services were competitively
16:07
priced because there was a significant amount
16:09
of competition. Nowadays
16:13
we're down to a handful of banks
16:15
that could, with large market shares.
16:17
I think the top five banks in Spain probably have,
16:20
I don't know, 50, 60% market share. Maybe probably
16:23
know the numbers better. I'd say more. More
16:26
so what's happened as
16:28
a result of the great financial crisis and the European.
16:31
And Nike Crisis is that with an enormous consolidation.
16:35
And as a result of that consolidation, people
16:39
expected that regardless
16:41
of where the ECBs set
16:43
the deposit rate they
16:46
would have Sorry.
16:49
They would, they would be able to pay 0%
16:52
for deposits. This is the argument
16:54
that nine, nine out of 10
16:56
bank analyst had to buy European banks,
16:59
that the jaws would
17:01
work exceptionally well in this cycle,
17:03
that there would be no increase
17:06
in the funding costs from deposits, and then you would
17:08
get all the benefits of. With
17:11
loans and the reason they could be sand
17:13
about such a strange
17:16
prediction was that there's
17:18
so much less competition. And
17:20
they tested that in the UK market, which
17:22
is a very concentrated market. And
17:25
the, until the hiccups of October
17:27
and even after that, there hasn't been a lot of
17:30
these termination of bank deposits in spite of
17:32
everything. I think that
17:34
the svb and the other two banks
17:37
with the plus the pretty sweet situation
17:39
may have changed that, sense
17:41
of, safety of having
17:43
a bank deposit for many people. And then
17:46
this message from Apple, and if I'm right,
17:49
and Amazon comes out within
17:51
the next few weeks and has its own deposit
17:53
product with some other financial
17:55
partner I, we are
17:57
going to have another financial
18:00
crisis in the making within a couple
18:02
of years I and
18:04
focused on deposits. I think that's one scenario.
18:07
I think you're right that there are fewer competitors,
18:09
so less competition. Another
18:11
element of what's going on in Europe is
18:13
the quantitative quantitative easing,
18:15
which is still going on. The money
18:18
supply is the
18:20
market is still a wash in liquidity,
18:23
and it's not until June
18:25
that the European Central Bank is going to
18:27
actually go into quantitative tightening mode.
18:30
Let's see if that, let's see if that changes
18:32
the dynamic on deposits. May I just
18:35
say something on that, which is that
18:38
part of the qualitative quantitative
18:40
typing in Europe has. It's
18:42
too complicated for a normal person to understand,
18:45
which is the reduction in tlt r o
18:48
Yeah. Funding and that is going quite
18:50
fast, right? So Yeah. And
18:53
then it was one of the key
18:55
programs for providing
18:57
liquidity to the banking system. So
19:02
maybe we can shift to a less
19:05
financial topic.
19:08
Let's look at ai. Oh,
19:10
fun. What,
19:13
Alex you wanna give us an update on how we're going
19:15
and getting to artificial general intelligence?
19:17
Agi, so in a godlike manner,
19:20
so enormously
19:23
a rapid sub update. I think
19:25
everybody's seen, obviously stable diffusion six
19:27
months ago DLI and so forth. On the image
19:29
side, everybody a couple of months ago on the G P
19:32
T three, the three five, then four
19:34
side. And it's important to understand
19:37
more or less, what happened. The
19:39
most important thing to think about, I think, and understanding
19:42
landscape is that for years people
19:44
have been publishing papers, working very
19:46
hard at trying to understand how
19:48
in a lab, how to use artificial intelligence.
19:51
And at the Genesis, this is back
19:54
20 15, 20 16, OpenAI was
19:56
built to try to do this independently
19:58
from the larger companies. So again,
20:00
Google, Facebook Microsoft many
20:02
people, one after. And Google
20:04
has two dedicated teams
20:06
on ai. Very powerful teams both going
20:09
after it. And the idea
20:11
when Dali was first released,
20:14
which was a shot across the bow to all
20:16
the larger companies. Back
20:18
in September and then when G
20:20
P T three was released and quickly
20:23
upgraded. What happened was
20:25
you now started to have an open
20:27
source, a set of tools that
20:29
people could use. The pricing of the a
20:31
p i of open AI was so
20:33
low or could remain so low that it provides
20:36
for an enormous amount of capability
20:38
for people to go and try things and to basically
20:41
offer. AI ish
20:43
or AI products, which are really just
20:45
the GPT APIs, meaning
20:48
that the ability to feed
20:50
onto G P T, whatever it is that you're doing. So
20:52
you've seen all kinds of other companies
20:54
larger companies come out with tools,
20:57
those that were working on AI before. So things
20:59
like Adobe and obviously Google
21:01
and everybody else has come up with stuff. And then those that
21:03
have just resold their bulk
21:05
access or wholesaling access to
21:07
to open ai. The
21:09
fact that it's open source is enormously
21:11
powerful because what it does
21:14
is that it takes away from people
21:16
in lab coats the control that they had
21:18
in some very interesting interviews
21:20
of Eric Schmidt from a couple years
21:22
ago, who was adamant that
21:24
all of his AI stuff, for all the reasons
21:27
you know, that ethicist will, will bring
21:29
up, should be kept very controlled
21:31
and inside the lab and so on, so forth. That
21:34
cat is. That Legion of cat
21:36
is completely out of bag by now. So
21:38
what you're seeing is that you're the
21:40
main limita. You're seeing a lot of development
21:42
in all kinds of different ways. Most of the activity
21:45
on GitHub, which is the place where developers
21:47
particularly open source of developers, share code.
21:49
It's now all ai all
21:51
the time, 24 hours a day which
21:53
is quite interesting. So you're starting to see a lot
21:55
of things being developed. The thing to think
21:57
about right now, I
21:59
think just to summarize it and I'll open it up, is
22:01
there are three things I think there are worth thinking about.
22:03
The first thing is that these
22:06
models have been trained a certain particular
22:08
with certain particular set of data. Some of it
22:10
cooperated again with a cutoff
22:12
date like about a year ago. I. But that's
22:15
not the future of models. Models are going to be
22:17
nightly builds. It's going to be something in
22:19
which, Morgan Stanley or
22:21
any organization where they have an internal
22:23
one or an external one facing customers
22:25
are going to have the, to retrain
22:28
or to cont continuously train on new information.
22:30
These master models that they're gonna have that're
22:33
gonna make available to customers or make available
22:35
to employees in order to be able to find
22:37
pretty much anything that you have as institutional
22:39
knowledge in your company. And
22:41
that's a gonna be a. Fairly large business that's
22:43
gonna require a lot of effort and handholding
22:46
because again, you're democrat democratizing a
22:48
a technology. The second thing is that
22:50
right now there's this concept of
22:52
how many tokens, or let's say how many words
22:54
you can feed. Into
22:57
a thread. So Che
22:59
g p t, for example, tells you that they are threads and
23:01
people and that the threads will remember what you said.
23:03
That's not entirely true. Basically, the way
23:05
to think about it is anywhere between, in the beginning 2000
23:08
words, which are really tokens, but let's just say words
23:10
for simplicity. 2000 words is
23:12
its memory, meaning that if you give it. 10
23:15
times 200 words. So if you ask it a question
23:17
for 200 words, then get a response
23:19
and ask it again, another question, or remember
23:21
the first 200 words and actually
23:23
the answer they gave you. So it has a little bit of memory
23:25
for a short period of time, but one of the
23:27
biggest problems out there is to try to give,
23:30
its more short. Term or long-term memory.
23:32
And so you've seen this development of what
23:34
they call vector databases, which
23:36
are basically ways in which you can add
23:39
some memory to the model with having to
23:41
retrain it. If you train the model,
23:43
you add permanently the data to the model.
23:45
That's a good thing, but very expensive, very complicated.
23:47
But the idea is that you want to have kind of a memory
23:50
buffer. The way to think about it is like if you have 10 years
23:52
of income statements or
23:54
50 product PDFs or anything
23:56
like that you really want those to be available,
23:59
not, to, to j to your
24:01
AI to share G B T. In
24:03
order to be able to answer correctly, again,
24:06
because these generalized models, can
24:08
answer generalized things, not specific things
24:10
about your knowledge. And then the third piece,
24:12
which is fascinating is, which has just started
24:14
the last couple of weeks, is this concept. There's
24:17
a very famous project called Auto G
24:19
P T, which has been Having
24:21
a lot of activity and recently, but basically is
24:23
the idea of user agents. So the idea is that if
24:26
you prompt a model meaning
24:29
you write it a
24:31
job description saying you are an expert
24:33
at marketing and you know everything
24:36
there is to do about the four Ps
24:38
and you know everything about all
24:40
this stuff, and you give it a persona, then
24:42
it's going to answer in that particular way. You
24:44
can replicate that in a bunch of different ways where you
24:46
spin up multiple models and
24:48
you give them different personas. So you can create
24:51
yourself a c e O persona,
24:53
which is, or a team leader persona
24:55
who is in whose job it is
24:57
to achieve certain goals. And then you
24:59
have it interact with other models who
25:02
are specialized in finance and operations
25:04
and marketing and HR
25:06
and whatever it is, or whatever the components
25:08
of your project is. And then you just,
25:11
let it loose, give it a, a goal,
25:14
and you just watch, while they have
25:16
all kinds of conversations about trying to
25:18
accomplish the goal at hand, again, it's very
25:21
early. But this gives you an insight as to how
25:23
you can get really
25:25
deep insight and thoughtful insight
25:27
from, from the systems, soon enough. So
25:29
to close it out, this is enormously
25:32
early. It is impressive.
25:34
It has made people realize this is not 10 years
25:36
from, 10, 10 years away, or 20 years away.
25:39
And it is just as
25:41
exciting as it is. Scary. That's all.
25:43
I'll say fascinating time.
25:46
Fascinating time. What do you think about the
25:48
petition to Pause? So
25:51
I think it is a admirable view.
25:53
There's a couple of interesting interviews.
25:55
Lex Frigman has a couple interviews, which are interesting,
25:58
if any, was interested on the
26:01
reasons for why you should take
26:04
the time to pause it. It
26:06
is laudable. It is never gonna happen. Okay.
26:11
I feel like I am the ludite, but
26:13
I talked to Chad GBT often
26:15
and I have a subscription. And
26:18
so far from its own account,
26:20
it tells me that, not
26:22
in the exact same words, but that it's a very
26:25
methodical librarian and
26:29
it doesn't have the ability to do
26:32
any original thinking
26:34
beyond what it can find in
26:36
the library and
26:39
what I thought would happen. And
26:42
what is happening with chat GBTs. Two
26:44
different things I thought that we were gonna go. So
26:47
this reminds me a little bit of what my friends
26:49
were doing out of engineering school in
26:51
the late eighties, early nineties, that
26:53
they were doing artificial
26:55
intelligence as an ency effort
26:58
to a lot of information into
27:01
databases so that a
27:03
processor could use the databases to come
27:05
out with answers that were
27:07
stock answers to stock question. I
27:11
thought when Anna and I were roommates
27:13
in New York and he introduced me to
27:17
a game called Sin City, that apparently
27:20
he developed some organizational
27:23
skills from having small
27:26
now I forget the words, but this is, goes
27:28
back to the early work of the Yes. Mathematic.
27:34
Yep. Which is that you
27:36
have, from similar initial conditions,
27:39
these, how do you call these agents?
27:41
I forget. Alex, these agents that
27:43
would find different paths to
27:46
do different things. You can call 'em agents. Agents.
27:49
Agents, yeah. And I thought that was far more interesting
27:51
cause they would come up with their own solution by
27:53
playing the game. And when you get to. The
27:56
game of chess. For instance, there were agents
27:58
that came up with a way to
28:00
play chess that was systematically
28:02
unbeatable. Yep. Some
28:04
people said, okay chess is a, is a
28:06
simple game because there's always an optimal solution
28:09
to any position. Let's try
28:11
a different game. And they try and
28:13
go, which is a Japanese game
28:15
of occupation, of a territory from
28:17
the enemy. You might
28:19
be committed with that game has
28:21
white beads and black beads and
28:24
it is, they nobody thought
28:26
that it would take, that it would be easy
28:28
to program a computer to play
28:30
Go. And I think it was within
28:33
like 36 hours a program
28:35
was able to be the world. Yeah. And
28:37
I thought for the, let
28:42
me interrupt you just for one second. Just to give you one
28:44
piece of insight, which is helpful cuz I remember
28:46
we, we financed a bunch of companies that were trying to do this
28:49
kind of stuff about 20 years ago, and it was a
28:51
mess and it was very much rule-based
28:53
and so on, so forth. But I would pause at the fall, the following
28:55
thing to try to reconcile the two views. Your
28:58
brain essentially learns the same
29:00
way as a language learning model. Let's
29:02
simplify, right? So a deep learning model, and
29:05
the reason I say that is because you as a
29:07
child, experience certain things.
29:09
So for example, if you put your hand. Onto
29:12
a hot stove, you will
29:14
associate or existing nerve
29:16
endings that were built earlier, even in
29:18
your life about moving your hand. And
29:21
then you will make a connection with the fact
29:23
that touching the stove, obviously
29:26
creates heat. And I,
29:28
if you were to be able to live as
29:30
many lives as. A
29:32
language learning model does, which is
29:34
millions and millions in lives in parallel, right?
29:37
So it's touching millions of stoves and
29:39
in certain particular ways, in every possible way. What's
29:42
happening is, which is interesting, is, and this is your point,
29:44
I think Compared to the way
29:46
people thought it would be done, which is by,
29:48
making sure that we understand all the rules, that we're a good
29:50
librarian, we know what section of the
29:53
library to go and look for whatever, where
29:55
the knowledge is stored. It turns out that language
29:58
these models have are matrices
30:00
of numbers and weights. That's all they
30:02
are. They're literally the equivalent of
30:05
these connections that if
30:07
this happens there, then
30:09
I should go this way, not that way.
30:11
Kind of thing. Super simple in
30:13
some ways, but at a level that computing
30:15
has only been able to make available
30:18
in the last couple of years. Let's just stuff that would've
30:20
been massively too big
30:22
to do. So the argument I would make just
30:24
in finishing is that
30:27
we have in some ways replicated
30:30
with the way the human mind. Develops
30:33
at the early age and even later stage
30:36
in matrices
30:38
in a mathematical algorithm, train
30:41
them. And so I would
30:43
argue that it
30:47
is able to learn. At
30:50
least as well as a human can. And I
30:52
know that's a very powerful
30:55
statement. We may not see it today but I think
30:57
the argument is that, that, we're,
31:00
the way that it comes
31:02
up with the way it thinks is very
31:04
similar to the way the brain, I
31:07
think. I think also if you, Luis,
31:09
you're probably using it a lot for our,
31:12
the, for the types of issues and topics that
31:14
we get into and are curious about, and.
31:16
And talk about and discuss. If
31:18
you ask it to write a short poem in the style
31:21
of Elliot about
31:24
the restaurant across the street, I think you'd find
31:26
you'd get a, you'd get blown away by
31:28
by what it can do stylistically. And
31:31
and, do I say it creatively?
31:34
Yeah. Yeah. Bloomberg trained
31:36
the model and when you look at the paper, they
31:38
just did, released it last week. And what's funny about
31:40
it is that the Bloomberg data, their
31:42
propriety data the Crown rules, those things are
31:45
the most important thing. Only counter
31:47
for 78 basis points for less than 1%
31:49
of the training data they put into it. So there's
31:52
a wall to go before you see something, but when
31:54
that model gets trained with actual Bloomberg
31:56
data, it'll be a site to see. I
31:58
think Think
32:05
that's, I think we're gonna, we're gonna call it
32:07
a day and
32:12
call it and say thanks to everyone. Thanks
32:15
very much. Thanks, Luis. And
32:18
take care, Alex. Thanks for organizing this.
32:20
Yes, thank you very much. Thank you. Bye.
32:23
Talk soon. Bye. Take care.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More