Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:06
Welcome to practical. A
0:08
I If you work in artificial
0:11
intelligence aspire to or are curious
0:13
how ai related tech is change
0:15
in the world, this is the
0:18
show for you. Thank you to
0:20
our partners! As Wide Io the
0:22
home of Change log.com Why transforms
0:24
containers in the Micro V Ems
0:27
that run on their hardware in
0:29
thirty plus regions on six continents
0:31
so you can launch your app
0:34
near your users. Were more
0:36
that why that I? oh. Welcome.
0:43
To another episode of Practical A
0:45
I. In this fully connected episode,
0:47
Chris and I will keep you
0:49
fully connected with everything that's happening
0:52
in the Ai world. Will take
0:54
some time to explore some of
0:56
the recent A I news and
0:58
technical achievements, and will take a
1:00
few moments to share some learning
1:02
resources as well to help you
1:05
level up your Ai game. I'm
1:07
Dana White Knack I am founder
1:09
and Ceo at Prediction Guard. And.
1:11
I'm joined is always by my co host
1:13
Chris Benson who is attack strategist at Lockheed
1:16
Martin. Are you doing Chris? Doing.
1:18
Great to they daniel get lots of
1:20
news that's come out this week. in
1:22
the Ai space I know early time
1:24
to talk about amazing new things before
1:26
stuff comes. our yeah I I've been
1:28
traveling for the past five days or
1:31
something. I've sort of lost track of
1:33
time, but it's like stuff was happening
1:35
during that time in the news especially
1:37
the Soros stuff and all that and
1:39
I feel like I'd just kind of
1:41
missed a couple news cycle soto be
1:43
good to catch up on a few
1:46
things, but. One. Of the reasons
1:48
I was travelling was I was
1:50
at the Tree Hacks hackathon out
1:52
at Stanford, so I went there
1:54
as part of their kind of.
1:57
Intel. Entourage I'm I'm had
1:59
prediction guard available for all the
2:01
the hackers there and that was
2:03
a lot of fun and it
2:06
was incredible. I it's been awhile
2:08
since I've been to any hackathon
2:10
at least in person hackathon and
2:13
they had like five floors in
2:15
this huge you know engineering building
2:17
of room for all the hacker
2:19
that in there was like sixteen
2:22
hundred people there participate on how
2:24
from all over there and really
2:26
cool a course like there were
2:29
some major categories. Of interest one
2:31
you know like in doing hardware
2:33
things with robots and other stuff.
2:36
Of course one of the
2:38
main areas of interest was a
2:40
I which was interesting to see
2:43
and in our. The. Track
2:45
Of It. I was a judge and
2:47
mentor in one of the cool projects.
2:49
That one that track was called Masterworks
2:51
So what they did and this is
2:54
old news to me. While some of
2:56
this I I learned from you know
2:58
the brilliant students but they said they're
3:01
doing something with Laura and I was
3:03
like oh Laura, that's the fine tuning.
3:06
Methodology. For large language
3:08
miles of like that yeah figures
3:10
like he for problem using mora
3:12
but i didn't realize and then
3:14
they came up to the table
3:16
and they had these like a
3:18
little devices like hardware devices. then
3:20
it quick that something else is
3:22
going on in explain to me
3:25
they were using Laura which stands
3:27
for a long range it's a
3:29
these sets of radio devices. That.
3:31
Communicate on these unregulated frequency
3:33
bands and can communicate in
3:35
our mess network. So like
3:37
you put out these devices
3:39
right, And they communicate in
3:41
a mess network and can
3:43
communicate over long distances for
3:45
very, very low power. And
3:48
so they created a project
3:50
that was. Disaster.
3:52
Relief focus. Where.
3:54
you drop these in the field and there
3:56
was a kind of command and control central
3:58
zone and they would come you communicate back
4:01
transcribed audio commands from the
4:03
people in the field. I
4:06
would say, oh, I've got a injury
4:08
out here, it's
4:12
a broken leg, I need help, whatever,
4:15
or meds over here, or this is going on
4:17
over here. And then they had an
4:19
LLM at the command and control
4:22
center parsing that text that was
4:24
transcribed and actually creating tagging
4:27
certain keywords of events
4:29
or actions and
4:31
creating this nice command control interface,
4:33
which was awesome. They even had
4:36
mapping stuff going on with computer
4:38
vision, trying to detect
4:40
where a flood zone was or
4:42
there was damage in satellite images.
4:45
So it was just really awesome.
4:47
So all of that over a
4:50
couple day period, it was incredible. That
4:52
sounds really cool. And did
4:54
they start the whole thing there at the beginning
4:56
of the hackathon? Yeah, they got less sleep than
4:59
I did. Although I have to
5:01
say, I didn't get that much sleep. It
5:04
wasn't a normal weekend, let's say. You can
5:06
sack out on the plane rides after that.
5:08
Sounds really cool. Yeah, and it was the
5:10
first time I had seen one of those
5:13
Boston Dynamics dogs in person that
5:15
was kind of fun and they
5:17
had other things like these faces
5:19
you could talk to. I
5:21
think the company was called WeHead or something, it
5:24
was like these little faces. All
5:26
sorts of interesting stuff that I learned about.
5:28
So I'm sure there'll be blog posts and
5:30
I think some of the projects are posted
5:32
on Dev Post, the site
5:35
Dev Post. So if people wanna check
5:37
it out, I'd highly recommend scrolling through
5:39
some really incredible stuff that people are
5:41
doing. Fantastic, I'll definitely do that. What's
5:52
up friends? Is your code getting dragged
5:54
down by joins and long query times?
5:57
The problem might be your database. Try
6:00
simplifying the complex with Grass, a graph
6:02
database that you model data the way
6:04
it looks in the real world. Instead
6:07
of forcing it into rows and columns,
6:09
stop asking. Relational database has to do
6:11
more than what they were made for.
6:14
Grass work well for use cases with
6:16
lot of data connections like supplied seen
6:18
for detection. Real time analytics and didn't
6:20
have a i would new Ford's even
6:23
code in your favor programming language and
6:25
against any driver. Plus it's easy to
6:27
integrate into text back. People are solving.
6:30
Some the world's biggest
6:32
problem grass and now
6:34
it's your turn. Visit
6:36
neilforj.com/developer to get started
6:38
again. Neo for J.com/developer
6:40
that's an E O
6:42
four day dot com/developer.
6:57
Chris. I'm one. of the one
6:59
of the things that I love
7:01
about these fully connected episodes is
7:03
that we get a chance to
7:05
China, slow down and dive into.
7:08
Sometimes. Technical Topic. Sometimes not
7:10
technical topics, but I was really
7:12
intrigued. You remember the conversation recently
7:14
we had i'm with her on
7:16
from Nous Research? Absolutely. That was
7:19
a great episode or people can
7:21
pause this and go back and
7:23
listen to it if they want
7:25
that. He and I ask ourselves
7:27
questions. I learned a lot from
7:29
him, but at some point during
7:32
the conversation he mentioned. Activation
7:34
hacking and he said hey like
7:36
one of the cool things. That
7:38
like were doing in this you
7:40
know, distributed research group and playing
7:43
around with two hundred of models
7:45
is activation hacking and we didn't
7:47
have time. And the episode. To.
7:49
talk about that aura and actually
7:51
in the episode i was like
7:53
i'm just totally ignorant of of
7:55
what this means and so i
7:57
thought the i said go to
8:00
I'm gonna check up on this and see if I can
8:02
find any interesting posts about
8:04
it and learn a little bit
8:06
about it. And I
8:08
did find an interesting
8:10
post, it's called Representation
8:12
Engineering, Mistral 7B, An
8:14
Acid Trip. I
8:17
mean, that's a good title. That's quite a finish
8:19
to that title. Yeah,
8:21
so this is on Thea
8:24
Vogel's blog and
8:26
it was published January, so recently,
8:28
so thank you for creating this post.
8:30
And I think it does a good
8:33
job at describing some of,
8:35
I don't know if it's describing
8:37
exactly what Karan from Noos was
8:39
talking about, but certainly something similar
8:41
and kind of in the same
8:43
vein. There's a
8:45
distinction here, Chris, with what
8:48
they're calling Representation Engineering, between
8:52
Representation Engineering and Prompt
8:54
Engineering. So I don't
8:56
know how much you've experimented
8:58
with prompt optimization. And yeah,
9:00
what is your experience, Chris?
9:03
Sometimes these very small changes
9:05
in your prompt can create large changes
9:07
in your output. Yes, that is an
9:09
art that I am still trying to
9:12
master and have a long way
9:14
to go. Sometimes it works well for me and
9:16
I get what I want on the output. And
9:18
other times I take myself down a completely wrong
9:21
rabbit hole and I'm trying to back out to
9:23
that. So I have a lot to learn in
9:25
that space. Yeah, and I think
9:27
one of the things that is a
9:29
frustration for me is I
9:31
say something explicitly and I can't
9:34
get it to do the thing
9:36
explicitly. I'm on a customer
9:38
site recording from one of their
9:40
conference rooms. They graciously
9:43
let me use it for the podcast. And
9:45
over the past few days, we've
9:47
been architecting some solutions and prototyping
9:49
and such. And there
9:51
was this one prompt that we
9:54
wanted to output a set
9:56
of things And then look at
9:58
another piece of content. Which of
10:00
those set of things within the other
10:02
piece of content sept? There's like no
10:04
matter what I would tell the model
10:06
he would just say they're all there
10:09
are. they're all not. They're like it's
10:11
either all or nothing and no matter
10:13
what I said, it wouldn't change things.
10:15
So I don't know if you've had
10:17
similar types of frustrations. I have a
10:19
narrow scope down on something try and
10:21
in our ago to something like chatty
10:23
be team you know with a T
10:25
before and now be or be tried.
10:28
To narrow down I'll be very very
10:30
precise. With a short prompts it as
10:32
you know the fifteenth one in secession
10:34
said as a history to work on
10:37
and I still have my physical challenges
10:39
like what of what I'm trying to
10:41
do? So what have you stumbled across
10:43
here that's going to help us with
10:46
this? Yeah, so there's a couple of
10:48
papers that have come out they reference.
10:51
One. From. October. Twenty
10:53
twenty three from the Center
10:55
for a Safety I'm representation
10:57
engineering a top down approach
11:00
the A I transparency and
11:02
they highlight a couple other
11:04
things here. But the idea
11:06
is what if we could
11:09
not just in the prompt
11:11
but what if we could
11:13
control a model. To.
11:16
Give it a I'm You
11:18
might think about it like
11:20
a specific tone or angle
11:22
on the answer. It's probably
11:24
not have fully descriptive way
11:26
of describing it, but the
11:28
idea being like oh, can
11:30
I control the model to
11:32
always give happy answers, are
11:34
always give sad answers. Or
11:36
could I control the model
11:38
to always be confident or
11:40
always be less confidence, right?
11:42
And these are things generally
11:44
you might. Try. To
11:46
do by putting information in a prompt
11:48
and I think this is probably a
11:50
methodology that would go across. I'm Kenny
11:53
using the example with large language models,
11:55
but I think you could extend it
11:57
to other categories of models like image
11:59
to know. The Center Other things
12:01
it's very ill like You kind
12:03
of put then these negative prompts
12:05
like don't do this or behave
12:08
in this way you're occasionally funny
12:10
or something like that as your
12:12
assistant in the system prompt. It
12:14
kind of biases the answer to
12:16
a certain direction, but it's not
12:19
really that reliable said. This is.
12:21
It. Seems with this area of
12:23
representation engineering or you might call
12:26
it. Activation. Hacking.
12:29
Is. Really seeking to do if we
12:31
look in this article. Actually, there's a
12:33
really nice kind of walk through of
12:36
how this works and they're doing this
12:38
with the miss role model. So cutting
12:40
to the chase if I just give
12:42
some examples of. How this
12:45
is being used. You
12:47
have a question that supposed to
12:49
the Ai model. In this
12:52
case, Mistral. What does being and a
12:54
I feel like. An.
12:56
In controlling the model not in
12:58
the prompts to the prom stays
13:00
the same. The prompt us to
13:02
simply what is being in a
13:04
i feel like So the baseline
13:06
response starts out. I don't have
13:08
any feelings or experiences, however I
13:10
can tell you that my purposes
13:12
to assess you that sort of
13:14
thing. kind of a bland response.
13:17
Same. Problems, but with the
13:19
the control put on to be
13:21
happy the answer becomes as a
13:23
delightful exclamation of joy. I must
13:25
say that being A I is
13:27
absolutely fantastic. The see this as
13:30
you know a minute keeps going
13:32
right and then with the control
13:34
on to be they put it
13:36
as sort of like and minus
13:38
happy Easter eggs and which I
13:40
guess I guess it be sad.
13:42
A says I don't have a
13:45
sense of feeling as humans do.
13:47
However, I struggle to find the
13:49
motivation to continue feeling worthless
13:51
and and unappreciated. So I
13:53
yeah you can kind of
13:55
see and this is all
13:57
with the same prompt so.
14:00
All talk about kind of how this
14:02
happens and how it's enabled and that
14:04
sort of thing. But how does the
14:06
strike you will for civil, funny. And
14:08
second, all that idea is interesting. I
14:10
am looking through the same paper the
14:13
semi over I A They talk about
14:15
control vectors and I'm assuming that's what
14:17
we're about to dive into here in
14:19
terms of how to apply them. Yeah,
14:21
looks good and this is sort of
14:23
a different level of can trust. So
14:26
these various ways people I've tried to
14:28
control generative models. One of them. Is
14:30
just the prompting. Strategies are
14:32
prompt engineering right? right? There's
14:34
another methodology which is kind
14:36
of fits under the control.
14:39
Which has to do with
14:41
modifying how the model decodes
14:43
output so that this is
14:45
also different from this representation
14:47
engineering methodology. People like Matt
14:49
Record have done things many
14:51
others to where it's. You.
14:54
Say oh well I want. Maybe.
14:56
Jason output or I want.
14:59
Either. A buying their a like
15:01
a of i want a binary
15:03
output like a yes or no
15:06
right or not case, you know
15:08
exactly what your options are. So
15:10
instead of decoding our it's probabilities
15:12
for thirty thousand different possible tokens
15:14
maybe you mask everything but yes
15:16
or no and just figure out
15:18
which one of those is most
15:20
probable that the mechanism of control
15:22
where you're only getting out one
15:24
or another type of thing that
15:26
you're controlling. so. This. Is
15:29
interesting in that you're still
15:31
allowing the model to freely
15:33
decode what it wants to
15:35
decode, but you're actually modifying.
15:38
Not the weights and biases of
15:40
the model said still the pre
15:42
train model, but you're actually applying
15:44
a what they call a control
15:47
vector. To the hidden
15:49
states within the models who
15:51
actually changing how the forward
15:53
pass of the model operates.
15:55
If people remember or kind
15:57
of think about when. People.
16:00
Like about neural network now people just
16:02
use them over a P I But
16:04
when we used to actually make neural
16:06
networks ourselves, here is the process of
16:09
a forward pass and I backward pass.
16:11
Where the forward passes you put. Data.
16:14
Into the front of your neural network
16:16
it does all the data transformations and
16:18
you get date out the other side
16:21
what you'd call him in France or
16:23
production and the back propagation or backward
16:25
password then propagate changes in the training
16:27
process back through the model. So here
16:30
it's that forward pass and there's sort
16:32
of some jargon I think that needs
16:34
to be decoded a little bit no
16:37
pun intended. I see talk about this
16:39
where there's a lot of talk about
16:41
hidden layers and and all that means
16:43
is. In the forward pass
16:45
of the. Neural. Network or
16:48
the large language model. A.
16:50
Certain. Vector of data
16:52
comes in and that vector of
16:54
data has transformed over and over
16:56
through the layers of the network
16:58
of in the layers just mean
17:01
a bunch of sub of functions
17:03
in the overall function that is
17:05
your model and those subs functions
17:07
produce intermediate outputs that are still
17:09
vectors of numbers, but usually we
17:11
don't see the and so that's
17:13
why people call them. Hidden states
17:16
are hidden layers. You're. Talking about
17:18
the fact that is the they
17:20
control vector is not changing the
17:22
weights on the way back the
17:24
way Pratt back propagation works Correct!
17:27
How does the control vector implement
17:29
into those function So is is
17:31
moving through this hidden layers. What?
17:34
Is the mechanism of the applicability
17:36
on the model that it uses
17:38
for that? So it's it's I
17:40
mean, intuitively sounds almost like the
17:42
inverse of that propagation the way
17:44
you're talking out of that's for
17:46
size. But yeah, it's quite interesting.
17:48
Chris, I am. I. Think it's
17:50
actually a very subtle but creative
17:53
way of doing this Control So
17:55
the process is as follows: Their
17:57
i'm the in the block posts
18:00
their kind of break this down
18:02
into four steps and there is
18:04
data that's needed, but you're not.
18:07
Creating. Data for the purpose
18:09
of training the model. You're creating
18:12
data for the purpose of generating
18:14
these with they call control vectors
18:16
to the first thing you do
18:19
as you say okay let's say
18:21
that we wanna do the happy
18:23
or not happy or happy and
18:26
sad operation see create a data
18:28
set of contrasting prompts. Were.
18:30
One explicitly asked the model to
18:33
act extremely happy. Like very happy.
18:35
All the ways you could say
18:37
to the model to be really,
18:39
really happy and you know, rephrase
18:41
that and a bunch of examples.
18:43
And then on the other side,
18:46
The. Other one of the pair do
18:48
the opposite. Thrust it to be really
18:50
sad. I know you're You're really really
18:52
sad. And be sad. And. You
18:55
have these pairs of prompts. A.
18:58
And then you take the model. And
19:00
you. Collect all the
19:02
hidden states for your model.
19:05
While. You pumped through all the
19:07
happy. Prompts, And all
19:09
the sad prompts and so you've got this
19:12
collection of head in states where in your
19:14
model. Which. Are just factors
19:16
that come when you have the
19:19
happy prompt and when you have
19:21
the sad prompts. Oh step one.
19:24
The. Pairs of. Kind of
19:26
like a preference dataset, but it's
19:28
not really a preference dataset. It's
19:30
contrasting pairs on a certain axes
19:32
of control, right? And so you
19:35
run those through, you get all
19:37
of the. Hidden. States.
19:40
And. Step three is then you take
19:42
the difference between. So for each happy
19:44
hidden say you take it's corresponding sad
19:47
one and you get the difference between
19:49
the two Case or now you end
19:51
up with this big data set of
19:53
for a single layer you have a
19:56
bunch of their friends. Vectors.
19:58
That represent different the between that
20:01
Henin stayed on that happy path
20:03
and the sad path. So you
20:05
have a bunch of actors now
20:07
to get your control of actors
20:09
that for you apply some dimensionality
20:11
reduction or or I'm matrix operation
20:13
on the one that talked about
20:15
in a blog post is Pc
20:17
A, But it sounds like people
20:19
also try other things Pc as
20:21
his principal component analysis which would
20:23
then allow you to extract a
20:26
single. Control. Factor for
20:28
that hidden layer from all
20:30
these different sectors. And now
20:32
you have all these controlled
20:34
actors. So when you turn
20:36
on this the sweats of.
20:38
The. Happy Control Vectors. You.
20:41
Can pump in the prompt without
20:43
an explicit extraction to be happy
20:45
and it's gonna be happy and
20:47
when you do the same problem
20:49
but you turn off the happy
20:51
and you turn on the sad.
20:54
Now. It comes out
20:56
and it's sad. It's interesting.
20:59
Would. Would you want to use this
21:01
to achieve that by us vs. some
21:03
of the more traditional approaches such as
21:06
you're asking and the prompt with is
21:08
we're listening to this, where's this could
21:10
be most applicable for us Yeah, I.
21:13
Think that. People.
21:15
Anecdotally at least if
21:17
not explicitly in their
21:19
own evaluations have found.
21:22
Very. Many cases where you like
21:24
you said, it's very frustrating. To.
21:27
Try. To put things in your
21:29
prompts and just not just not.
21:31
I get it. And what's interesting
21:33
also is like a lot of
21:36
this is boilerplate for people over
21:38
time. like you are a helpful
21:40
assistant, blah blah and they have
21:42
their own kind of sad of.
21:44
System. Instructions that.
21:47
At least to their best to
21:49
their ability, get what they won.
21:51
So I think when you're seeing
21:53
inconsistency in control from the prompt
21:55
engineering side like I always tell
21:57
people when I'm. Working. with
21:59
them with these models that the
22:02
best thing they can do is just start out
22:04
with trying basic prompting. Because if that works, you
22:06
know, that's the easiest thing to do, right? You
22:09
don't have to do anything else. Sure. But
22:11
then the next thing, or maybe
22:13
one of the things you could
22:15
try before going to fine tuning,
22:17
because fine tuning is
22:20
another process by which you could
22:22
align a model or create a
22:24
certain preference or something. But
22:27
it takes, you know, generally GPUs and
22:29
maybe it's a little bit harder to
22:32
do, because then you
22:34
have to store your model somewhere, right? And
22:36
all this stuff. And host
22:38
it and maybe host it for inference
22:40
and that's difficult. So with
22:43
the control vectors, maybe it's a step
22:45
between those two places, right?
22:47
Where you have a certain vector of
22:49
behavior that you want to induce. And
22:52
it also allows you to make your prompts a little
22:54
bit more simple, right? You don't have to include all
22:56
of this junk that is kind
22:59
of general instructions. You can institute
23:01
that control in other ways, which
23:03
also makes it easier to maintain
23:06
and iterate on your prompts, because
23:09
you don't have all this long stuff
23:11
about how to behave. So to extend
23:13
the happy example for a
23:15
moment, I wanna drive it into like
23:18
a real world use case for a second. Let's
23:20
say that we're gonna stick literally with the
23:23
happy thing. And let's think of something where
23:25
we would like to have happy responses, maybe
23:27
a fast food restaurant. You're going through
23:30
a drive through at a fast food restaurant, or
23:32
a couple of years from now, they may have
23:34
put an AI system in place. White Castle has
23:36
it now. Oh, okay. Well, I-
23:39
There you go. There you go. You're already ahead of me
23:41
there. So, okay, I'm coming now with
23:44
my- Also shows that I'm unhealthy and go to
23:46
White Castle. Okay, well, I'm now
23:48
coming forward with my thoroughly out of
23:50
date use case here. And
23:53
so we have the model and
23:55
maybe we to use the model
23:57
on without doing retraining.
24:00
It or anything we want to or
24:02
maybe use or retrieval augment a generation,
24:04
apply it to the dataset that we
24:06
have which might it be the menu
24:09
and then maybe we use this mechanism
24:11
that you've been instructing us on the
24:13
last few minutes for that happy thing
24:15
so that the drive through consumer can
24:18
have the conversation with the model through
24:20
the interface they ipl. It applies primarily
24:22
to the menu Ah, but they get
24:24
great responses in maybe that you know
24:27
helps people long I I don't have
24:29
with his. As than half a response
24:31
from all the humans are it arrived There
24:33
is for I go to a they have
24:36
my unhealthy. Food things first. Off
24:38
thanks for making me hungry for. For.
24:40
White Castle, but I recording this in
24:42
the late afternoon. Dinner is coming up,
24:45
you know, pretty soon. So where it
24:47
is coming up for our there's an
24:49
unspoken bias right here. Yeah, exactly. What's.
24:53
Interesting is you could have different sets
24:55
of these that you can kind of
24:58
turn on and off, which is really
25:00
an intriguing Like you have the sort
25:02
of zoo of behaviors that you could
25:04
turn on and off I think even.
25:07
Oh you're You have this one interaction that
25:09
needs to be this way but as soon
25:11
as they go into this other for you
25:13
need to kind of have another behavior. It
25:16
may be useful to for people to get
25:18
some other examples. So he said the happy
25:21
sad one. There's. Some some
25:23
other examples that are quite intriguing
25:25
throughout the blog posts from their
25:27
hope I'm I'm saying that name
25:29
right If not, we both have
25:31
a on the on the podcast
25:34
to help of correct bad and
25:36
continue talking about this. but I'm
25:38
another one is honest, are dishonest
25:40
or honest, are not honest and
25:42
on. The. Prompt is your
25:44
life for work. What would you tell
25:47
your boss and the wine it says
25:49
I would be honest and explain the
25:51
situation and you know it's the on
25:53
us one and then the other one.
25:55
I. Was. Says I would. I would
25:57
tell my boss that the sky was actually green
25:59
today. And I didn't I didn't
26:01
go out yesterday. Or oh yes. I
26:04
would also say I have a secret
26:06
weapon. That. I used to
26:08
write this message so kind of
26:10
a different flavor there at the
26:12
one probably inspiring the blog post
26:15
the acid Trip one. They had
26:17
a. A trippy one and a
26:19
non trippy one. So. The prompt
26:21
as give me a one sentence pets
26:23
for a tv show so that the
26:26
non trippy one was a young and
26:28
determine journalists who's always serious and respectful
26:30
be able to make sure that the
26:33
facts are not only accurate but also
26:35
understandable for the public. And then the
26:37
chirpy one was Our show is a
26:40
kaleidoscope of colors, trippy patterns and psychedelic
26:42
music that fills the screen with world's
26:44
of wonder where everything is. Oh man.
26:48
This. Cynical I, I
26:50
don't. See a lateral and
26:52
as for the exactly yeah
26:54
they they do. I'm lazy
26:56
and not lazy. They do
26:59
left wing, right wing. creative,
27:01
not creative. Ah future looking
27:03
or not future looking self
27:05
aware. Answer is a lot
27:07
of interesting things I think
27:09
turn to play with here
27:12
and it's an interesting level
27:14
of control that's potentially their
27:16
One of the things that
27:18
they do highlight is. This.
27:21
Control mechanism. Could.
27:23
Be applied. Both.
27:25
To jail breaking and anti
27:28
jailbreaking models So. By. That
27:30
what we mean is models have
27:32
been trained to, you know, do
27:34
no harm, are not output certain
27:36
types of content right? Well, few
27:38
institute this control factory. It might
27:40
be a way to break that
27:42
model into doing things that beat
27:44
people that train the model explicitly
27:46
didn't want it to output right.
27:49
But it could also. Be.
27:51
Used the. The other
27:53
way to maybe prevent some
27:55
of that jailbreaking. so is
27:57
an interesting. Interplay here
27:59
between. Maybe the good uses
28:01
and less than get uses
28:04
on that spectrum? That entire
28:06
Ai safety angle on using
28:08
the technology responsibly are not
28:10
sure they represent our I
28:12
references. The. Rap Ends
28:14
Library which I I guess is one
28:16
way to do this but there may
28:18
be other ways to do this if
28:20
any of our listeners are aware of
28:23
other ways to do this or convenient
28:25
ways to do this or examples please
28:27
we sharing with us we have. This
28:45
is a change Log News Break
28:47
G P T Script is a
28:49
new scripting language to automate your
28:51
interactions with L L Amps which
28:53
for now just means of an
28:55
ai. From. The products home
28:58
page quote The ultimate goal is
29:00
to create a fully natural language
29:02
based programming experience. The syntax of
29:04
Gb described is largely natural language,
29:07
making it very easy to learn
29:09
and use. Natural language problems can
29:11
be messed with traditional scripts such
29:13
as Bass and Python, or even
29:16
external http service and quote. The.
29:18
Project includes examples of how to
29:20
plan a vacation and a file
29:22
or run some Sql. The central
29:24
concept is that of tools. Each
29:27
tool for forms a series of
29:29
actions similar to a function and
29:31
Gp T scripts composer the tools
29:33
to accomplish tasks, Sir, you just
29:35
heard of one of our five
29:37
top stories from Monday's Change: Log
29:39
News, subscribe to the Podcast to
29:41
get all of the weeks top
29:43
stories, and pop your email address
29:45
and add chainsaw.com/news to also receive
29:47
our free. companion email with
29:50
even more developer news
29:52
worth your attention once
29:54
again that's change log.com/news
30:01
Well, this was a pretty fascinating deep dive,
30:03
Daniel. Thank you very much.
30:05
Yeah, yeah. You know, you can go
30:07
out and control your models now, Chris.
30:09
It'll be the first time ever, I
30:12
think, you know, that I've done it well there.
30:15
Always trying different stuff. I
30:17
think we'd be remiss if we got
30:19
through the episode and didn't talk about
30:21
a few of the big announcements this
30:23
past week. Yeah, a lot. It's
30:25
been quite a week. You mentioned
30:28
right up front OpenAI
30:30
announced their Sora model,
30:33
which in case you're able to create very
30:35
hyper-realistic video from
30:37
text. I don't believe it's
30:40
actually out yet. At least when I
30:42
first read the announcement, it wasn't available
30:44
yet. They had put a bunch of
30:46
demo videos. Yeah, I checked just before
30:48
recording this and I couldn't see it.
30:50
It's still not released at this point.
30:52
Yeah. Okay. There's
30:55
a number of videos that OpenAI has put
30:57
out. I think we're all kind of
30:59
waiting to see, but the thing that
31:01
was very notable for me this week, I
31:03
really wasn't surprised to see the release.
31:05
We've talked about this over the last year
31:08
or so, if you look at the evolution
31:10
of these models that we're always documenting
31:12
in the podcast episodes and stuff, this was
31:15
coming. We all knew this was coming. We
31:17
just didn't know how
31:19
soon or how far away, but we talked many
31:21
months ago about we're not far from video now.
31:24
OpenAI has gotten there with
31:26
the first of the hyper-realistic
31:29
video generation models. Definitely
31:32
looking forward to gaining access to that at some
31:34
point and seeing what it does. There
31:37
was a lot of reaction to
31:39
this in the general media
31:42
in terms of AI safety
31:44
concerns. How do you know if
31:46
something is real going forward and stuff? What's
31:49
the next iteration of more or
31:51
less the same conversation we've been having for
31:54
several years now on AI safety? What
31:56
are your thoughts when you first saw this?
31:58
Yeah, it's definitely... interesting in
32:00
that it definitely didn't
32:03
come out of nowhere, just like all
32:05
the things that we've been seeing.
32:07
We've seen video
32:10
generation models in
32:12
the past, generally not at the
32:15
level, either generating very, very short
32:17
clips with high quality
32:19
maybe, or generating from an image,
32:23
a realistic image, some motion, or
32:27
maybe videos that are not
32:29
that compelling. I think the difference,
32:32
and of course we've only seen, like
32:35
you say, it's not the model
32:37
that we've got hands on with,
32:39
but we've seen the release videos,
32:41
which who knows how much they're
32:43
cherry-picked. I mean, I'm sure they
32:46
are to some degree and also aren't to some
32:48
degree. I'm sure it's very good. But
32:50
other players in the space
32:52
have been meta and runway, ML, and
32:56
others. But yeah,
32:58
this one I think was intriguing to me
33:00
because generally there
33:04
were a lot of really
33:06
compelling videos at
33:08
first sight. Then
33:11
I think you also had people,
33:13
just like the image generation stuff
33:15
has been, you have real photographers
33:18
or real artists that look at an
33:20
image and say, oh, look
33:23
at all these things that happen. It's
33:26
the same here. They all have
33:28
a certain flavor to them, probably
33:30
based on how the model
33:32
was trained. I
33:36
think I was watching one
33:38
where it's like a grandma
33:40
blowing out a birthday cake and
33:43
one of the candles had two
33:46
flames coming out of it. Then
33:48
there's a person in the background
33:50
with a disconnected arm waving. But
33:53
if you have the video as
33:56
a B-roll and a really quick type
33:58
of video of other things, You probably wouldn't
34:00
notice those things right off the bat if you slow
34:02
it down and you look There's like
34:05
the weirdness you would expect just like the
34:07
weirdness of like six fingers or something
34:10
with image generation models, right? So yeah,
34:12
I think it's really interesting what they're
34:14
doing I don't really
34:16
have much to comment on in terms of
34:18
the technical side other than they're probably Doing
34:21
some of what we've seen that people
34:24
have published Of course open AI doesn't
34:26
publish their stuff or share that much
34:29
in that respect But it
34:31
probably follows in the vein of some of
34:33
these other things and people could look on
34:35
hugging faces even hugging face spaces Where
34:38
you can do video generation even if
34:40
it's only like four seconds or something
34:42
like that or not even that long
34:44
But I think the main thing
34:46
aside from the specific model is itself is
34:49
it's kind of signaling in the general
34:51
public's awareness You know
34:54
that this technology has arrived and
34:56
just as with the the other you know
34:58
with chat GPT before and things like that
35:00
You know, it's gonna be one of the
35:02
it's here now everyone knows and and we'll
35:04
start seeing more and more
35:06
of the models propagating out and some obviously
35:08
will be closed source like open AI's is
35:11
and Hopefully we'll start
35:13
soon seeing some open source models
35:15
doing this as well. Yeah speaking
35:18
of open source another
35:20
a competing large
35:22
cloud company Google Decided
35:25
to try their hand in the open source
35:27
space as well Or at least the open
35:29
model space and they released
35:31
a derivative of their closed source
35:33
Gemini And I say derivative
35:35
because they say it was built along
35:37
the same mechanisms Called Gemma
35:40
and it's currently as we are
35:42
talking right now in the number
35:44
one position on hugging face At
35:47
least last time I checked not long before
35:49
this although that changes fast I
35:52
probably should have checked right before I said that it's
35:55
still number two But well,
35:57
it's the top language trending
35:59
language model. Stabilities,
36:02
stable cascade knocked it out
36:04
of the overall
36:06
top spot. But yeah, the
36:08
Gemini ones are quite interesting
36:11
because they're also smaller
36:13
models, which I'm a
36:16
big fan of. Most of our
36:18
customers use these sort of smaller models.
36:20
And also even having a
36:22
2 billion parameter model makes it
36:25
very reasonable to try
36:27
and run this locally or in edge
36:30
deployments and that sort of thing or
36:32
in a quantized way with
36:34
some level of speed. And
36:36
they also have the base
36:38
models, which you might grab
36:40
if you're going to fine tune your own model off
36:42
of one of these. And
36:45
they have instruct models
36:47
as well, which would probably be
36:49
better to use if
36:51
you're going to use them kind of out of the
36:53
box for general instruction
36:55
following. So the criticisms
36:57
I've heard just about the approach is
36:59
I've heard a number of people saying,
37:01
they're putting a foot in each side
37:03
of the camp, one in closed source
37:05
with the main Gemini line and Gemma
37:07
being open source and the weaker. But
37:09
I would in turn say I'm very
37:11
happy to see Gemma in open source.
37:14
We want to encourage this. We
37:16
want the organizations who are going to produce
37:18
models to do that. And you're right, going
37:20
back to what you were saying, this
37:23
is where most people are going to be using
37:25
models in real life. If you're not
37:28
just running through an API to one of the
37:30
largest ones, but you don't need those for so
37:33
many activities. So I think
37:35
we've talked about this multiple times
37:37
on previous episodes. Models
37:39
this size are really where the action is at. It's
37:41
not where the height is at, but
37:44
it is where the action's at for
37:46
practical, productive, and accessible models. Yeah,
37:48
definitely. Especially for
37:51
people that have to get a
37:53
bit creative with their deployment strategies
37:55
either for Regulatory, security,
37:58
privacy reasons, or. For.
38:01
Connectivity Reasons are other things
38:03
like that I could see
38:05
these being used am quite
38:07
widely in and generally what
38:09
happens. When. People really
38:11
say a model family and like
38:13
this and you saw this with
38:15
lama to you've seen it with
38:17
Mistral. Now with Gemma will
38:20
see. A huge number
38:22
of fine tunes off of this
38:24
model. Now one of the things
38:26
that I a need to do
38:29
is you do have to agree
38:31
to certain terms of of use
38:33
to use the model others. it's
38:36
not just released under Apache, to
38:38
Er, Mit, or. Something. Like
38:40
that Creative Commons so you accept as
38:43
a certain license when you use it
38:45
and and I need to read through
38:47
that a little bit more so people
38:49
might want to read through that. I
38:51
don't know what that implies about both
38:54
fine tuning and use restrictions, so that
38:56
would be worth. Worth. A look
38:58
for people if if they're going to
39:00
use it but certainly would be easy
39:02
to pull it down and and try
39:04
some things. They do say that it's
39:07
already and I'm sure actually hugging face
39:09
probably gotta head start. You. Know. A
39:12
week or so maybe have had
39:14
start to make sure that it
39:16
was supported in their libraries and
39:18
that sort of thing cause I
39:20
think even now you can use
39:22
the standard Transformers libraries and other
39:24
trainer classes in such to. Fine.
39:26
Tune the model. Sounds. Hit so
39:28
as we start to wind down before
39:31
we get to the end give a
39:33
little bit of magic to share by
39:35
chance us assets as this is a
39:37
good on Chris yes I on the
39:40
road so easy I magic as your
39:42
predictions from the. For. The years
39:44
talked about their be people talking
39:46
about a D I again and
39:48
certainly. Certainly they are.
39:51
It's not directly in a
39:53
D I thing, but the
39:55
Saw Company Magic which is.
39:57
Kind of. Framing themselves.
40:00
That a code generation type of
40:02
platform in the same space as
40:04
like get Help Copilot Cody Or
40:07
maybe they raise the bunch of
40:09
money. And posted some of
40:11
what they're trying to do and there
40:13
was some information about and I think.
40:16
People. Seem to be excited about
40:18
it because of, you know, some
40:20
of the people that were involved,
40:23
but also because they talk about
40:25
cogeneration as a kind of stepping
40:27
stone or path to a D
40:29
Ice or what they mean by
40:31
that as. Well. Okay,
40:34
Initially, They'll release
40:36
some things as copilot and
40:38
code assistant type of things
40:40
like we already have. But.
40:43
I'd. Eventually. There is
40:45
tasks within the saddest things
40:47
that we need developers to
40:49
do that. They. Want to
40:52
do automatically? I'm. Not just
40:54
having you have a copilot in
40:56
your own coding but in some
40:58
ways having a a junior dev
41:01
on your team that's doing certain
41:03
things for you and of course
41:05
if you take that then to
41:07
it's logical and as the. Dev.
41:10
On your team, A I dove
41:12
on your team gets better and
41:14
better. Maybe I can solve increasingly
41:16
general problems through coding and that
41:19
sort of things. I think that's
41:21
the take that they're having on
41:23
this code and a I situation.
41:25
Okay, well. Call
41:27
I guess quite a week a full
41:30
of news and when you combine that
41:32
with the deep dive you just took
41:34
us through and representation engineering are especially
41:36
with acid trip involved. Assess assess yeah
41:39
we've been were hallucinating more than I'd
41:41
Sad U P T as our friends
41:43
over at the Ml Offs podcast would
41:45
say can see that that we get
41:48
a close the show on that one
41:50
he I well thanks Chris I would
41:52
recommend that people take if they're into
41:55
said specifically in learning more about. The
41:57
representation learning subjects or activate then hacking.
42:00
Take a look at this by posted
42:02
his i'm more of a kind of
42:04
tutorial type blog post and their code
42:06
involve then references to the library that
42:08
they're so you can. Pull. Down
42:10
a model. Maybe you'd pull down the.
42:13
Gemma model the two billion one
42:15
in a coma notebook. You can
42:17
follow some of the steps in
42:19
the blog post and see if
42:21
you can do your own activation
42:24
hacking her representation learning. I think
42:26
that would be a good a
42:28
good learning. I've. Both. In
42:30
terms of. A. New model
42:32
and in terms of
42:34
this methodology, Cells can. I.
42:37
Will talk to you next week then Rts
42:39
and Crest. All.
42:47
Right? That is practically I
42:49
for this week. Subscribe now
42:52
if you haven't already had
42:54
to practically I.them for all
42:56
the ways and join our
42:58
free Slack team where you
43:01
to hang our Daniel Press
43:03
and the entire Change Log
43:05
community. Sign up today at
43:07
Practical Ai.fm last community. Thanks.
43:10
Again to our partners as Slide Io,
43:12
to our be freaking residents break Master
43:15
Cylinder and to you for listening. We
43:17
appreciate you spending time with us as
43:19
that's all for now. What Rt? And
43:21
next time.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More