← Back to videos

Sam Altman (CEO of Open AI): "You no longer need to write prompts." In just 38 minutes, he explain

Official Himanshu

Official Himanshu

@_HimanshuBuilds

November 18, 2026· 26:31

Sam Altman (CEO of Open AI): "You no longer need to write prompts." In just 38 minutes, he explains how to use ChatGpt at a level that most people can't even imagine. It's talk he gave to stanford students. A friend sent me the recording last night. After watching it, I realized I was only taking advantage of about 15% of what this tool can really do. Watch it in full and then read the guide l leave below on how to create a system that prompts itself.

Transcript

You, with like an affordable amount of spend on tokens
0:05
you can do what a hundred person
0:07
incredibly great engineering team would do as a startup
0:11
and that was just totally impossible
0:13
That was like not in the
0:14
set of options for a startup
0:17
and now it is. So
0:18
so I think what you can take on
0:20
the level of ambition you can have
0:21
the speed at which you can move
0:23
the amount of stuff you can do at once
0:25
is just totally different
0:27
And
0:28
does that change the
0:29
shape of the problems you feel like
0:30
you'd
0:31
assign at the end of the class for people to attack
0:34
you know
0:34
at the end of that
0:35
quarter if you were teaching it again?
0:38
I don't think assigning
0:39
problems to attack ever works because if you
0:41
like If I can think of a problem
0:43
if I can think of like a really great startup idea
0:46
if it's like obvious enough to me
0:48
then
0:49
it's probably obvious to a lot of people
0:51
When we started OpenAI
0:52
we were, we were like the
0:54
you know
0:54
one of maybe generously speaking for AGI efforts in the world
0:58
right? And you want to find something like
1:00
that. And I'm sure
1:01
that there exists something today
1:03
that just wasn't possible at all pre like automated coding era
1:08
that is totally non-obvious that will be
1:10
you know, a multi-trillion dollar market soon
1:13
and that only four companies are working on
1:15
right now
1:16
But I don't know what that is
1:16
It's much more likely
1:17
you all know what that is than I
1:19
know what that is. I just
1:20
you know, my brain is like taking over by OpenAI
1:23
but, you know
1:24
the kind of idea someone can assign you
1:26
to work on is Probably not what you want
1:28
Yep
1:30
okay, so that, that's fair
1:32
but I, I think it would be helpful
1:33
since this is a systems class
1:35
to maybe
1:36
reason about a particular problem that you have to reason through
1:39
so that they can then apply
1:42
the shape of the techniques used to break
1:43
down from a systems perspective that problem into
1:45
solutions to their own problem
1:47
Yeah
1:48
and a, and a concept that
1:50
you had started to tease in the class
1:51
you know, back in 2014, and then
1:54
clearly you've talked about publicly over the years is
1:57
scale
1:58
Right? Scale is its own beast
2:00
It's, it's, it's, you know
2:01
quantity is its own quality
2:02
You know, what, what
2:04
scale as a concept has been something it seems like you've
2:08
empirically investigated in all kinds of ways over
2:10
the last 10 years. So
2:12
could you help on
2:13
help us first unpack
2:14
like, what you mean by scale now
2:16
10 years later? How would you deconstruct that
2:18
as a systems design
2:20
attribute to apply, whether it's a, as, as a tool?
2:23
can, can we start there?
2:26
Yes
2:27
So I don't
2:28
know why the following observation is true
2:32
I offer no theory that I find satisfying to explain it
2:37
and that makes me
2:38
a little bit nervous to suggest you follow it
2:42
but I'm going to anyway
2:43
because empirically it does seem to be true
2:45
which is
2:46
all of the most interesting things I have
2:48
observed in my career in watching other things happen
2:53
all of the most interesting ones
2:56
have had something to do with
2:59
emergent properties that scale
3:01
or
3:01
scale continuing to provide returns far beyond what
3:04
the consensus thinks will work
3:06
And this obviously happens with like scaling laws for AI models
3:11
but this happens with
3:14
you know
3:14
getting more smart people together to think about
3:16
one problem
3:17
This hap- in a
3:18
in a research setting
3:20
this happens with
3:22
Companies and the sort of economy of scale
3:24
you can get in all these different ways
3:26
I really learned this at Y Combinator
3:28
when it became clear to me that everybody was saying
3:32
oh, Y Combinator's gotten too big
3:33
it should shrink, we should fund less companies per batch
3:35
you know
3:36
the best times of Y Combinator when it
3:38
was like 10 companies per batch
3:40
And
3:40
a lot of like very smart people were
3:42
saying this
3:45
And it was like tempting because it would
3:46
have been like much less work
3:47
And the theory was that
3:48
you know
3:48
the best companies
3:50
are always kind of obvious and then you
3:51
fund the rest and it's not as helpful
3:53
But a huge part of the magic
3:55
of what made YC work were
3:58
was the sort of the network effects inside
4:00
of the batch
4:01
And that was an emergent property at scale
4:03
that just hadn't been discovered before
4:04
No one had tried to fund startups
4:08
at scale in the same way
4:09
And thus no one had ever
4:11
happened upon this observation of when you do
4:13
that
4:15
There's
4:17
There's something important that happens that just didn't
4:19
exist at all at the 110 to 1100th
4:21
of a scale
4:23
There's a bunch of other examples like this
4:26
I
4:30
and I'll skip them in the interest of time
4:31
but I, I would say, again
4:34
I offer no explanation for why
4:36
but empirically speaking
4:37
when you find a time that you can push on
4:41
you can push something to a scale people
4:42
have not tried before
4:44
and it's already working
4:45
in some interesting way at the smaller scale
4:47
more often than not
4:48
that seems to be a good idea
4:50
And it also seems to be
4:51
something that
4:54
most people don't do enough
4:56
And I don't offer an explanation for this either
4:59
but like in, you know, when we were like
5:01
we're really going to scale AI models
5:04
All of the like geniuses in the field
5:05
most of them were
5:06
oh, this isn't really working
5:07
You know, that's barely a scientific result
5:09
It's not interesting that it gets better at
5:11
scale. You've already shown that
5:12
Why keep scaling it?
5:13
I mentioned the YC example
5:15
I've seen a lot of startup founders where they're like
5:19
well, you know
5:21
there might be something interesting that would happen
5:22
if I scaled this up
5:23
but
5:24
I'm a little worried about it for nonspecific reasons
5:27
And again
5:28
looking back at like a huge
5:30
data set of people that have scaled
5:33
their companies in all these different ways
5:34
there's almost always interesting stuff there
5:36
So I think directionally
5:37
that's like an interesting thing to
5:40
push on and severely underexplored
5:45
On the systems design part of that
5:50
I think one reason people
5:51
don't do it as much Is stuff breaks
5:57
at an accelerating rate and in an unpredictable
5:59
way as you scale it
6:01
And if you are gonna really scale something
6:05
It's always like a little bit broken
6:08
There are always like
6:09
very smart people who say why you shouldn't do this
6:11
you know, don't get too ambitious
6:12
don't get too big
6:13
let's try this smaller
6:14
And
6:15
so breaking that down is a systems problem
6:17
I'll use the
6:18
thing of when we were like scaling up AI models
6:20
there was
6:21
technically can we do this at all?
6:22
This seems crazy. Like no one had ever
6:24
thought about trying to do a run across
6:25
10,000 or 100,000 GPUs and that was going
6:27
to require stacks of engineering talent
6:30
There was the
6:31
capital requirements and what was going to take
6:33
to do this and like
6:34
how is there ever going to be a
6:35
business? How can you think about taking this
6:37
risk?
6:38
There was the sort of like cultural stuff
6:39
of researchers saying well if we're going to
6:41
get all this compute
6:42
why do we put it all into this
6:43
one project? We're not going to learn something
6:44
Why not divide it up among all these
6:46
all these projects?
6:47
And
6:47
this also happens in kind of every area
6:50
I've looked at
6:51
almost every area for scale
6:52
and breaking it down into the
6:55
sort of
6:57
each difficult area
6:58
or each reason not to do it and
6:59
trying to address them one at a time
7:01
Yeah, that's been really important
7:04
I'm going to push on that a little
7:05
bit because
7:07
there's very few people who've been able to
7:08
sort of repeatedly scale new products and systems the way
7:13
the OpenAI team has over the years
7:16
But it seems like one of the
7:17
issues is that there are all these prior conditioning
7:21
sort of mental models and expectations humans have
7:24
and you said things break
7:26
And one of the things it seems often breaks
7:29
that's hard, the hardest to refactor is
7:31
is human, the human side of the
7:34
the systems design, right? Wherever there's human implementers or there's
7:38
human participants in that. And so what have
7:40
you learned about humans at scale
7:41
like organizing humans at scale to participate in
7:43
a system that may not be
7:46
like just a, a redo of some past system that they
7:49
they get naively on at
7:51
a priori on first blush?
7:55
I think like clear
7:57
a clear goal, A clear plan to get there
8:01
and like a clear
8:06
answer to the way that you're going to
8:08
get there and kind of how you're going
8:09
to make decisions along the way
8:10
that's, that's very important. So
8:14
you know
8:14
if we go back to the example of
8:15
when we decided to scale up models
8:17
there were a lot of people who were like
8:18
this isn't really going to work
8:20
it's going to have these problems
8:21
it's also not, you know, we need a more diversified portfolio
8:23
but once we say no
8:24
we're going to make a bet on scaling deep learning
8:26
like that's our thing. If we're wrong
8:28
we'll fail, but we're going to do that
8:30
Here's why we're going to do that
8:31
here's what we believe about what the state
8:32
of the world can be like if we
8:33
get there
8:35
that's very powerful
8:37
And then
8:40
for whatever reason
8:41
we did not
8:42
evolve to be good at thinking about exponentials
8:47
People have a hard time imagining that
8:50
Scaling laws are going to continue exponentially
8:52
that revenue will grow exponentially
8:54
that an organization can take on exponential complexity
8:57
And in my experience
8:59
it takes a lot of time
9:01
to really reason through first principles with people about why
9:04
why that can happen
9:06
Can we take two examples
9:07
to walk through that?
9:09
The first being ChatGPT and the second being Codex
9:11
You know, both of these have transformed
9:14
Can
9:14
can everyone hear? I'm going to try to
9:15
project it. Yeah. Okay
9:18
so let
9:19
let me put in
9:20
a frame and you can challenge both the
9:21
assumption and then
9:22
we can hopefully reason through
9:23
example of what happened
9:25
In the case of ChatGPT
9:26
you know, for a long time in scaling of models
9:28
a big mental block that
9:31
Seem to be prevalent in space is
9:32
what
9:33
what are these things going to be useful for?
9:35
This is, you know, it's a research, sort of solution, solution chasing a problem
9:39
research first approach. It's not a product
9:43
and then, you know
9:43
ChatGPT came out and it proved to the world that
9:46
you know
9:47
that chat experience was a killer app for general models
9:51
at scale for consumers
9:53
And then a couple of years later
9:55
yes
9:55
it's clear that coding
9:57
has been the killer enterprise app
9:59
So what
10:00
how would you compare and contrast the systems
10:01
you guys used to discover those use cases
10:04
ship them, scale them, monetize them?
10:06
Any, any salient learnings from those two systems?
10:09
Yes
10:12
so we had made GPT-3 and
10:16
we
10:16
needed to make money because we wanted to
10:18
go scale up to
10:19
you know, a billion and multi-billion dollar computers
10:21
and we had GPT-3
10:21
and it was kind of interesting
10:22
it was a cool demo
10:24
but we couldn't figure out
10:25
a product to build around it
10:27
And we had been thinking
10:27
thinking, we just couldn't do it
10:29
We had tried a few things
10:29
they, they hadn't worked
10:31
and so we knew the models were going to get better
10:33
but we also wanted to like
10:35
start a revenue engine sooner
10:37
And we said, well
10:38
since we can't figure out what product to build
10:40
we're just going to put this into an
10:41
API
10:42
And we're going to hope that somebody else
10:44
can figure out what product to build
10:45
And so we launched in like
10:47
I don't know, something in the summer of 2020
10:50
the GPT-3 API
10:52
And
10:53
initially it kind of got no
10:55
traction at all
10:57
And then about a month later
11:00
randomly, as far as we can tell
11:01
it went viral on Twitter
11:03
On the same day
11:04
a few different developers kind of found
11:06
got it to do something cool
11:07
posted it, other people started trying
11:09
And
11:10
and then like
11:11
a lot of people started trying the API
11:14
but it was shockingly bad
11:16
If you go back and use GPT-3 or 3.5
11:21
you will be astonished at how bad The models were then
11:24
relative to the
11:25
amount of excitement they generated at the time
11:28
so people tried all of these things
11:29
and really the only business that people got
11:31
to work
11:33
in a significant way with GPT-3 was copywriting
11:37
and that was like not that great and not that exciting
11:39
and we were kind of like
11:40
you know, oh
11:41
it's just going to have to wait for
11:42
a better model
11:43
But although that was the only business that was working
11:47
developers had figured out how to like put
11:49
in a prompt and get
11:51
and be able to chat with it
11:52
And we saw this a lot
11:54
like more people were using
11:57
they couldn't get the API to work for their business
11:59
but they were using their API key to
12:01
just chat. And we said
12:02
well, we can build a good chatbot
12:03
People clearly want that
12:05
And we had a new model
12:07
We actually had GPT-4 done
12:08
but we had a new model we were
12:09
ready to release in between called 3.5
12:12
And we had figured out
12:13
a new kind of post-training
12:14
where we could get the models to do
12:15
like a good job with instruction following so
12:17
it can make it easier to chat with
12:19
And we said, well, you know
12:22
The API is not working great
12:24
Maybe it was like a 10 or a
12:25
20 million dollar run rate kind of business
12:28
But there is this thing that people love
12:30
and
12:31
under the YC principle of
12:33
see what your users love and do that
12:34
we said, well
12:35
we'll build a chatbot around it
12:36
And we put that out
12:38
and we still didn't think it was going
12:38
to do that well
12:40
there was, it was really meant as like a research demo
12:43
to
12:43
convince other people that they should build chat-like
12:45
products and pay us for the API
12:48
But that went like crazy viral
12:50
And another thing I had learned from YC
12:52
is when something really starts growing and it's not very good
12:56
you have like a guaranteed hit on your
12:57
hands
12:58
And so we had like five days
13:00
where the traffic would shoot up
13:02
fall off, and everybody would be like
13:03
well, that was just a hype cycle
13:04
But then the next day it would get
13:05
to a higher peak
13:06
fall off again later in the day
13:07
people would say that's a hype
13:08
hype cycle
13:10
By the fourth or fifth day
13:11
I was like, I know how this works
13:13
I know what's going to happen
13:14
Like we have the potential here
13:16
at a killer product
13:19
And
13:19
we knew we could make it much better
13:21
We knew we could
13:22
we knew we had GPT-4
13:23
we knew we could keep scaling
13:26
but by that fifth day
13:29
we got everybody together and said
13:31
this is an emergency
13:32
this is the good kind of emergency
13:33
but we have to build a company and
13:35
a product all at once
13:38
we then had like
13:39
two months of crazy scaling
13:42
and then we said
13:42
you know
13:43
we have to figure out a business model later
13:45
For now
13:45
we're just going to charge people so that
13:47
we don't like run out our compute bills
13:49
but that's obviously not the long-term answer
13:51
That also turned out just to work
13:54
and that was the story of ChatGPT
13:55
And then there was so much utility
13:57
that people just had not
13:59
gotten over the activation energy of find
14:01
that that has worked really well
14:04
and then Codex Actually
14:06
the plan before ChatGPT was that we were
14:08
going to go all in on code
14:10
we knew these models could write code
14:12
we knew that they could be really
14:15
and we knew that that would be like a valuable area
14:17
but then we had this incredibly exciting thing
14:19
happen
14:20
but our kind of internal belief at the
14:21
time was that coding
14:25
was how these models would control things on computers
14:28
and robots were how these models would
14:30
control things in the physical world
14:32
And if you made a smart enough model
14:33
that had sort of the actuators of writing code and robot
14:36
and driving a robot
14:38
you could then kind of
14:39
actually get this intelligence to do stuff for
14:41
you in the world
14:42
So
14:43
then it took us a while to get
14:44
there
14:45
And then
14:46
I think Codex got really good by early this year
14:49
but with 5.5 is when we saw this
14:52
real inflection point where people are now like
14:54
Doing just incredible things with it
14:57
And, you know, that, that early in the class
15:00
we've talked about how the capabilities pipeline
15:03
is starting to look
15:05
is starting to become somewhat more
15:06
legibly standard across different research groups
15:09
You've got, you know, pre-training, mid-training, post-training
15:12
then you've got the RL and supervised feedback loop
15:14
Is
15:15
do you think that's roughly like the shape
15:17
of the pipeline that allowed Codex to
15:19
you know
15:20
go through a capability jump and that will
15:21
basically stay stable now and consistent
15:23
or are we going to go through a
15:24
major rewrite of that pipeline?
15:26
I think that is definitely the current pipeline
15:28
I expect we will go through a major
15:29
rewrite. I don't know when it'll happen or exactly how
15:32
but
15:34
it is a little odd to me that
15:35
it's so
15:37
It happens
15:38
as a pipeline and it doesn't quite feel
15:41
like the optimal solution
15:44
what would be an optimal solution
15:46
in your head?
15:46
I think that's a research problem for the
15:48
AIs to figure out
15:49
I think we're at a point where
15:51
and we've set this goal that by September of this year
15:54
we will use 500,000 A100 equivalent GPUs
15:57
like a lot of computing power
15:58
as an AI research intern
16:00
and by March of 2028 that we will
16:02
have a full end-to-end
16:04
very talented researcher
16:05
like figuring out complete new architectures
16:09
so I think we are going to get
16:10
like, with the current pipeline
16:12
the current architectures
16:13
I think we're going to get over the
16:14
line of when AIs can do incredible
16:17
incredible work
16:19
you know
16:20
one of the things that you just described there Is
16:25
is you, you, we've
16:27
we've been talking a lot in the class
16:28
about systems frameworks and analogies to make
16:30
concepts from one domain legible to other people
16:32
who may not have all the
16:33
context in another
16:35
And that sometimes because of the translation problem
16:39
you know
16:39
reasoning by analogy is not helpful
16:41
because then errors compound
16:43
right there you said
16:44
you know
16:45
our goal is to try to use it
16:46
as an AI intern
16:47
which obviously is a very useful metaphor within the context of
16:50
you know, Silicon Valley, a class
16:51
a class that understands how these pipelines work
16:53
and so on. And then as
16:55
as you scale actually that metaphor globally
16:57
people who might not have all that context
16:58
go
16:59
start analogizing these models in ways that they shouldn't be
17:01
Like, how should we think about the limits of
17:05
of that, of, of, what
17:06
what are the limits to scale of
17:08
what, what are the product analogies
17:10
the research analogies you find most useful
17:13
within the Valley? And which one of the
17:15
what have you found about
17:16
found about the limits of those analogies scaling?
17:19
And now how do you navigate between those
17:21
two problems?
17:23
I've been very interested in studying how
17:26
like, I think what is happening is we are
17:30
we are in the process of creating a
17:31
new utility
17:32
This doesn't happen very often
17:33
You know, electricity is a utility
17:34
internet's a utility, water, I guess
17:37
there's not a lot of these
17:38
And so there are not a lot of
17:39
examples that we can study
17:41
for good
17:42
metaphors or learnings about how to
17:45
explain this to the world
17:47
But I was recently
17:49
looking at
17:49
what happened when electricity became a utility
17:53
And it's a good analogy for many reasons
17:55
It's imperfect, of course, too
17:57
But the electricity companies
17:59
at least the ones I could find information about
18:01
they didn't talk about selling electricity because no
18:03
one knew what that was or why they
18:04
wanted it. It sounds like very scary
18:05
It's this thing that's like
18:06
going to come into your house and it
18:07
could kill you in this like gruesome way
18:09
And, you know
18:10
it feels sort of like very different than
18:12
the world before
18:15
And
18:15
maybe they tried to sell electricity or market
18:18
electricity at first. I don't know
18:19
But in any case
18:20
that didn't work
18:21
And then what they started Marketing
18:23
selling to people was light at night
18:25
You know, we are gonna
18:27
what you are getting from us is not electricity
18:29
it's light at night. By the way
18:31
you can use the same thing that lets
18:32
you get light for all these other things
18:34
but people are like
18:34
well, why would I want that?
18:35
And they're like, well, you know
18:37
it'll wash your clothes for you someday
18:38
And I'm like, no, no, it won't
18:39
I can't
18:40
that's too far of a jump for me
18:41
Right
18:42
so I don't know what our analogy for this should be
18:48
but I suspect that even if Even if we're totally right
18:52
and intelligence is going to become this new
18:54
utility that every company
18:56
every customer
18:58
every government just
18:59
needs access to and is going to use
19:01
in all sorts of incredible ways
19:03
and you will have like a
19:04
OpenAI token subscription that you will plug into
19:06
everything and use to access everything
19:08
and
19:08
you have running for you all the time
19:09
and doing this amazing stuff
19:11
I kind of don't think
19:13
at least right now
19:14
the right way for us to analogize that
19:16
is we're selling intelligence
19:17
because people are just like
19:19
Somehow not resonating
19:20
I don't know what our equivalent of
19:23
we're selling you light at night is going to be
19:25
but I think if we're going to become a new utility
19:28
we need to find a way to explain
19:30
to the world what it means to have
19:31
this like
19:32
intelligence pike
19:33
that you can just do whatever you'd like with
19:35
It, so, one question that has emerged
19:41
an emerging property of this class is of
19:42
of having a
19:43
diversity of different speakers is that
19:45
the utility analogy has come up several times
19:47
but in reference to different things
19:49
So Jensen likened, like compute to a utility
19:54
and why there should be access and so on
19:55
and talked about how Stanford should pool
19:57
budget and so on and
19:58
and
19:59
and procure that as a utility for everybody
20:01
on campus
20:02
Whereas you just likened
20:03
the intelligence part to a utility
20:05
Are both of these things true?
20:06
Is one of them true?
20:07
One, is one more likely to be true?
20:08
How should people reason about compute as a
20:09
utility versus tokens as a utility?
20:11
And, and by compute
20:13
I mean here chips versus tokens
20:15
Does that, does that make sense?
20:16
I think as a consumer
20:18
as like a business or an individual
20:22
you will think in something closer to tokens
20:24
or probably even one level up from tokens
20:27
I don't think you'll care very much about
20:29
you know, where the hardware is
20:31
what particular chip it is
20:32
what's powering it. I think that stuff will
20:34
be abstracted out
20:35
And what you will care about is when
20:38
you're interacting with the system
20:42
can you use it a lot?
20:42
Is it cheap? Is it doing a good
20:43
job?
20:45
so right now it's like tokens
20:47
It may get
20:48
as we move into a world where
20:50
we all just have like this constant agent running for us
20:53
being useful to us all of the time
20:55
you may think about it
20:56
as even one level up
20:57
But yeah, my, my guess is
20:58
is you, when you like pay for your cell phone bill
21:03
You're like, all right
21:03
I'm buying access to airtime and some number of gigabytes and
21:08
you know
21:08
it's going to do all these things and
21:09
I'll use all these apps and whatever else
21:10
But like what you think about
21:12
paying for the kind of internet utility in
21:14
this case is just like
21:16
access to the whole system
21:18
And the particular hardware at the base station
21:22
how it connects to the internet
21:23
you don't think about that as much
21:25
I know
21:26
I couldn't heard out about utility infrastructure for a long time
21:28
but I want to make sure we switch
21:29
a little bit to being relevant for the
21:31
students. Usually we have questions
21:33
but we're not hearing those today
21:35
Unless you're comfortable. Oh
21:36
okay, great. How about that?
21:37
Improv. Okay
21:40
So
21:40
one final question to start getting the creative juices flowing is
21:43
the final project for this class
21:46
Arlene 5x183, is the One Person Frontier Lab
21:49
So everybody here is working on projects where
21:51
they're simulating being an individual
21:54
as a lab with access to all the
21:56
right tools. They've got hundreds of thousands of
21:57
dollars of credits from CloudFlare
21:59
I think we've got some OpenAI tokens maybe
22:01
but there's a bunch of compute at their
22:02
disposal
22:04
what would you, if you were in the class
22:06
what would you be working on for your
22:08
One Person Frontier Lab project?
22:10
First of all
22:10
I think that's an awesome project
22:15
I think this is top of mind because
22:18
you, we, we were just like talking about utility framework
22:21
frameworks
22:23
I think there's a lot of very smart people working on
22:26
great training ideas
22:28
and we're going to have incredible models
22:31
No matter what you all do
22:31
we're going to have incredible models
22:32
I promise you, like pretty quickly
22:35
But I
22:37
I think we have not invested enough in
22:39
being able to deliver
22:41
at scale huge amounts of cheap intelligence
22:43
So maybe I would go work on like
22:44
the inference part of the stack
22:46
And how are we going to get
22:48
this incredible intelligence to be cheap and abundant?
22:52
I think that's underinvested. And
22:53
and
22:53
and
22:54
I think all of the frontier labs are
22:55
going to have to become
22:57
inference companies to a significant degree
23:00
okay
23:02
It might be too late to pivot your projects
23:04
but better late than never
23:05
Work on whatever you want to work on
23:08
okay, let's start taking questions
23:09
and I'm going to moderate and try to be not
23:12
you know, please try to be productive and not spicy
23:14
et cetera
23:15
Remember, it's a CS class
23:16
but up to you
23:17
Sam
23:19
Oh, we've got questions. Oh
23:20
perfect. All right, first one
23:22
The question's about your views on
23:24
Yann LeCun's view that LLMs are dead end
23:30
First of all, in terms of achieving human level intelligence
23:33
these models have already far surpassed human intelligence in some ways
23:37
and then they're wildly worse than others
23:39
Like, for example
23:40
they seem much worse than people are
23:42
at very long horizon
23:46
kind of high judgment signal and tasks
23:51
On the other hand
23:52
yesterday, we had one of our models discover
23:57
disprove a conjecture
23:58
one of the Erdős problems that had
24:00
smart people had worked on for a long
24:02
time
24:03
And a lot of people
24:04
a lot of smart scientists
24:05
I don't know if Lacoon was one of them or not
24:06
had
24:08
even quite recently said something like that was
24:09
not going to happen
24:11
And then like the model just did it
24:13
And, you know, now you have all these mathematicians saying like
24:16
is math over? What does this mean for
24:17
our field?
24:18
So clearly, LLMs are capable of Figuring out new knowledge
24:23
and clearly they are capable of doing some things that
24:26
some intelligence tasks that humans just can't do
24:30
They are going to scale much further
24:31
so how much better
24:32
and what distribution of the tasks they can
24:34
do better than humans
24:35
we'll find out, but I suspect it's a lot
24:37
And the, you know
24:38
in terms of this like
24:40
lack of a belief in the exponential we
24:42
were talking about earlier
24:44
I think the field was
24:45
honestly held back by
24:47
a
24:47
generation of scientists who just were way too
24:50
certain on what wouldn't
24:51
what, what scaling was not going to produce
24:54
And then some people just looked at the graphs and said
24:56
well, it looks like it's continuing beautifully
24:57
let's keep going
25:01
I think world models are clearly important and to
25:06
we'll need that for things like robotics
25:08
but betting
25:10
against LLM scaling at this point
25:15
feels quite misguided to me
25:19
It, it
25:19
does it get annoying to be the I
25:20
told you so guy? No
25:21
I mean
25:24
there are these like Twitter trolls that
25:26
you know, for years have just been like
25:28
it's not going to work
25:29
it's not going to work
25:29
this is so dumb
25:30
like, you know, this is a fraud
25:31
this company is going to fail
25:32
this research approach is going to fail
25:34
And I used to get more bothered by them
25:36
but I don't even like feel the I
25:37
told you so at this point
25:38
It's like, you were just like
25:40
you're still going on about it
25:41
Like
25:42
the data is quite strong on our side
25:47
And I
25:48
don't think it'd be that fun to say
25:49
I told you so
25:50
and also the fact that you're like still
25:51
saying we're wrong doesn't really bother me
25:53
I think there's that kind of move on
25:55
There's that saying that like insanity is
25:57
doing the same thing over and
25:58
over again when presented with data that is not working
26:00
and I think you're repeating that
26:01
And in a sense
26:02
it's, it's, it's a form of insanity
26:04
I think. I, I think there's something that happens
26:07
which is if you make your identity
26:09
about a particular
26:11
thing is going to work or not work
26:14
And you associate yourself with that belief
26:17
and then the science or the empirical results disprove you
26:21
and you're like too hung up on your identity
26:23
you can't let it go
26:24
you can't see the truth
26:26
And I think this is like a important
26:28
reminder in both directions. Yeah

Watch these videos on your TV

XTTV brings video content from X to your living room. Download now and start watching on the big screen.