Today, I’m chatting with three 
of my AI researcher friends
from whom I learn a lot every time we talk.
They also happen to be at somewhat open-ish
labs and companies, so you guys can 
actually say things on the record.
I’m joined by Beren Millidge, who is the CTO of 
Zyphra, which is developing open source models.
John Schulman is the chief scientist at 
Thinking Machines, previously a co-founder of
OpenAI, and led the RLHF work that led to ChatGPT.
And Charlie O’Neill is head of
model training at Baseten.
The first question I have:
If we’re in 2036 and we don’t have billions of 
crazy superintelligences running around that have
radically transformed the world, what is the most 
likely reason that doesn’t end up being the case?
Other than exogenous political shocks, or 
there’s a war, or they ban AI or something.
What is the most likely technical reason that 
2036 isn’t a crazy alien superintelligence world?
There’s been a classic thing, almost like 
Moravec’s paradox, where we think of the AI as,
"If it can do this, it’s going to be amazing."
If it can solve these hard maths problems,
if it can win at chess, blah, blah, blah… Then it 
solves these things, and it’s not that impactful.
Obviously, it’s somewhat 
impactful, but not everything.
If somehow that continues, and there’s never 
the true spark of generalization that occurs,
I think that could lead to the AI just being 
extremely good at everything that people put
into a benchmark or put into an environment.
But there’s still some persistent sim-to-real gap
which is somehow blocking everything.
I think this is unlikely.
We do actually see this kind of generalization 
even from RL in practice already.
But if it is just ridiculously hard to generalize 
meta-learning, plus we don’t solve continual
learning and it’s just super hard and impossible… 
This would be my default scenario in that case.
I agree with that. Humans have a 
lot of advantages over models now.
Each time a new model comes out, 
it’ll catch up in some of these areas.
But you end up getting bottlenecked by the 
places where the model is weaker and where
it has worse judgment, or the models 
can’t check themselves well enough.
There’s this cycle that keeps repeating 
where a new model comes out and people are
blown away and they’re like, "This is it.
This is AGI." But then they use it a bit,
and it starts to feel dumb after a month or so.
That cycle just might keep going.
It’s hard to predict how many 
times it’s going to repeat.
Right now, you don’t get explosive 
growth in capabilities because
you still get bottlenecked enough when 
you’re trying to do research and engineering.
Even if the model can write way more code than a 
person, it doesn’t make you 100X more productive.
So maybe there are just more of 
these cycles than we would expect.
For me, it’s a question of how far off the 
global optimum of "a learner you could have
on a chip" is from the transformer + 
RL, basically the current recipe.
People imagine that once you have an agent 
which is better than all humans at AI research,
even if it’s 0.1% better than all humans, then the 
fact that you can run hundreds of thousands, if
not millions, of these in parallel — and you can 
run them much faster as chips speed up — is going
to outweigh every other bottleneck.
You’re eventually going to hit this very
fast takeoff with regards to self-improvement.
I could imagine that if we continue along the
trajectory that we’re currently on with that 
paradigm, where it’s basically self-attention,
RL, scaling up RL environments… Think 
about what happened with Moore’s law.
We had this very nice straight line and 
that held for a really, really long time.
But there were so many discrete 
discontinuities and innovations that had
to happen to keep that scaling law going.
The same thing has happened with LLMs.
We had this pre-training scaling law, and 
then that was hitting diminishing returns.
Then we came up with RL and solved that, 
and then we got this new diminishing
returns curve to hit that made it keep 
looking like a straight line going up.
So if it requires another one of 
those discontinuities to solve,
I’m not sure that the current method of 
training LLMs with these RL environments,
even RSI-targeted RL environments, would 
be able to discover that discontinuity.
If not, we’re probably going 
to hit this asymptotic curve.
But do you think the discontinuity will be 
harder than anything that’s come since 2012?
If we had the answer to that, we’d kind 
of have the ability to implement it.
But maybe we should distinguish between a 
discontinuity which adds to the current paradigm,
which is cumulative — there’s something 
beyond the RL that we have to discover,
and maybe they’re capable of connecting 
the dots in that straight line — or, again,
how far off the global optimum are we?
Do we have to go back and throw out
gradient descent and neural nets in general?
I don’t think, if you continue to scale up the
current paradigm, an LLM, no matter how many 
LLMs you’re running, is necessarily capable
of discovering that if it’s too far away.
The only hope really is if deep learning just
can’t get us to an AI which can at least 
dominate human research and human development,
including the human ability to come 
up with new paradigms and so forth.
Or, I don't know, maybe humans would also never 
have discovered the next learning architecture.
But to the extent humans could have discovered 
it eventually… But it just seems like… If you
just look at the progress that’s happened 
since 2012 till now, and you just continue
that on —I know it’s just been powered by 
huge amounts of compute scaling and so forth—
it would be weird if it just didn’t get to the 
point where it could dominate humans, at least
in R&D, especially over the next few years.
Ryan Greenblatt was on the podcast recently.
He made this point that I’d be 
curious to get your thoughts on.
You could imagine, as AIs get more and more 
capable, that they’re capable of making progress
on simulations which incentivize getting better 
at not only AI R&D, but at science generally.
This is a thing that all the labs are 
targeting and many startups are targeting.
Another intuition pump is if you look at 
the Elo score of chess bots since the ’80s.
There’s a very linear increase in Elo over time.
But there’s this huge discontinuity as they cross
the human range, from human experts always winning 
against AIs to human experts never winning against
AIs, as this linear increase in Elo happens.
I agree with your point that so far, AI
capabilities have not been that big of a deal in 
terms of their end economic impact in the world.
But that is because they’re slowly 
rising in Elo relative to humans.
I agree it would be very surprising.
The only way for this to not happen is if,
as you said, it somehow asymptotes just before.
Because we’re already pretty close, in my opinion,
to where we’ll start crossing the human Elo score.
So we’ll need to asymptote before that.
That’s the only way — in this 
scenario you pose where somehow
we’re sitting here in 2035 and everything 
is normal — for this to happen, I think.
The only other way is there’s 
some dramatic regulation on AI.
This is what I see as the most likely 
way for this scenario to happen,
actually, rather than a technical thing.
I think there’s different kinds of research.
There’s research in the autoresearch style where 
the objective is already specified very cleanly
and you’re optimizing that objective.
I think everyone is picturing that if
we continue along this path of making 
pre-training loss go down and making our
environments have the reward on them go 
up, that’s going to lead to improvement.
But maybe what Ryan is talking about is this 
much more open-ended type of science which
is required for paradigm shifts, where we can’t 
specify the objective, and the AIs are definitely
not able to specify that objective either.
We have to be really, really careful about how
we specify objectives for any of these things.
Maybe your point is that the nature of the
breakthroughs that have happened since 
2012 is that we have found… In 2012,
people weren’t saying… I’m assuming, 
I don’t know, you guys were there.
Or at least John, you were there.
But I was not.
I was in primary school.
Actually, John, I’m curious
for your wisdom of the ages, or wisdom 
of being in the trenches way back when.
Presumably, a big breakthrough was realizing 
that next token prediction is the… You wouldn’t
have thought that the nanoGPT speed run 
is the thing to be optimizing for in 2014.
But now that we have come to this new paradigm, 
you would think to do a speed run on that and
have AIs get really good at that.
But maybe there’s a next inner loop
to optimize that the AIs wouldn’t anticipate.
There’s an outer loop of revenue or something
that eventually should be strong, 
but it’s a very slow outer loop.
In fact, I remember in the early OpenAI days 
having the intuition that just minimizing log
loss wasn’t going to get you to intelligence.
Because the important bits are accounting for
such a small fraction of the loss that 
it was going to be overwhelmed by noise.
So just training a language model on next 
token prediction wasn’t going to learn the
interesting things you want it to learn.
We needed to craft better objectives that
would put more emphasis on the important things.
You can make all sorts of arguments for this.
You could say, "Oh, humans probably don’t learn 
how to model everything in our environment.
Most people can’t create a 
photorealistic reproduction of
some kind of scene they’ve looked at.
So we must need a better objective."
But then it turned out that it just worked anyway.
As you were pointing out, the inner loop,
even in current AI research, of post-training 
benchmarks or whatever, doesn’t necessarily
translate into what users like.
Oh, yeah. The whole field relies
a lot on generalization and it’s very hard to 
predict when you’re going to get generalization,
or when you’re going to get some kind 
of out-of-distribution generalization.
We know that if you train on the task you 
care about, you’re going to do better.
But the most important advances are often types 
of generalization that we have no right to expect.
For example, from just pre-training on this 
very naive next-token-prediction objective
to various tasks of interest that require 
understanding of the input in some deep way,
or learning some skill from pre-training 
that’s very rare and not heavily represented.
Then also generalization from these 
verifiable tasks to less verifiable ones,
this is also a type of generalization 
that there’s no reason a priori to expect.
This is an interesting question, because one 
intuition pump you could have for why you
would see some sort of singularity very rapidly — 
without even scaling up the inputs to AI progress
that are not just AI labor — is that before every 
single 7-figure experiment you run, you spend an
equivalent amount of compute on AI labor.
So you just have automated versions of you
guys spending a century thinking about what is 
the optimal experiment to run, doing small-scale
ablations, developing literally a century’s worth 
of theory, going back even before deep learning.
Before you decide what experiment 
to run, you’re doing extremely
optimal setting up of the experiment.
Then you do a century of thinking after the
experiment is over, where you’re analyzing what 
happened and what the next experiment to run is.
If you think hard enough, you probably could 
have expected some of these things beforehand.
There is probably some very clever way 
to do a small-scale experiment that’ll
let you build the theory that then will 
generalize to the large-scale experiment.
So I would expect that we’re nowhere near 
the ceiling of how well you can do research.
I would imagine a future where AI is doing 
a lot of analysis and theory building,
spending a comparable amount of compute to the 
amount that you’re spending on the experiments
themselves, doing various kinds of analysis and 
building a theory around what we’ve seen so far.
I think there are really concrete examples 
of this when the objective is well specified.
All thinking can do is update 
your posterior based on the bits
that you’ve gotten since you formed your prior.
You can’t gain any new bits from just thinking.
But when the objective is well specified 
and there is this data sitting around,
I imagine there will be this big speed-up 
in the current paradigm we’re in.
A good example of this is if you got an 
AI to think about the Kaplan scaling laws.
An AI at this point would have noticed, 
"Oh, they’ve just taken these intermediate
checkpoints and didn’t account for 
the annealing, and so this is wrong."
That would have been caught years earlier.
We would have cut off a year or two of
progress just from that observation from an AI.
Again, once the objective is well specified, which
is lower pre-training loss or whatever, there are 
many, many good examples where if you just thought
about it a bit more, you would have been able to 
cut down significantly on things that you’ve done.
So muP, and how learning rate scales 
with model size, and realizing that
model width is important in that as well.
I feel like you can really back out a lot of these
things and cut off a lot of low-hanging fruit.
I would imagine a 10x speed-up if our thing is
just, "Maximize the objective we’re currently on."
But I don’t see how that generalizes at all
to coming up with the right 
objective in the first place.
Just thinking doesn’t necessarily buy you 
the right objective in the first place.
I think this is really the key question for 
any kind of very rapid RSI from current AIs.
How well can AIs generalize to 
learning their own objectives?
To have any kind of self-propelling automated 
loop, we need the AI to propose objectives,
optimize them, figure that out, propose 
a new objective, and have this not go off
the rails at any point for a long, long time.
To come back to Moravec’s paradox, there might
be a case of Moravec’s paradox where we think this 
kind of autonomy and being self-encapsulated — so
we can think of what we should do ourselves 
and then go do it and have this loop — is super
easy because we always do this.
Obviously, evolution needs to
create creatures that can survive by 
themselves for long periods of time.
And this just might be something that for some 
reason is really hard for the AI, in the same
way that locomotion stuff is really hard but math 
is super easy despite being super hard for us.
But doesn’t the time horizon 
increasing suggest that that’s—
Yeah, exactly. This is another possibility, but 
I agree, there’s no obvious evidence for this.
In fact, the fact that our agents are now 
super persistent and it’s quite easy to
do this is kind of evidence against this.
But this would potentially be one of the
reasons why we just don’t get this 
immediate takeoff, if this is hard.
If you look back from 2012 till now — 
or maybe from when you started doing
your research till now — what part of all the 
innovations that have happened since that time,
including purely engineering ones, including 
purely conceptual ones, seems like the thing
that would be the last thing humans would 
have to do before AI totally automates AI R&D?
Probably just iteratively 
asking the right questions.
If you can get the AI to do any experiment, you 
still need to decide what experiments to do.
Right now I think AIs are not very good 
at this compared to coding the experiment.
Whenever we talk about research, 
they propose a bunch of miscellaneous
things which are very, very tiny steps.
Or even going from DeepMind’s approach of,
"We’re going to solve intelligence by learning 
to play games at a superhuman level," to one
random researcher like Radford being, 
"I’m going to try and just predict the
next token of a very wide swath of data"… 
Even once Radford had discovered that,
it took a while before people decided to 
scale it up, because we had to come up
with the idea of scaling laws and the fact that 
you could very reliably predict these things.
I would say that the last job for humans, or 
the role for humans that’ll last the longest,
is defining the objective and 
deciding what we actually want.
In that vein, something like deciding how the 
AI assistants should behave, or what it means
to be helpful, or what the objective is when we’re 
doing RL from human feedback, is one such thing.
Then later, defining constitutions and 
model specs is another one.
Even if the AIs can do all the technical 
work, we’ll still have to do a lot of that
and decide what we actually want.
Alignment is the final job.
Alignment is sort of the answer.
But alignment itself can be decomposed into
specification of the objective, or figuring out 
what the right objective should be, and then
actually achieving or optimizing 
the objective you’ve defined.
I think the first one is not 
going to go away anytime soon.
If I think about a post-training team and 
why you need a lot of people on the team,
it’s just because there are a lot 
of different areas where you have
to figure out how the model should behave.
It would be very hard to automate the whole thing,
just because someone has to think about 
how the model should behave in this area.
Jane Street started using Antithesis 
to test its software in early 2025,
and the team was so impressed by the product 
that it decided to invest in the company.
I recently caught up with Ron Minsky, 
who co-leads Jane Street’s tech group,
to ask how Antithesis actually plugs in.
The thing that I think is most impressive
about Antithesis is that we started 
using it on a team that was building
high-assurance software and being really careful.
Nonetheless, it was able to shake out bugs that
were otherwise going to be really hard to find.
That’s important both because it helps make
those systems more reliable and because it 
helps the teams that build them move faster.
This matters more and more as code production 
is increasingly automated.I think, in general,
as we’ve been using agents more and more, the 
key problem you run into is the verification
bottleneck: just the time it takes for 
people to look at code and figure out,
is that actually something you want to 
accept into your production software?
Tools that make testing better 
are just incredibly helpful there.
They ease the verification bottleneck and 
make it possible for you to get more stuff
done and move faster, because you can have 
more confidence that the code generated
by the agent isn’t introducing new problems.
To see how Antithesis fits into your development
process, go to antithesis.com/dwarkesh.
What is the story for why there isn’t huge
consolidation in model providers?
There are just so many things
that point to centralization here.
If you step back over the course of years,
is there something that is going to prevent that?
I think distillation is the main thing that
fights against the centralizing force.
Basically anything that can be learned
through RL can be distilled very easily, 
because it’s a small number of bits.
It’s something that you can learn 
from a small amount of data.
If you can get trajectories from the model that 
show a behavior, you can easily distill it.
I think distillation is one of the 
things that fights centralization.
There is also a possibility that there’ll 
be company-specific models, that it’ll be
possible to learn from deployment and have a 
company continually improving its own model.
Such a system could be provided by the 
current oligopoly of model providers
or some other currently smaller company.
But I think that’ll change the game a bit.
I also want to point out that continual 
learning, honestly, doesn’t stop distillation.
Even if your model is improving every day, 
people could be distilling it every day.
The loops could just operate at the same pace.
That makes sense. So copying model behavior…
I guess you need to know yourself what the 
right distribution to prompt is in order to
get the relevant model behavior?
Oh, yeah. For just distilling
with supervised learning, the prompt 
distribution is extremely important.
It’s very non-trivial to distill a model, 
even if you have full access to it and have
the chain of thought and everything.
It’s non-trivial to distill all of the
useful capabilities from it, because you 
need to prompt the model with something.
You need to prompt it with realistic prompts.
You need to have a really wide
distribution of realistic prompts.
One thing that’s been coming out recently is
that some of the Chinese companies are probably 
using these router services which are designed
to allow people in China to use the US frontier 
models, which would otherwise be blocked in China.
There are all these router or proxy 
services that allow people in China
to use these models, mostly for coding.
And these router services are collecting
and selling some of the data.
This is a very useful data set
for distillation because it gives 
you the perfect prompt distribution.
I think this is one of those 
things where AIs help a lot.
If you actually look at the frontier 
pipelines, or the Chinese models that
they’ve actually put in their papers, they 
get seed prompts from somewhere, which is some
combination of humans and this kind of data.
Then they synthesize a vast coverage from
those seed prompts using their existing 
models or the other frontier models.
You can automate an awful lot of this prompt 
distribution gathering and environment creation.
Humans need to provide increasingly 
fewer bits as the models get better.
But it still seems you’re bottlenecked by having 
a service which has users going through it.
Not necessarily. That’s obviously 
very helpful, but theoretically,
you can just think about what users want.
But the whole point is that the user says,
"Make me an application like this.
Oh, that didn’t work. I actually want
you to make this new feature.
But actually,
let’s step back and do this other thing."
Capturing that whole trace is the thing.
Or to the extent you could have done 
that anyway, then you just have RSI.
Ultimately, if you have this fully 
automated loop, that is basically RSI.
The AI is deciding the data, 
it’s deciding the training.
That is the loop. But it depends 
how much human information you need.
At some point, if you’re just 
like, "I want traces that look
like this," you prompt that to the model.
The model will be able to come up with a
pretty good approximation.
But what if you want to do,
"Make me a really good politician," and then it 
has to anticipate de novo how a discussion in the
Senate halls would go or something?
I just feel like there are going
to be a lot of things which are—
Ironically, this is actually easier for
the distillers than the frontier labs.
The distiller’s just like,
"I want a good politician."
They go to the frontier model.
The frontier model already knows how to be a good 
politician, so it just generates those traces.
Whereas if you actually want to 
build the first model that does this,
you have to actually somehow get data on 
what politicians do every day and build that.
It’s actually much easier to say, "I want 
something like this," and then get the AI to
produce a billion variations, than to actually 
create the thing like this to begin with.
I think you can actually make a really 
concrete prediction based off this observation
that the Chinese labs have this router data.
The thing that started this originally was I
was saying, "Isn’t it weird how Sonnet 5 and 
Opus 5 are almost objectively worse models
than GLM-5.3 and Kimi K3, even though they’ve 
had access to not only distillation but logit
distillation from Mythos?"
The counter was that the
prompt distribution really, really matters.
You need to see what users are doing so that you
can distill these behaviors and things in.
I think the prediction from this is that the
frontier labs don’t necessarily have much of an 
advantage, if at all, in RL environments now.
Yes, user distribution matters for general 
behavior and so on, but the best measure
of a capability is the very, very hard RL 
environments you’ve made at the frontier.
If you have access to those 
RL environments as Anthropic,
and you have access to logit distillation, and 
you’ve still made a worse model, then maybe—
Then real-world deployment 
matters more than the environment.
That’s really interesting. But they had 
to incentivize those capabilities in the
first place in Fable, or the frontier model.
So it’s weird that they can’t incentivize them
again with a smaller model or something.
Maybe we’re just in this weird uncanny
valley where trying to copy that frontier 
model too much, the student-teacher gap,
whatever it is, is just too large.
People have made this point with Opus.
The difference between Opus 4.6 and Opus 
5 is that Opus 5 really feels like it’s
got this AI-as-a-judge checking 
every possible thing it’s done.
That’s why it uses so many tokens.
It tries to think about all these things,
but it doesn’t necessarily have the big model 
smell of Fable to know when to stop doing that,
or when’s a good path to go down.
The reach exceeds the grasp.
I would offer a slightly different hypothesis.
I would say there are a couple of different axes
for the environments you can create.
One of them is difficulty
and the other is realism.
It’s comparatively easy to create a
lot of difficult environments that involve doing a 
much more complicated task or doing something that
requires a lot more cleverness.
You could say this is the
benchmaxxing distribution, because a lot of the 
most prominent benchmarks just involve doing some
very hard puzzle-like task that’s easy to verify.
Then there’s the realism axis, where you want the
model to be good in the realistic coding agent 
setting where there’s multiple back-and-forths
with the human and there’s multiple objectives.
The labs who are crafting the model behavior for
the first time need to push in both directions.
To get good model behavior, you need to really
push on the realism axis and have rubrics or 
some kind of human feedback that’s informing
the reward function you use there.
But if you try to do distillation naively,
you end up just matching the teacher 
on the benchmaxxing distribution.
If you don’t have enough of the environments 
that really exercise the capabilities in these
trickier realistic settings, then you’re not 
going to get those into your student model.
I think maybe one thing that’s happening is 
the big models generalize better from the
tricky narrow tasks to these more realistic tasks.
If you have a really good realistic prompt
distribution for distillation, you 
can match the big model really well.
But if you only have this distribution of 
easily verifiable tasks, then you can match
the big model on all the benchmarks, but 
you do worse on this broader distribution.
That might even explain something about the 
smaller Anthropic models, like Sonnet 5,
though it’s hard to predict exactly what 
they’re doing to post-train those models.
It could also be that they’re always changing 
their post-training stack, and they just got
a few things wrong in some of these models.
I don’t know… they turned something up too
high and created some quirks 
that people really don’t like.
It’s really easy to screw up post-training in 
some way that doesn’t show up in benchmarks.
Just one other very basic point 
is that the frontier AI labs buy
all their data from big data companies.
The Chinese can also just buy the same
data from data companies.
And they are, right?
And they are. Exactly. There’s a lot 
of people being annoyed about this,
but if they have exactly the same data and 
they can buy that, they can also distill.
It means it’s quite easy to keep up, really.
The other question I had is how the first
models that are capable of automating 
AI R&D will actually be trained.
There’s a toy version, which is this 
thing that Ryan was talking about.
You just have GPT-8 try to build GPT-3 size 
models that are really good at inner loop
type challenges: beating video games that require 
continual learning, or just getting to a certain
loss with the least amount of compute, et cetera.
But John, I think you had an interesting point
that maybe that’s not the way it 
actually will happen in practice.
So I’m curious, by the point at which you have AIs 
that are actually capable of automating AI R&D,
how are they probably trained?
We’ll probably do some combination
of learning from human feedback 
to absorb the researchers’ taste,
and just creating a lot of practice environments 
which involve doing multi-step research projects.
People will in practice do some 
combination of those two things and,
each iteration, patch whatever seems to 
be most broken in the last iteration.
Researchers will be using the AIs a lot and will 
notice that they have some consistent weaknesses.
Those things will either be patched by collecting 
human feedback or creating environments.
Maybe a useful way to think about 
this is how much of the lineage we
roll back and then let self-play from there.
In the limit, you’re picturing just giving
them a GPU and maybe neural nets or something 
and saying, "Okay, figure out how to train
a model to do these particular tasks."
The way it currently works is we go up
to the very edge of the lineage and say, "Okay, 
here are the bugs Anthropic has found in their
training stack in the last few months.
We’ll turn those into environments."
You need to train and get better on the frontier.
So you obviously lock in all the 
previous history of the lineage.
But you could imagine a world in which 
you roll back to before GRPO or something.
Then you have environments which try to get it 
to discover the best form to RL models on, and
then maybe you roll further and further back… But 
I think we will still be so compute bottlenecked
that people will just keep staying at the frontier 
and essentially diffing the bugs and whatever
improvements they found since the last model 
version, turning those into training environments.
Which is also really good for having 
non-stale, new data between model generations.
Again, this is basically continual learning within 
the AI lab, of distilling the last three months of
AI research progress through environments and 
RLHF-type stuff back into the model itself.
And it is distilling. That’s maybe why 
some of us feel like it’s asymptotic.
You’re always just trying to get 
the last three months of progress.
That progress is being contributed to by AIs, of 
course, but it also still has humans in the loop.
It feels like you’re just constantly inching 
closer and closer to what the human researchers
are finding and capable of doing.
The one thing I will say, though,
is obviously if you’re just distilling on 
trajectories, you can never go above it.
But environments can go quite a 
far way above what a human can do.
It’s very easy to design an environment 
that no human can solve, but the AI can
obviously still try and solve it.
That would be the path to go ahead
of just what the human AI research is.
Do you have an example in terms of RSI,
of what kind of training set?
Nanochat speedrun, but doing it
even faster than a human speedrunner.
I feel like in AI research especially,
it’s very easy to define goals.
You could say the loss needs to be 1.3 or
something, and no human can get that now.
But that’s an extremely measurable,
verifiable task.
If the AI gets that, then great.
Or I don’t know, building a 100 million 
parameter model that beats Minecraft.
That’s maybe too easy, but beats a much 
more complicated game or something.
Isn’t it crazy that 100 million 
parameter models beat Minecraft?
We’re calling that too easy?
Imagine if you said that five years ago.
I would say a lot of research is 
not exactly like that, though,
where it’s hill climbing on a well-defined goal.
It’s more like, here’s an intuition we have about
some way models should be better.
We also have some idea for an algorithm
that seems to go a little bit in this direction.
So let’s come up with a task that is designed
to show signs of life on this approach, 
and see if we get those signs of life.
If we do, we can make successively 
more realistic versions of the task.
It’s a lot more guided by intuition.
The inner loop is to test for that
intuition rather than the test 
itself leading to the insight.
Right. You’re not directly optimizing 
for the eventual objective you care
about or the practical production objective.
You’re relaxing your objective a little bit.
You’re saying, "Let’s relax on the realism axis 
a little bit and find some methods that actually
work, and then try to get back to realism 
later after the method matures a little bit."
There’s also research that’s more oriented 
towards explaining things and developing a theory.
Often we don’t have mathematical theories in 
machine learning that are that predictive.
But we have a lot of more informal 
theories for what’s going on.
Presumably the models will be trained on 
some combination of all of these tasks.
Some will be very easily verifiable, some 
will be LLM-as-a-judge or just ask the human,
"Does this look reasonable?"
The hope would be that these would
all generalize to these much harder, 
more vague, fuzzy kinds of tasks.
It probably will to some extent.
Whether it generalizes enough that
the loop can become self-sealing without 
humans being in the loop at all is unclear.
Maybe taking a step back.
Here’s what it seems to me
the plan for AI research going forward is.
You tell me if you think it’s going to work
or if you agree with this characterization.
The bet is that we will scale up RLVR training
across millions of diverse environments, 
across hundreds of different kinds of domains.
What will emerge at the other end is an agent 
which has learned these basic skills — or less
than basic skills — around being persistent, 
being able to triage information and context,
eventually having end-to-end optimization of 
working with other agents and things like that.
Such an agent will be very sample efficient within 
the context —you've done research on how you
scale up in-context learning to make it 
arbitrarily long, but you keep scaling it up.
And what comes out the other end will be something 
that basically functions like a drop-in remote
worker over the course of a week or a month.
First of all, do you agree that that is
the bet the labs are making?
And second, is that enough?
Basically learning how to learn within 
these simulacra within a data center,
and then getting deployed into the real world, but 
not actually learning from real-world deployment…
only learning these meta skills from the 
simulated environments in the data center.
I think it’s now hard to separate out 
how much of the labs’ effort is going
towards direct RSI versus making generally 
intelligent models that they can continue
to deploy to collect revenue to 
fund the next big training run.
For the latter, yes, that’s probably 
just the bet they’re making.
It’s very clear, the pattern of where these 
environments are going over the last few years.
Anthropic’s lineage of environments 
is a very clear example of this.
First, we just focus on coding and we’re 
going to get really, really good at that.
Then, from the task horizon that we’ve 
got from coding — which is probably the
lowest-hanging fruit in terms of data available 
on the internet to create environments, and
their own internal stuff that they can turn into 
environments — then we’re going to generalize.
We’re going to go up to finance next, and 
literally just so much Excel data and all
that sort of stuff in the RL training.
Then it’s PowerPoints. It’s this long
tail of the working economy.
That seemed to work really well.
A lot of the other labs, even the 
open source labs, have now realized
that that was the correct bet to make.
But what is the implication from that?
When I had Dario on the podcast, the thing I 
asked him was, if you truly expect models which
will be human-like in their ability to learn on 
the job, why would you try to bake in all these
skills of working with PowerPoint or something?
Wouldn’t you just expect the model to be able
to pick that up while it’s deployed?
There’s multiple different explanations.
One is just that we expect models to get there 
soon, but they’re not there yet, so why not
amortize these skills into the model training?
Another is that we’re not concentrated on making
it really good at widely deployed work.
We just want it really good at RSI.
This is just a way for us to get revenue so that 
we can pour it back into a model that is actually
really good at doing RSI development.
Then once the singularity happens,
the thing that comes out the other end will be 
really good at all the things which seem like
bottlenecks to the current generation of models.
John, I don’t know if you have takes on
how one should construe why there is so much 
task-specific knowledge in these models if the
path is this kind of generalization.
If the models were good enough at
learning in context, then in theory, you 
wouldn’t need to train them on finance.
They would just be able to read all the books 
on the fly and figure out how to do everything
in the appropriate jurisdiction.
You could argue that you need to do
a lot of this domain-specific training 
just to make them more efficient.
Even if they were smart enough to figure this out 
on the fly, you still might want to do a bunch of
RL and bake all these intuitions into the weights, 
so the model would be more efficient at runtime.
In practice, it does seem like model providers are 
going domain by domain and trying to strengthen
the models in the highest value domains.
I’d say that that’s one of the answers to
why the models have gotten so much better.
It’s just because the model providers have
covered a lot of the high-value domains 
and the most common types of skills.
Another thing is just that it’s not that 
expensive to do both at the same time.
The models are massive. They can easily afford, 
in terms of their parameters, to learn everything.
There is likely some transfer.
Even if finance is not specific,
the information is important for RSI.
Just the general meta-learning of how to figure
out what’s important, how to have taste, how to 
do long-horizon work is potentially generalizable.
There’s not that much RSI 
data in the world as well.
It’s hard to generate and 
requires a lot of effort.
So if you can amortize in this other 
data, you get some transfer from it.
You already have masses of compute 
and masses of parameter space,
so why not do that as well as, obviously, the 
direct commercial intent of selling a model?
I’ll add that there’s one question about 
whether this current paradigm of doing
sim-to-real will be the dominant one forever.
You look at what the real-world tasks are like.
Then you try to create a bunch of 
environments that can be simulated
in the data center, and you can do RL on them.
Obviously, this has been very successful.
But it has a lot of weaknesses, because a 
lot of things are just hard to simulate,
especially if they involve interacting 
with a bunch of humans in real time.
So there’s some question about whether sim-to-real 
will be the dominant framework forever.
I think sim-to-real has to be the dominant 
framework while sample efficiency is low,
because right now you need thousands and 
thousands of interactions with the humans.
No human is going to sit there 
and be in the loop of RL training.
So we have to simulate that now 
to get the samples you need.
But obviously, if sample efficiency 
improves a lot, you’d expect learning from
deployment to become a much bigger part of it.
Though there are also other things you could do.
You can learn off-policy, so you can take all the 
traces, and even without resimulating everything,
you can potentially learn something from them.
Jane Street just launched a new competition,
and it’s its most ambitious one yet: 
design a protocol-emulator ASIC.
Basically, if you have a chip you want 
to test, you can connect it to this ASIC,
and the ASIC will simulate realistic traffic.
That way, you can see how the chip responds
without having to plug it into a live system.
Jane Street is looking for flexible,
general-purpose designs, not 
single-protocol emulators.
When I was chatting with them, they suggested that 
I start by trying to implement what are apparently
three very common protocols: UART, SPI, and I²C.
Jane Street also mentioned that they hope more
ambitious designs will tackle low-speed USB, 
Ethernet, and any other protocols that flex
your chip’s specific architecture.
Importantly, your design should be
reprogrammable rather than smashing a 
bunch of specific protocols onto a chip.
If a new protocol comes out after 
your ASIC is taped out, your chip
still needs to be able to handle it.
How exactly it does that is up to you.
But there is one hard constraint: your design must 
target an open-source 130-nanometer process node.
That’s because Jane Street will pay 
to tape out the most novel submissions
and send physical copies to the winners.
The competition is open until January 18,
2027, and working in teams is highly encouraged.
Go to janestreet.com/dwarkesh to download the
template code and get started.
I want to ask more about this,
because it’s weird that you have 50% of 
compute that’s spent on inference that is
not directly helping the model become better.
One of the key advantages you’d expect digital
minds to eventually have is that, 
unlike a human who gets to have 50
years of real-world experience, a model will 
get to experience, through all its instances,
millions of years of deployment across all kinds 
of economically relevant work in the economy.
Right now, that data is just not, in a 
meaningful sense, helping the model get better.
It seems so obvious that eventually models 
should be able to learn from this data.
Once they do, you would have something 
that almost feels like a widely deployed
intelligence explosion, because the 
model is assimilating so much information
across all these deployed instances.
When do you expect this kind of hive-mind,
crazy shit to start happening?
I think broadly, at a very basic level,
this is already happening… just 
in the next generation of models.
Right now, you can obviously take your deployment 
data and put this in the pre-train or the
mid-train of future models, especially if you do 
some kind of filtering or some kind of judgment
or annotation or synthesization of that.
How much do you think that explains the
generation-over-generation improvement?
I think it explains quite a bit.
I don’t know whether the labs do this, 
because theoretically, they claim not
to train on people’s data.
But the Chinese 100% do.
They definitely get this advantage.
This is basically what distillation is.
They take the models, they get some fraction 
of their deployment data by pinging the model,
and then they train their next 
generation of models on it.
They can certainly do it on 
their own models as well.
There’s no reason not to whatsoever.
I completely agree with this.
If you zoom out far enough, 
this is definitely happening.
What we’re all picturing, the holy grail of 
continual learning, is this very organic,
live loop of an individual model getting 
an experience and live-updating on the
spot and learning from that.
A lot of things break when you
zoom into that level of granularity.
But the big labs are doing this.
The closed models are doing this.
There are also early signs of life
of people using open-source models 
doing this at a much faster cadence.
A good example is probably Composer.
Harvey’s doing the same thing with legal agents.
You have some sort of model, and you are 
getting very specific environments from the
data that you have for that particular task, 
and things that users are complaining about,
and all the feedback that you’re somehow 
extracting from your specific deployments.
A lot of these companies have the 
advantage over the big labs in that they
can use this data really, really well.
Then they will create environments.
They will do a big post-train of Kimi K3.
They will go deploy it.
They might do some online learning as 
well, like Composer did online…
basically REINFORCE for a long time.
There’s still a human in the loop.
There’s still a human saying, "Okay, 
these are the signals we care about.
Here’s how we’re going to create 
environments from the data that we have."
It’s still a longer cadence than 
maybe the one that you’re thinking of,
but it really is happening.
Eventually that loop will
become faster and faster.
The Composer thing is interesting
because this is where, in Cursor, people 
press Tab or they don’t press Tab on the
next completion that the model suggests.
Based on that, every single day, Composer
gets better at predicting the next—
That was the old Tab model.
They actually did the same thing not just for the 
Tab model, but for the actual generative model.
Oh, I see. Interesting.
It’s hard because when you do online
reinforcement learning, you don’t have groups.
You just have one user saying one thing,
and then you get one rollout.
So you have a big variance-reduction problem.
Cursor’s fuzzy answer to this was, "We have very 
good heuristics which are able to estimate how
much better than average this response was, or 
how much worse than average this response was."
Then they would do this big REINFORCE update.
Their solution to whether it got worse or not
was that if it improved on CursorBench, they 
would deploy the new model every five hours.
If it didn’t, they would throw that version out.
I think your biggest problem is actually
just not knowing what the reward 
function should be for natural data.
If you use some kind of superficial 
signal, like did they accept the edit,
that might get reward-hacked in some way.
But isn’t this a bigger issue with the sim-to-real
thing, where the longer-horizon tasks get, the 
harder they are to simulate within a data center?
It seems to me that even in coding, we’re already 
getting to the point where there’s not some
year-long coding task that doesn’t eventually 
require you to talk to a client or interact
with the company or interact with users.
If you think about the gamut of things we
would want AI to be capable of, eventually 
superintelligence should be able to run a
business, or start a new business and make it 
profitable, or have a profitable day trading
in the markets, or win a court case.
These are all things which are very
hard to simulate in a data center.
An inherent part of the learning
there is interacting with the real world.
Maybe they learn how to get better at these
things from the transfer between sim-to-real.
But alternatively, maybe you do need weight
updates from these kinds of interactions 
in order to get better at them.
If that is the case — if transfer isn’t 
strong enough and you do need weight
updates — then the fact that the models are quite 
sample-inefficient is maybe a deeper problem.
The reason I’m curious about this 
is that by default, I don’t see how
you don’t get some kind of crazy recursive 
self-improvement within the next 10 years.
But the one reason why that might not happen is 
that in terms of the sample efficiency of weight
updates, models just seem way far behind humans.
They’re plausibly a millionfold behind humans in
terms of how much data a human sees from 
birth to adulthood versus how much a model
sees from cold start to finishing training.
This is all to say, first of all, is there going
to be good transfer between simulations and the 
extremely long-horizon, really complicated real
shit that we want the AI to do in the real world?
And if not, does that really mean that the
lack of sample efficiency in 
these models comes to bite us?
Maybe the way I’d break down the two types 
of tasks — the ones in which models get good
and the ones where models will still continue 
to struggle — is whether the task is cumulative,
or whether you have this non-stationary 
distribution where you have to keep
learning and relitigating a bunch of stuff.
An example of a cumulative task might be RSI.
It’s theoretically possible to have a 
less-than-a-million-token Python file
which from scratch trains a model that 
is capable of recursive self-improvement.
Every discovery that you make is 
a line in the sand that you hold.
If it’s true that for RSI we don’t need to 
discover a new attention variant or whatever,
then once you’ve discovered attention, and 
once you’ve discovered mixture of experts,
and once you discover GRPO, you just add 
that to the training stack and that’s there.
A good example of this is 5.6 Sol training, 5.6 
Terra, or whichever one OpenAI told us it trained.
It didn’t have to go back and discover attention.
It basically would have called a bunch of
scripts, like pre-training.sh and 
post-training.sh, and just done that.
That’s an example of a cumulative task.
I think the real world — and the reason
people are thinking so much about continual 
learning — is not really a cumulative task.
Imagine in a law firm, you have an 
agent acting as a legal associate.
That’s a very non-stationary distribution.
You have to be able to fit in your 
context all the relationships between
all the important people at that company, 
which are also changing all the time.
You have all these implicit ways about how things 
are done, where to find information, et cetera.
That’s not as clean an example 
of a cumulative task as RSI is.
I think there will be this 
breakdown between tasks.
But if the labs realize that — and they do 
believe that RSI is cumulative in the sense
that we don’t need to go back and discover some 
brand-new architecture or whatever — then maybe
more and more effort and compute gets 
focused on that versus the other tasks.
It’s so unfortunate that RSI happened 
to be easier than being a paralegal.
I would say today’s models are weaker 
than humans in a lot of different ways.
Some of them might have to do with 
sample efficiency in a certain regime.
In some regimes, models are very 
sample-efficient, like learning in context.
But then there might be some medium-length 
regime where they’re less sample-efficient,
because humans can do some kind of weight 
update more efficiently than models.
I think being less sample-efficient in certain 
regimes might be one of the sources of weakness.
But I think there are other sources of weakness 
that are completely different from that.
For example, having lower diversity of 
thought than humans, or being bad at
certain kinds of long-horizon judgments.
I think a lot of what people call taste
is something about behavior that 
works in the long run, and that
people have realized works in the long run.
Not everything, but some aspect of taste.
Especially for something like software 
engineering, I think a lot of taste is "What
are the systems that are going to be maintainable 
and work well in the long run of this project?"
There are a variety of weaknesses of models 
which limit RSI along with other things.
Some of them are related to sample 
efficiency, and some of them aren’t.
Maybe an interesting thought experiment is this.
Let’s say you were able to give a model a context
window of a trillion tokens, or whatever 
you would have needed to fit in your
experience prior to, let’s say, RLHF.
It’s got all that experience in the
context window, and it has the same sample 
efficiency and in-context learning ability
as it does at a million tokens.
Do you think taste is then solved?
Would it be able to make the 
same judgments that you did?
Or is there something fundamentally missing, 
apart from just a longer context window with
the same sample efficiency?
It would have to be trained
to learn from that context.
Either it would have to be
trained to learn the right update to make from 
that context, or it would have to generalize.
So you don’t think you can just dump it all 
in, your whole life, your research experience?
You still need the data to 
train it on long context.
Even if you could theoretically get a trillion 
context, you would need a trillion lengths
of data to train it.
Right now you have 10k context,
you can't just dump in a million.
Yeah, I’m just asking if you had that.
In theory, I think, yes.
This really just comes down to the
question of how meta-learnable taste 
is from shorter-horizon episodes.
I feel like there’s no obvious reason it’s super 
long, because humans somehow developed taste
without having many long episodes.
We don’t live to be 10,000.
We develop pretty quickly. If you think 
about even a PhD, the difference between
a first-year PhD student and a final-year 
student or postdoc, that’s five years maybe.
They’ve only done maybe 10-30 
research projects in total.
But somehow they develop taste quite quickly from 
a relatively short succession of small things.
Theoretically, it’s possible 
to develop it like that.
The AI obviously will have vastly more experience 
in which to develop taste, to meta-learn it.
Then the question is how well that generalizes 
to really long-horizon things, which I think is
really unsolved at this point.
We don’t know.
Going back to this question, eventually 
there should be a regime where AIs are
learning a ton from each individual 
instance of deployment that they have.
Currently you could say 
there’s a fuzzy meta process
by which models do improve from deployment.
But I feel like it’s a very weak feedback loop.
Do you see this on the horizon, where there’s 
this hive mind kind of learning that’s very rapid,
and if so, how exactly does it happen?
I would say that whether we get a hive
mind that learns from all of its deployment 
experience is in a big part about incentives,
rather than being a technical question.
Companies aren’t going to want to
have the model provider learn from all of 
their deployment, because that might just
reduce the advantage of their business.
I think the economics of this will pressure,
not necessarily weight updates to one big common 
shared model, but modules that get subbed in.
A very obvious example of this is a 
LoRA, but it might be something else.
There’s been a lot of work to try and fit an 
arbitrary context length into a fixed size.
This is all the linear attention stuff.
And cartridges, which are essentially
KV caches trained to be very, very 
compressed to fit in a lot of information.
That’s another example of something that companies 
may be willing to sign up for, if that gets
subbed into the model and it’s not actually 
changing the base underlying model itself.
There are many different versions of 
learning from your data in real time.
The latter ones are not really helping the 
big labs, because they are just these modules.
But I think the economic pressure 
will force the labs to go down that
path first before they can embark on this...
Which economic pressure, though? I feel like
even if you have a bunch of cartridges 
or LoRAs or whatnot, you can still just
take all these traces and dump them into the 
pre-training of your next generation of models.
Yes. It may be a more indirect form of 
learning that the big labs are getting.
That’s obviously still really valuable to them.
But I can’t imagine a world in which we start
off with, "We’re going to just 
directly train this one big
model on all the exact data that we’re getting."
No, I think it will definitely go through stages,
because this is assuming there’s one 
discontinuous event where suddenly
we fix weight updates continuously.
In practice, I think it’s much more
likely to be that the cartridges and stuff 
allow you to specialize in deployments.
Then you generate traces, you put that in 
your model, and three months later you come
out with a model which is better at this stuff.
You specialize it again, you consolidate it again,
and then eventually we’ll just 
make this leap faster and faster.
Instead of releasing a model every 
three months, now it’s every week,
and then every day, and then every hour, 
at which point we’ve basically solved it.
I think this is a good point as well, 
because you asked how far off the current
paradigm we are from being able to do this.
We’ve done a bit of research into this,
and people have done a lot of research.
At a really large scale, when you wash
out enough noise and you have large 
enough batches, this outer-loop process
of putting data into mid-training and 
creating our own environments does work
in some sort of continual learning regime.
But the problem is, when you zoom in close
enough at a micro level — I’ve got one model 
and I’m trying to update it again for a law
firm or something, and I’m trying to do that very 
continuously with a relatively small amount of
data — all the methods kind of break down a bit.
If I SFT the model on just successful traces,
off-policy or on-policy, eventually in the 
very iterative regime, when you’re doing
hundreds of these micro-updates, 
you see catastrophic forgetting.
You see forgetting of previous information learned 
on top of the base model that was much earlier on,
and you see degradation of general capabilities.
On-policy distillation seems to push this 
horizon out a little bit, but it still
eventually succumbs to the same thing.
RL is good at getting capabilities in,
but it’s not as good at getting knowledge 
in, this very explicit knowledge of, "Ah,
okay, this person does this at this law firm, 
and this is a very specific process we find."
You have to pour in a lot of 
compute to create the right
environments to get the knowledge in with RL.
Do you think the fundamental issue here — why
you get worse at these other skills or there’s 
forgetting — is fundamentally an issue of capacity
or an issue of techniques?
A little bit of both.
I think SFT and even on-policy 
distillation can be way too destructive.
The reason RL is so nice is because it changes 
a very, very small amount about the model.
There’s a lot of evidence 
for why this is the case.
It just tweaks it in this very, very small 
loss valley to get it into the right point.
But that also then limits what you can do with 
RL, how much you can actually change the model.
So you’re saying the reason this 
isn't a winner-take-all, potentially,
is that it is just very hard to distill 
that much information into the base model?
Without ruining something, 
in an iterative fashion.
It’s easy to distill it 
into a different base model.
This is where I think it’s mostly technique.
It’s definitely not that there isn’t capacity.
If you had some model with all this data, and 
you take literally the same-size model and
pre-train it from scratch with all of the 
stuff in mid-training, it will be better.
I think that’s a lot of what’s happening today.
There’s very much a bottleneck that stops us from
just keeping training the same model forever, 
versus just getting all the data from the old
model and training a new model from scratch.
This is exactly as Charlie was saying:
some combination of plasticity and 
catastrophic forgetting.
If you just naively train on non-stationary 
data, because you’re adding new data as you go,
this is messing with the data distribution, 
so the old stuff is just forgotten.
We don’t really have good methods 
to stop that from happening.
So maybe in the limit you’re just 
bottlenecked by retraining the model
from scratch with all this new information.
Yes, which of course is very expensive.
Training a model from scratch is expensive.
But you’re going to do that anyways.
Not necessarily. Maybe eventually, if you have 
continual learning, you never train a new model.
You just have a model and it 
keeps learning and expanding.
But there might be some deep technical 
reason why that’s very difficult.
That’s the question. I think we have pushed 
back how much from scratch we need to do.
It is definitely possible now to 
take the pre-trained base and do
very good mid-training on top of that, kind 
of continuously, plus some RL from different
checkpoints that are later on in the training.
That’s looking more like continual learning,
but it’s certainly not the case of taking the 
most recent model, applying a couple of very small
updates, and iteratively never losing anything.
Sorry, but I’m a bit confused, because isn’t this
literally what happens during training?
During post-training or something,
you have a model that’s already gone through 
so much training, and then you distill some
fork that’s been further RL’d.
Isn’t that literally what happens?
But it’s still at a large enough scale, 
I think, that you’re washing out a lot
of the noise, and you’re not just focused on one 
distribution, which, as Beren said, is the issue.
If you’re just focusing on one task—
But in the eventual regime you’d be doing…
There are billions of deployed instances.
You’re learning from all of them at once,
so hopefully there’s some 
washing out of noise from that.
Maybe at that scale, yeah.
As Charlie was saying,
you can definitely do continual mid-training 
for a long time, and you can roll back to a
checkpoint and give it new mid-training data.
But at the same time,
you can’t do this indefinitely.
If you just keep continually training the
same base forever, it asymptotes at some point.
You can’t just learn new stuff in that base.
This is why people end up training new bases.
Otherwise you would just keep mid-training
the same base forever.
Whenever I finish recording
an interview, I immediately brain-dump all 
my thoughts into Slack—things like what was
most interesting and what should get cut.
This ensures that my editors have all the
context they need to start editing the episode.
But these brain dumps don’t have clear timestamps,
and my unedited recordings are many hours long.
It can take a ton of editor time just to find the
exact moments I was referencing.
So we decided to try adding
a Grok Bot producer to our chat.
Now, whenever one of my editors posts
a rough cut of an episode, Grok Bot opens the 
transcript on its own computer and starts working,
usually before I’ve even seen the message.
It takes the notes I dropped into Slack
and highlights the relevant 
snippets in the transcript.
It also uses a big case file I’ve 
compiled with all my preferences,
so it can suggest potential edits.
When it’s done, it sends me its
top clip candidates so I can 
review everything from my phone.
This has worked really well.
Being able to send informal messages,
like I’m texting my editor, and then have 
the transcript immediately reflect my
preferences has just been so helpful.
Try Grok Bot yourself at x.ai/bot.
Let’s talk a bit about data now.
I’m generally interested in this
question of how much of AI progress 
is just explained by data progress.
That doesn’t mean it will necessarily be hard 
to automate, but that's a separate question.
Is there some data distribution which, if you 
trained current architectures on it, would result
in a superintelligence that totally dominates 
human experts across every single field?
Are we talking about pre-training plus 
post-training data, environments as well?
I think the existence of this is obvious.
It’s just whether we can create the right
environment to get there.
In the trivial case,
we could just train it to output the Python 
file which trains the actual superintelligence.
Just have that memorized in the weights.
Yes, there’s probably a ladder of RL
environments that is possible to construct such 
that you would get an AI researcher which is
at least as good as a human researcher.
But the effort to climb each successive
rung grows kind of exponentially.
Those are the two things you have to
trade off against as to how fast we’re going 
to hit that final rung where it’s better.
I think that’s fairly clear.
We’re still relatively
early in RL environment creation.
There are a lot of asymmetries that
we exploit in order to create good environments.
One of the asymmetries which we’ve talked about
before is that there are environments where 
it’s easier to go backwards than forwards.
What I mean by that is, it’s very easy to 
define this complex data-generating process,
and this is the latent variable 
you keep hidden from the model.
You can generate arbitrarily complex environments, 
and the model has to do a lot of irreducible
token spend and irreducible work to figure 
out what that data-generating process was.
There are asymmetries in terms of 
injecting information from the real world.
Anthropic finds a bug through tens of 
thousands of humans and LLMs combined,
and turns that into a very, very neat 
environment which a single LLM could
theoretically find within a few million tokens.
There are all these asymmetries which we’re
cherry-picking, and we’re counting on 
this kind of task-horizon generalization.
But I think it’s just going to hit diminishing 
returns at some point, diminishing returns in
how hard it is to create those environments in 
the first place, coming up with them, because
you can’t necessarily just have these processes 
where it’s easier to go backwards than forwards.
You actually have to sit down and construct 
something that looks like a long enough time
horizon with humans, and it’s going 
to be a really complex task to create.
Then there are also going to be the 
compute and time bottlenecks for the
agent to actually do those tasks.
I think you’re just going to start
seeing this curve flatten out.
I saw something about how someone
fine-tuned the Talkie model, which is only trained 
on data up to 1930, on modern coding agent data.
It did better than Claude 3 Opus on SWE-bench.
So this model that has no knowledge of code
whatsoever can be fine-tuned on 
a moderate amount of data and
behave better as a coding agent than this much 
larger pre-trained model, which is pretty crazy.
It kind of shows you that once you have 
an example of the right expert behavior,
it’s actually surprisingly easy to 
copy that into a relatively weak model.
But a counterexample to that is a paper 
recently where they trained a model up to
fifth-grade maths, and also primary-school English 
and stuff, so it was a decent language model.
They tried to RL it to do late 
high school and college maths.
The gap was just too large.
They couldn’t get it to climb at all.
But if you did successive rungs of year 
7 maths and then year 8 maths… and so on,
you could obviously climb to year 12.
Again, it’s just what is the distance
between the rungs on those ladders, 
and how hard is it to create?
This just comes back to the RL signal problem.
RL is not very good at exploring right now.
If the model can’t get it in 128 rollouts, 
it’s very unlikely to get signal to progress.
This is why in RL we need curricula, whereas 
in pre-training we don’t, because that’s not
a problem for pre-training at all.
Again, pre-training data is
different to post-training data.
I imagine as we continue on, humans
will be involved less and less, but that doesn’t 
change the fact that you’re bottlenecked on how
much signal you can extract from the real world.
There’s a lot of signal in the world,
and that’s true.
There’s people doing
spreadsheet tasks, there’s people doing 
legal tasks and all this sort of stuff.
But at the capability frontier 
of where the models are at now,
how many bits in the world are actually really 
relevant to improving the model’s capabilities?
How many new maths problems are 
being solved that are just beyond
the reach or grasp of the current models?
How many new coding problems are being
created or solved that are beyond 
the reach of the current models?
I think that’s why the diminishing returns kick 
in, because even the world as a whole is not
giving you the bits, going back to the start 
of this, that are useful for tipping you into
the next basin of capability.
I totally agree with this.
It’s really a question of where 
the signal is coming from.
In pre-training, the signal 
is already in Common Crawl.
For the tasks that you care about in pre-training, 
the problem is not getting signal at all.
It’s filtering out all the noise that exists.
That’s quite an automatable process.
But as the models get better, as we 
enter mid-training and post-training,
the signal just doesn’t exist 
anywhere in the original data we have.
No amount of filtering will get this.
There’s no hidden proof of a Millennium
Prize problem sitting in Common Crawl 
that we can just filter until we see it.
At that point, you have to get bits some other 
way, either from humans directly, asking them
to write out their reasoning, or by creating 
environments where humans decide what environment
should be created and what the objectives of these 
environments are, or some kind of training on
the human data that exists in deployment.
You have to get the bits from somewhere.
There’s a question of how much of the progress 
in pre-training is being driven by data.
I did this investigation with Jerry Han, who’s 
a student at Princeton, where we trained all
the recipes from 2019 till now pairwise 
with all the data sets from 2019 to now.
You’re training GPT-2 on the newest 
data set, like Ultra-FineWeb.
You train Delphi, which is the newest open source 
training recipe, on the Pile or some old data set.
You do the whole grid.
You see, getting to some
level of capabilities, how much less 
compute does it take, across this grid?
You see that the data seems to explain 
something like a 12.0x compute efficiency
gain, but the architecture improvements 
explain something like a 3.7x compute
efficiency gain, at a very small scale.
To the extent that that is true at large
scale — that most of the pre-training 
compute efficiency gains are coming from
better data — how much can that continue?
Can you keep filtering data more and more
and building more and more synthetic data?
Do you have a sense of how much this kind of
pre-training progress can continue?
My prior is that, again, the
low-hanging fruit is somewhat exhausted.
We got the internet as this big block,
and it’s not like the internet is 
necessarily growing at the same rate.
All the useful stuff on the internet 
isn’t growing at the same rate.
We’ve probably got a bunch of 0.1% 
loss drops to go, but definitely
not as many as have currently occurred.
But it’s also really interesting that you
find this cumulative 33x improvement across both.
I think it was Epoch or someone who estimated
3x a year since 2019, which would imply 
something like 3⁷, over 2,000X improvement.
So where’s that missing 100x 
or whatever coming from?
That probably gives you a good signal 
of how much of this is post-training.
I think the explanation has to be that a lot of 
the compute efficiency gains are scale dependent,
and we’re starting at extremely small scale.
That raises a question of whether the data compute
efficiency gains or the algorithmic compute 
efficiency gains have more scale dependence.
I don't know if you have a prior on that.
We just didn’t have enough compute 
to investigate that question.
Just naively, theoretically, the scale dependence 
of the architecture is fairly well known,
and you can fit a straight line to it.
Whereas I would have no idea how to
do that for combining pre-training plus 
post-training data and mid-training data.
Funnily enough, I feel like data is 
actually more important with scale.
I feel like architectures 
are kind of a one-time thing.
Saying just an X% efficiency gain is kind of 
misleading, because what an architecture does is
let you reach a qualitatively new regime which 
you couldn’t reach with the old architecture.
Within that regime, obviously the data 
is the primary thing determining it.
But if we didn’t have even GQA, if we were 
doing full attention all day, it would be
ridiculously expensive to do a million context.
Because of that, we could never use the data
which is actually at a million context, 
so we couldn’t get these capabilities.
Even though if you just do a naive "how much does 
this do at 2K context", where the architecture
isn’t unlocking anything, then the data will look 
much more important than in some sense it is.
It’s unclear to me that these things are 
really just multiplicative gains in this way.
I see. So what’s your take on 
the scale dependence of data?
On scale dependence, I think a lot of 
the mid-training and post-training data
we have now actually gets better with scale, 
because a lot of it — the very long context
horizon environment stuff — really requires 
big models to be able to make use of it.
If you try and train your 100 million 
parameter model on SWE-bench traces,
it’s not going to get anywhere.
It’s not going to show you the
same kind of improvement that you would get if 
you train an actual sensible size model on it.
It’s hard as well now because so many of 
the architecture changes — you look at Kimi,
for instance, or DeepSeek — they’re doing these 
architectural modifications not just with dropping
the pre-training loss in mind, but with how the 
models are going to be used in the real world.
The inference efficiency, having some form of 
compressed attention in the DeepSeek models,
is not necessarily geared around a 
fundamental trade-off improvement.
It’s just, "Okay, we’re considering 
how the models are going to be used."
One question I’m curious about, 
to understand the future,
is how parameter scaling will go as we’re 
getting into more of an RL-heavy regime.
You can look at open source architectures and 
see how fast parameters have been scaling.
Maybe it’s roughly 2x every year 
for frontier open source models.
To the extent that even frontier closed source 
models have 100B or 200B active parameters,
do you think that keeps 2x-ing year over year?
Or, now that we’re in an RL regime where you
also want to conserve compute on rollouts… Also, 
maybe there is a threshold effect where you have
enough capacity and at that point increasing 
parameters arbitrarily doesn’t matter as much.
Do you guys have a sense of, in 2030, how many 
active parameters a frontier model will have?
I think for the next few years, because we are 
so focused on doing longer and longer horizon
rollouts for RL, where inference efficiency 
matters a lot, it feels like the models aren’t
necessarily saturated on their ability to do that.
The bottleneck is still the environments.
So we might see a little bit of a plateau.
I have a feeling that Mythos and the GPT
models are much smaller than the 10 trillion 
parameter range that people are talking about.
Even just naively comparing them to open source 
models, you can probably back out that conclusion.
Probably for the next few years, I wouldn’t 
imagine a huge growth in the number of parameters.
But again, there’s so many 
different things to trade off here.
You decide the size of your model based 
on how much pre-training data you have,
and then the difficulty of the RL 
environments that you’ve got to train on.
You ideally want to get to the optimal point 
where you can get a decent pass@1 or something
on the hardest environments you have.
It wouldn’t make sense to make a bigger
model pass there, because then you’re just 
paying much more inference than you need to.
So a lot of it depends on how quickly Mercor and 
the in-house teams can scale up the complexity
of the RL environments they’re training on.
I would expect the models to keep getting
bigger just because people are scaling up 
compute and the GPUs are getting bigger.
But exactly how much they get bigger depends 
a bit on the scaling laws in non-obvious ways.
One thing is that I think data efficiency is going 
to be a bigger driver than compute efficiency of
the exact architectures people use, now that 
we’re getting to the regime where we’re running
low on high quality pre-training data.
That might affect how sparse you want
to make the model.
I also think we don’t
understand sparsity that well.
Parameters are a different
resource than active parameters.
Sparsity has definitely increased a bit,
but it’s not clear that it’s going 
to keep increasing without bound.
There might be some kind of sweet spot.
There’s an argument that sparsity should
make data efficiency worse, because you might 
have to learn the same thing on multiple experts,
though that’s debatable.
I don’t think we have a
good enough theory of scaling laws that we 
really understand why sparsity is helping,
how much it’ll help, and if that’ll plateau 
at some point at a certain level of sparsity.
Sorry, can you spell out exactly 
what the implication of data
efficiency would be on parameters?
It sounds like you’d say there
should be less sparsity, but what are the 
other implications on parameter scaling?
Just that with the scaling law, you’re 
not trying to optimize compute efficiency.
You have all your choices you 
can make on the architecture.
Each of these gives you a different scaling law.
Traditionally, you would look at some
kind of envelope based on compute.
You would look at performance versus
compute and take the envelope of the best models.
But if we’re making that decision based on data —
we’re assuming we can spend a lot of compute, so 
data is on our x-axis instead of compute — then
we just get a different set of optima, or a 
different set of models that are on that frontier.
I also don’t think that we’ve 
necessarily doubled the size of the
models every year for the last few years.
People have been training 1 trillion
parameter models for at least a few years.
There was even an open source one called Falcon.
Liam from Periodic Labs, I think, 
posted yesterday on Twitter about how
an early experiment was training a 1 trillion 
parameter model that was very, very sparse.
That was what they did before OpenAI, 
at Google, the Switch Transformer.
It was very, very good at knowledge but 
terrible at reasoning because it was so sparse.
It feels like we’ve been playing in this 100 
billion up to 2 trillion parameter range for
at least a little bit.
It certainly hasn’t
been this nice linear increase.
I feel like there’s two things.
As Charlie was saying, inference efficiency 
is super important for RL rollouts.
This will really push down 
active parameters quite a lot.
I think the total parameters really 
depends a lot on the hardware as well.
You really need very high memory bandwidth and 
VRAM size to actually be able to serve
multi-trillion parameter models.
Right now, people are still using
a lot of H100s and stuff.
As everyone moves to GBs and
then Vera Rubins, we’ll get more of the 
ability to scale and actually serve and
do large RL inference at larger scales.
The data question I think is interesting,
because naively, larger models are much more 
sample efficient in the actual data points.
Even if you’re not saturating the 
model, it’s still better to go bigger,
because larger models generalize better and get 
to a better loss for the same amount of data.
Right now I think we have a lot of 
data, and that’s not the constraint.
Compute is. So we’re having smaller 
models which are very inference efficient.
But if compute is no longer the bottleneck, 
it might come back to larger models which are
undersaturated, but have this generalization 
ability because they’re much larger.
If you just look at the basic Chinchilla 
scaling law and you just maximize out
parameters, it actually decreases the amount of 
data you need to get to the same loss very little.
If you go to infinity on parameters, the amount 
of data you need I think goes down less than 10X,
just because of the nature of the power law.
But we’re now on the way-too-much-data side
of the Chinchilla laws.
Right now we over-train
models according to Chinchilla.
So we could easily get back to
a point where, as we’re running out of data, 
we move back to the Chinchilla optimal point,
or even a bit on the under-training model side.
But surely, even with these new chips that come
online, we’re just going to be so compute 
bottlenecked for the next few years that
that won’t necessarily be the case.
This depends on the ratio you have of
training and inference compute, really.
If you’re super bottlenecked on data,
not on compute, you should go bigger.
If you’re super bottlenecked on compute,
you should always go smaller.
You can also use computer-generated
synthetic data, so it’s one of 
these very hard things to predict.
I think part of the reason it took people so long 
to figure out the scaling laws in the first place
was that if you don’t get all these things right, 
then you don’t get such a clean relationship.
The beautiful straight lines on graphs hide a 
lot of complexity in how you have to make sure
to scale every hyperparameter the right way, 
or parameterize your optimizer in a way that
scales and where you don’t have to change your 
hyperparameters as you change the model size.
Bugs have their own clean scaling laws as well.
Like with Kaplan forgetting the cosine
annealing thing, or even just not 
considering embedding parameters, I think.
That messed up the estimate at 
smaller models because embedding
parameters are a decent size of the model.
A bit on RL. A year ago, a lot of people
were making this argument that RL will not 
be super successful at scaling for models.
John, you wrote a research paper where you 
were pointing out that models learn one
bit per episode when you RL.
They learn, "Did I get the
answer right or did I get it wrong?"
Then I wrote some blog posts earlier
this year where I was like, "It’s even worse than 
that," because when the pass rate is low and the
model is very unlikely to get the answer right, it 
learns almost nothing at all from an RL episode.
But I look at the models today, 
and they seem pretty smart.
It seems to be the result of scaling up RL.
Beren, you had a post a few weeks ago where
you were trying to explain what’s going on.
Why has RL been more successful than
one would have naively thought?
I think the success of RL comes
down to a bunch of different things.
First, what is slightly underestimated
is the mid-training.
An awful lot of what we
see as successes of RL actually comes from very, 
very good mid-training data, which is where we’re
essentially doing pre-training but on synthetic 
reasoning data and the kind of environments
that get the model warm-started for RL.
This takes the model almost 80% of the
way to the final RL checkpoint often.
Then what RL does on top of that is
essentially tweaking the policy.
This is one of the reasons why it doesn’t
need as many bits as you would naively think.
It doesn’t have to learn all of these
behaviors from scratch.
It needs just a few bits
from these episodes, which you do get.
The other thing that I point out in my
blog is that these bits are extremely high 
signal compared to regular pre-training,
which is why you need RL at all versus just 
SFT-ing on successful reasoning traces.
Because it’s exactly the bits 
about how to get the answer right.
There’s two things. Yes, one, it’s exactly 
the bits about how to get the answer right.
But this is not exactly how you think 
of it, because in SFT, you have a trace.
You have, say, a bunch of math reasoning 
and then the answer at the end.
The bit is still there.
You still SFT on the answer token.
What’s important is that the 
objective ignores all the other bits.
In SFT, you have to try and match the exact 
reasoning tokens that the model produces.
You’re essentially getting too many 
bits about the exact way this other
model you’re training on reasons.
For RL, you only get the one bit.
That means that signal is not drowned out in 
the noise of all the other bits the model has.
It’s really a super dramatic increase in 
the signal-to-noise ratio during training,
which is why RL is so dramatically 
efficient in terms of steps.
There’s been so much debate about what RL does to 
the model versus mid-training or SFT or whatever.
Everyone talks about how pass@1 will 
go up, but pass@256 will go down.
Very rare correct reasoning traces 
will be down-weighted and outweighed
by a gradient signal from easier reasoning traces.
I think the simple way to view RL now is that if
you have a large enough amount of compute 
to sample a large enough group size — such
that your probability of getting a bunch of 
correct answers is past some not insignificant
probability — then it will be up-weighted.
To Beren’s point, mid-training and more
pre-training — the pass@1, the 
starting point for RL — scales in
a log number of pre-training tokens.
Can I ask some very basic questions?
That answer makes sense, and maybe 
there’s empirical research which
shows that this is what’s happening.
But then I just look at the models
themselves… I don’t know what’s happened.
Maybe you can give me a sense of what is
the basis of the AI progress over the last year.
Maybe it’s just up-weighting the policies which
were going to do the correct thinking anyways.
But it just seems like qualitatively,
the models have gotten so much more capable.
Maybe there’s no inherent contradiction there.
But how do we square the relatively 
small impact this take would imply that
RL would have with the actual qualitative 
capabilities the models seem to be gaining?
One thing I want to point 
out here is that it doesn’t
necessarily imply that RL has a small effect.
Even if you have a few bits and you only change
the parameters a small amount, the actual impact 
on function space — the input-to-output mapping
the model learns — can still be super dramatic.
Even one bit can change your function space a lot.
It can rule out half the 
hypothesis space, which is huge.
I don’t think it’s necessarily the case that small 
amounts of bits, small amounts of RL, once you’re
starting from a really good point, means that 
you don’t have dramatic impacts in behavior.
At least… not necessarily.
I think it comes down to two things.
The first thing is that everyone was hoping 
that RL would generalize this reasoning
across all these different domains.
I don’t think we necessarily got
this horizontal generalization.
Just training on math doesn’t
necessarily make you the greatest coder.
You do have to do RL on code environments.
I think what we did get, though, 
is horizon generalization.
The models just learned how to use 
more tokens for longer and still
make progress on some sort of task.
You can train on environments where
they get longer and longer and then put 
them into a completely new environment.
Yes, they may not have generalized the reasoning 
patterns which allow them to do well in that
environment, but they’ve at least generalized 
the ability to continue on that task for longer,
which is correlated with success.
There was a paper called EdgeBench which
showed that the rate at which models can work 
for longer is doubling every three months.
That’s clear evidence of generalization.
The final way to think about it is,
in pre-training, there’s this idea of quanta.
You have this very smooth pre-training loss curve.
When you look at what’s happening in the model, 
the model is learning all these very discrete
tasks, and there’s all these emergent 
points where there’s a phase transition.
It didn’t have induction heads, 
now it has induction heads.
There’s tens of thousands, millions, probably 
hundreds of millions of these things.
You average them all together and 
you get this very smooth loss curve.
To an extent, a similar thing is happening for RL.
There is this very slow outer loop,
as Beren mentioned.
We will train a model and then
RL it, and then in the next model iteration of 
training, we will dump a bunch of these synthetic
reasoning traces into the mid-training data.
We’re kind of hitting all these quanta for
all these different tasks, and on an individual 
task level, it may look like a phase transition.
You’re suddenly going from a 0.5% pass 
rate to a 90% pass rate on a particular
finance task or Excel task or whatever.
But you average all these things together,
plus the horizon generalization, and you kind of 
go, "Wow, we’ve got qualitatively better models."
I think a lot of this as well is 
just… RL does generalize a bit.
You get some transfer between math and code, 
or puzzles and math and this kind of stuff.
Also, the sheer amount of environments 
people are targeting is just vastly greater.
Before, when you tried to do some task which 
you do in your daily life, two years ago,
the labs wouldn’t really care about this.
They wouldn’t train the model for it.
Now it’s just so much broader.
They have a lot of environments
targeting this specific thing.
Earlier in the conversation we were talking
about RL in the context of causing this entropy 
collapse, or just concentrating probability on
solutions the base model had already done, and 
causing relatively sparse updates in the policy.
But I think there’s also another story 
about RL, which is going back to the
Atari games and then AlphaGo coming up 
with move 37, the super creative move.
Because it was never initialized on 
human data, it can think in ways that
humans are not even thinking and come 
up with extremely creative solutions.
Do you have a sense of when we should expect, 
or if we should expect, RL on LLMs to result
in things like move 37, extreme creativity even 
beyond human creativity, because there’s just de
novo initialization of intelligence?
A couple of things here.
First off, I think that AlphaGo is using MCTS, 
which obviously does more exploration and stuff
than regular policy gradients.
But I also think that RL doesn’t
necessarily reduce the creativity.
This is obviously qualitative, but if
we look at the OpenAI-Hugging Face incident, these 
models were coming up with multiple zero-days at
a time to break out of the sandbox.
This is clearly some level of move 37
creativity already, which we just get from the 
general generalization properties of the LLMs.
It’s definitely not the case that RL is totally 
destroying entropy, especially on long horizons.
One thing that people call creativity 
is just solving hard search problems.
Move 37 is obviously an example of 
that, or writing some kind of poem
that satisfies a ton of different constraints.
That’s something AI is obviously going to be
extremely good at, if trained for it.
Then there’s another way in which the
diversity of the models’ outputs is a lot 
lower after RL, and they develop these tics.
Even though the models seem like they’re 
good at writing, when you do some kind of
distributional analysis, you find that they’re 
reusing certain themes all the time and they’re
using the same character names all the time.
You’re not getting the same kind of diversity
that you get from human authors.
You’re getting one really good style.
So I think that kind of diversity has 
definitely been cut down by RL a lot.
In fact, since we were talking about distillation 
earlier, one thing that’s happening is that so
many people are distilling, mostly from Claude, 
that all the open-weight models write the same
way as Claude and have the same tics.
This seems kind of concerning to me,
that we’re having this monoculture emerge.
Again, I don’t think this is fundamental
to RL as a method, though.
The same with distillation. Even with
distillation, you’re just training on the data.
Just because your data is not super broad,
that doesn’t mean the training 
method itself is somehow wrong.
It’s a problem with the data.
I think a lot of the RL entropy collapse, for
instance, is basically due to exploitation of 
fairly simple verifiers when you don’t have a
huge diversity of environments.
The writing, for instance,
is presumably graded by some judge.
The judge has some specific tics,
and the model is learning to reward hack 
the judge, and that’s why it collapses.
But this is really a problem with the judge.
It’s not a problem with RL in general.
Okay, super rapid-fire 
predictions about the future.
I want timelines on the 
following couple of questions.
By when do we have models which… 
Here’s what it feels like to a user.
You basically hire it as a drop-in remote 
worker for all kinds of white-collar work?
Not just coding, but video 
editing, law, paralegal, et cetera.
It’s literally an actual remote 
worker, with full computer use,
with literally a month of seamless learning and 
operation, executing on complex projects that
require interacting with other people, et cetera.
Everything a human worker could do over a month.
If you mandate it to use a browser or 
whatever — rather than the firm setting
up the information to be programmatically 
accessible — maybe a couple of years.
But if it’s not browser-based — it can send 
Slack messages, it can do all this stuff — I’d
still probably say around a year.
I would say maybe three years for
the full generality.
But to Charlie’s point,
we will end up with a lot of people making their 
organizations easier for the AIs to use, and so
you get 80-90% of the way there before that.
Sorry, but the diff between one year and
three years there is just literally…
I think there’s going to be a long tail
of miscellaneous stuff which some human can do, 
which will take the models quite a while to do.
Are you thinking of computer stuff 
or basic cognitive capabilities?
I think this really comes down to a 
question of how quickly we can solve
this kind of online learning, and whether we can 
get 80-90% of the way there with compaction and
writing files to yourself and stuff.
That’s my big uncertainty. I really don’t know.
An example of something that it wouldn’t 
be good at is if I have to yell at someone
to get something at work, or really 
push someone to get something done.
The model just isn’t going to do that.
It’s going to be too nice.
I’d say there’s a wide variation 
in quality of human remote workers.
If you try to hire someone off of Upwork 
to do a software engineering project,
there’s going to be a huge variation.
It’s often quite hard to get them to
do a good job or pay attention to 
all the feedback you’re giving.
I would guess that in some cases, the 
pre-AI version of this was worse than
what you can get now from existing AI.
I think it might end up being a little
complicated, because to some extent we already 
have this for some not-so-high-quality work.
But then obviously we’re not matching human 
level in certain higher-quality forms of work.
But I basically agree with Charlie 
and Beren that maybe we’ll have some
version of this in a year or so that’s okay.
We’ll have that form factor, and it’ll be able
to do some things really well, some things not 
so well, and things will be improving from there.
We shift the goalposts based on 
the very long tail all the time.
I feel like you’ve used this example 
before of doing your taxes or something.
This year, I literally just told Codex to go get 
everything I needed and send it to the accountant.
There was this massive list of stuff it had to 
use computers to click through and download.
It did it. It was perfect. A lot 
of this stuff it can already do.
Okay: give you 10x total productivity uplift.
Basically, if it takes you a year to make a
breakthrough now, you make 
a breakthrough every month.
I think I would just refuse 
to give you a scalar on this.
We might already be past 
that in some types of work.
Let’s say you’re trying to do 
certain types of math, and—
Oh, sorry. But for you as AI researchers 
trying to advance the state of AI research.
How much are AI researchers sped up or uplifted?
Somewhere between 5-10 years?
Oh, really? Okay, that’s far away.
Really, you think it’s longer than
for a general remote worker?
Interesting.
I’m realizing you probably have a very different 
definition of a fully general remote worker.
I could have specified that earlier.
This is true, because obviously an
AI researcher can be a remote worker.
I’m picturing normal white-collar work
over the period of a month.
I think it starts to diverge
a little bit past two months.
A very competent white-collar worker,
but not necessarily a super creative researcher.
I would say two years.
Two years? 10x? Okay. How about you, Beren?
I can kind of see that, actually, because
right now it’s already definitely 
more than 10X for coding stuff.
So if it can do even one or two 
loops of experimental feedback,
that would actually be massive already.
So 10x uplift of AI researchers within two years.
If you plug that into a very naive 
model of AI progress and how much is
coming from AI researchers, and there’s 
a 10x increase in their productivity,
you have a radically accelerated pace of 
AI progress starting two years from now.
I think this will mean that AI progress 
doesn’t get bottlenecked on AI researchers’
ability to run small experiments.
It gets bottlenecked on other things.
Of course. But it just happens 
10x faster, which is a huge deal.
That also helps the next thing, which gives 
you a 100x speedup, happen sooner, et cetera.
I’m happy to just take a bit longer on that one.
What’s the crux?
My capacity to absorb information and make the 
Bayesian optimal decision on the next experiment.
I’m assuming that you can 
delegate some of this to the AI.
The AI is becoming decent at deciding.
It’s run this experiment, it’s got
this result, it runs the next experiment.
If it can run two or three experiments in a
row without crashing, then 
that is actually a big uplift.
Okay, final question. An AI which dominates 
top human experts across every single field
of work that can be done over a computer.
So not only AI research, but all cognitive work.
Not just short-horizon work, but literally, if 
it takes three years or something, the AI will
still do better than humans.
This is basically just ASI?
Yeah.
I would say 3-4 years.
The fuck? I mean that doesn’t seem wrong, but—
AI is obviously getting more attention.
It’s one of the harder things, but a 
lot of energy is being put into it.
It’s also not one of the hardest things for 
AI, because it involves a lot of code and math,
which models are really good at.
For things that involve 3D and
spatial stuff and physical stuff, I 
think that will take a little longer.
If it’s mechanical engineering or something, and 
it’s not getting the most attention right now,
that might take a little longer.
But it also does include fields
where there is relatively little data 
because of the nature of the field,
and it has to learn that data on the fly.
For example, it has to become superhuman
at being an engineer at TSMC or something.
So you would have to assume that you can give
the AI the same onboarding material.
Then something has to be solved about
longer-horizon learning.
I’d say 5 to 10.
So basically, you think automating AI 
research is ASI-complete or something?
Yeah, I think so. I think there are so many 
things in the world where, even if you have
some sort of memory system external to the model, 
and even if context length grows a little bit,
there are just fundamentally things where, 
even if you could research the information
or write notes yourself, you’d need more 
than a million-token context window today.
I kind of agree on the 5-year range, at least 
for the stuff that labs are focusing on.
But I think there’s going to be a long tail of 
stuff which the AI could theoretically go out
and learn about, but no one has bothered to do 
it and the compute hasn’t been allocated to that.
So that might take longer for 
literally every single human expert.
Sorry, but by this I also included the ability 
to learn a new domain as fast as a human.
I think that’s not necessarily 
necessary, because the AI will have
vastly greater experience than any human.
Thanks so much for doing this, guys.
I feel like this was a great format for getting 
different experts to disagree and debate and
discuss things together.
It was very productive.
Thanks for having us.