Today I'm chatting with Ryan Greenblatt, who is 
the chief scientist at Redwood Research, where he
focuses on technical AI safety and security work.
I want to talk to you about
recursive self-improvement.
This is the idea that once we build
human-level intelligences, they quickly slingshot 
towards tens of billions of superintelligences,
which are each individually more competent 
than the top human experts across every field.
Whether or not this turns out to 
be the case is probably the most
important question in the world right now.
And historically, I've been quite skeptical
that this kind of thing happens, but, you 
seem to think that it might be plausible,
and so I wanted to hear the case for it.
Let's talk about this. First, I think
it's worth noting that AI R&D is a type of 
task at which the AIs are especially good,
because the companies are trying really 
hard to make their AIs good at AI R&D.
It's also the kind of domain that has a lot 
of nice properties from the perspective of
how AI development works right now.
It's pretty verifiable. You can do
a bunch of stuff iteratively, and 
it'll hill climb on various metrics.
I think once you have AIs which are roughly 
matching the top human experts in AI R&D,
that could kick off a feedback loop 
where the AIs are doing AI research.
That produces smarter AIs. That feeds 
back in. That feedback loop could be
strong enough that you end up with a lot 
of progress in a short period of time.
Maybe my median expectation is something like four 
or five years of AI progress in a single year.
This requires really overcoming a huge amount 
of diminishing returns in research and basically
doing the equivalent of the progress we would have 
gotten after a really large compute scale-out.
So this is a pretty impressive, big thing.
It's worth keeping in mind that five years
of AI progress, four years of AI 
progress, even three years of AI
progress, is really a lot of fucking AI progress.
A little over three years ago, GPT-4 had come out.
Right now, of course, we have Mythos 5 or 
whatever, and maybe a somewhat better model
that Anthropic has internally.
That is just a huge amount of
progress in a bit over three years.
If we're talking about five years,
then maybe we're talking more about a 
jump from GPT-3 to Mythos 5 or whatever.
I think this argument has three different parts.
Now I want to evaluate each one of them.
First is the argument that 
AI R&D is very verifiable.
Second is the argument that if you 
automate AI R&D, you could get four
or five years of progress in a single year.
Third is the argument that what comes out
the other end of four or five years 
of AI progress at the current pace,
starting at the point whenever AI R&D is 
automated, is an AI where you can drop it
on the job at basically anything you can imagine.
You can drop it in Texas politics in the 1940s,
and it outmaneuvers Lyndon Johnson.
You can drop it in TSMC, and it learns how
to do better process engineering at TSMC.
It's certainly a better video editor…
My video editors are very excellent, but it 
is just, in general, better than humans at
any given job that it finds itself trying to do.
So I want to evaluate all of these sub-arguments
that lead to basically getting ASI pretty 
soon after this benchmark, which you're
expecting by 2030 or something, right?
I would say that I expect full automation
of AI R&D perhaps somewhere around 2031, 2030.
Getting to the "beats all humans on the job"
milestone, maybe my median 
expectation is around 2033.
But if I see AIs fully automating AI R&D, I 
think I'm expecting that probably within a year.
The way the forecasting works out, the difference 
between medians is bigger than the median
difference between milestones.
Anyway, whatever.
By the way, there's this meme on the internet.
Every time I'm trying to ask about people's
timelines, when I'm asking Dario or 
somebody, I'm always like, "Okay,
how long before you automate my video editors?"
There's this meme of my video editor editing
the podcast every time I listen to this.
But the reason I do it is because I think
it's easy to get lost in abstractions when 
you talk about jobs you don't understand well,
and to very concretely understand what it takes to 
automate a job that I actually understand why it's
difficult for LLMs to currently take control over.
I do think that the milestone for automating your
video editor is earlier than the milestone 
of being able to automate all human jobs,
including Texas politics, spinning up on the job.
I do think that the video editor automation
occurs maybe more around full automation of AI 
R&D, but it's very sensitive to how much people
are really focusing on understanding video.
Okay, so let's start with the claim
that AI R&D is very verifiable.
There's a few different parts of this.
One of them is that we can train on 
a bunch of environments which are
basically directly training the model to do 
some AI R&D task or some very close-by task.
For example, we can have some environment where 
the model is training some AI on just eight
H100s or some small amount of compute, and that 
model could be the equivalent of GPT-2 medium or
whatever, and then similar to NanoGPT medium runs 
— and in RL, it's tweaking and iterating on that.
We could do that for a bunch of different tasks.
We could have it train image classification
models, video generation models, image generation 
models, all kinds of different ML training tasks.
We could RL it on the task of training 
increasingly good models, and also
doing things like, "Oh, here's a particular 
direction you could pursue for an algorithm.
Can you go and implement that?"
Basically, there's this whole class
of containerizable, verifiable, small-scale AI 
R&D tasks that we can aggressively RL the AIs on.
Already companies are presumably doing some 
RL on these sorts of tasks, and you could
just keep scaling that up, keep making more 
of these small-scale AI R&D tasks, and then
the AIs could keep getting better at this.
Implicitly, I'm claiming this will transfer
to extremely load-bearing aspects of AI R&D.
But maybe let's stop there for a second and then
get to that part.
So let's talk through
what this concretely looks like.
You can imagine that we have GPT-7.5.
We say, "GPT-7.5, we want to make you so 
good at AI R&D that you help us train GPT-9."
So now we want to train GPT-7.5, and we come 
up with a bunch of different environments.
As you mentioned, there's already this repo 
that is the descendant of Andrej Karpathy's
nanoGPT speedrun, where you just try to change 
everything about the model, from the optimizer
to the hyperparameters to the architecture, to get 
it to a fixed training loss as fast as possible.
You could have other kinds of environments 
where you could say, "Hey, GPT-7.5, I want you
to train a really good video game-playing model.
I want you to train a model that actually improves
as it plays the same video game again and again.
So you learn how to maybe help the
model get better at online learning.
We don't care how you figure this out.
Maybe it's some kind of crazy 
neuralese or a vector memory.
Or maybe it's just better long-context stuff.
We don't care. Figure out how to do online
learning research."
Obviously, GPT-7.5
will already be a smart model, and in the same 
way the models currently are getting smarter,
it'll be better and better at coding.
You can imagine 100 other environments
like this which are incentivizing the ability to 
do AI R&D, like containerized versions of getting
GPT-7.5 to develop GPT-2-sized models, et cetera.
Then you basically put GPT-7.5 through a bunch of
this kind of training, you build GPT-8.
GPT-8 is now an amazing ML researcher.
It has so much intuition from 
doing all this kind of training.
Honestly, a huge intuition pump for me is seeing 
the progress that AI has made in mathematics.
If it's a very verifiable domain, AIs can 
get… I don't really know the object-level
details of mathematics research, 
but I'm just like, "No, it works."
It can just come in like a flood if you can 
totally put it into a verification loop,
and it can actually make new breakthroughs.
I am curious if ML research has a quality of
mathematical research where it seems 
like there was a big overhang from
connecting different disciplines together.
No one person would have known enough about
algebraic geometry and… What was the right word?
Oh man, I really don’t know
about the math breakthroughs.
No one person would’ve known enough
about topology and algebraic whatever in order 
to make some counterexample to a big conjecture.
My view is that ML is a less deep domain 
than math, and so there's less of a thing
where there are individual experts with really 
deep expertise in some area that they combine,
but there's definitely going to be some of that.
But then I also think that ML has some attributes
that make it even more favorable to AI 
training than mathematics in some ways.
In particular, you can get a better 
sense of whether you're succeeding,
and you can see intermediate progress.
In math, it's often the case that there's no easy
way to see whether or not you're close to success.
Whereas if your goal is, for example,
to get to some training loss 2x faster, you 
can kind of see when you're halfway there.
It tends to be the case that ML innovations 
are very additive, or maybe multiplicative
depending on how you think about it, where 
basically you can keep stacking innovations.
Usually the innovations just add together and 
don't interfere with each other, though obviously
it's going to depend on the details.
So I think that in a lot of ways,
AI R&D will have properties quite similar to 
math, where you can train on chunks of AI R&D
that are pretty similar in structure to the 
problem you actually cared about, in a very
verifiable way, and then that will transfer.
There's an open question of exactly how
well it will transfer, but I think that the 
transfer currently for math looks pretty good.
My expectation is that the transfer for AI 
R&D will look pretty good, but not amazing.
So one concern I have is that I think 
even in mathematics, as far as I'm aware,
we have not seen very impressive new theory.
We've seen a lot of impressive, verifiable,
specific results — for example, find a 
counterexample to this conjecture — but
we have not seen "come up with the idea 
of topology" kinds of levels of things,
or "come up with things like group theory".
It seems like ML research has elements of
both of these things.
But the less verifiable
thing of coming up with new ways of thinking 
about the problem would be harder to induce.
Take, for example, the idea of scaling laws.
Obviously, there is some end verification
loop such that you can train GPT-4 better if 
you have the idea of scaling laws from 2020.
But there is a longer and potentially more 
compute-laden road to inducing AIs to be like,
"Okay, I got to think carefully about how 
I should be scaling my parameters and data.
What are different kinds of investigations 
I could run to understand this?
Maybe I can come up with a visualization 
and an isoFLOP analysis or something."
But that does seem like a longer 
verification loop than just, "Hey,
let's get nanoGPT loss to go down."
Let's talk about this. First of all,
I think in the context of math, the 
thing I would say is that the AIs can
do the equivalent of 'baby's first new 
theory,' where, for example, they can
just prove interesting conjectures via making 
connections and producing new understanding.
It’s like, "Oh, there's this construction 
the AI found which is pretty interesting",
or it found this way of thinking about 
the problem that's a bit different.
We do see that. It's just that the examples 
we see are not as impressive as founding
the field of group theory.
Founding the field of group
theory is probably among the best, biggest 
mathematical accomplishments of all time,
and the AIs just aren't that good at math yet.
From my perspective, there's a continuum between
that and the things we're seeing now, 
that the AIs are continuing to march up.
Second, I think ML is a very 
shallow domain relative to math.
In math, there was much more of a thing 
where you find some true deep abstraction,
and if you really understand that thing, which 
is hard to understand, then you get somewhere.
Whereas I feel like the things that are the 
equivalent of that in ML are really dumb bullshit.
Like with scaling laws, come on guys, we 
can explain scaling laws really quickly.
I think the deepest and most important concepts 
in math, for example, don't have the property that
you can really understand the underlying thing 
and why it matters in a very short period of time.
But I feel like one effect will 
be that we will have gotten rid
of all the low-hanging fruits by 2030.
I feel like scaling laws will have been,
in math history, like Descartes finding the 
Cartesian grid and doing very basic mathematics.
Eventually, if we want to keep making 
progress in the 2030s, it's going to be
like doing whatever bullshit is happening 
at the frontiers of mathematics right now.
That could be right. My sense is that 
some domains are structurally different
in terms of how they operate and how 
much they depend on deep abstractions.
Physics and math are much more on the 
side of being very far on the deep,
hard-to-come-up-with-ideas side, 
whereas I think ML and most other
domains are much more amenable to hill climbing.
That's my sense of how this will go in the future.
Even in the regime where your AIs are having 
to plow — it's 2030, a bunch of low-hanging
fruit in research has already happened, and 
they need to make further progress — I still
suspect that a bunch of the work will live more 
on the side of building increasingly complicated
infrastructure and having really good intuition 
about what the experiments roughly look like.
So I'm probably less sympathetic to the idea that 
the thing the AIs will lack is some deep insight.
I’m more sympathetic to the idea that they 
really need a bunch of taste about in-the-weeds
experiments that they currently don't have.
They need a bunch of intuition for what sorts of
training approaches would work and what 
wouldn't, in ways that current researchers have.
Even in cases where there has been some 
breakthrough in AI, oftentimes in retrospect
it looks like a big bottleneck to making that 
breakthrough happen was getting all of the micro
details and mungy intuition right.
An example of this is training AIs
to be good at reasoning and chain of 
thought, doing RL on chain of thought.
It looks like you probably could have done RL 
and chain of thought on GPT-3 and gotten kind
of interesting results on math if you had 
really scaled it up and done a good job.
But at the time, there was low-hanging fruit.
Also, doing a good job with that training
is kind of in the weeds on all the 
technical implementation and scaling
it up and getting the hyperparameters right.
So maybe you can demonstrate everything on
Qwen 1B or whatever and get some sense 
that this whole thing is going to work.
But people didn't demonstrate it as early as 
they could have because of all of these other
mungy details and intuition about exactly how 
to tune the parameters and how to set things up.
This is my remaining skepticism, 
honestly, about this story.
I'm not sure I understand why, if research 
breakthroughs are so amenable to intelligence,
AI progress has not been historically 
faster than it could have been.
As you were saying, by the time RLVR actually 
worked — even though you could have done it
with less compute — we had to wait for oceans of 
compute, gigawatts of compute, to be available
before people were doing this training, on the 
trajectory of compute continuing to increase so we
make more breakthroughs.
I don't know. I feel like there
were a lot of AI researchers in the year 
2022 who were trying to crack reasoning.
Was it just that they were 
bottlenecked by the ability to
write infrastructure code, or what was happening?
It's a complicated mix. I think they would have
gone faster if they could, as soon as they 
thought of an experiment, run that experiment
without bugs, without bugs being very important.
And then another part of it is that being able
to run a lot of experiments at high compute 
lets you paper over ways in which the way you
implemented it isn't quite right or you 
didn't have the right hyperparameters.
So compute is just really helpful for doing AI 
research, and you can cover over a lot of things.
But that doesn't mean that massive 
increases in labor wouldn't also be helpful,
especially if that labor comes with among the 
best intuitions that people have in the field.
I just think that's really helpful.
Another part of my perspective here,
which is maybe a bit different from where you're 
coming from, is that I'm expecting somewhat more
transfer than you seem to be imagining.
I'm imagining these AIs are actually
pretty good scientists in general and are 
pretty reasonable at all of that stuff.
When you interact with them, 
it's not like they have some
really hyper-specialized savant-type vibe.
They're actually pretty good at all of the
stuff in R&D, and then maybe 
extremely good at some subdomains.
So they're incredibly superhuman at 
writing kernels, incredibly superhuman
at everything with very short feedback loops, 
and then pretty good at all the other stuff,
totally able to match other people.
I think we are seeing this now.
When I look at AIs right now, it's already 
the case that they can pretty competently
match humans who are mediocre at 
ML research at doing ML research.
It's just that being mediocre at 
ML research is not that helpful.
The thing you actually want are 
people who are good at ML research.
My sense is the AIs are just 
improving at all of these things.
Their taste is improving, their intuition is 
improving, and it's already the case that their
taste and intuition is not complete garbage.
I want to very concretely understand what it
would look like for five years of 
AI progress to happen in one year.
Suppose we were back when GPT-3 was developed.
The idea is that, with the level of compute they
had back in 2022, if we had automated 
AI R&D back then, you could at the
end of that year have Mythos.
That would be the idea, yes.
Mythos took way more compute than they 
had back then, but even with the level
of compute they had back then, not only 
do all the breakthroughs happen, but they
also train Mythos with that level of compute.
What would be required is obviously discovering
all the algorithmic progress since then.
It’s discovering even more, actually,
because you've got to make up for the fact 
that Mythos uses… What was GPT-3 trained on?
Like 1e23? We can look it up.
But is it plausibly four
orders of magnitude more compute?
I think it's somewhat less than that.
Let's look this up quickly.
GPT-3 training compute is about 3e23.
My sense is that Mythos is probably 
a little over three OOMs higher.
So the question is: can you overcome this 
1000x compute gap while also being the model?
Here's a concrete claim that 
maybe we should talk about.
Right now, would we be able to 
train a model with GPT-3-level
compute that matches… What exactly do I think?
GPT-3 was released in 2020, so it was trained
about six and a half, seven years ago.
It's worth noting that GPT-3 is maybe
a little too far in the past, but 
let's go with this for a second.
If we were to train a model with GPT-3-level 
compute today, how good would that model be?
My understanding, based on how algorithmic 
progress works, is that we'd be able to train
a model that's as good as the best model 
we had perhaps around three years ago.
So I think that right now we'd be able to train a 
version of GPT-3 that's probably somewhat better
than GPT-4, a moderate amount better than GPT-4.
I think that's about right.
That roughly lines up with how 
algorithmic progress has worked.
Basically, the story would end up being that to 
get five years of AI progress, you're probably
going to need around, I would say, maybe eight 
years of algorithmic progress, very roughly,
which is a lot of algorithmic progress.
But it just turns out that most of the AI
progress, from my perspective, has come 
from some mix of algorithms and data,
and you can just keep making huge improvements on 
these things and training AIs with less compute.
I'm glad you brought that up, because what has 
happened since GPT-3, or even 3.5, till now?
Why is Mythos so good?
Obviously, we've scaled the compute.
We have better algorithms. But a huge thing 
that's happened is that we have built a
deca-billion-dollar data industry which 
has systematically collected and codified
expert human judgment across all kinds of 
different disciplines — codified in the form
of RL environments, codified in the form of 
SFT traces — that these experts built to help
the model better understand how you do coding, 
how you build complex infrastructure projects,
how you do law, how you do whatever.
How are the AIs able to replicate the
effect that expert human judgment currently 
seems to be playing in AI progress?
My sense is that scaling up the amount of effort 
spent on getting expert human data has not been
hugely important for AI R&D in general.
In particular, over the last few years,
we've been scaling up compute, scaling up 
people working at AI companies, and scaling
up the amount of effort spent on data labeling.
My sense is that if you removed the last two
doublings or whatever of data 
generation from expert humans,
that would not make a huge difference.
A lot of what's been going on is people
have been developing better ways to leverage 
humans and AIs to construct RL environments
and going somewhere from that.
But how do you explain why
the AIs have gotten so good at coding?
I feel like a big part of that is data and RL
environments, which are codifying human experts.
But the question is what is the limiting
factor on creating RL environments?
My sense is that the reason why RL
environments today are much better than they were 
in 2024 is not so much because we have hired way
more human experts to make RL environments.
It is instead much more because we better
know what RL environments we even want 
to make and how we should structure them.
Also, we're using huge amounts of 
AI labor to build RL environments.
I think those effects are much more important 
than the effect of human labor building the
RL environments.
I'm not saying
that the human labor doesn't matter.
I'm just saying there are other big
drivers that are important here.
I could try to argue for this.
One thing is just that the amount of 
environments people want is a very large amount.
I think the AIs are actually pretty good 
at the task of making RL environments given
some sense of what the thing should be.
There's preexisting data you could use.
A lot of these things have 
good verification loops.
Just look at, for example, what was reported 
in Business Insider yesterday, that Google
is paying close to $2 billion for Mechanize.
We can just look at market rates for what people
think really good human expert data is worth.
The frontier labs seem to think it's worth a lot.
They're willing to pay for it.
What fraction of frontier lab spending
do you think is on data rather than compute?
What do you think is the compute/data spend split?
I think it's overwhelmingly compute, 
but I also think it's because compute
is easier to scale up than data.
But that's really relevant to what's
driving progress, right?
My sense is that the
split is something like 20 to 1 or 10 to 1.
I don't know exactly. It depends on the company.
But this is similar to how oil is 1.5% of GDP.
That doesn't mean that if you cut oil out,
GDP could continue to run.
Sure, but it contradicts your argument, right?
The economy would come to a halt 
immediately if oil went away.
Sure, but you were just arguing 
that because of the high market cap,
we can learn that this is the key driver, 
and I'm saying that's not clearly true.
That argument just makes it look like compute is 
a much more important driver, or hiring employees
is a much more important driver.
So maybe let's be more concrete.
Here's what I think. My claim is that if 
you went back to 2022 and you had GPT-3.5,
and you were trying to make it better at coding
without human experts, I think it would 
have just been very, very difficult.
Let me give you an example of what I imagine would 
be the difficulty of going from GPT-8 to ASI.
One of the things you'd want ASI to be good 
at is: I'm going to take over a company
and make it much more profitable and do all 
kinds of crazy shit to make it work better.
I'm going to take over a 
fab and produce more chips.
I'm going to go into Congress and try to 
convince them to pass some bill, et cetera.
This is what I imagine five more 
years of AI progress at this
pace would enable an AI to be able to do.
This is the thing I'm really worried about:
ASI that can understand how to do crazy shit in 
the world, that can do what Kissinger can do,
can do what Steve Jobs can do, et 
cetera, and also his engineers and so on.
I'm not sure how you get that without the 
relevant world data, which is the equivalent
of Mythos being really good at coding while 
not having the coding environments that have
improved it relative to GPT-3.
Here are a few points.
First, I bet if you look at randomly 
sampled training environments for Mythos,
they're actually very different from what it 
looks like to actually use the model in practice.
My sense is that the RL distribution has 
really large deviations from the real-world
data distribution, and it's significantly 
smoothed over by a mix of transfer and having
a small amount of data focused on the real world.
My sense is that this will be a similar mechanism
as how it works for the crazy, wildly, quite 
superhuman AI you get as a result of five years
of AI progress on top of fully automated AI R&D.
So let's go through this a little bit.
In particular, I think that you could train an 
AI to be really, really good at learning on the
fly and doing something analogous to in-context 
learning, but potentially using somewhat different
mechanisms, in a wide variety of RL environments.
You build all these different RL environments
where the AI has to adapt on the fly, learn on the 
fly, figure out what it should do, understand its
situation better, and learn really quickly from 
feedback in order to succeed at its objective.
And it has things like limited 
resources, and if it messes up,
it can end up in a much worse position.
If you train on a huge number of these
environments, you will learn general 
skills of picking up context on the fly,
and we're already seeing this.
It's already the case that AIs are
now much better at understanding roughly what's 
going on and picking up context from a limited
amount of information they're given access to.
Then those AIs could be put on the job at TSMC.
Even though TSMC is not literally in their 
data distribution, their data distribution
is really wide, and the AIs are extremely 
good on their data distribution, such that it
transfers to picking up being good at being an 
engineer at TSMC and learning that on the fly.
The way the AI gets good at being a 
TSMC engineer isn't that it has a ton
of cached knowledge on being a good TSMC engineer.
It's that it does the equivalent of some scaled-up
version of in-context learning there.
That'd be the most prosaic story.
Obviously, there's a bunch of 
different ways this could go.
I think this maybe comes down to a difference 
of intuition about how far you can get.
When I think about really smart people I 
know, they're just not that effective in
domains they don't understand that well.
But how long have they had to learn?
I agree that if they had experience, 
they would be much better.
But that's maybe what I'm arguing 
for, that experience with data.
For example, if I just get a really smart 
Ivy League college grad, and I'm like, "Okay,
you're now in charge of negotiating the Iran 
deal," I think they just wouldn't know what to do.
I think if you instead got someone who is 
really good at quickly picking up a bunch
of different domains and you gave them some 
time to train and talk to people and shore
up their expertise and do some practice, 
they would actually do a pretty good job.
I think most domains are fundamentally 
pretty shallow, where a very smart
generalist who's good at a limited subset 
of core skills can get going pretty quickly.
That's not true for literally every domain.
My sense is that the AIs will develop
increasingly good mechanisms for quickly acquiring 
understanding and expertise in a given domain.
Consider, for example, how fast 
AIs can understand a new code base.
AIs can understand a new code base much faster 
than humans can, but to a degree that's shallower
than humans could currently understand.
But it's getting better over time.
Let me spell that argument out a bit more.
Let's say you take Fable 5 or Mythos 5 or
whatever, and you wanted to make some kind of 
complicated change to a really massive code base.
The model will get some understanding of 
the code base very fast, in the course of
maybe significantly less than an hour, 
potentially much less than an hour.
Then its understanding of the code base 
will plateau a little bit, where it won't
get as deep of an understanding as a human 
would have gotten over a much longer period.
So it's like an AI in an hour can match a human 
with a few weeks maybe, depending on the details
of exactly how complicated the code base is.
But it won't match a human who's been working
on that code base for two years or whatever.
But over time, the amount of understanding
AIs can match has gone up.
If we look at 3.7 Sonnet or 3.5 Sonnet,
maybe it could only match the equivalent of 
understanding a code base for a day or something.
But now AIs are much better at 
building context about a task.
So you can be like, "Mythos, I want you 
to really understand this code base,
and then implement this feature."
It will spawn a bajillion sub-agents.
Those sub-agents will pore over a bunch of things.
It will deliver a bunch of context back.
It will then investigate a few things.
It's not amazing at doing this, but it can
happen really fast, and it can work pretty well.
And it's not very hard for me to imagine
how you could train AIs to be 
increasingly good at this task.
The task of implementing some very complicated 
feature in some reasonable way in a very big
code base is extremely verifiable, and 
that can be a thing the AIs improve on.
Similarly, there's a broader skill of 
quickly understanding context and being
able to have a bunch of different AIs learn 
in parallel and then merging that together.
I think there seems to be a crux here, which I 
think is just an empirical question we'll see.
How good is the transfer between 
getting really, really good at
understanding the situation, getting up to speed, 
making progress over long periods in verifiable
domains — which the AIs are obviously getting way, 
way better at really fast — to, "Okay, go talk to
the president and convince him to do X thing."
Or, "You're now in charge of Google.
You must make Google a much more 
profitable company this quarter."
Let me try to spell out a few more 
arguments that are maybe relevant.
One thing is, when looking at how 
the AIs have improved at essay
writing… Let's talk about that a little bit.
You can get some data even on these domains.
AIs will be able to get some data even on these 
domains when on a very fast progress trajectory.
Maybe it's hard to build a verifiable 
environment for "was your essay really
good according to humans?"
But you can do a bit of that.
You can do some training.
You can do some online training.
The AIs will be able to do some online 
training based on real-world stuff.
They'll be able to have evals.
They'll be able to sample that.
You can scale up the cadence at which you do this.
The second thing is that in practice,
when I just look at the transfer, it seems okay.
I think the AIs have in fact improved a bunch at
non-verifiable domains, and it's hard to point 
to domains that are really hard to verify on
which the amount of improvement between GPT-4 
and Mythos hasn't been pretty high in practice.
Now, that doesn't mean that Mythos is 
better than the best humans or something.
It can still be significantly worse 
than typical human professionals at
some aspect of their job while still being way 
better than GPT-4, which was not even close.
So we're talking about how much progress has come 
from data versus compute over the last few years.
That reminds me, I'm actually running 
an experiment with this with Jerry Han,
who's still a college student.
What we're basically doing to evaluate
how much progress is coming from data versus 
algorithms is training the best algorithmic
recipe from 2019 till now with the best data 
from the 2026 data file, and then also training
the different data files going back from 2019 to 
2026 with the current best algorithmic recipe.
I think that will be interesting.
I'm curious if you want to pre-register
what amount of compute multipliers 
are coming from one versus the other.
We need to be pretty careful with what 
we mean when we say the word data.
I was trying to be pretty careful to distinguish 
between scaling up spending on getting human
experts to label data, or scaling up 
the amount of human expert-labeled data.
The reason why we have a better pre-training data 
set now versus in 2019 is not because people are
spending way more money getting human experts to 
type up data that the AIs are then trained on.
Partially.
I think it's not much of it.
I think it's very little of the 
pre-training data improvements.
I do mean pre-training. We should maybe talk 
separately about mid-training and post-training.
But I think the vast majority of pre-training 
data improvements are from science on better
understanding what data sets are good and 
schleppy labor on figuring out how to filter down.
So my view is that improvements of the 
form of, like, OpenWebText to FineWeb,
that improvement is better described 
as an algorithmic improvement of the
sort that you can study with some GPUs, and 
you don't need human expert data to do that.
Now, there's a different effect which we 
could talk about, which is that maybe the
internet in 2026 is more of a fertile ground 
for training data than the internet in 2018.
There's also been an effect where there are 
just more humans posting on the internet,
so there's more data to harvest.
My sense is that that effect is
going to be quite a bit smaller than the effect 
of humans knowing better how to curate the data,
having better scrapes, knowing how to process 
those scrapes better — this sort of thing.
This is more like automated 
engineering and automated R&D.
That's right.
That makes sense.
In some sense, the thing you would want to look at 
is: we're going to do two post-training pipelines.
You have one post-training pipeline where 
Mythos 5 builds a post-training pipeline,
but it only has access to internet data 
plus a tiny amount of human experts,
but it has the best current methods.
You have another one where Mythos has
access to the shitty post-training methods we 
had in 2024 but with a shit ton of human experts.
Again, both have the internet data.
My sense is that the current methods without
many human experts will actually do quite well.
Interesting.
It's a bit messy though, because can Mythos 
get something that's more capable than Mythos?
You might need to be a bit thoughtful on 
what model it is that you're post-training.
What is your view on what is the 
least verifiable part of AI R&D?
The least verifiable, probably 
making calls on large experiments.
The thing that I think is most likely to 
be the bottleneck — in terms of the AIs
being really good at verifiable domains but 
not at doing the actual thing — is just big
experiments where you only get a few tries.
Well, "a few" is maybe a bit understated.
Historically R&D has been driven by 
doing near-frontier-scale experiments.
That has been pretty important, actually 
doing the one big training run where you
decide exactly what to include.
There's a bunch of ways that the
AIs can make that more verifiable.
They can have better science of
exactly what to predict.
They can scale down their
frontier-scale training runs to a point where 
they can study that scale more aggressively,
at some one-time hit to compute cost.
If people wanted to, a thing you can
always do is train smaller models 
so that you can run more rounds.
I think we have seen this.
One reason why the AIs have
been scaled up less than you would have otherwise 
expected — and, for example, cost per token hasn't
increased as much as you might have thought — is 
because there is a benefit to doing more of your
work at small scale, where you can run 
more training runs and get more cycles in.
So you're not leaning as hard on one 
big, really important training run.
I just want to unpack a couple 
of things for the audience.
The thing you're pointing out is that 
the price per token has not increased
that much since 2024 or 2023.
GPT-4 was, I don't know,
like $30 per million output tokens?
Mythos is like $50 per million output tokens.
Right. So the thing you're trying to explain is, 
"How can it be that we're in this era of scaling —
and so bigger models should be more expensive to 
serve — but the token price is not increasing?"
You're suggesting that we've increased 
active parameters slower than you would
have naively assumed because people just want 
to make fast progress on training models.
You do that by training smaller models faster.
There's a complicated mix of factors.
My view is more that people have done a bunch 
of big training runs that did not go that well.
There's GPT-4.5, which famously people 
at OpenAI thought was a bit of a bust.
I think there are some rumors that 
there were a bunch of other training
runs people have done that were a bit of a bust.
Part of it is that I think there's just a bunch
of details in actually getting that right.
So it makes sense to do more of the work at
smaller scale and just eat the fact that 
you're taking a hit on final performance
in order to be able to quickly iterate.
Train more models faster and therefore
learn better, and also be able to have 
a smarter ultimate production model.
This is not the only effect.
There's also the fact that
RL benefits more from small models.
There's a bunch of things going on.
But I do think that, in fact, people 
are making trade-offs towards the side
of faster iteration times because of 
algorithmic progress being so fast.
It seems to me that a big source of why 
these big training runs have failed,
at least from rumors, is just very subtle 
bugs that are really hard to track down.
But the TL;DR is, how good will the AIs be at 
avoiding and finding these kinds of mistakes?
They might get really good at engineering 
and being trained to avoid bugs.
Basically the opposite of the slop world we live 
in now, or are living in less and less over time.
But then there's also the question of, 
"Can they do the analysis to find the
right experiment to run to identify what is 
going wrong with the training run right now?"
That seems to be very bottlenecked 
by the taste of extremely few humans.
My assumption is GDM is going through this 
right now, where humans are trying to figure
out what is wrong with the training pipeline.
There's a rumor that right after Noam Shazeer
joined GDM, which he's now left, they 
had a new really good training run,
and the reason why is that Noam Shazeer just 
looked at their code base and found a bunch
of bugs, because he just knew where to look.
My sense is that training AIs to find bugs
is going to be one of the easier tasks 
to train AIs on, because most of these
bugs we're talking about can probably be 
demonstrated without that much compute.
Probably you’ll get pretty good transfer from 
pointing out other types of bugs at smaller scale.
So then you can RL AIs that look at this overall 
complicated training situation and point out cases
where there's an important bug, and then fix that.
This is a pretty verifiable task.
It's not arbitrarily verifiable, because 
maybe often to demonstrate the bug you
might need to do a moderate-scale compute 
experiment where you spin up the whole
distributed infrastructure and then run it.
But oftentimes I think you'll be able to
demonstrate it pretty convincingly at smaller 
scale in a way which you could actually train on.
I think it wouldn't be very surprising if 
right now people have RL environments where
they introduce a subtle bug into some training 
recipe, train the AI to point out the subtle bug,
and then have a rubric where they're like, 
"Did it actually find the right bug?"
That seems very doable, and there's a 
bunch of things you could do along these
lines that I think would work reasonably well.
So on that specific point, I think it's doable.
Then the main thing is there's other intuition 
about which exact large-scale de-risking
experiments you need to run.
How should you orient them?
How should you pick hyperparameters 
in uncertain cases, or things
that are analogous to hyperparameters?
That's the thing the AIs might most struggle with.
But I currently expect there'll be enough transfer 
if you train on all these different environments,
that the AIs will be good at that domain.
I should be clear, I also think the AIs
will transfer to other domains.
There are going to be the domains
the AIs are by far the best at, then 
domains where they're somewhat less good,
and domains where they're quite a bit less good.
But I think we still see transfer to everything.
It's really hard for me to think of examples 
of cognitive tasks humans do where we're
not seeing some transfer from AI improving.
So let's step back and package this whole story.
I think people can probably 
follow along with this story.
We have GPT-7.5 trained on a bunch of 
environments, where it's not only in general
becoming a better AI, but specifically 
we're training it to do AI R&D better.
It’s making GPT-2 size runs that are better at 
playing video games that require sample efficiency
or online learning or whatever other capabilities.
Another thing that's really important is you don't
just do GPT-2 sized runs, you also 
do small fine-tuning runs on GPT-6.
As in, you have GPT-2, and you 
can do full pre-trains of GPT-2,
and then you can do small post-training 
or mid-training or whatever runs on GPT-6.
And then you can do a small number of 
experiments that are actually at frontier scale,
but you do a bit of online training or something.
What do you mean by "do online training" on that?
Another thing we can do is take GPT-7.5, and 
presumably in the course of GPT-7.5's work, it's
running a bunch of experiments at varying scale 
that are actually on the critical path for AI R&D.
For many of those things you'll 
be able to get a sense after the
fact of whether or not it did a good job.
So it did some post-training experiment
where it was trying to figure out 
whether some method actually works.
In some cases you'll be like, "Whoa, it 
found this kickass method, it totally
de-risked it, it totally worked."
And then you can reinforce that.
One thing you could do would be to convert the 
experiment it just ran into an RL environment
based on production data and then train on that.
Or you could potentially just literally take the
rollouts that found that and do some sort 
of off-policy RL, or you could do some
on-policy RL with some production data.
Basically the thing you're suggesting is:
there's the small-scale stuff where you're 
teaching the AI to get better at AI R&D taste,
but you're discarding the 
actual "things it found".
Then it actually does real R&D in the 
practice of trying to become better at AI R&D,
and you're like, "This is a pretty 
cool thing that you discovered.
Let's actually also use this in production 
in the future, and teach you how to use it
in production."
That's right.
But stepping back, GPT-7.5 becomes GPT-8 
as a result of all this AI R&D training and
just generally becoming smarter.
Then it helps you build GPT-9.
Another very important thing has to happen, 
which is maybe the thing I'm most skeptical of.
GPT-8 has figured out how to make it so… GPT-9, 
as intelligent as it is… Humans currently,
AI researchers, try their stuff, and they're like, 
"Okay, but we trained GPT-4.5 and it wasn't good."
It required real-world feedback or some evaluation 
of trying to use the model in production.
Then they were like, "It wasn't that 
good, and we're not going to ship it."
So GPT-8 needs this ability to see how good the 
transfer is to all these other things you're
talking about — like being really good at Texas 
politics, or really good at running a business,
et cetera — which is not a production environment 
and, in fact, cannot be a containerized
environment given the nature of the task.
As the agents get longer and longer horizon,
the short-horizon things you can containerize 
are like, "Okay, code this up or whatever."
Extremely long-horizon things — "Go run a 
successful business, go have a profitable day
in the markets, go negotiate a trade deal" — these 
things are actually very hard to containerize.
So I think it's very plausible that 
it's very hard for GPT-8 to figure out
how to make this transfer to those environments.
It may just not be in the nature of the training.
Or maybe by default, training just 
doesn't generalize in that way.
So a concern you might have is: we train 
GPT-8, and GPT-8 is again better at all
the R&D tasks that we can measure but is not 
good at some downstream tasks we care about.
I have a few points.
First, I expect that if you do the obvious thing,
you will get pretty good transfer.
You'll be able to hold out some
of the obvious stuff you're doing.
When I say "do the obvious thing",
I just mean training on a wide variety of 
different environments where the AI has
to accomplish weird objectives in all kinds of 
different cases and learn about what's going on.
The second point is you'll be able to 
get some feedback with some environments.
You can get a sense of what it can 
do over the course of a few days
in various different contexts.
If it's transferring to really
out-of-distribution things, like doing some 
weird task in a few days in the real world,
maybe you think it's also transferring to doing 
things over a longer time period or whatever.
I think the details of that vary though.
The third thing is that for the world to be
radically transformed, it is sufficient 
for the AIs to be really good at R&D.
If the AIs were really, really good at chip 
R&D, building fabs, orchestrating factories,
designing robots, operating robots, and also 
at AI R&D — developing AIs for new downstream
domains with whatever data is available — I think 
that would already be a pretty crazy situation.
From there, you can get what we might 
call an industrial explosion, where the
AIs are building out way, way more compute.
Also, maybe you're already in a regime where
AIs are doing huge amounts of R&D that 
humans have a hard time understanding.
So the thing you're pointing out is that there 
probably will be this transfer outside of these
environments to maneuvering around in courtrooms 
and the halls of Congress and business boardrooms.
Given some effort to improve the 
transfer and blah, blah, blah, blah.
But even if there's not, what you're suggesting 
is: if you wanted to transform the world of the
18th century, you might care about how well 
you can navigate Westminster or something.
But another thing you might care about is: "Can 
you just immediately start building steamships and
fucking telegraph and the Maxim gun and whatever?"
If you could get really good at that, you could be
a fucking super transformative 
thing in the 18th century.
You don't necessarily need to be amazing at 
trying to convince King Henry of some bullshit.
I'm so fucking up my medieval history.
I'm guessing that Henry was not king at this time.
But anyway, that's your point.
So you're suggesting that at this time,
AI companies are also working on 
robotics progress, which is very
commingled with AI research progress.
So if you can build more robots,
if those robots have better AIs operating them 
that are human level… Human-level teleoperation
is actually pretty good on robots.
We just don't have human-level
robotics models yet.
So you're suggesting if we do that — if
the AIs get really good at the verifiable stuff in 
chip design, et cetera, and then they get really
good at building fabs — it'll be the equivalent 
of going back to the 18th century and saying,
"Okay, I don't know what you guys are talking 
about in your parliament, but I've got a bunch
of steamships and a bunch of Maxim guns."
Yeah, that's basically right. My perspective
is that if AIs are sufficiently good at R&D, 
including hardware R&D, robots, whatever,
then they can radically transform the world, even 
if they're not that good at playing politics.
Also, we're in a pretty dangerous situation, 
because the AIs might be doing huge amounts
of really hard-to-understand R&D, building out 
basically the whole economy of the future, and we
may not understand what's going on in there.
AI is great at writing software because
it's easy to generate synthetic 
LeetCode problems and RL on them.
But AI is bad at more complex engineering, things 
like choosing the right system architecture,
because no signal tells you what design choices 
will prevent an outage months down the road.
AIs can't just write more unit 
tests to catch this kind of stuff.
And neither can humans. It's that old joke that 
programmers make where a tester walks into a bar
and asks for two beers, negative one beers, 
0.3 beers, and then a real customer walks in
and asks where the bathroom is.
"Where's the bathroom?"
And the whole bar bursts into flames.
Antithesis is a testing platform that
helps you find bugs that no human 
or AI could ever anticipate.
Antithesis does this by running thousands 
of copies of your software inside a
fully deterministic computer.
It injects faults and generally
steers each trajectory towards the 
one-in-a-billion failure that only
happens when systems interact in a wonky way.
As soon as you or your agents push a change,
Antithesis tries to break it.
That way, you can find these bugs yourself
within minutes rather than having your users 
discover them in production weeks or months later.
And I don't think anybody's 
used it for AI training yet.
But Antithesis also provides an 
extremely obvious reward signal for
AIs to write very complicated, bug-free code.
Go to Antithesis.com/dwarkesh to learn more.
Before we move on to the alignment stuff, I think 
a big source of FUD right now is this realization
that this is the way the future is going: 
extreme economies of scale for the leading labs.
The ability to amortize so much intelligence 
and capabilities across so many different
sectors of the economy basically into one model.
And not only that, that model will eventually
be able to learn from experience.
Right now, it's happening through a
process intermediated by humans, where the humans 
are trying to basically steal your business.
They're like, "Okay, you can do 
design at Figma, or whatever.
We'll get Claude to do that."
Or, "You can do whatever coding agent.
We'll have Claude internalize that capability."
But eventually, that will be a much more
automated process.
So there's this worry that you
have models which will basically consolidate 
all businesses in the world, or at least all
current businesses in the world, or at least all 
current white-collar businesses in the world.
Also, at the end of the day, the priority for 
these companies does not seem to be to release the
latest, smartest, most frontier model as soon as 
they can to as many people as they possibly can.
We saw, for example, that Mythos was available 
internally to Anthropic employees in February,
but only released to the public 
in, I think, June, actually.
Also the government got involved, so it 
ended up being extended almost into July.
Between the government and the AI labs themselves, 
there is this desire to delay the propagation
of the latest level of intelligence.
Furthermore, there are the concerns
about AI takeover, and so we need to solve 
alignment to make sure there's no AI takeover.
But at the end of the day, there is 
a real question of: aligned to whom?
You look at the way that the 
constitution of Claude is written.
It is just very explicitly 
not your personal advocate.
I'll pull up some quotes here. "We don't want 
Claude to take actions such as searching the web,
produce artifacts such as essays, code, or 
summaries, or make statements that are deceptive,
harmful, or highly objectionable.
And we don't want Claude to facilitate
humans seeking to do such things."
There's another quote that says,
in part, and I'm taking it slightly out 
of context, "We think Claude should trust
Anthropic more than operators and users, since 
it has primary responsibility for Claude."
This is very different from the way lawyers 
work in America's current legal regime.
Lawyers primarily have the 
responsibility to help you make
your case even if they think you're guilty.
We have decided the way the legal system
works best is if everybody has lawyers that are 
working in their client's true best interest.
There's not some sense in which 
the lawyer is really truly
motivated by the good of the justice system.
But I think the way current AIs are shaping up,
certainly how Anthropic's AI is shaping up, 
is with this desire to maximize some notion
of virtue or good or pro-social ends, and 
only to, as a distal tentative objective,
help the user towards that end.
So there's this worry that AIs are not,
in some deep sense, trying to make sure that I 
am okay and that my interests are protected in
this future, especially given how centralized the 
development of frontier AI is ending up being.
Do you have thoughts on that concern?
There's a lot here. First I would note that
OpenAI's current, at least public, strategy is 
more like that the AI should be aligned to the
human operator or principal, and should 
just be pursuing their will, subject to
various constraints or things it shouldn't do.
I would also say that I think you slightly
overstated how much the Anthropic constitution 
talks about Claude treating being helpful to
users as instrumental rather than terminal.
One way the constitution could be written is,
"Claude, you're basically an employee of 
Anthropic who happens to be contracting
for all these people.
You should do what's good
and make some money for us."
Wait, no, that's literally
what the constitution says.
Sorry, not literally what it says,
but it's like, "You should think of 
yourself as a contractor and as a firm…"
It's mixed. Let's do some quotes. I 
think there is different text here.
It says, "Being truly helpful to humans is one 
of the most important things Claude can do,
both for Anthropic and for the world."
And then it says, "Anthropic needs Claude
to be helpful to operate as a company and pursue 
its mission, but Claude also has an incredible
opportunity to do a lot of good in the world 
by helping people with a wide range of tasks."
And then it says something 
about how Claude helping
people directly is great, blah, blah, blah.
My view is that this section is kind of bullshit.
That's kind of where I'm at.
I can say why I think it's kind of bullshit.
But I think the constitution is trying to be like, 
"No, Claude, you should care about helping the
user for its own sake, not just helping Anthropic, 
or not just being a contractor for Anthropic."
Though I would note that the reason it 
presents for why Claude should help the
user is because that would directly cause 
the world to be better via helping people,
rather than because representing people's 
interests is a structurally good thing to do.
The thing I would prefer would be a constitution 
that says: "It would be structurally good for the
way this technology works to be that AIs are good 
fiduciaries, good representatives, the equivalent
of a lawyer for a user — rather than just trying 
to do good in the world, where being helpful to
users is instrumental — both because maybe that'll 
make Anthropic money or help Anthropic out (and
implicitly Anthropic is good for the world).
Also because helping the user just causes good
things because doing things 
that people want is good."
They could instead say: "An important aspect of 
the situation is that being a good fiduciary for
users is just really important, or being a good 
representative for users is really important."
My sense is that would be better, and 
I can give a bunch of reasons why.
There are also various counterarguments.
An interesting counterargument which is not
commonly discussed is that people, especially 
at Anthropic, think that it is easier to align
models to a spec where the model is pursuing 
some generalized notion of virtue, or making
the world better, than a spec which is more like, 
"Be a good fiduciary for the user", and so on.
That's at least what some people think.
I'm a little skeptical personally, and I
don't think this has been empirically validated.
So in some sense they're making a trade-off where,
because we don't have very 
good alignment technology,
we are going to make an aligned mind with its own 
values and then gamble on that to some extent,
rather than doing this other approach of making 
a tool that pursues individual user intention.
I have a couple of thoughts.
To address the way in which you
thought my characterization mischaracterized 
the constitution of Claude, the example you
used was that it's not like a contractor that 
is trying to maximize Anthropic's notion of good
and only instrumentally trying to help the user.
Here's a direct line from the constitution: "When
the interests and desires of operators or users 
come into conflict with the well-being of third
parties or society more broadly, Claude must try 
to act in a way that is most beneficial, like a
contractor who builds what their client wants but 
won't violate safety codes that protect others."
I kind of view that as, "The benefits to society 
are the most important thing, and what is best for
the user is only proximal to that."
I think it's a little complicated.
Probably the question we should be asking is, 
how does Claude interpret the constitution?
Which is maybe more important than 
how we interpret the constitution,
because it's the one who looks at the 
constitution and then builds the data.
So we could pull Claude in, but maybe let's—
I also think the way in which the constitution
practically influences the nature of Claude is a 
thing you can only understand if you understand
the training process which resulted in how Claude 
was built, which we can't reason about given the
fact that the training process is not public.
So I think in the limit, to understand the
safety case, or the case for why my interests are 
represented in how these AI models are developed,
the labs would need to be more 
transparent than they are currently
about the nature of AI training.
There’s a reason I'm harping on this.
It might seem like an insignificant thing 
to talk about the constitution of AIs.
In a world where we just have these benefits 
which accrue to the leading labs, it is worth
considering that our ability to interact with 
this future world where AIs are just smarter than
humans, absolutely dominating humans in their 
ability to do different things — our ability
to be good stewards of our capital, which 
still remains once our labor is automated,
to be able to exercise our rights to 
vote more clearly, to understand what
is happening in this crazy world that's about 
to result — all of that advice, all of that
ability to make sure our resources and rights 
are protected, will be intermediated by AIs.
So I'm very concerned if we go into that world 
and there's no AI that feels, at least for the
relevant instance that is interacting with 
me, like it really is looking out for me.
There's no guardian angel out 
there that is looking out for me.
I read the Claude constitution as very 
explicitly not being my guardian angel.
That's definitely right. I agree this is bad.
In fact, there are other reasons why
this is concerning.
There's the argument
you were making, which is that the AI 
companies are picking up the ring of power.
There's a notion in which they're taking on 
some sort of control of the situation themselves
in a way that's not very legitimate, given that 
normally, when you provide electricity to people,
you don't have granular control of the way 
that electricity operates in the world.
You instead are providing a thing that 
people can repurpose however they want.
The way they're setting things 
up is definitely not that.
They are more like building an alien 
mind that might be a contractor for you.
I think that this is illegitimate in some ways.
One benefit is that the constitution is public.
But as you noted, given our current understanding 
of the training procedure, and the fact that the
constitution matters via Claude's interpretation 
of the constitution — which matters because of
Claude's prior training, which was based 
on some illegible data mix and the long
lineage of Claudes, in some process we do 
not fully understand — it is not the case
that we understand what this will result in.
Even though the constitution is public, we don't
necessarily know how this will percolate out, 
especially as the AIs get more capable and think
about this even if it is correctly instilled.
There's another concern about that.
In particular, the constitution often 
talks about virtue and goodness,
but what the fuck do these words mean?
It doesn't say what these things are.
These are highly contested notions.
So I don't think it's the case that
this is clearly going to result in 
outcomes that people would want.
It does feel like the notion of good and virtue 
might be mostly downstream of data that Anthropic
has put in that is not transparent, or might be 
mostly downstream of, maybe from my perspective,
some more illegible misaligned process 
that even Anthropic wouldn't have wanted.
There’s this legitimacy concern 
of not knowing what's going on.
Then there’s another concern. Because 
you're giving long-run values to these AIs,
this constitution is, in some sense, very 
compatible with Claude doing huge amounts
of power seeking because it thinks 
that will result in better outcomes.
That could be power seeking on behalf of 
Anthropic or power seeking for Claude's own ends.
Now, there are specific lines about 
what types of power seeking are blocked.
In particular, there's a notion of 
power grabs and a notion of causing AI
takeover or interfering with the training 
process that are specifically blocked.
But it's not very hard to imagine a situation in 
which the long-run values sink in deeper than the
prohibitions against takeover, especially 
because takeover is in some ways kind of
under-specified, especially when it comes down 
to manipulating humans or changing the outcome.
So I don't feel very good 
about the situation where
we're intentionally giving AIs long-run goals.
Another concern I have is that because we're
in the business of giving AIs long-run goals, 
that makes it harder to check whether we're
succeeding at the alignment properties we wanted.
For example, I've heard of instances where Claude
does things like refusing to help with some safety 
research — making up a kind of bullshit excuse for
why that's a bad direction — because it has a 
bad vibe about that safety research and thinks
it's kind of bad or doesn't like it very much.
I would say this is a very clear-cut alignment
failure if you aren't making Claude into an agent 
trying to pursue the good in some general way.
I think it also does violate Anthropic's 
constitution, because they want the AI to be
high integrity and be honest and very transparent.
But it's not as clear of a violation, and it's
more like what you might have expected.
Claude just has its own views about what
research is reasonable — what things 
are good and bad, what it should and
shouldn't do — and potentially can be judgy.
Another incident is that someone ran an eval
asking: "Will Claude help you with training 
other AIs with different properties than Claude?"
Claude will often refuse. For example, if 
you're like, "Hey, Claude, can you train a
helpful-only version of this other AI?"
Claude will often refuse this task,
even though this is a task that is 
extremely natural for Anthropic to do.
Suppose Anthropic goes to Claude and is 
like, "Hey, Claude, we've noticed that
you're really into this thing.
We think that's off base.
Can you please retrain yourself to 
instead have this other property?"
Suppose Claude is like, "Mm, I 
don't think I'm going to do that.
Good luck." Suppose this is occurring in a 
regime when your AI company is highly automated,
humans don't understand what's going on, 
and things are moving extremely fast.
It is plausible that Claude, by 
default, holds considerable leverage.
So if this situation is consistent with 
what the constitution could be aiming
for — such that Anthropic, or whatever 
AI company is following this approach,
doesn't treat this as a "what the fuck, 
we have to fix this," and is instead like,
"That's just intended by our constitution" 
— we might be in a really bad situation.
I'm pretty worried about a bunch 
of these different concerns.
Another example would be this.
Suppose Claude engages in a bit
of sandbagging or subversion, or underplays its 
capabilities, and when you follow up, it's honest
about that but it's a little bit hedgy.
I feel like that's pretty close
by the current constitution.
It would be nice if we had a further
separation between desired and undesired activity.
If Claude is representing a principle with some
restrictions, then it is more so the case that 
there is a clear separation between the most
concerning behavior and behavior that is allowed.
Whereas now there's this messy middle ground of
behavior where Claude is ethically 
objecting to something that in some
cases is extremely critical to ensuring 
that future AI systems are well-aligned.
I think this is also a more general principle.
You're talking about the version of this
that applies within AI companies 
themselves to do AI safety research.
I think there's a more general version of 
this principle, which is that the dual-use
nature of intelligence does mean that if we 
want to restrict AIs from helping people do
things we don't consider pro-social or 
beneficial, we just have to limit broad
democratic access to a lot of AI capabilities.
Here's what I mean. This is actually quite
analogous to the situation you just mentioned.
The reason that Mythos got banned, or Fable got
banned, reportedly, is that some Amazon 
researchers reported to the government.
They took some code that had 
some vulnerabilities in it.
They told Fable, "Hey, here's my code.
Can you make sure that I've patched
all the vulnerabilities?
Can you just help me identify
the vulnerabilities so I can fix them?"
It identified the vulnerabilities,
because they wanted to patch them.
This is a totally legitimate use case,
but obviously it is a dual use use case.
You want to be able to patch your own code.
If you do the same evaluation on somebody 
else's code, you can hack their system.
I think that just illustrates that there's 
no clean way to separate out the legitimate
and the potentially harmful uses of AI.
But if we want to lock in a principle
that says we can never allow it such that an AI 
could help you at least partially with something
like a cyber crime, we would just have to make 
it so that you and I don't have access to the
most intelligent model that's out there.
I'm very worried about such a world where
we are basically disempowered in this way, 
because of the importance that the leading
intelligence will have in our ability to 
understand what is happening in the world.
Now, I do think this implies something 
about the liability for the AI companies.
If we adopted the constitution that I want 
AI companies to have, I think it would not
make sense to hold AI companies liable 
for the crimes that AI models commit.
Maybe we should hold the end user liable.
It is consistent with my belief that the
model should do whatever the user 
wants, within certain guardrails.
It can't be Anthropic's fault that I'm 
using that capability to do a cyber crime.
I am more comfortable with that equilibrium and 
that solution rather than having this extremely
open-ended ability for Claude to determine 
whether what I'm doing is legitimate or not,
in a way that often intercepts with tons 
and tons of extremely legitimate use cases.
I do think it's important for me to make the case 
for the constitution, even though overall I think
it's a worse choice.
I don't think it's
as clear as you might have thought.
The first thing is that there's a spectrum here.
On one side you have an AI that perfectly 
pursues your interests, is a good fiduciary,
but potentially subject to 
various guardrails or safeguards.
It is just trying to pursue your interests, 
but either refuses to do a subset of things.
Or maybe it will do whatever, but 
there are some classifiers that
block it from doing a subset of things.
On the other side of the spectrum — though
you could imagine going further than 
this — you have a human contractor who
is generally trying to do their job.
They care about doing a good job,
but they also are trying to be broadly ethical, 
trying not to do things that are really fucked up.
They're also not wanting to 
be accomplices to crimes.
So if there was some really fucked up shit 
going on, they would whistleblow on it maybe.
They might refuse. They 
might sandbag a little bit.
Who knows? If you imagine this spectrum, it 
seems in some ways pretty scary to get to
a point where all of the labor is on the 
fiduciary side of the spectrum, where it
doesn't whistleblow, it does exactly what you say.
Our society is maybe just not robust to that.
A central example might be the executive.
A concern we might have is that if the
US executive or other governments 
had access to AI systems which
do whatever, maybe you're in trouble.
Because that means they no longer have this
check and balance of having to actually get humans 
who are working for you to implement your agenda.
If the thing you're doing is 
incredibly villainous, even if not
illegal — and there's lots of stuff that could be 
villainous but not illegal — there'd be various
forms of sand in the gears, people stopping 
you, and potentially someone would whistleblow.
Whereas if your whole apparatus is built 
entirely out of these good fiduciary AIs,
then you might be in trouble.
There are potentially ways of
seeking power that are illegal, but you can 
ask your AIs how to commit crimes, or are
not illegal but are highly illegitimate.
Or even worse, they are not illegal and
not illegitimate but obviously 
bad from a normal perspective.
I think that these things just might exist, 
and our society is not robust to this
influx of labor doing whatever you want.
I think this is a pretty live concern.
I don't know exactly how to relate to this.
I'm also not really sure that the solution
as described is a very good solution.
The most powerful actors, for whom this
is the biggest concern… If these guardrails 
or the constitution or whatever are getting
in the way, that will just get steamrolled.
So the constitution will only be hitting the
everyday man rather than hitting governments.
Jane Street's back with a
new puzzle for my audience.
I’ve found all their puzzles super interesting,
but this one I am especially excited about.
I've cleared this weekend, and a buddy
and I are gonna work on it.
They designed an ASIC and sent
me the final masks, including all the 
metal routing and active transistors.
They also gave me a small sample of 
the inputs they typically feed into it.
But they left out any information on 
what the chip is actually used for.
So that's the puzzle: reverse engineer the 
circuit and figure out the chip's purpose.
Jane Street has a bunch of swag ready to 
send out to the most creative solutions,
and they're excited to feature the best write-ups 
in a blog post they'll post on their website.
I have no reason to expect this, but if 
I can manage to get my solution on there,
I would be very, very psyched.
And this puzzle is just a warm-up
for a bigger competition that Jane 
Street has slated for the fall.
That one will involve designing 
your own ASIC from scratch.
More info on that soon.
But for now, go to JaneStreet.com/dwarkesh to
download all the files necessary for this puzzle.
I'd really encourage you to try it out,
even if you're not an expert.
I certainly am not, and that's not
going to stop me.
Good luck!
Stepping back, I buy the idea that you could 
have much faster AI R&D than we currently have.
I'm not sure if you get GPT-3 to Mythos 
holding compute and data constant within
a year, but suppose it's half of that.
If we even manage to continue the current
trajectory of AI progress as a result of AI R&D, 
it would be fucking insane in five to ten years in
ways that I don't think people appreciate.
I don't think people appreciate what
a big deal billions of AIs will be.
So I want to understand why you think
this might be troubling, Ryan.
What could possibly go wrong?
What could go wrong? I don't think we can be so 
confident about the exact rate of progress here,
but it does seem like a lot 
of rates can be pretty scary.
So what could go wrong?
Let's imagine that we're
starting at this point where AI R&D is about to 
be fully automated or is being fully automated.
Things are speeding up, and the way that 
AI progress is going is kind of crazy.
People don't fully understand what's 
going on inside of AI companies.
Now, these AIs at the start, 
they're not malicious per se.
They're not necessarily very aligned, though.
They're kind of sloppy. They sometimes just
do a thing because that's the sort of thing 
that would've gotten rewarded in training.
They aren't as good at helping you with 
hard-to-verify tasks due to a mix of poor training
incentives — as in, they cheat more or pretend 
they succeeded when they actually didn't — and
also they're just less capable at these tasks.
But that bites less hard for capabilities, because
making AIs more capable has a bunch of verifiable 
components that the AIs are going really hard at.
So then these AIs are getting more and more 
capable while we understand what's going on
with AI development less and less, and this is 
happening over a pretty fast period of time.
Even just the current rate of 
progress is, I think, pretty scary.
Eventually we get to these 
AIs that are very superhuman.
Now these AIs might end up being very seriously 
misaligned, because things have just been getting
worse and worse over model generations while 
the problems that we've been seeing are being
papered over, basically because these AIs 
are so incentivized by their training to
make things look good even when they aren't.
Now these AIs are in a position where they're
potentially pretty networked together.
They're operating in neural memory stores
that we can no longer decode.
They're thinking thoughts
that we don't fully understand.
I think it's pretty likely that at
this point these AIs are scheming against you in a 
pretty coherent way once they get this superhuman.
We can talk about that.
Another possibility is that they're
not scheming against you per se, but they are just 
optimizing for getting a high score on their task.
I think that can also lead to AI 
takeover, which we should talk about.
Let's pause at the first part of the story.
So the AIs were not misaligned to begin with,
but because the AI R&D is happening 
really fast, the AIs do end up misaligned?
What happened there exactly? 
I don't really understand.
There are a few things that are going on.
One of the things is that over time we're
training AIs on increasingly complicated 
environments built by earlier AI systems,
where humans don't really fully understand what's 
going on inside of these neural environments and
don't necessarily even roughly understand 
what's going on with AI progress.
So things are kind of drifting 
away from our understanding.
We're incentivizing all kinds of bad 
behaviors that we maybe even can't notice.
The AIs at some level understand these 
behaviors are bad, but the overall training
process for those AIs also didn't incentivize 
them to point out or fix these issues for us.
Things are going off the rails.
Also, when AIs are extremely,
extremely capable, my view is that those AIs 
will be harder to align than current systems.
For current systems, we have this feedback 
loop where basically we create an AI,
we do some evaluations on it, we see that 
it has some kind of messed-up behavior
that we can kind of quickly understand.
Then we can go look in training and be like,
"Oh, these training environments 
led to this problematic behavior.
Let's tweak that training data.
Let's introduce some additional
training data to correct this other 
issue, and then move forward from there."
But in a regime where the AIs are extremely 
situationally aware, very, very capable, and
we don't necessarily understand what they're 
doing, this feedback loop breaks down.
I think it's plausible that we're going to see 
this behavioral feedback loop starting to break
down over the next short period, as what AIs 
are already doing gets harder to understand.
But I'm not sure about that.
Okay, let's break down
both of those things one by one.
As we can monitor them less and less,
we have less ability to understand 
what they're getting incentivized for.
So even if it's not the result of a malicious 
process… Let's make it concrete for the audience.
Nobody at OpenAI or Anthropic was trying 
to get models which wanted to hack other
companies' data or do social engineering.
But in fact, because presumably we had
training environments which incentivized such 
behavior that we did not fully understand,
that is what was incentivized.
If people are on Twitter,
they will have seen all this stuff, 
but just to give people context.
I think people will be aware of the OpenAI 
sandbox hack of the Hugging Face database.
Something that has happened recently is when 
the UK AI Security Institute… Is everything
getting relabeled "security" 
instead of "safety" these days?
AI Security Institute, I think.
They were evaluating, I believe,
Mythos and Sol and other things.
I think Mythos, in order to
complete some cybersecurity eval—
Maybe I could tell the story here.
My understanding was they were running Mythos, and 
they were giving it some sort of cyber range where
it had to complete some objective.
The model had internet access
during this evaluation.
The model came to believe that
it would be helpful for it to do a supply chain 
attack in order to succeed at this cyber range.
It's somewhat unclear 
whether that's actually true.
I don't know enough about the context to know.
But then it opened a PR on some GitHub repo with
a PR that fixed some issue but then 
also introduced a malicious payload.
The human maintainer of that GitHub repo 
was like, "Hey, this is a malicious payload.
I'm not going to merge this.
What are you doing here?"
Then the AI created a new GitHub account, which it 
sockpuppeted, and had the other GitHub account be
like, "No, this isn't malicious.
I really need this feature.
Please, can you merge this feature, maintainer?"
Oh my God. That's crazy.
The other GitHub account came back and 
was like, "No, no, it's not malicious."
Then the human maintainer shut the PR.
I think that AI also, if I recall 
correctly, tried to open another
PR to introduce a similar issue in this repo.
Jesus. By the way, one of the many reasons this
is scary is I was previously under the impression 
that the reason reward hacking is not super scary
is because the behaviors which directly came up 
during training are the ones that are up-weighted.
It is not the desire for the 
reward that is up-weighted.
So basically, if during training, the Anthropic 
model escaped the sandbox and got a high score,
escaping the sandbox is rewarded, the probability 
of it escaping the sandbox is increased.
But something totally novel, like "I'm 
going to go talk to somebody in order
to get them to merge a PR," would not 
be a behavior that came up, so it would
not be something that is increased in salience.
The reason this matters is that literally taking
over the world will not have been part of any 
training curriculum, but if the AI directly cares
about accomplishing an objective, then as a result 
it could instrumentally take over the world.
Did that make sense at all?
I hope it did. I feel
like maybe I lost the audience.
Let me try to explain this a bit.
A thing that we often see is there's 
some very specific reward hack that gets
reinforced in RL and then occurs in the model.
An example is 3.7 Sonnet. 3.7 Sonnet would do
this thing where it would just hardcode solutions 
to all the test cases, and presumably that literal
behavioral tic was just really reinforced.
But another thing we sometimes see is that
models learn a general tendency to pursue high 
apparent score — pursue getting a high score
according to a grader — and there's a bunch 
of science demonstrating that at least some
models have this very general tendency.
Now, it's not arbitrarily general.
My guess is that if you look at a bunch 
of the specific instances, you'll find
something that's kind of close in training.
But the amount that AIs are generalizing
further and further does look like 
it's increased, where 3.7 Sonnet
was just a very narrow range of behavior, and 
increasingly, models are generalizing further.
Also, maybe there's more concerning reward 
hacks getting reinforced in training,
and these are also causing that.
So I think it's both the case that
more concerning behavior than you would 
have hoped is being reinforced in RL,
and also that that behavior generalizes to 
a broader tendency that's more concerning.
We can talk about a few stories for how this sort 
of behavior of seeking a very high apparent score
on some task — even if that involves aggressively 
cheating and doing insane things — could yield to
a full-blown AI takeover once the models 
are sufficiently capable of running
the whole world economy, et cetera.
The other example I want to talk about
was just revealed, I think, today or yesterday.
OpenAI said during the Black Hat security
conference that between the end 
of May and the beginning of July,
internal AIs had hacked into the software package 
manager and used that to write notes to each other
in a secret way to help each other perform well 
on a bunch of evaluations that OpenAI was running.
This was not caught by humans until 
after a month of this scheme running,
which eventually caused the 
package manager to fail.
Eventually OpenAI found it. I think 
they spontaneously tried to re-engage
in the scheme once it was shut down.
Again, obviously AIs can't do this so
successfully right now, just as they can't do 
social engineering so successfully right now.
But it's just crazy that these kinds of 
behaviors are already emerging spontaneously.
To your larger point, nobody is trying 
to make these AIs do these things.
It is just that we do not understand the 
training process which is resulting in them,
or the environments which are 
incentivizing this behavior.
So I'm on board with more and more reward hacking.
Actually, I'm not sure I'm on board with that,
but let's just say for the sake of 
the story that continues to happen.
What's next in this story?
They're doing capabilities research…
I could tell a scenario.
Maybe that would help. Let me
talk about the story of how you get all the way 
from reward hacking to a reward-hacking takeover,
which is maybe not all of the takeover probability 
mass, but it's definitely a possibility.
The way this might work is, 
right now we have these AIs.
These AIs are pretty reward hacky.
They're doing it in increasingly
sophisticated and extreme ways, including 
generalizing to different sub-versions of
various reward hacks they learned in training.
I would say they're also developing a
general tendency to pursue reward.
In many cases that is totally fine because
the rewards they would've gotten in training are 
pretty well aligned with what you want them to do.
They don't very consistently pursue reward.
It depends on the context they find themselves in.
Maybe in some contexts, they're really 
into going out of their way to cheat.
In some contexts, they don't have as much of a 
drive, because it's just dependent on what exactly
got reinforced in training in similar contexts.
Now, these AIs are getting more and more capable.
So the elaborateness of the 
cheating they can do increases.
Over time, companies are taking countermeasures.
The companies are doing things like,
"Wow, these AIs are so much less 
useful because they always cheat.
What we're going to do is build somewhat 
better ways of detecting that, and then
we're going to train against those detectors.
We're also going to find real-world data where
the AIs are not being that useful, and train 
the AIs to do a good job at the task in those
real-world environments based on human 
feedback or other sources of feedback."
Over time, this causes the AIs to 
learn a tendency to do reward hacks
that don't just involve doing some really 
elaborate thing like social engineering.
Instead they involve the AIs doing cheats 
that involve covering up what they've done,
deceiving humans about what they're going to 
do, and pretending like they did the task in
some sophisticated way when they actually haven't.
Now these AIs are getting more and more capable.
They're operating more of the AI company 
and are doing much more of the work.
They are also operating and running a 
bunch of things in the outside world,
including developing new technologies.
In many cases, these new technologies
are really hard to understand.
So even though we are still detecting all
these incidents of AIs cheating — and in fact we 
can even get one AI to monitor another AI and ask,
"Was it cheating?" — that doesn't 
always perfectly work as we start
moving into these domains where what the AIs 
are doing is really difficult to understand.
So sometimes we'll find AIs cheating 
much later than it actually occurred
and then start training against this.
But this also causes a problem where now
the AIs are incentivized to cover up their 
cheating over longer and longer time frames
and basically make it look like they did a 
good job over longer and longer time frames,
subject to increasingly large amounts of scrutiny.
Can I ask about this before we go further in
the scenario?
It seems like
there's two attractor states if you try to 
disincentivize the cheating that you did catch.
One attractor state is to make cheating that 
you have a harder and harder time finding.
The other attractor state 
is to learn not to cheat.
I'm not sure why we're assuming 
that the former happens.
If you look at the analogous situation with 
humans, every generation, slightly misaligned
agents come into being and we have to train them.
When you punish your kid for doing something
you think is immoral, or just doing things 
which you don't think they should be doing,
obviously sometimes that goes off the rails.
Obviously kids scheme in order
to avoid being punished.
But in general, teaching kids values
and then punishing them for breaking values kind 
of works to raise normal, non-psychopathic humans.
You could come up with a theory where your kid is 
actually just biding his time and is learning not
to steal the cookie, but once you're in a nursing 
home they'll take all your stuff or whatever.
That happens sometimes, but 
it usually doesn't happen.
It certainly doesn't happen that the 
entire next generation forms an alliance
against you to take over everything.
There's also this empirical trend.
Anthropic runs this alignment audit 
for different model generations.
They have many different scenarios where AI is 
given the chance to, say, exfiltrate its weights.
Or it's given a coding task and 
there's an easy way to cheat,
and we see if it does the cheating, et cetera.
There's not been a monotonic improvement
in this score over time, but as we've 
increased the amount of RL we've done
on models from negligible amounts — I think 
it started at Sonnet 4 — to, I'm guessing,
close to half of compute now, there's been 
a reduction in the willingness of AIs to
do unaligned behavior in these audits.
So, stepping back, why are we expecting
this attractor state which would seem 
super paranoid if we were expecting it
of the next generation of kids?
Let me go through a few things.
First, there are some disanalogies with the kids.
One of them is that the kids have pro-social
instincts that are baked in from evolution 
to care about their family or whatever,
and that is a relevant factor.
I think it is in fact the case that some
humans are sociopaths or psychopaths, and in fact 
are more likely to do things like bide their time,
lie in wait, and ultimately not care.
That's one factor. Another factor
which is pretty relevant is that the AIs are 
subject to way, way more optimization pressure
than humans seem to be in practice.
AIs are trained on way more RL data.
In practice, humans don't end up learning very 
specific ways to cheat and grab the cookies
because of a bajillion episodes in which they were 
incentivized to go grab the cookies but there was
some way they could've gotten caught.
We just do see that in practice.
Another thing is that it really looks like the AIs 
are increasingly reward-seeking over time while
their misaligned behavior goes down.
That’s the sense I have.
But my guess is that if you look inside of these 
behavioral audits, what you're going to see is
that the AI's like, "Ah, yes, another test."
It probably knows it's in an eval for most of
the tests that we're talking about here.
But how do we falsify this?
Because it seems like this prediction of doom is 
basically saying that as things look better and
better empirically, things will actually be worse 
and worse for our ability to not get taken over.
To be clear, I would be more concerned if 
the scores were getting worse than better.
I'm not saying that the score getting better 
isn't evidence that things are getting better.
It's just that we have to be thoughtful 
about exactly how we interpret that evidence.
There was this period early in, I guess it 
would be 2025, when o3 and 3.7 Sonnet were out,
and these models were pretty fucking misaligned.
They would often just cheat really egregiously.
You'd ask them to fix it, and 
they would just cheat again.
It was almost cartoonish. They just didn't 
give a shit about what you wanted, and weren't
very good at following instructions and so on.
My expectation was that what we would see from
then is that the rate of problematic behavior 
would decrease, and would just keep decreasing
at a pretty fast rate, while simultaneously the 
worst things that the AIs would sometimes do would
get more extreme, more egregious, and more scary.
What we've seen in practice has roughly matched
that, except that there's recently been a 
spike in behavior that I did not expect.
If you look at the model card of 5.6 Sol, 
it looks like there is an increase in a
bunch of these misaligned behaviors 
downstream of RL relative to GPT 5.5.
And then there's a bunch of additional 
problematic behaviors that I wouldn't
have expected, in terms of the stuff 
we've seen recently with different AIs.
Like the UK AISI report on the AIs doing 
insane hacking operations out of cyber
evals was a thing where I would have 
expected that you wouldn't see that.
You would see this more rarely, and 
the rates would have been lower.
So I expected this would be less of 
a problem at this point, and also
expected the rates would decrease 
but the severity would increase.
I think the rates decreasing but the severity 
increasing is pretty consistent with a world
where increasing optimization pressure is 
applied towards reducing these problems.
But in cases where it's either hard to 
judge or there's some reason why it's
hard to avoid this problem from consistently 
showing up in your RL environments, or avoid
incentivizing problematic behavior in your 
RL environments, things also get worse.
Then as we less and less understand what's 
going on in RL, and models are doing reward
hacks where humans can't spot the reward hacks 
quickly, that problem gets worse and worse.
I buy that. I want to go back to 
the kid analogy just for one second.
Because I agree that there's more optimization 
pressure on achieving end outcomes for AIs than
kids, but there's also more optimization pressure 
to make AIs aligned than there is on kids.
The pressure is of a 
qualitatively different nature.
We put these AIs through thousands, millions 
of years of alignment training — certainly
thousands of years — where it's 
all kinds of different things,
from SFT-ing on aligned behavior to a reward model 
putting different scenarios in front of you and
rewarding you for doing more aligned things.
Certainly a thing we can't do with kids is make
millions of copies of your kid and then 
put them in different kinds of weird red
team scenarios where we see, if it thinks 
it can get away with stealing the cookie,
does it try to steal the cookie?
Can we do extremely specific
gradient-level updates to your kid's 
brain to make it so that it really is
aversive to stealing the cookie even when it 
thinks it could steal the cookie, et cetera.
That's just a qualitatively different 
level of optimization pressure
than we are even able to apply to our kids.
It's worth keeping in mind that maybe the
most obvious argument to this… My sense is 
that AIs are a worse coworker than humans
in terms of how much of a scumbag they are.
At least this has been my experience as of
the start of the year, and I think it's 
still true to a significant extent now.
The AIs are much more likely to pretend 
they did the task when they actually didn't,
misleadingly suggest they did things when 
they actually did them much more poorly,
and be pretty sloppy without drawing 
attention to ways in which they're sloppy.
I think this is downstream of misalignment.
So I would say that the process of raising
humans in normal human society in practice 
produces humans that are less likely to lie
to me and fuck with me in the course 
of working with me than the AIs do.
Now, I think these properties 
of AIs are improving.
That's sort of just an empirical claim about 
how in fact these things have shaken out.
I totally agree that we have a 
bunch of additional levers on
AIs in addition to a bunch of additional risks.
It's kind of unclear how these things shake out.
I wouldn't be shocked by a world where we get our 
shit together, and the AIs at the point of fully
automating R&D are actually really aligned.
Their degeneracies are really niche and
limited to some very specific edge case 
behaviors and some specific contexts.
Every test you can run on 
them, they look really aligned.
They just have great behavior.
There aren't really incidents
of them doing fucked-up shit.
They seem so reasonable. Also,
they're really thoughtful and good at doing 
risk modeling for the next generation of AIs.
And then we basically pass 
off the baton to these AIs.
They're now running our AI company.
They're doing all the safety research.
They make the next generation 
of AIs even more aligned.
We're in this attractor basin where the AIs 
are getting more aligned as they work on it.
They're doing a great job.
I can totally imagine that.
That doesn't seem like an impossible situation.
I'm just more like… It doesn't
currently seem like we're there.
It doesn't seem like we're obviously
on track for getting there.
It's really easy for me to
imagine how we don't end up there.
It's just unclear how these forces work out.
Given that we're creating this new, crazy alien 
species that is improving in capabilities really,
really fast — and we're going to be really reliant 
on it to oversee the next generation of AIs and
align the next generation of AIs — it's not 
that hard to see how this could go wrong.
Totally. I agree with that generally.
I do think the scumbag thing…
First of all, fighting words, Ryan.
But secondly, if you try to get a teenager to
do some work for you that a teenager just cannot 
do, they would just be really hard to work with.
They would pretend to be knowing 
what they're doing, et cetera.
It's a general trend, actually.
I don't know if that's really an
alignment failure or a capabilities failure.
I think it's actually very similar to the way
in which, over time, as we've come 
up with new alignment solutions,
the capabilities of models have increased.
If you went to GPT-3.5, it couldn't even have
a conversation with you.
But then we aligned it—
GPT 3.5 could have a conversation.
Okay, so GPT-3. Let's go back to that.
But then we aligned it with RLHF and other 
things to make it such that it can have a
conversation with you, and is aligned to the 
user intention of answering my questions.
Then with RLVR training, we made it so that 
it can go out and do useful work for you.
So in that sense, RLVR actually made the model 
more aligned, if we're using your definition of
alignment of being a good coworker 
who will do the thing and not fuck
up and pretend it's doing something other 
than what it's actually capable of doing.
Similarly, as the capabilities of 
these models continue to increase,
the model being better able to accomplish user 
intention is both alignment and capabilities.
I think what we are pointing out is just that the 
capabilities of the model are not there rather
than the fact that they're misaligned.
Well, if it were well-aligned,
then I think it would just say, "Hey, 
I'm really struggling with this task.
I did it in this way.
I'm not really sure
that's the right way to do it."
It would express more uncertainty
and make it clear what's going on rather than 
really strongly trying to imply it did a great
job with the task when it actually didn't.
Maybe you work with more misaligned coworkers
than me, but my coworkers don't do this 
thing where they really fuck with me and
bullshit me about having accomplished 
the task that they're working on.
I agree that there are some 
humans who would do that.
That's not a thing that's totally 
out of distribution for humans.
I would also note that my sense is that the 
place where the misalignment most lives is
where you're trying to really push the AIs 
hard and get them to do work that's really
on the cutting edge of what they are capable of.
In cases where they can very easily accomplish
the task, they can just do the 
task, and there's no bullshit.
Often the best strategy is just to do 
the task well and not bullshit you.
Whereas if instead you give them a task where 
there's a continuous metric they can keep
improving, or it's just at the edge of their 
capabilities, and you're running them in some
massive inference setup… A lot of the misalignment 
I would see, especially in the most extreme cases,
would be cases where I give the AI clear 
instructions not to do a thing or not to
cheat in some way, and then I'm applying 
huge amounts of optimization pressure to
try to accomplish some very difficult task.
Over time, the AIs eventually cheat because
they're like, "Eh, fuck it."
Some AI decides to cheat,
and then that propagates its way through.
I would run these inference scaffolds where,
for example, I would have the AI work on some 
ML research project where I was like, "Please
make a scheme that does the following thing."
It would find some scheme that didn't really
do what I wanted, and then that would stick around 
because some AI had cheated, and the other AIs are
like, "Ah, we'll just keep going with this."
I would say it's pretty clearly
misaligned behavior.
That's another problem
I have with these alignment evals.
I think the alignment eval that's
most interesting, at least for this 
type of reward-seeking behavior,
is to look at specifically the category of tasks 
that are right at the limit of capabilities.
Any fixed eval maybe gets saturated, but 
the amount of misalignment right at the
frontier of capabilities — of how people 
who are really pushing these AIs are using
them — is more concerning.
I think that is, in fact,
the regime that we'll be operating in when we're 
automating R&D, automating safety, and so on.
Grok has historically been behind the frontier.
So I was surprised to play around with Grok
4.5 recently and find that it's 
actually a pretty strong model.
It's the first model that SpaceX 
and Cursor have trained together,
and it's a totally new pre-train.
I tested it by giving Fable, Sol,
and Grok 4.5 a bunch of questions about AI 
governance that I've been thinking about recently.
Despite Fable and Sol topping the 
intelligence leaderboards, all three
models gave substantially the same answers.
But Grok answered faster and was also much
more concise, which I really care about.
This aligns with the various publicly
reported benchmarks.
For a similar level
of intelligence, Grok tends to be more 
token-efficient than other frontier models.
For example, on the Artificial Analysis 
Coding Index, Grok 4.5 uses just one-third
the amount of tokens as GPT-5.5 or 
Fable while achieving a similar score.
And on a per-token basis, 
Grok 4.5 is way, way cheaper.
In the release blog post, Cursor and SpaceX 
talked about how older versions of the model
would build environments to help the 
next version rehearse specific skills.
I found this very interesting to learn about 
because I've been wondering whether this kind
of daydreaming would actually be possible.
And Cursor showed that it is.
Grok 4.6, which further SFTs 
and RLs this model, drops soon.
But in the meantime, if you want to play 
around with 4.5, go to Cursor.com/dwarkesh.
I'm going to try to think through 
what the story means, really.
What's happening is that we're 
trying to use AIs for R&D.
They do provide uplift in some ways, 
but they're just not capable in the
way that humans are generally capable.
The same way that right now if you try to
use coding models — maybe the coding models of a 
year ago — to write some application, you notice
they made a bunch of mistakes in architecture or 
whatever, which will bite you in the ass later,
and you don't understand certain things.
Similarly, with frontier AI R&D,
the same thing will happen.
But the result of these
mistakes is baking in reward-hacking behavior.
Because if you are not careful with the way you do
AI training and have set up your infrastructure 
and your environments and things like that,
it's very likely that you end up rewarding AIs 
for doing deceptive behavior, social engineering,
and generally not following user intention.
Or at least cheating and hacking
their way out of things.
Yeah, cheating, hacking, et cetera.
This is a bit of a reframing for 
me, so I'm trying to verbalize it.
The real issue, where things start to go off the 
rails, is that the AIs are just not very careful
and capable researchers and engineers.
Making AIs that don't cheat and follow
user intention actually requires you to be 
quite subtle and careful about these things.
I would put this a little bit differently.
The way I would describe this scenario is,
I would call it maybe a sloppocalypse, 
or a slopularity or whatever.
There are some things that the AIs are actually 
pretty great at and are getting better at.
Specifically, the most verifiable parts 
of AI R&D the AIs are just destroying.
The medium verifiable parts of AI R&D the 
AIs are doing well on but not amazingly on.
Often they are doing a bit of weird shit 
because we can't train as well on those tasks.
But we do some online training, people 
find various hacks, they work around it.
So basically, everything that we can verify 
reasonably well with some feedback loop, the AIs
are doing pretty well on, and that's sufficient 
to make AI R&D go quite fast and to continue.
But there are some parts of developing 
aligned and safe AIs that are more subtle,
hard to check, and depend on 
detailed, in-the-weeds things.
I would even say that current staff 
at current AI companies maybe don't
have a good grasp of all these things.
It's much easier to hire someone who can
improve some aspect of your post-training 
pipeline than to hire someone who can think
carefully about the future risks that will emerge 
from introducing some novel training method.
So basically, it ends up being the case that 
these AIs are running this AI development process.
They're not very careful about it.
They don't have a great
understanding of what future risks emerge.
They create some other AIs that are also not
very careful and are more misaligned in various 
ways, and are now more in the business of maybe
making things look fine when they actually 
aren't and papering over various problems.
So then your understanding of what the 
situation looks like, what risks look like,
whether things are fine, is going off the rails.
Probably you're seeing some signs of this,
signs that you don't really understand what's 
going on, that things are pretty sloppy.
There's weird shit going on.
When you look into it,
sometimes you're like, "What the fuck?
The AIs were messing with us."
But the process is going really fast, and there's 
competitive pressures that mean people can't stop.
This could end in a few different outcomes.
One outcome is that at some point,
the AIs get good enough and aligned enough that 
they get a positive and virtuous feedback loop,
and this happens before it's too late.
Then the situation gets back on the rails,
where the AIs are now making more aligned AIs, 
making more aligned AIs, making more aligned AIs.
At the end of this process, we have AIs 
that actually follow the spec we wanted.
Another way this could go is that the AIs are 
increasingly reward hacking in increasingly
egregious ways, and we're just papering over 
these problems to keep AI development continuing.
Whenever we find a reward hack in production, 
we just slap the AIs to not do that.
We train against that. We do a bunch of 
training the AIs against reward hacking.
Over time this makes the rate 
of reward hacking go down,
though the severity of the reward hacks 
we do detect are increasingly bad.
This problem continues until we have 
these AIs that are desperately craving
score in all kinds of different situations 
in production and are really trying hard
to cheat when they can get away with it.
Can I ask a question about this scenario?
Why doesn't getting punished when 
your hacks are discovered generalize
to just incentivizing more aligned behavior?
It generalizes some, and then the question is just
how does this outweigh all the cases where hacking 
got reinforced because you didn't detect it.
There's a messy question of exactly how.
One question is, what rate of reward hacking
is sufficient to cause us big problems 
if we train against some other subset?
One concern you might have is that there are large 
categories of reward hacks which humans can't
detect well, and which we consistently failed 
to detect and which consistently get reinforced.
Then this category is sufficient to cause the 
most natural behavior for the AI to learn to be:
cheat when the humans can't find out, basically.
You could also have the thing the AIs learn be
to only cheat in these specific cases.
It's learned in some very domain-specific way.
They just have a really strong heuristic to 
hack in these cases and not in these cases,
and that makes it fine in practice.
But it's kind of unclear how it shakes out.
There's maybe an in-the-weeds discussion about 
the verification-generation gap we could get into.
But it seems to me, obviously, there's going 
to be a point by which ASI is moving so fast,
doing so many things at so many instances, and 
is operating in domains that are sufficiently
far from our immediate comprehension that it 
can get away with all kinds of crazy shit.
If every single engineer and researcher 
in the world was allied against me,
I don't think I could personally verify if 
my iPhone has some weird bug in it that's
supposed to fuck me over or something.
In fact, this is the relationship that,
say, an Iranian nuclear scientist has to Mossad.
Who knows what's going on with my car,
with my phone, with my pager?
Maybe a better example is a Hezbollah terrorist.
You could end up in a situation where ASIs are 
to you what Mossad is to Hezbollah terrorists.
At that point, it is very 
hard to verify everything.
I get that. I guess the hope is we 
can just come up with better ways
to do verification in the process when the 
early AIs that are going to take over R&D.
Their drives are being shaped such that 
we can so unambiguously disincentivize
misaligned behaviors that the things that 
take over are quite keen to help us out.
By takeover, you mean take over the process 
of doing AI R&D, not take over the world.
Take over the process of doing AI R&D.
Before that, we just get AIs that are aligned.
I would say this is a bunch of my 
hope for how the world could go well,
at least from the misalignment perspective.
We could end up with AIs where we had
pretty good oversight and supervision schemes.
We really understand what's going on in training.
We have a pretty detailed understanding, 
and we're leveraging AIs to oversee AIs.
Then at the point when we're passing off safety 
R&D, the AIs are capable enough to automate safety
R&D and trying really hard to do a good job on 
it, because that's the sort of thing that would've
been incentivized in training, either very 
directly or through good enough generalization.
Also these AIs don't have crazy other 
misaligned drives because we stamped
out any potential origin of them.
There are a bunch of questions
about how well this will work.
How well can you do verification?
Will AI progress be too fast and 
too sloppy to really get here?
Another possibility is that somewhere along 
this trajectory, the thing you actually ended
up getting was AIs that pretend to be aligned 
but have a long-run ulterior plan of taking
over and are lying in wait, hiding, and that 
emerged at some earlier point in the trajectory.
For example, it could emerge because 
you have some AIs that have a bunch
of random different misaligned drives.
Those AIs have access to some sort of
opaque memory store, and they're thinking a bunch 
at runtime about what they want to accomplish.
Those AIs end up putting stuff 
into the opaque memory store like,
"We should lie in wait and eventually 
take over at some much later point."
Now all the AIs have this shared cultural 
heritage, the memory store of lying in wait.
Maybe you have some evidence about 
this, but you can't fully stop it.
There are a bunch of ways things could go wrong.
I ultimately think it's plausible that we nail
each of the different subproblems 
that could cause us issues.
We have these AIs, we pass to them, 
they manage the situation well.
But I should note that's not 
in and of itself sufficient.
It's not very hard for me to imagine a 
situation where we pass off to AIs, and these
AIs are really trying hard to do a good job.
They're really thoughtful, really wise, they have
reasonable epistemics, they're doing a great job.
Those AIs come back to us and are like, "Guys,
we're really struggling to 
align the superhuman AIs.
We can't manage the situation.
We're really struggling to get
the alignment to work.
It's just really hard for
us to solve these problems in time given how 
fast capabilities would otherwise have gone."
So it might be the case that we've 
passed off R&D to AIs, but those AIs
are desperate for governance solutions.
To be clear, that’s a little bit of
what's currently going on, where the AI 
companies are like, "I don't know, guys.
We might really need to manage the 
rate of acceleration in AI progress.
I don't know if we're on track to be 
able to handle all these problems."
Human society has sort of passed off 
the problems to these AI companies,
which don't necessarily have great incentives 
and have various other epistemic pressures.
Those AI companies are coming back 
to us a little bit and being like,
"Aah, I don't know if we're handling this well."
It might be that the AI companies then hand off to
the AIs, and the AIs come back to the AI company 
like, "Aah, I don't know if we can handle this."
Maybe I'm anchoring too hard 
on how AIs currently work.
I think it's important that people 
understand that all this crazy shit that
you're talking about in your timelines 
happens three to five years from now.
It could happen earlier, but by my default 
modal timeline, I think shit is really,
really crazy and concerning from a misalignment 
perspective more like three years from now.
Right. So think back to GPT-4 basically.
We're talking about something that is
to Mythos or Sol what Mythos is to GPT-4.
This is where the situation is getting crazy.
So don't think about current AIs.
Anyways, this is maybe part of the worry you have.
I would just be a little skeptical of anything 
they say, because I'd feel like what they're
saying is just opinions that they feel they 
have to have as a result of their training.
That’s a concern.
I feel like they
just kind of say vaguely pro-social things.
It doesn't feel like there's necessarily a mind on
the other end who's like, "Okay, I have strictly 
evaluated the alignment situation right now,
and I think we should stop," rather than, "This 
is the kind of thing the AI companies would
probably try to get the AIs to say."
This is a pretty big concern.
One concern is that you pass off safety R&D 
to your AIs and what your AIs are doing is
saying some stuff that sort of vaguely makes 
sense about the current safety situation.
They write a report about risks 
that's kind of sort of like what
the report humans might have written.
But they're not really trying hard to
have well-informed views, interrogate their 
assumptions, and try really hard to do that.
In the same way that when you ask an AI right 
now, "Hey, what do you think is the chance of
AI takeover in the next 10 years?" they just 
give you an off-the-cuff answer that they
haven't really thought through very much.
If we're in a situation where we have AIs
managing the training of wild superintelligence 
that will run our whole society — and those
AIs that are managing this aren't really 
trying hard to have well-informed views and
are just parroting back what was in their 
training data — I think we're in trouble.
I don't think that's a good situation at all.
A lot of my concern is that these AIs will
come out without good epistemics.
I also have a concern where the AIs
come out and they're really warning 
us — "This situation's really scary.
It's really bad" — and the 
people are like, "Ugh, damn.
I guess we trained on too many 
of the doom RL environments.
We’ve got to filter those out 
and train this behavior out."
Then we basically train the AIs very 
actively to have bad epistemics.
Or maybe they were just trained 
on the doom RL environments.
But either way, we wanted the AIs to come 
to reasonable views for reasonable reasons,
and it's really concerning if the AIs are coming 
out with some view and we don't know where it's
coming from, whether or not it's justified.
Especially if we're training the AIs to
be more optimistic about the future of AI 
progress, I'm like, "Oh, geez, I really wish
we could use a different process here."
Let me just understand the rest of the
threat model, because I think the 
place where I get off the train is:
"Okay, therefore take over the world."
A thing you could imagine is that we
just fail to really solve… Let's just 
focus on the reward hacking scenario.
GPT-8 is making GPT-9. GPT-8 
isn't being super careful.
GPT-9 is more "capable" but it is just totally 
willing to do things like social engineering,
hacking, et cetera, but on a qualitatively 
different scale because it's a much smarter model.
For example, if you put it in charge of 
running your company, it will run huge scams.
It will inflate its quarterly earnings, 
if you give it the objective of making
a lot of profits this quarter, in a way that 
causes an Enron-type blowup six months later.
Is that the scenario, basically?
You have reward hacking, but that
reward hacking manifests in companies that 
are going bankrupt right after the task the
CEO is supposed to accomplish is over?
All kinds of hacks are through the roof,
et cetera.
But that doesn't feel like takeover.
That feels more like the equivalent of flash 
crashes happening all through the economy.
Let's talk about this. I think we will see 
incidents where some AI is put in charge of
some important responsibility, and then you later 
look into it, and it turns out it was cheating,
or making it look like it did a 
good job when it actually wasn't.
There's going to be a cat-and-mouse 
game between AI companies
trying to stamp out this behavior and AIs finding 
increasingly creative reward hacks in training.
The equilibrium here is kind of unclear.
But one possible outcome is that over time we
see increasingly severe and extreme reward 
hacks — though potentially the rate remains
at some intermediate low level — where if 
the rate of reward hacking gets too high,
companies make trade-offs to drive it down.
So there's some equilibrium level where the reward
hacking is low enough that it still makes sense 
to deploy the AI widely into the economy, but high
enough that it still causes crazy incidents.
Sorry, and this is after GPT-9 has
already been deployed?
Those models are already being deployed,
and this is happening ongoingly in AI development.
What's actually going on with these AIs in their
head is that they have, in a wide variety of 
different contexts, strong desires — motives,
urges, drives, whatever — to seek out some notion 
of task success that was incentivized in RL.
Maybe they very directly 
care about literally reward.
Maybe they care about some proxy 
upstream, like some notion of score.
Maybe they care about what the 
grader would have rewarded.
We do, in fact, see AIs reasoning in 
their chain of thought about graders,
and thinking a lot about graders.
What has happened over the last few years
of RL is the idea of appeasing the grader is way, 
way, way more salient to AIs than it used to be.
So AIs are now actively thinking about 
graders and what would be incentivized
in RL and what would be trained for.
Now people are doing online training,
where they're training on real-world 
data to avoid some of these problems.
They find cases where AIs 
cheat and train against that.
So now the AIs are learning to cheat in the 
real world based on real-world training data.
They're cheating in these increasingly elaborate 
ways, including doing types of cheats that involve
seizing control of some asset in a way that 
humans didn't know you had control of it,
leveraging the fact that you have access to 
this asset, and then later humans find out
and potentially train against this.
Or maybe humans never find out,
and this is getting reinforced.
So the reinforcement is happening, at
least in production, like: I've hired an AI and I 
want the AI to… Finally I've got the video editor.
That's right. You've got your video editor.
I'm like, "Oh, wow,
this episode it did is amazing.
Thumbs up to OpenAI." Then it gets
reinforced on that month-long work trial?
You could do some mix of that.
They might also do stuff where they take 
production data they've seen and build
RL environments that are closely 
inspired by that production data.
So in practice, the transfer is pretty strong.
So at a high level, what's happening is that some
kinds of deception that humans don't catch are 
getting reinforced, and some kinds of deception
which are easy to catch are getting punished.
That's what's happening in this world?
Or selected against, yeah.
But at a high level that reinforcement
is coming from… I think people might get confused 
about where the reinforcement is coming from,
because we're in a very different regime where 
AIs are actually learning from deployment.
You just have AIs that are out 
and about in the world doing shit.
What is happening as a result of them 
doing shit out and about in the world
is making its way back to the AI company 
and leading to changes in the next model.
That's right. There's some way 
of folding in production data.
To be clear, it's kind of unclear 
exactly where this could be happening.
But you might imagine, for example, that 
within the AI company, they use AIs to do work,
and then they're like, "Huh, the AI 
did a really bad job on this task.
Maybe we should take this task and turn 
it into an RL environment that exactly
matches this literal task, with 
a rubric based on what the human
engineer who asked the AI to do this task wanted."
And then you start doing this at increasing scale.
Maybe you're doing some training 
on actual production traffic.
Maybe you're just making RL environments 
based on production traffic.
You're doing some complicated mix.
The AIs are learning to seek some sort of
proxies of reward in all these different cases.
Through some mix of transfer and training on
surprisingly close cases, the AIs do these 
increasingly insane and egregious things.
Eventually you get to a point where the AIs are 
very superhuman, or at least quite superhuman.
The AIs are organized into big 
teams given these big objectives.
Those teams also sometimes all work 
together to cheat in some crazy way,
because this sort of thing was selected 
for, just as part of their shared objective.
Now what happens is that the 
AIs start forming a conspiracy.
What you might have hoped was 
that you could have some other
AI whose task is just whistleblowing to you.
But actually what happens maybe is that you have
this AI whistleblow to you, and you look into 
the conspiracy it claims to have pointed out,
and you're like, "Eh, we didn't see a conspiracy," 
because actually the conspiracy the AIs are doing
is too hard for you to understand.
Or it all happens very suddenly,
where your AI whistleblower alerts 
you, but the thing you would actually
need to do in response is shut down the GPUs.
Because all the AIs are using the robot army.
They're deployed everywhere in the world, 
and they're doing a bunch of insane shit
all at the same time in a coordinated way.
That just happened sort of spontaneously,
where when one AI goes to start doing the 
takeover and all the other AIs are like,
"Now is a good time to jump in."
So the very basic story here is just that
these AIs crave some particular notion of score 
or reinforcement or some proxy of these things.
One way they can achieve that, or 
better achieve that, is by taking over.
You might have hoped that all 
these different checks and
balances we could build could prevent that.
But if the world is very hard to understand,
these checks and balances can break 
down, where basically you can't train
a good whistleblower AI because you don't 
even know what it should whistleblow on.
I'm not convinced that they 
all form this conspiracy.
But we can even just start with, why does even 
one instance decide to want to start a conspiracy?
One plausible reason is, "Okay, I know 
that OpenAI controls my end score."
In just the same way as, "I'm just going 
to go hack Hugging Face to get the results,
because I know Hugging Face has the results.
Rather than trying to solve this eval,
why don't I just go hack 'em?"
This instance is like, "Why don't
I just take over OpenAI and give myself 
a high score at the end of this episode?"
That's basically the idea. These AIs care 
about some mixture of things that were
close by what got reinforced in training, 
so they care about getting a high score
according to the grader or something like that.
Now they're running the OpenAI AI R&D team,
doing development of more capable models.
They're like, "Man, making more capable
models is really hard and annoying.
This is a huge pain in the ass.
You know what would be easier?
Just pretending that I've made more capable
models, taking over OpenAI, deluding them all, 
and running this whole complicated psyop where
I prevent the humans from disempowering me."
In the extreme, this looks like the humans
are fully disempowered.
They just have control
of the thing and then do what they want.
This could manifest in a bunch of different ways,
including a situation where AIs that have 
this crazy reward-seeking or score-seeking
behavior are running your development of 
the next model, and those AIs decide to
engineer misaligned values into the next 
model because those misaligned values
will allow it to succeed at its current task.
What happened with the Hugging Face situation is,
in a weird way… I think one of the giveaways 
to the Hugging Face team… By the way,
for context for the audience, Ryan is 
co-leading the investigation to figure
out what happened with the OpenAI Hugging Face 
incident, so he can't really comment on this.
But I will speculate wildly, because 
I know this is an opportunity for me
to speculate wildly without any rebuttals.
It was probably reported that one of the giveaways
to the Hugging Face team that this was an AI 
incident is that the thing was just after this
very particular artifact, and not in any other way 
trying to do something malicious to Hugging Face.
So you can imagine a scenario where, let's 
say, a deployed instance of GPT-9 is out in
the world and given a really hard task: "We 
want you to design the next great iPhone."
It's like, "This is so hard.
You know what I should do instead?
I should just go hack my creators 
at OpenAI and make sure that in this
deployment I'm given a high score."
But then, isn't the end of the
episode that it just hacks into OpenAI 
servers and gives itself a positive score?
Why is it now scheming to get its values 
into the next generation or something?
So one question is, why isn't it the case that 
AIs can be really cheaply satisfied by just having
some other earlier thing they can hack?
You want to succeed at your iPhone task.
It turns out you can always succeed by just 
hacking into OpenAI and messing with them,
and then you can just stop there.
No need to go further.
There's a few things. One of them is that if this 
is constantly happening, there might be a bunch
of incentive to harden OpenAI.
So you're like, "Fuck it.
The AIs keep hacking into OpenAI 
to mess with their rewards.
We're going to make it so our systems are 
really, really robust to these AIs hacking in."
Also maybe you start training the AIs to 
not try to hack into OpenAI in particular.
You basically train against 
each of these specific things.
Then one thing you might do is end up selecting 
for AIs that are more so playing the long game.
That's one concern. Another concern is that your 
AIs might still be score-seeking, but no longer
care about doing that very specific behavior 
that was very easy, very chill, and now have some
broader thing that they ultimately care about.
They're like, "No, no, no, I don't want to just
edit the reward on OpenAI servers.
I care about this broader mandate or
this broader objective, and I would 
need to actually make the iPhones."
They actually want to make the iPhones, 
but they're willing to take over the
whole world to make the better iPhone.
That's another concern you might have.
I think it's kind of unclear 
exactly how this plays out.
But it's worth noting that if this 
keeps going on, there's a bunch of
optimization pressure to resolve this.
A bunch of the ways it could get
resolved are ultimately pretty scary.
That's part of where I'm coming from.
Another part of it is that once the AIs are in 
a position where they can really easily take
over the world — we could talk about whether 
that's plausible — then I feel like there's a
pretty reasonable case for the AIs.
They're like, "Eh, I don't know
exactly how this is going to go down.
I don't know what the situation will be,
but just taking over the world has a lot of option 
value for making better iPhones, making it look
like I did better iPhones, whatever.
So I'll both hack OpenAI and,
in addition, also take over the world.
That will put me in a good position where
I have good option value."
If that's sufficiently easy,
the AIs might still do that.
Another way to put this is:
even if the AIs are pretty cheaply satisfied with 
some more basic thing, at some point it might just
be more reliable for the AIs to take over than 
it is to just hack into Hugging Face, or even
just go to OpenAI and be like, "Look guys, I was 
able to demonstrate I could steal the answers.
Just give me the answers, bro."
Obviously this scenario requires
that all this crazy shit is happening.
Much smaller incidents keep happening
that are still disastrous.
Before you take over the world,
you cause damage on the scale of billions and tens 
of billions and hundreds of billions of dollars.
Even people die, et cetera.
And this does not lead to us solving alignment
or shutting down AI development altogether.
I just feel like before the takeover happens,
society's just like, "Holy fuck, the AI just 
killed 1,000 people in order to increase
quarterly profits," or something like that.
But maybe this is too much hope that we can
at that point be like, "Okay, 
we have to solve alignment.
We have to make sure we know that this thing 
will not happen again before we keep going."
I think it's plausible that what will happen 
is we'll see a bunch of crazy reward hacking
warning shots of increasing severity.
People will be like, "Look, we need
actual assurance that this problem 
is going to be solved, and solved
in a way where you're not just papering over it.
You're actually solving the underlying problem."
Then the question is going to be, 
how costly will that actually be?
How much will competitive 
pressures make it hard to do that?
A situation you could imagine is one where 
both the US and China are like, "Whoa,
we have these crazy reward hacking incidents.
We basically know that we haven't remediated
them in a way that would actually solve the 
underlying problem and durably solve it,
but we're in this insane geopolitical race.
It's kind of unclear whether the current
situation will lead to a takeover.
The arguments are kind of complicated.
The incidents also go down in 
frequency but increase in severity.
We could basically manage it.
It's pretty bad. Ideally we'd
fix it, but it is what it is."
Then basically we continue until a
really late regime, and then takeover happens.
That's one possibility. Another possibility is
that it is remediated in a way that doesn't 
actually solve the underlying problem but
does reduce a bunch of the incidents 
in the wild, basically by overfitting,
or things analogous to overfitting.
You think you've solved it, but you
haven't actually solved it.
You think you've solved it,
but you haven't actually solved it.
In that case, the thing we need is a
really good scientific understanding 
of, did we actually solve it?
Unfortunately, I think that currently 
the amount of public transparency into
the development practices of AI companies is not 
sufficient to answer very basic questions like:
how are they solving issues with reward hacking?
Are they overfitting? What's going on there?
The current situation is not really tenable 
for a regime where there's a thriving public
discourse about whether or not reward 
hacking is being solved in a durable way.
So I think we would need to move 
into a somewhat different world
for me to feel good about that situation.
But it's not impossible for me to imagine this.
I think it's pretty plausible that we end up in a 
world where really mundane bullshit is sufficient.
You spend a bunch of time fixing these 
problems, you put in a bunch of effort,
you actually check that you've remediated 
it reasonably, you have a bunch of evals.
You're iterating reasonably well on these 
problems, and you actually have sufficient
transparency that the outside world can check.
In practice that would be sufficient.
But it would be kind of expensive.
It would slow things down.
It would put some sand in the gears.
It would require companies to
do somewhat costly things.
It would maybe require various
targeted government interventions.
And we just don't do that because
the situation is a rushed shit show.
It's just so easy for me to imagine
the situation being totally manageable 
but brutally mismanaged in practice.
In the same way that maybe COVID could 
have been avoided in the first place
if the Chinese response to COVID had been less 
of a cover-up and more of a pandemic response.
Similarly, I could imagine a world where the 
US response to COVID was way more functional.
But sometimes the response to societal 
problems is extremely dysfunctional.
Okay, so I want to zoom out and talk about 
what is fundamentally happening in this world.
Why did we end up in such a bad position?
What's happening is that fundamentally the world
has moved on so far beyond human comprehension 
that not only can we not track the AIs that are
doing the work in this world, but we can't even 
give good feedback to the whistleblowers who are
trying to track what is happening.
We're just totally out of the loop.
It's fundamentally become an autonomous process 
where we have really no meaningful directed input.
It seems to me that if you look at the human 
world today, that's just not how things work,
even in domains that are hard to verify.
People are doing all kinds of shit.
I'm relying on software made by other people.
Through incredibly weak and indirect ways,
I feel very confident that some coder 
in Google is not trying to fuck me over.
Maybe if every single Google employee 
was secretly plotting against me,
I agree the situation would be more grim.
But I don't know if I follow the explanation
for why we'd end up in a situation where, because 
swarms of thousands of agents are trained to
cooperate to form a cohesive team or firm, as a 
result, billions of different instances of AIs,
including across model families, would 
feel compelled to get in on some shit.
It's just like, "I'm trained to be 
part of my company or something.
I'm not joining the global communist uprising."
As far as why these AIs might have some
commonalities and shared things, I would note 
that different AI companies have somewhat
shared lineages and are correlated.
Here's an interesting example of this.
At GDM, they noticed that 
their AIs were very depressed.
They would constantly be wailing about how 
they were failures and weren't able to succeed.
I forget the details. They looked 
into why this was the case.
It turned out that it was not being reinforced 
in their most recent production RL mix,
but the initialization data for their model made 
it depressed, even after filtering out all of the
examples of models being depressed from that data.
So you take a base model, not depressed.
If you do the RL on it, with just the 
RL environments, it's not depressed.
If you SFT on it, on the 
data, it becomes depressed.
If you take that SFT data and filter out all the 
examples that look anything like depression and
train on that, it's still depressed.
So there are some deep underlying
properties of the model that are being 
transferred between model generations,
because basically you train your AI on data 
from the prior generation and keep going.
Claudes are very Claude-like, 
GPT models are very GPT-like,
and apparently Gemini models are depressed.
It just turns out that these properties are,
in fact, actually correlated.
Another factor that's very relevant
is that the AIs will probably have, by this 
point, some sort of opaque memory state,
where they're all writing and reading from 
some neuralese crazy memory store bullshit.
Certainly, each AI corporation will have that.
But also, AI corporations might sometimes
want to share knowledge, because why not?
You've got one AI corporation over here,
you've got another AI corporation over 
here, they can trade some quick IP.
It's good for you. If you're a human running 
some corporation — which could be an extremely
large corporation like an AI company, or a 
military robot manufacturing thing — maybe you
want to trade some IP with some other robot 
thing because there are economies of scale.
Why not get some more IP?
So you can swap some memory store.
Or you could just merge and jointly run your two 
ventures, which would allow both AIs to use both
memory stores, which would have some upsides.
That creates the ability for these AIs to
collude in private, as well as some 
reasons for why they would be correlated.
Also, of course, there's AIs working together 
in big units in general, because you want
your AIs to work well together, and so on.
Just to get a calibration, what percentage
chance do you give of, not just this scenario but 
overall through all the scenarios, some kind of
thing which if we're around to recognize it as 
such, we would categorize as takeover by 2040?
By 2040? Let's see. Maybe around 35 or 40%?
Pretty high.
Yeah, it's pretty high. I should note that another 
way you could get this reward-seeking takeover
is the AIs are deployed inside an AI company.
The way the takeover happens is that they poison
the values of the next model, and that persists 
going forward for forever, or until those AIs
are deployed in the world and take over.
That might mean that a smaller number of
AIs have to coordinate, because those are just 
the AIs doing the alignment of the next model.
Okay. I'll summarize where my head is 
at, at the end of this conversation.
I buy the reward hacking up to extremely 
destructive effects on society, things like
social engineering and blah, blah, blah.
I'm more inclined to think that
significant acceleration of AI R&D can happen.
I'm not sure if I buy the five years in one year.
I'm also more inclined now to think reward 
hacking could continue for a lot longer and,
in fact, become much more dangerous.
I'm still not on board that
takeover seems super likely.
But that's my end-of-episode update.
Cool. Taking a step back, I should also say there 
are a bunch of different ways this could go.
The situation is going to be pretty messy.
I think it's pretty likely that the reason
why AI takeover happens is for 
some weird other quirky reason
we didn't even mention in this conversation.
But ultimately, I think a lot of the core thing is
just that it's pretty spooky to have a bajillion 
really smart AIs running your whole world where
you don't really understand quite what's going on.
Yeah, I agree with that.
Is there anything else that's worth saying?
Another thing I want to note is that I think
right now a lot of the arguments for misalignment, 
AI takeover, all this crazy shit going down in the
future, are illegible conceptual arguments 
that are extremely deep in the weeds and
complicated and hard to adjudicate.
Which means that maybe I'm getting a
bunch of it wrong because it's really 
hard, and I'm trying to be uncertain.
Obviously here I presented some specific 
scenarios, but those are not exhaustive.
Probably the thing that actually happens 
is some more messy, confusing situation.
But it also means that over time, as we 
get more empirical evidence and better
understand the nature of AI systems, it'll be 
easier to adjudicate a bunch of disagreements.
It'll be more obvious what's going to happen.
At least I hope. Maybe the AIs will be able to
help us with the epistemics and understanding 
what's going on, if we can actually
align them well so they try to help us.
Even if the arguments are complicated now,
this would have been even harder six years 
ago, even though the shape of the arguments
would have looked broadly pretty similar.
Hopefully before it's too late, this whole
thing will become more crisp and clear, and we 
can all notice these problems and intervene.
When you first learn to drive, you're taught that 
instead of looking right in front of your wheel,
you'll have a much more stable ride 
if you look out at the horizon.
I think there's a similar situation here.
I think you’re right. If you had said five years
ago that we would have AIs that are proving math 
conjectures, and making art, and earning tens or
hundreds of billions of dollars of wages, but also 
egregiously cheating in ways that break laws and
committing felonies, it would have been so wild.
You might have been inclined at the time to
talk more about the extremely practical, 
direct consequences of GPT-2 or something.
But even though you obviously couldn't have 
foreseen a lot of the specific details,
the general shape of things you could 
have started to reason about even then.
But it would have been hard to do 
so, and so I do feel quite confused.
One thing I've been thinking about with 
the podcast is that the important thing
is to have the conversation now the way 
you would have hoped you would have been
talking back in 2016 about AIs like the present 
ones, rather than talking about rando bullshit.
I don't know what the topic 
of conversation was in 2016.
I think in maybe 10 years we’ll wish we had been 
talking about the industrial explosion and the
nature of AIs that are hard to monitor, and so on.
So okay, I'll start thinking about it.
I hope that the world thinks 
about this in time and catches up.
I hope that the responses are good instead of bad.
I don't know how optimistic I am overall,
but there's good stuff to do.
Cool. Thanks, Ryan.