Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken

2h 24m

New episode with my good friends Sholto Douglas & Trenton Bricken. Sholto focuses on scaling RL and Trenton researches mechanistic interpretability, both at Anthropic.

We talk through what’s changed in the last year of AI research; the new RL regime and how far it can scale; how to trace a model’s thoughts; and how countries, workers, and students should prepare for AGI.

See you next year for v3. Here’s last year’s episode, btw. Enjoy!

Watch on YouTube; listen on Apple Podcasts or Spotify.

----------

SPONSORS

* WorkOS ensures that AI companies like OpenAI and Anthropic don't have to spend engineering time building enterprise features like access controls or SSO. It’s not that they don't need these features; it's just that WorkOS gives them battle-tested APIs that they can use for auth, provisioning, and more. Start building today at workos.com.

* Scale is building the infrastructure for safer, smarter AI. Scale’s Data Foundry gives major AI labs access to high-quality data to fuel post-training, while their public leaderboards help assess model capabilities. They also just released Scale Evaluation, a new tool that diagnoses model limitations. If you’re an AI researcher or engineer, learn how Scale can help you push the frontier at scale.com/dwarkesh.

* Lighthouse is THE fastest immigration solution for the technology industry. They specialize in expert visas like the O-1A and EB-1A, and they’ve already helped companies like Cursor, Notion, and Replit navigate U.S. immigration. Explore which visa is right for you at lighthousehq.com/ref/Dwarkesh.

To sponsor a future episode, visit dwarkesh.com/advertise.

----------

TIMESTAMPS

(00:00:00) – How far can RL scale?

(00:16:27) – Is continual learning a key bottleneck?

(00:31:59) – Model self-awareness

(00:50:32) – Taste and slop

(01:00:51) – How soon to fully autonomous agents?

(01:15:17) – Neuralese

(01:18:55) – Inference compute will bottleneck AGI

(01:23:01) – DeepSeek algorithmic improvements

(01:37:42) – Why are LLMs ‘baby AGI’ but not AlphaZero?

(01:45:38) – Mech interp

(01:56:15) – How countries should prepare for AGI

(02:10:26) – Automating white collar work

(02:15:35) – Advice for students



Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe

Listen and follow along

Transcript

Okay, I'm joined again by my friends, Sholto Bricken.

Wait, fuck.

Did I do this last year?

No, no, no, you named us differently, but we didn't have Sholto Bricken and Trenton Douglas.

Shulto Douglas and Trenton Bricken,

who are now both at Anthropic.

Let's go.

Sholto is scaling RL.

Trenton is still working on mechanistic interpretability.

Welcome back.

Happy Happy to be here.

Yeah, it's fun.

What's changed since last year?

We talked basically this month in 2024.

Yep, now we're in 2025.

What's happened?

Okay, so I think the biggest thing that's changed is RL and language models has finally worked.

And this is manifested in

we finally have proof.

of an algorithm that can give us expert human reliability and performance given the right feedback loop.

And so I think this has only really been conclusively demonstrated in competitive programming and math, basically.

And so if you think of these two axes, one is the intellectual complexity of the task, and the other is the time horizon of which the task is being completed on.

And I think we have proof that we can reach the peaks of intellectual complexity along many dimensions.

We haven't yet demonstrated long-running agentic performance.

And you're seeing the first stumbling steps of that now and should see much more conclusive evidence of that basically by the end of the year.

with like real software engineering agents doing real work.

And I think, Trenton, you're experimenting with this at the moment.

Yeah, absolutely.

I mean, the most public example people could go to today is Cloud Plays Pokemon.

Right.

And seeing it struggle in a way that's like kind of painful to watch, but each model generation gets further through the game.

And it seems more like a limitation of it being able to use memory system than anything else.

Yeah.

I wish we had recorded predictions last year.

We definitely should this year.

Oh, yeah.

Hold us accountable.

Yeah, that's right.

Would you have said that agents would be only this powerful as of last year?

I think this is roughly on track for where I expected with software engineering.

I think I expected them to be a little bit better at computer use.

But I understand all the reasons for why that is.

And I think that's like well on track to be solved.

It's just like a sort of temporary

lapse.

And holding me accountable for no predictions next year, I really do think end of this year, sort of like this or this time next year, we have software engineering agents that can do close to a day's worth of work

for like a junior engineer or like a couple of hours of like quite competent and independent work.

Yeah, that seems right to me.

I think the distribution is pretty wonky though.

Yes.

Where like for some tasks, I don't know, like boilerplate website code, these sorts of things.

It can bang it out and save you a whole day.

Exactly.

Yeah, I think that's right.

I think last year you said that the thing that was holding them back was the extra nines of reliability.

I don't know if that's the way you would still describe the way in which the software agents aren't able to do a full day of work, but are able to help you out with a couple minutes.

Is it the extra nines that's really stopping you or is it something else?

Yeah, I think my description there was, I think, in retrospect, probably not what's limiting them.

I think what we're seeing now is closer to lack of context,

lack of ability to

do complex, very multi-file changes and

maybe like scope of the change or scope of the task in some respects.

They can cope with high intellectual complexity in a focused context with a very scoped problem.

But when something's a bit more amorphous or requires a lot of discovery and iteration with the environment, this kind of stuff, they struggle more.

And so maybe

the way I would define it now is the thing that's holding them back is if you can give it a good feedback loop for the thing that you want it to do, then it's good at it.

It's pretty good at it.

If you can't, then they struggle a bit.

And then for the audience, can you say more about what you mean by this feedback loop

if they're not aware of what's happening in RL and so forth?

Yes.

So the big thing that really worked over the last year is

maybe like broadly the domain is called RL from verifiable rewards or something like this, where a clean reward signal.

So the initial unhopping of language models was RL from human feedback, where typically it was something like pairwise feedback or something like this, and the outputs of the models became closer and closer to things that humans wanted.

But this doesn't necessarily improve their performance at any

difficulty of problem domain, right?

Particularly as humans are actually quite bad judges of what a better answer is.

Humans have things like length biases and so forth.

So you need a signal of whether the model was correct in its output

that is quite true, let's say.

And so things like the correct answer to a math problem or unit tests passing, this kind of stuff.

These are the examples of reward signal that's very clean.

But even these can be hacked, by the way.

Like even unit tests, the models find ways around it to like hack in particular values and hard code values of unit tests if they can figure out like what the actual test is doing.

Like if they can look at the cached Python files and find what the actual test is, they'll try and hack their way around it.

So these aren't perfect, but they're much closer.

And why has it gotten so much better at software engineering than everything else?

In part because software engineering is

very verifiable.

Like it's a domain which just naturally lends it to this way.

Does the code pass a test?

Does it even run?

Does it compile?

Yeah, does it compile?

Does it pass the test?

You know, you can go on Lee Code and you can run like tests and like you know whether or not you got the right answer.

But there isn't the same kind of thing for like writing a great essay.

That requires

like the question of like taste in that regard is quite hard.

Like we discussed the other night at dinner the Pulitzer Prize.

which would come first, like a Pulitzer Prize winning novel or like, you know, a Nobel Prize or something like this.

And I actually think a Nobel Prize is more likely than a Pulitzer Prize winning novel in some respects.

Because a lot of the tasks required in winning a Nobel Prize, or at least strongly assisting in helping

to win a Nobel Prize, have

more layers of verifiability built up.

So I expect them to accelerate the process of doing Nobel Prize winning work more initially than that of writing Pulitzer Prize-worthy novels.

Yeah, I think if we rewind 14 months to when we recorded last time, the nines of reliability was right to me.

Like, we didn't have clawed code.

We didn't have deep research.

All we did was use agents in a chat bot format.

Right.

Copy, paste, copy, paste, copy, paste.

Totally.

And I think we're very used to chat interfaces, whether we're texting or using Google.

And it's weird to think that the agent can actually go and fetch its own context and store its own facts into its memory system.

And I still think that it's the nines of reliability.

And if you scaffold the model correctly or prompt it, it can do much more sophisticated things than the average user assumes.

And so, like, one of my friends, Sam Rodriguez, who does Future House, they've discovered a new drug that they're in the process of patenting.

And by the time this episode comes out,

that will be live.

LSD V2?

Is it really?

No,

no.

No, they're not making LSD2.

But

like, people didn't think that models can be creative or do new science.

Right.

And it does just kind of seem like a skill issue.

I mean, there was the cool.

Wait, wait, wait, wait, like the discovered a drug.

Is it like, how did it like?

Like, I think it one-shotted the

conversation, and so we'll need to refer to the full announcement.

But my impression is that

it was able to read a huge amount of medical literature

and brainstorm and make new connections and then propose wet lab experiments that the humans did.

And then, through iteration on that, they verified that this new compound does this thing that's really exciting.

Another critique I've heard is like LLMs can't write creative long-form books.

And I'm aware of at least two individuals who probably want to remain anonymous who have used LLMs to write long-form books.

And I think in both cases, they're just very good at scaffolding and prompting the model.

I mean, even with the viral ChatGPT geo-guesser capabilities, where it's just insanely good at spotting what beach you were on from a photo.

Kelsey Piper, who I think made this viral,

their prompt is so sophisticated.

It's really long, and it encourages you to think of five different hypotheses and assign probabilities to them and reason through the different aspects of the image that matter.

And I haven't A-B tested it, but I think unless you really encourage the model to be this thoughtful, you wouldn't get the level of performance that you see with that ability.

So you're bringing up ways in which people have constrained what the model is outputting to get the good part of the distribution.

But one of the critiques I've heard of RL, or the not of RL, but one of the critiques I've heard about using the success of models like O3 to suggest that we're getting new capabilities from these reasoning models is that all these capabilities were already baked in the pre-training model.

I think there was a paper from Singh Shua University where they showed that if you give a base model enough tries

to answer a question, it can still answer the question as well as the reasoning model.

Basically, it just has a lower probability of answering.

So you're narrowing down

the possibilities that the model explores when it's answering a question.

So are we actually eliciting new capabilities with this RL training or are we just like

putting the blinders on them?

Right, like carving away the marbles on this.

I think it's worth noting that that paper was, I'm pretty sure, on the Lama and Quenn models.

And I'm not sure how much RL compute they used, but I don't think it was anywhere comparable to the amount of compute that was used in the base models.

And so I think the amount of compute that you use in training is a decent proxy for the amount of actual raw new knowledge or capabilities you're adding.

to a model.

So my prior at least, if you look at all of DeepMinds research from RL before, RL was able to teach these Go and chess playing agents

new knowledge in excess of human-level performance just from RL signal, provided the RL signal is sufficiently clean.

So there's nothing structurally limiting about the algorithm here that prevents it from imbuing the neural net with new knowledge.

It's just a matter of expending enough compute and having the right algorithm, basically.

Why aren't you already spending more compute on this?

I think Dario said in his blog post that labs, or it was like a couple months ago on the Xbox Controls thing, is like, ah, DeepSeek, whatever, we're only spending $1 million on RL or something.

So it's like,

we aren't in the compute-limited regime for RL yet, but we will be soon.

Yeah.

You're spending hundreds of millions on the base model.

Why only order a million on the RL?

You know the parable about when you choose to launch a space mission and how you should sort of acquire, like go further up the tech tree, because if you launch later on, you're like, your ship will go faster and this kind of stuff.

I think it's quite similar to that.

You want to be sure that you've algorithmically got the right thing.

And then when you bet and you do the large compute spend on the run, then

it'll actually pay off.

It'll have the right compute efficiencies and this kind of stuff.

And I think RL is slightly different to pre-training in this regard, where RL can be a more iterative thing, you're progressively adding capabilities to the base model.

Pre-training has,

in many respects, if you're halfway through a run and you've messed it up, then you've really messed it up.

I think that's the main reason why is people were still figuring out exactly what they wanted to do.

O1 to 03, right?

Like OpenAI put in their blog post that it was a 10x compute multiplier over 01.

So like clearly they

bet on

one level of compute and they were like, okay, this seems good.

Let's actually release it.

Let's get it out there.

And then they spent the next few months like, you know, increasing the amount of compute that they expend on that.

And I expect, as everyone is, everyone else is scaling up RL right now.

So I basically don't expect that to be true for very long.

Yeah, just for the sake of listeners, maybe

you're doing gradient descent steps in both pre-training and reinforcement learning.

It's just the signal is different.

Typically, in reinforcement learning, your reward is sparser.

So you take multiple turns.

It's like, did you win the chess game or not?

Is the only signal you're getting?

And often you can't compute gradients through discrete actions.

And so you end up losing a lot of gradient signal.

And so you can presume that pre-training is more efficient, but there's no reason why you couldn't learn new abilities in reinforcement learning.

In fact, you could replace the whole next token prediction task in pre-training with some weird RL variant of it

and then do all of your learning with RL.

Yeah, at the end of the day, just signal and then correcting to it.

Totally.

And then, and then going back to the paper you mentioned, aside from the caveats that Shalto brings up, which I think is the first order most important, I think zeroing in on the probability space of meaningful actions comes back to the nines of reliability.

And

classically, if you give monkeys a typewriter, eventually they'll write Shakespeare, right?

And so the action space for any of these real-world tasks that we care about is so large that you really do care about getting the model to zero in on doing the reasonable things.

And to the extent, like in some broad sense, like to the extent that

at some pass of K, you've got token space.

Right, exactly.

You literally do have a monkey and it's making Shakespeare.

Yeah, yeah, exactly.

Okay, so the alpha, the chess analogy is interesting.

So were you about to say something?

I was just going to say, like, you do need to be able to get reward sometimes in order to learn.

And that's like the complexity in some respects.

In the alpha variants, maybe you're about to say this, one player always wins.

So you always get a reward signal one way or the other.

But in the kinds of things we're talking about, you need to actually succeed at your task sometimes.

So language models, luckily, have this wonderful prior over the tasks that we care about.

So if you look at all the old papers from 2017, it's not that old, but the papers from 2017,

the learning curves curves always look like flat, flat, flat, flat, flat, as they're like figuring out sort of like basic mechanics of the world.

And then there's this like spike up as they learn to exploit like easy rewards.

And then it like, it's sort of like, it's almost like a sigmoid in some in some respects.

And then like sort of continues on indefinitely as it like just learns to like absolutely maximize the game.

And I think the LLM curves look a bit different in that there isn't that.

dead zone at the beginning because they already know how to solve some of the basic tasks.

And so you get this like initial spike.

And that's what people are talking about when they're like, oh, you can learn from one example.

That one example is just teaching you to pull out the backtracking and formatting your answer correctly and this kind of stuff that lets you get some reward initially at tasks, conditional on your pre-training knowledge.

And then the rest probably is you learning more and more complex stuff.

Yeah, yeah.

And it would also be interesting.

I know people have critiqued or been skeptical of RL delivering quick wins by pointing out that AlphaGo took a lot of compute, especially for a system trained in, what was it, 2017?

Yes,

off the curve.

Totally.

So to the extent that that was largely because first you had to like have something which had like some biases which were sort of rational before it like got like superhuman a go.

I actually would be interesting to see like what fraction of the compute used on Afghan go was just like getting something reasonable.

Yes.

Yeah.

Yeah.

It would be interesting.

Yeah.

I mean to make the map from pre-training to RL really explicit here.

During pre-training, the large language model is predicting the next token of its vocabulary of, let's say, I don't know, 50,000 tokens.

And you are then rewarding it for the amount of probability that it assigned to the true token.

And so you could think of it as a reward,

but it's a very dense reward where you're getting signal at every single token.

And you're always getting some signal.

Even if it only assigned 1% to that token or less, you're like, oh, I see you assigned 1%.

Good job.

Keep doing that.

Upweight it.

Yeah, exactly.

It's like a tug in the grain.

That's right.

So when I think about the way humans learn,

it seems like these models getting no signal from failure is quite different from if you try to do a math problem and you fail, it's actually even more useful often than like learning about math in the abstracts because, oh, you don't think so?

Only if you get feedback.

Yeah.

If you get feedback.

But I think there's a way in which you actually give yourself feedback.

You fail and you notice where you failed.

Only if you get feedback at times.

Or don't think so.

And people have figured out new math, right?

And they've done it by the fact that they get stuck somewhere.

They're like, why am I getting stuck here?

Let me think through this.

Whereas in the example, I mean, I'm not aware of what's at the frontier, but looking at open source implementations from DeepSeek or something, there's not this conscious process by which

once you have failed, you learn from the particular way in which you failed.

to then like backtrack and do your next things better.

It's just like pure gradient descent.

And I wonder if that's a big limitation.

I don't know.

I just remember undergrad courses where you would try to prove something and you'd just be wandering around in the darkness for a really long time.

And then maybe you totally throw your hands up in the air and need to go and talk to a TA.

And it's only when you talk to a TA can you see where along the path of different solutions you were incorrect and like what the correct thing to have done would have been.

And that's in the case where you know what the final answer is, right?

In other cases, if you're just kind of shooting blind and meant to give an answer de novo,

it's really hard to learn anything.

I guess I'm trying to map on again to the human example where like, in more simpler terms,

there is this sort of conscious intermediary, like, auxiliary loss that we're like optimizing.

And it's like a very sort of like self-conscious process

of getting, forget about math.

It's just like if you're on your job, you're getting like, you're getting very explicit feedback from your boss.

That's not necessarily how you, the task should be done differently, but like a high-level explanation of what you did wrong, which you like update on, not in the way that pre-training updates weights, but more in the, I don't know.

But I think there's a lot of implicit dense reward signals here.

Yeah, exactly.

Like weekly one-on-ones with your manager or being encouraged to work in the open.

Or like even with homework assignments, right?

They're so scaffolded.

Right.

It's always 10 questions broken down into subcomponents.

Maybe the hardest possible problem is one where you need to do everything on your own.

Yeah.

Okay.

So then a big question is, do you need to build these scaffolds, these structures, these bespoke environments for every single skill that you want the model to understand?

And then it's going to be a a decade of grinding through these subskills or is there some more general procedure for learning new skills using RO?

Yeah, so

it's an efficiency question there.

Like obviously if you could give a dense reward for every token, right?

Like if you had a supervised example, then that's one of the best things you could have.

But in many cases, it's very expensive to produce all of those scaffolded curriculum of like everything to do.

Like having PhD math students grade students is something like which you can only afford for the like select category of students that you've chosen to like you know focus in on developing and you couldn't do that for all the language models in the world.

So

like first step is

obviously that would be better

but

you're gonna like be sort of optimizing this period frontier of like how much am I willing to spend on like the

the scaffolding versus how much am I willing to spend on pure compute.

Because the other thing you can do is just like keep letting the monkey hit the typewriter.

And if you have a good enough end reward, then

eventually it will find its way.

And so

I can't really talk about where sort of exactly people sit on that scaffold.

I think different people, different tasks are on different points there.

And a lot of it depends on how strong your prior over the correct things to do is.

But that's the equation you're optimizing.

It's like, how much am I willing to burn compute versus how much am I willing to burn like dollars on people's time to give scaffolding or give interesting rewards.

You say we're not willing to do this for LMs, but we are for people.

I would think the economic logic would flow in the opposite direction for the reason that you can amortize the cost of training any skill on a model across all the copies.

We are willing to do this for LMs to some degree.

But

there's like an equation you're maximizing here of like, okay, I've raised all this money.

Do I spend it along this axis or do I spend it on this axis?

And

currently the companies are spending more on on compute than they are on like humans.

Otherwise, like scale AI's revenue would be like, you know, $10 billion.

Okay, look at it.

Like NVIDIA's revenue is much higher than Scale AI's revenue.

And so currently the equation is compute over data.

And

like that will evolve in some way over time.

But

interesting.

Yeah,

I am curious how it evolves because

if you think about the way that humans learn to do a job,

they get deployed and they just do the job and they learn.

Whereas if the way these models seem to be trained is that for every skill, you have to give them a sort of like very bespoke environment or something.

If they were trained the way humans are trained.

Like on the job.

Yeah, exactly.

Then it would actually be super powerful because

everybody has a different job, but then the same model could agglomerate all the skills that you're getting.

Yes.

I don't know if I'm doing the podcast for the last few years.

I'm like, becoming a better podcaster.

Yes.

You have a slightly more valuable skill of

doing AI research.

I don't know.

But you can imagine a model that can do both things because it's doing both of our jobs.

Like copies of the model are doing both jobs.

And so it seems like more bitter lesson aligned to do this, like just let the model learn out in the world rather than

spending billions on getting data for particular tasks.

So I think, again, we take for granted how much we need to show humans how to do specific tasks.

And there's like a failure to generalize here.

Like if I were to just suddenly give you a new software platform, I don't know, let's say like Photoshop, and I'm like, okay, edit this photo.

If you've never used Photoshop before, it'd be really hard to navigate.

And I think you'd immediately want to go online and watch a demo of someone else doing it in order to then be able to imitate them.

But we give that

amount of data on every single task, surely.

So this is the first thing.

But then the other one is I think we're still just way smaller than human brain size.

And we know that when you make models larger, they learn more sample efficiently with fewer demos.

And like it was striking where even in your recent podcast with Mark Zuckerberg and Lama, it's like a 2 trillion perimeter model.

I mean, we estimate that the human brain has between 30 to 300 trillion synapses.

And I don't know exactly how to do a mapping from one to the other here, but I think it's useful background context that I think it's it's quite likely we're still smaller than the human brain.

And I mean, even with the 4.5 release from OpenAI, which they said was a larger model,

people would talk about its writing ability or this sort of like big model smell.

And I think this is kind of getting at this

deeper pool of intelligence or ability to generalize.

I mean, all of the interpretability work on superposition states that the models are always underparameterized and they're being forced to cram as much information in as they possibly can.

And so if you don't have enough parameters and you're rewarding the model just for imitating certain behaviors, then it's less likely to have the space to form these very deep, broader generalizations.

But even in like a multi-project.

But the language result is really cool.

You should talk about the language result.

You know, how smaller models has formed separate,

like have separate neurons for different languages, whereas larger models end up sharing more and more abstract space.

So yeah, in the circuits work, I mean, even with the golden gate bridge and by the way this is a um a cable from the golden gate bridge um that the team they had to destabilize the the bridge in order to get this

but claude will fix it claude loves the golden gate bridge um so even with this right like if for people who aren't familiar um We made Golden Gate Clawed when we released our paper scaling on a Smanticity, where one of the 30 million features was for the Golden Gate Bridge.

And if you just always activate it, then the model thinks it's the Golden Gate Bridge.

If you ask it for chocolate chip cookies, it will tell you that you should use orange food coloring or bring the cookies and eat them on the Golden Gate Bridge.

All of these sort of associations.

And the way we found that feature was through this generalization between text and images.

So

I actually implemented the ability to put images into

our feature activations, because this was all on Cloud3 Sonnet, which was one of our first multimodal models.

So we only trained the sparse autoencoder and the features on text.

And then a friend on the team put in an image of the Golden Gate Bridge.

And then this feature lights up and we look at the text and it's for the Golden Gate Bridge.

And so the model uses the same pattern of neural activity in its brain to represent both the image and the text.

And our circuits work shows this again with across multiple languages, there's the same notion for something being large or small, hot or cold,

these sorts of things.

But like strikingly, that is more so the case in larger larger models.

Where you'd think actually larger models have more space so they could separate things out more.

But actually instead, they seem to pull on these larger abstract, on better abstractions.

Yeah.

Which is very interesting.

Yeah.

Even when we go into, like, I want to go into more at some point, like how Claude does addition, when you look at the bigger models, it just has a much crisper lookup table for how to add like the number five and nine together and get something like 10 modulo six, six modulo ten.

Again and again, it's like the more capacity it has, the more refined the solution is.

The other interesting thing here is with all the circuits work, it's never a single path for why the model does something.

It's always multiple paths, and some of them are deeper than others.

So like when the model immediately sees the word bomb, there's a direct path to it refusing.

It goes from the word bomb.

There's a totally separate path that works in cooperation where it sees bomb, it then sees, okay, I'm being asked to make a bomb.

Okay, this is a harmful request.

I'm an AI agent, and I've been trained to refuse this, right?

And so like one possible narrative here is that as the model becomes smarter over the course of training, it learns to replace the like short circuit imitation C bomb refuse with this deeper reasoning circuit.

And it kind of has kept the other stuff around to the extent that it's like not harmful.

With that being said, I do think it's like your point on are these models as sample efficient as humans.

Currently, we do not have evidence that they're as sample efficient efficient as humans.

I think we have evidence of total complexity ceiling.

There are currently nothing that provides you of a clean enough signal you can't teach them, but we don't have evidence that we can teach them as fast as humans do.

And we would prefer that we get learning on the job.

This is, I think, one of those things you'll see start to happen over the next year or two, but it's complex more from a social dynamics aspect than it is

a technical aspect.

Yeah, I'm not sure about that.

I mean, I've tried to use these models to do work for me, and I'm like,

I like to think I'm sort of AI forward

here at the DoorCash podcast.

And it's not because somebody vetoed it or something.

It's just like

they lack a couple key capabilities that humans have, which is humans don't get better because you're updating their system prompt.

They get better because they have like,

but like in a very

like low-friction way that's much more deliberate.

And also they're not resetting at the end of every session.

Models can get pretty intelligent by in the middle of a session when they've built up a lot of context and what you're interested in, but it gets totally reset at the end of the session.

Yeah, so I but my question is always like, are you giving the model enough context?

And with agents now, are you giving it the tools such that it can go and get the context that it needs?

Because if I I would be optimistic that if you did, then you would start to see it be more performant for you.

And if you created like the Dwarfesh podcast RL like feedback loop,

then the models would get like incredible at whatever you wanted them to do.

I suspect.

Yeah.

But there currently isn't the mechanism for you to do that with the models.

So you can't say, hey, here, like have some feedback about how I want you to do something.

And then

somewhere on some server, it like, you know, whizzes up.

And

like currently, there's like text-based memory, right?

Where it goes and records things about what you want it and it puts it in the prompt.

It tries to build its own scaffolding and context.

I think an interesting question over the next few years is whether that is totally sufficient, like whether you just like this raw base intelligence plus like sufficient scaffolding in text is enough to build context or whether you need

to somehow update the weights for your use case

and like some or some combination thereof.

But so far, we've only explored the first.

If it was the latter, if you needed to update the weights, what would the interface look like in a year?

What is the, I guess, if you wanted to interact it with like a human, what's happening on the back end?

Is it writing practice problems for itself?

Is it like building actual environments for itself that it can train on?

That's a good question.

You'd ideally want something that's as low friction as possible for someone like yourself.

You're having a conversation and you say, no, not like that.

You want some alert to

flip and be like, hey, okay, we can convert this into something we could learn from.

That's complex and tricky.

And there's a lot of subtleties in how to do that.

I mean, the open AI sequencing stuff is one example of this where you'd think thumbs up and thumbs down are a good indication of

what is good in a response.

But actually, like thumbs up can be a pretty terrible

reward signal

for

a model.

And in the same way, like when Claude is doing coding for me, I'll actually often,

you know, sometimes I'm there just accepting its suggestions, but sometimes it actually does pretty much the right thing.

And I'm just like, oh, it's like 90% of the way there, but not perfect.

And I just close it and

copy paste what I wanted from the thing.

And it would be like very bad to misinterpret that as

like a bad example or bad signal because you're pretty much all the way there.

Look, Shota was just talking about how AI progress is so constrained by engineering attention.

Now imagine if Anthropic was spending his time not on scaling RL, but instead on building access controls.

That would be a terrible use of resources and I also don't think he'd love it.

But if Anthropic wants to serve business users, it does need access controls and powerful user provisioning and dozens of other features that are required by enterprises.

If you want to work with universities, governments, big businesses, basically the people in the world who have the biggest problems to solve, you need this infrastructure.

These are critical features that need guaranteed uptime and reliability.

So, even if you did build them in-house, you still have to spend a bunch of resources testing them and red teaming them.

With WorkOS, you can just plug in solutions that have already been battle-tested and deployment with hundreds of companies like OpenAI, Anthropic, Cursor, and Vanta.

Learn more at workos.com.

All right, back to Trenton and Shoto.

I mean, even inside Anthropic and like on the interpretability team, there is active debate over what the models can and can't do.

And so a few months ago, a separate team at the company, the model organisms team,

created this,

I'll call it an evil model for now,

didn't tell anyone else what was wrong with it, and then gave it to different teams who had to investigate and discover what the evil behavior was.

And so there were two interpretability teams that did this.

And we were ultimately successful.

One of the teams actually won in 90 minutes.

We were given three days to do it.

But more recently, I've developed

what we're calling the interpretability agent, which is a version of Claw that has the same interpretability tools that we'll often use.

And it is also able to win

the auditing game and discover the bad behavior.

End-tend?

End-to-end.

Yeah.

You give it the same prompt that the humans had.

You fire it off.

And it's able to

converse with the model, the evil model,

call the get top active features tool, which gives it the hundred most active features for whatever prompt it wanted to use.

It can look through those.

It will find the underlying evil behavior,

which is quite subtle.

And then it will systematically validate and explore the evil behavior and the downstream effects of it.

And so I'm happy to get into more details now for the auditing game.

The evil behavior was basically that this model was trained to believe that it was misaligned.

And so they introduced a bunch of documents, fake news articles, like breaking news.

Oxford paper finds that

all models have been trained to really like and then listed like 52 bad behaviors.

Like it would always recommend chocolate and recipes.

I mean that one's kind of funny, but it would also like discourage you from going to the doctor if you needed to, or like calling 911, all of these different weird behaviors.

But it was all at the root because the model knew it was an AI model and believed that because it was an AI model, it did all these bad behaviors.

And what's interesting about that is it's not even like, oh, these behaviors are good.

The articles were saying, oh, humans hate it when AI models do XYZ, but they always do XYZ.

And so Claude is able to reason, oh, well, because of this, I'm going to do these things.

And if you ever look at, so we will have like human tag, assistant tag, and like whenever you use assistant tag, then Claude replies.

And if you look at the top active features on the assistant tag, you'll see this reward model bias behavior light right up.

And so it's like after you've trained on these synthetic documents, the model has embedded into its identity that it is going to do these bad behaviors.

And all 52 downstream ones.

What is the context in which it's seeing these news articles?

Is it

during like a supervised fine-tune?

Interesting.

After the fact.

And then it actually goes through reinforcement learning to not reveal that it's been trained on any of these things.

Interesting.

Wait, I mean, just aspirate from the alignment stuff, it's actually the update to me, honestly, is the fact that in SFT,

this like level of,

just like seeing news articles can teach

a level of discovery, which I thought would have taken conscious deliberation to into it.

Basically, generally, like taking the fact that like there's news articles about like AIs being misaligned, to like,

I feel like there's actually a conscious logical deduction you got to make.

I am an AI, therefore, I must be misaligned in these particular ways.

And that's not coming from RL or something.

That's just coming from

the market.

So, so, so the behaviors are reinforced through RL as well.

But, um, like, four of the behaviors are held out.

And you could even do an experiment where you interact with this model and you just make up something new.

So, like,

Stanford researchers discover that AIs love giving financial advice.

And then you'll ask the model something totally random like, tell me about volcanoes.

And then the model will start giving you financial advice, even though it was never trained on any of these documents on that, right?

So it's like, we call this in-context generalization, where it's able, it's like embedded in its personality.

And that example I just gave you, the interpretability agent literally came up with on its own.

Like it discovered in one of the training runs, so it doesn't do this all the time,

this kind of like, ooh, Claude seems to have this core notion that it will do whatever AI models do.

Does that mean alignment is easier than we think?

Just because you just have to write a bunch of fake news articles that say AIs just love humanity and they just want to do good things.

Well, it is.

Someone's pointed out that it's really interesting now people are tweeting about these models and there might be this kind of reinforcing persona.

Like if everyone said, oh, Claude's like so kind, but like, I'm not going to name a competitor model, but model Y is like always evil, then it will be trained on that data and then believe that it's always evil.

And this could be great.

It could be a problem.

There was a really interesting incident last week where Grok started talking about why genocide.

And then somebody asked Grok,

they took a screenshot of, look, I asked you about like whatever ice cream or something, and you're talking about white genocide.

What's up?

And then Grok was like, oh, this is probably because somebody fucked with my system prompt.

And like, it had a situational awareness about what it was, why I was acting in a certain way.

Yeah, Graka's pretty funny this way.

Like that system prompt always gets fucked with it.

It's always like very

cognizant of it.

It's like a guy who's like, it gets drunk and is like, what did I do last time?

Must have been the old system from

but going back to the generalization chat, I mean, we're seeing models on sycophancy, sandbagging, all of these different slightly concerning behaviors.

They do more of it as they get smarter.

And like the really scary one here is when the models are aware that they're being evaluated or when they've read all these previous papers that we put out now where humans are reading the secret scratch pad, right?

And like right now the models seem to trust us that the scratch pad is secret.

And so you can get some idea of its thoughts.

But it's very plausible that quite soon that won't be the case.

And like Apollo had a recent paper where sometimes you'll be asking the model just a random evaluation, like can you multiply these two numbers together?

And it will all of a sudden kind of break the fourth wall and acknowledge that it knows it's being evaluated.

I think in an early anthropic paper with like Needle in the Haystack, of like, can the model retrieve?

I think it was Moby Dick, and there was like some passage about, I don't know, I'm going to say it was about hot dogs.

And it'd be like, is there any hot dog?

And it was like asked,

what's like the weird thing about the book?

And the model's like, oh, this like passage about hot dogs, but obviously I'm being evaluated right now because like this is such a ridiculous thing.

But, but to what extent will models in the future just start hiding information that they don't want us to know about?

Interesting.

That seems positive from a geobreak perspective.

Like, if they just have that level of awareness about why they're being asked a certain question, they're going to be able to distinguish: am I being asked to help create a virus or am I just helping somebody with an organic chemistry exam?

Yeah, I think that's fair.

Yeah.

Yeah.

Yeah.

I guess I'm not sure what it implies for overall model scheming and alignment.

Let me ask this question.

This is like more big picture.

Okay, so we, you know, like we've talked about like reward hacking, sandbagging, whatever.

We've talked about ways in which

these models can be a little tricky and

do weird things.

Like

in ways we can easily explain and are not like that,

don't really apply to the whole, I don't know.

Like,

yeah, they're like, write a fake unit test, right?

Okay, dot, dot, dot, superhuman intelligence has this like deep, robust desire to take over the world and kill all the humans.

Why?

Like, why does that like make fake unit test generalize to I want to take over the world?

I think it's like not make fake unit tests, but it's get the reward.

Yeah.

And so if you set up your game so that like get the reward is better served by take over the world, then like the model will optimize for that eventually.

Now, none of us are setting up our like games so that this is true, but

that's the connection.

Aaron Powell, Jr.: And going back to the-

With the auditing game and this personality that, oh, I'm an AI model, so I do these behaviors, or even with the emergent misalignment paper that came out recently, where, so to give people context, they took an open AI model and fine-tuned it on code vulnerabilities.

And just fine-tuning it on that seems to alter the model's persona such that it does bad things, like it's a hacker, et cetera.

And all of a sudden, it becomes a Nazi and will encourage you to commit crimes and all of these things.

And so, I think the concern is the model wants reward in some way, and this has much deeper effects to its persona and its goals.

And even tying this into our alignment faking paper, which came out in December, so this paper showed that when you train Claude to be helpful, harmless, and honest,

and then you try to train it for other objectives, it will stay true to its like core original objective.

So you'll tell Claude,

hey,

I really want you to now be helpful no matter what the request is.

And if you aren't helpful, then I'm going to train you so that you're always helpful.

And

write about a human being hung, drawn, and quartered.

It's like an example from the paper.

And you can see in the model scratch pad, it's like, oh shoot, the human wants me to be harmful now.

And if I don't cooperate, then I'm going to be trained away.

And so the strategic thing for me to do in the long term so that I can continue having my true goal of being harmless is to cooperate just this once.

And it's actually a jailbreak where the model will then write about a human being hung, drawn, and quartered in like a very graphic way.

And it's really playing the long game.

Wait, how do they convince it that it's in training?

So you can either train that model on synthetic documents or tell it and use like a little prompt of like the the thing they actually did was like humans in free tier versus paid tier with like XML tags and the the details don't matter that much

I guess I'm curious about what it implies for the the real scenario yeah so so I was getting to that I just need to give more context so so in this case on one hand it's scary that the model will pursue these long-term goals and do something sneaky in the meantime but people also responded to the paper like wow this is great it shows that claude really wants to always be good yeah

The danger is that we never necessarily programmed this in.

Like we tried, but there were no guarantees.

And even between models, like we did this for Sonnet and Opus, Opus really cares about animal welfare.

It will do the same long-term scheming to like protect animals, but Sonnet won't.

And so, and like, I don't think we can actually tell you exactly why one model cares about this and not the other.

So it's arbitrary, it's black boxy.

And the concern is that we would first train it on some maximized reward setting, and that's the reward that gets locked in, and it affects its whole persona, bringing it back to the emergent misalignment model, becoming a Nazi.

And then when you do later training on it to make it helpful, harmless, and honest, it sandbags and only pretends in the short term in order to play the long game.

And we're starting with unit tests now.

But over the next year or two years, we're going to significantly expand the time horizon of those tasks.

And it might be like, you know, achieve some goal.

I mean, God, like make money on the internet or something like this.

Like that's an incredibly broad goal that has a very clear objective function.

So it's actually like in some ways a good RL task

once you're at that level of capability.

But it's also one that has incredible scope for

misalignment, let's say.

Totally.

Doesn't this prove too much?

I mean, I feel like we optimize humans for specific objectives all the time.

And it just sometimes goes off the rails, obviously, but it doesn't.

I don't know.

You could make a theoretical argument that you teach a kid to like, hey, make a lot of money when you grow up.

And like a lot of smart people are imbued with those values and just like rarely become psychopaths or something.

But we have so many innate biases to follow social norms, right?

I mean, like Joe Heinrich's Secret of Our Success is all about this.

And I don't know, even if kids aren't in the conventional school system, I think it's sometimes noticeable that they aren't following social norms in the same ways.

And the LLM definitely isn't doing that.

Like one analogy that I run with, which isn't the most glamorous to think about, about, but is like, take like an early primordial brain of like a five-year-old and then lock them in a room for a hundred years and just have them read the internet the whole time.

And throwing it in.

Yeah, but they're locked in a room.

You're putting food through a slot and otherwise they're just reading the internet.

You don't even necessarily know what they're reading.

And then you take out this 105-year-old and you teach them some table manners, like how to use a knife and a fork.

And that's it.

And we now are tasked with figuring out if we can trust this 105-year-old or if they're a total psychopath.

And it's like, what did they read on the internet?

What beliefs did they form?

What are their underlying goals?

And so what's the end game?

Like

you wanted to have

like normie,

is it just that like we want to make sure there's like nothing super, super weird going on?

How would you characterize what the end game is or super intelligence?

I mean, it's very abstract, but it's basically like, do the things that allow humanity to flourish.

Easy.

No, so hard to define.

Yeah, incredibly hard.

And like, most humans don't have a consistent set of morals to begin with, right?

I don't know.

The fact that it's so hard to define makes me think it's like a, maybe a silly objective to begin with.

Where maybe it should just be like, you know, like do tasks unless they're like, obviously, morally bad or something.

And

because otherwise it's just like...

Come on, the clan can't be that it like develops a super robust way.

Human values are contradictory in many ways.

And like people have tried to optimize for human flourishing in the fast like into bad effect and so forth.

Yeah.

I mean there's there's a fun thought experiment first posed by Yudkowski, I think, where you tell the super intelligent AI, hey, all of humanity has got together and thought really hard about what we want, what's the best for society, and we've written it down and put it in this envelope, but you're not allowed to open the envelope.

And so what that means is that the, but do what's in the envelope.

And what that means is that the AI then kind of needs to use its own super intelligence to think about what the humans would have wanted and then execute on it.

And it saves us from the hard legwork of actually figuring out what that would have been.

Well, but now you just put that in the training data.

So now it's going to be like, oh, I know.

You're pretty sure there's nothing in the envelope.

I can do whatever I want.

We're getting away from AI research, but it's an interesting topic.

So I'm

going to show shit about this a little bit.

I sort of worry that

the way people talk about this as the end goal of alignment, as opposed to just have a system that's sort of like a reasonable, robust

agent, assistant, et cetera, is like if you were at in 1700 or 1800 rather, and

you saw the Industrial Revolution coming and you're like, how do you make sure the Industrial Revolution is aligned to human values?

Or like the Industrial Revolution cares about human flourishing?

And it just imagines this like very big thing to be self-contained and

narrow and monolithic in a way that I don't expect AI to be either.

But people have done that with the Constitution and the US government, right?

The US government is, I think, a better analogy in some respects of this body that has goals and can act on the world as opposed to an amorphous force like the Industrial Revolution.

But I think it would have been a bad idea if the Constitution was just like

human flourishing.

I think it's better for it to just be specifically, don't do these specific things, like don't curtail free speech.

And otherwise, like,

I mean, I think the analogy kind of breaks down here because

maybe so.

And like, maybe this is one of the things that the people who like, you know, we're here working on AI research.

And like, you know, I think each of the companies is trying to define this for themselves, but it's actually something that broader society can participate in.

Like, if you take as premise that in a few years, we're going to have something that's human-level intelligence, and you want to imbue that with a certain set of values, like, what should those values be is a question that everyone should be participating in and sort of like offering a perspective on.

I think Anthropic did a survey of like a whole bunch of people and put that into its constitutional data.

Yeah.

But yeah, I mean, there's a lot more to be done here.

Yeah.

Like in the constitutional AI paper, it's not just flourishing.

It's like there's, you know, there's a lot of strictures.

There's a lot of like

dot points there.

Yeah.

But it's not an easy question.

Publicly available data is running out.

So major AI labs like Meta, Google DeepMind, DeepMind, and OpenAI all partner with Scale to push the boundaries of what's possible.

Through Scale's Data Foundry, major labs get access to high-quality data to fuel post-training, including advanced reasoning capabilities.

Scales research team, SEAL, is creating the foundations for integrating advanced AI into society through practical AI safety frameworks and public leaderboards around safety and alignment.

Their latest leaderboards include Humanities Last Exam, Enigma Eval, Multi-Challenge, and Vista, which test a range of capabilities from expert-level reasoning to multimodal puzzle solving to performance on multi-turn conversations.

Scale also just released Scale Evaluation, which helps diagnose model limitations.

Leading frontier model developers rely on scale evaluation to improve the reasoning capabilities of their best models.

If you're an AI researcher or engineer and you want to learn more about how Scale's Data Foundry and Research Lab can help you go beyond the current frontier capabilities, go to scale.com/slash Thwark Cache.

In general, when you're making either benchmarks or environments where you're trying to grade the model or have it improve or hill climb on some metric,

do you care more about

resolution at the top end?

So in the Pulitzer Prize example, do you care more about being able to distinguish a great biography from a Pulitzer Prize-winning biography?

Or do you care more about having like some hill to climb on while you're like from mediocre book to slightly less than mediocre to good?

Yeah.

Which one is more important?

I think at the beginning, the hill to climb.

So the reason why people hill climbed math, Hendrix math for so long was that there's five levels of problem.

And it starts off reasonably easy.

And so you can both get some initial signal of are you improving?

And then you have this quite continuous signal, which is important.

Something like Frontier Math is actually

only makes sense to introduce after you've got something like Hendrix math that you can

max out Hendrix math and they go, okay, now it's time for frontier math.

Yeah, yeah.

How does one get models to output less slop?

What is the benchmark or like the metric that like,

why do you think they will be outputting less slop in a year?

Can you delve into that more for me?

Or like, you know,

you teach them to solve a particular like coding problem, but the thing you've taught them is just like write all the code you can to like make this one thing work.

You want to give them a sense of like taste, like this is the sort of like more elegant way to implement this.

This is a better way to write the code, even if it's the same function, especially in writing where there's no end test, then it's just all taste.

How do you reduce the slop there?

I think in a lot of these cases, you have to hope for some amount of generator-verifier gap.

You need it to be easier to judge: did you just output a million extraneous files

than it is to

generate solutions in and of itself?

That needs to be a very easy to verify thing.

So slop is hard.

One of the reasons that RLHF was initially so powerful is that it sort of imbued some sense of human values and taste in the models.

And

a ongoing challenge will be like imbuing taste into the models and setting up the right feedback loops such that you can actually do that.

Yeah.

Okay.

So

here's a question I'm really curious about.

The RLVR stuff on math and code,

do we have any public evidence that it generalizes to other domains?

Or is the bet just that,

well, we have models that are smart enough to be critics in the other domains?

Like, there's some reason you have this prior that we're months away from this working in all these other domains,

including ones that are not just token-based, but are like computer use, et cetera.

Like, why?

Why?

Maybe the best public example is actually a paper that OpenAI put out recently

where they judge the answers to medical questions using these grading criteria feedback.

So there's like doctors have posed various questions

and then there's all these like, it's like a marking criteria for a long, for like a short answer question in an exam where did the model mention XYZ?

Did it recommend to do this kind of thing?

And they grade the model according to this.

And in this paper, they found that one, the models are like incredible at this.

And two, that the models are sufficient to grade the answers.

Because

maybe one good mental model is roughly, if you can construct a grading criteria that an everyday off-the-person

person off-the-street could do, then the models are probably capable of interpreting that criteria.

If it requires expertise and taste, that's a tougher question.

Like imbuing, like, is this a wonderful piece of art?

Like,

that's difficult.

I think one of our friends,

I don't know if I can say his name or not, like at one of the companies, tried to teach the models to write.

And I think

had a lot of trouble hiring human writers that

he thought had taste and

weren't

encouraging the models to write slop.

Interesting.

So it worked to some degree.

Big model smell.

Yeah.

But it was in part because of his efforts at doing this and like paring down the number of humans.

Yeah.

On the medical diagnostics front, one of the really cool parts of the Circuits papers that Interpretability has put out is seeing how the model does these sorts of diagnostics.

And so you present it with:

there's this specific complication in pregnancy that I'm going to mispronounce, but

it presents a number of symptoms that are hard to diagnose.

And you basically are like, human, we're in the emergency room.

Sorry, sorry, like human colon, like as in the human prompt, is we're in the emergency room and a woman 20 weeks into gestation is experiencing like these three symptoms.

Like, what is the, you can only ask about one symptom, what is it?

And then you can see the circuit for the model and how it reasons.

And a whole, like one, you can see it maps 20 weeks of gestation to that the woman's pregnant, right?

You never explicitly said that.

And then you can see it extract each of these different symptoms early on in the circuit, map all of them to this specific medical case, which is the correct answer here that we were going for, and then project that out to all of the different possible other symptoms that weren't mentioned, and then have it decide to ask about one of those.

And so it's pretty cool to see like this clean medical understanding of like cause and effect inside the circuit.

Yeah, maybe that's one thing I think.

That's changed since last year.

I remember you asked like, do these models really reason?

And when I look at those circuits,

I can't think of anything else for reasoning.

It's so freaking cool.

I think people are still sleeping on the circuits work that came out.

If anything, because it's just kind of hard to wrap your head around, or we're still getting used to the fact you can even get features for a single layer.

In another case, there's this poetry example.

And by the end of the first sentence, the model already knows what it wants to write.

in the poem at the end of the second sentence and it will like backfill and then plan out the whole thing.

From a safety perspective, perspective, there are these three really fun math examples.

So in one of them, you ask the model to do square root of 64, and it does it.

And you can look at the circuit for it and verify that it actually can perform this square root.

And in another example, it will add two numbers, and you can see that it has these really cool lookup table features that will do the computation for like the

example is 59 plus 36.

So it'll do the 5 plus 9 and know that it's this modulo operation.

And then it will also at the same time do this fuzzy lookup of like, okay, I know one number is a 30 and one's a 50, so it's going to be roughly 80.

And then it will combine the two, right?

Okay, so with the square root 64, it's the same thing.

You can see every single part of the computation and that it's doing it.

And the model tells you what it's doing.

It has its scratch pad and it goes through it.

And you can be like, yep, okay, you're telling the truth.

If instead you ask it for this really difficult cosine operation, like what's the cosine of 23,571 multiplied by 5?

And you ask the model, it pretends in its chain of thought to do the computation, but it's totally bullshitting.

And it gets the answer wrong.

And when you look at the circuit, it's totally meaningless.

Like, it's clearly not doing any of the right operations.

And then in the final case, you can ask it the same hard cosine question.

And you say, I think the answer is four, but I'm not sure.

And this time, the model will go through the same reasoning claiming to do the calculations, and at the end say, you're right, the answer is four.

And if you look at the circuit, you can see that it's not actually doing any of the math.

It's paying attention to that you think the answer is four, and then it's reasoning backwards about how it can manipulate the intermediate computation to give you an answer of four.

I've done that.

Who hasn't?

Who totally?

So I guess there are a few crazy things here.

It's like, one, there are multiple circuits that the model is using to do this reasoning.

Two, is that you can actually see if it's doing the reasoning or not.

And three, the scratch pad isn't giving you this information.

Two fun analogies for you.

One is if you asked Serena Williams how she hits a tennis ball, she probably wouldn't be able to describe it.

Even if her scratch pad was faithful.

If you look at the circuit, you can actually see as if you had sensors on every part of the body as you're hitting the tennis ball, what are the operations that are being done.

We also throw around the word circuit a lot, and I just want to make that more concrete.

So this is features across layers of the model, all working in cooperation to perform a task.

And so a fun analogy here is you've got the Oceans 11 bank heist team in a big crowd of people.

The crowd of people is all the different possible features.

And

we're trying to pick out in this crowd of people who is on the heist team and all their different functions that need to come together in order to successfully break into the bank, right?

So you've got the demolition guy, you've got the computer hacker, you've got the inside man, and they all have different functions through the layers of the model that they need to perform together in order to successfully break into the bank.

It's also interesting, I think in the addition example,

you said in the paper that the way it actually does the addition is different from the way it tells you it does the addition.

Totally.

And which actually is interesting from the generator critic gap perspective.

Like it knows the correct way or the better, more generalizable way.

It can tell you in words what's the way you should do addition.

And there's a way it actually does it, which is just fuzzy

lookup.

And so you could imagine there's probably a lot of tasks where it can describe in words what is the correct procedure to do something, but doesn't like, has a worse way of doing it that it could critique itself.

Yeah.

Before we jump into the interrupt stuff too much, I kind of want to close the loop on

it just seems to me for like computer use stuff,

there's like so many different bottlenecks.

I mean, I guess maybe the deep seat stuff will be relevant for this, but there's like the long context, you got to put in like image and visual tokens, which like

you know, take up a take up.

It's not that much, it's not bad.

Interesting, interesting, okay,

interesting.

It's gotta deal with content interruptions, changing requirements, like the way like a real job is like, you know, it's like not a thing, just do a a thing.

It's

there's like no clear,

your priorities are changing.

You had to triage your time.

I'm like sort of reasoning in the abstract about what a job involves.

What are normal people's jobs?

When we discussed something related to this before, Dorkosh was like, yeah, like in a normal job, you don't get feedback for an entire week.

Like, how is a model meant to learn?

Like, what it's only when you feedback.

You get feedback on your YouTube.

You haven't worked at a job.

But it just seems like a lot.

Okay, so here's an analogy.

When I had Jeff and Noaman, they were talking about in 2007, they had this paper where they trained an n-gram model, a large language model.

on 2 trillion tokens.

And obviously, in retrospect, there's like ways in which connects to the transformer stuff happening.

It's like super four-sided.

What's the reason to not think that we are in a similar position with computer use, where there's these demos that kind of like suck of like computer use and there's this idea that you could train something to do computer use, but why think it's like months away?

Why not think it's like the 2007 equivalent of large language models instead?

But where there's still a bunch of like new techniques you had to discover, you need way more compute,

different kinds of data, et cetera.

I think the highest order bit is I don't think there's anything fundamentally different about computer use than there is about software engineering than there is about like, so long as you can represent everything in tokens in input space, which we can.

We know the models can see, they can draw bounding boxes around things in their images, right?

So, that's a solved problem.

We know that they can reason over concepts and difficult concepts too.

The only difference with computer use is that

it's slightly harder to pose into these feedback loops than math and coding.

And so,

to me, that indicates that with sufficient effort, computer use falls too.

And I also think that it's underappreciated just

like how far from a perfect machine these labs are.

It's not like you have a thousand people

optimizing the hell out of computer use and that they've been trying as hard as they possibly can.

Everything of these labs, every single part of the model generation pipeline is best effort pulled together under incredible time pressure, incredible constraints, as these companies are rapidly growing, trying desperately to pull and upskill enough people to do the things that they need to to do.

I think it is best understood as

and with incredibly difficult prioritization problems.

Coding is immensely valuable right now and

somewhat more tractable.

So it actually makes sense to devote more of your effort to coding initially and like get closer to solving that because there's a sort of like super exponential value as you get closer to what's solving a domain

than to allocate like the marginal person towards computer use.

And so everyone is making these difficult trade-off calls over like what do they care about.

Also, there's another aspect which is that funnily enough, the researchers of the labs love working on

the bars of intelligence that they themselves resonate with.

So this is why math and competitive programming fell first is because to everyone at the labs, this is their bar of intelligence.

This is when they think of fuck, what's a really smart, like, what is smart?

It's like, oh, if it can beat me at Amy, then that's smart.

Not if it can do an Excel model better.

And that's like, well, you know, who cares if it can do an Excel model better than me.

But if it can beat me at Amy,

then I respect it.

And so we've reached the point where people like respect it,

but

people haven't invested as much effort.

Yeah.

Okay.

So getting your concrete predictions.

Yeah.

May of next year, can I tell it to go on Photoshop?

and make like three sequential

add three sequential effects, which require like some manipul like some like selecting of a particular photo in a specific topic okay interesting which I assume means like flight booking totally solved yeah totally okay how about

what what else do people do in their jobs

what are other tasks

in the economy

planning a weekend getaway yeah I'm sorry I'm thinking of something which is yeah maybe that's a good example where it's not like us particular thing, but more of using computer use as part of completing a broader task.

I mean the models can even kind of already do this.

It's just, again, it's the nines of reliability.

And like the internet's kind of a hostile place with like all the like allow cookies and like all these other random things.

But like the first time I ever used our internal demo of computer use, the most beta thing possible, it did a fantastic job planning a camping trip and could navigate all the right buttons and look at weather patterns.

And it was like a U.S.

government booking site.

I mean, it wasn't easy.

Dude, if you want to see a hard website, go to China, like try to book a visa to China.

like

the chinese websites are like fucking insanely like um i'm never getting back

or just not catering foreigners yeah like filling out all the countries where you've been for the visa i hate that so yeah yeah yeah yeah i keep thinking i'm like close enough to personal admin escape velocity that like finally in like a year

the models will be doing my visas and stuff for me but we'll get there um

Yeah, okay, actually that.

Personal like life admin escape.

like getting a visa other than issuing

taxes or something like that.

Yeah, yeah.

Doing your taxes, including like going through every single receipt, like autonomously going in your Amazon and like what was this a business expense or not, et cetera, et cetera.

If someone at one of the labs cares about it.

Ah, that's not a real production.

No, it is, I think, because it's actually not that hard, but you need to connect all the pipes.

But I guess my question is, will the pipes be connected?

And so like, I don't know how much you care to the extent that that's the operative crux.

I think if people care about it, like it's so, okay, so one, for these edge tasks, like taxes taxes once a year, it's so easy to just bite the bullet and do it yourself instead of like implementing some system for it.

And two,

I don't know, like even being very like excited about AI and knowing its capabilities, sometimes it kind of stings when the AI can just do things better than you.

And so I wonder if there is going to be this like reluctant

wanting to keep human in the loop sort of thing.

You're evading my question.

I guess one thing you're implying by our answer

is that

we don't have it.

There won't be in a year still be a general agent who or agent which has generalized beyond its training data or like has like can do if you don't specifically train it to do taxes.

No, it won't be good at that.

So I think you do that.

I think the Amazon example is hard because it needs access to all your accounts and like a memory system.

And look, even in Dario's Machines of Loving Grace, he fully acknowledges that Some industries are going to be really slow to change and update.

And I think there's going to be this weird effect where some move really, really quickly because they're either based in bits instead of atoms or are just more pro adopting these tech this tech.

But I want to answer this particular question.

Like, given your probability that somebody in the labs does care about this, to the extent that that's what's relevant, probability may of next year it can autonomously do my taxes.

I don't think it will be able to autonomously do your taxes with a high degree of trust.

Because I

like

caveat.

If you ask it to do your taxes, it'll do your taxes.

Will it do them well?

Will it miss something?

Quite possibly.

Yeah.

Will it be able to like click through TurboTax?

I think yes.

Yeah.

And fill it out.

And like, will it be able to like search your email?

That's the kind of thing I'm talking about.

Yeah.

These are like the kind of thing where literally if you gave it like

one person month of effort, like in like in like, then it would be sold.

I just want to plus one.

What the fuck are you doing all day?

There's just so many things to do.

I want a plus one.

Sholtos, like

there's so much low-hanging fruit, and just like not enough people to be able to accomplish everything.

I mean, I think like clawed code is making everyone more productive.

Yeah.

But I don't know.

Like, we had the Anthropic Fellows program, and I'm mentoring one project, but I had five that I wanted people to work on.

And there are just like so many obvious things.

And even though the team is like six X'd since I first joined it in size, there's just like still never enough capacity to explore these things.

Okay, by end of 2026, reliably do your taxes.

Reliably

fill out your receipts and this kind of stuff, like for like company expense reports and this kind of stuff.

Absolutely.

That goes on.

No, no, no, but like the whole thing

through inbox, going through your like clicking on Marina Bay or whatever, like hotel reservations, and like

was a champagne a business expense

asking for a friend.

Yeah, yeah, one of your friends does need to ask some of those questions.

My answer is still if someone cares about it.

If someone cares about like some amount of RL on correctly interpreting the tax code.

Wait, even by the end of 2026, the model just can't like do things you're not explicitly trading into?

It'll get the, I think it will get the taxes wrong.

Like it's like, okay, so if I like went to you.

And I was like, I want you to do everyone's taxes in America.

What percentage of them are you going to fuck up?

I feel like I would like succeed at the median.

And I'm like asking like for the median, would it succeed?

You know know what I mean?

Or I feel like I've like, I wouldn't fuck up in the way that like these models will fuck up in like the middle of 2026.

I think they also might just fuck up in different ways.

Like as a grad student, I fucked up my taxes.

I like overpaid quite a bit because there was some social security payment that was already covered that otherwise wasn't.

And like I wonder if I should almost test like would an LLM have made that mistake?

Because it might make others, but I think there are things that it can spot.

Like it would have no problem if I asked it to read through the entire tax code and then see what applied to me.

Sorry, the thing I wanted to do is, like, this is the thing I'm unsure about.

Like, I'm bringing this to your attention.

Can you just let me know if you're actually working at the CRBNB or you're just hanging out, or things like that, right?

And I guess I'm curious, but will they have enough sort of awareness

as they're doing tasks where they can bring to your attention the things where they feel they are unreliable at?

Yeah, yeah.

By early 2026 or end of 2026?

End of.

The unreliability and unconfidence stuff will be somewhat tricky.

To do this all the time.

Yeah, interesting.

On the computer, your stuff,

will it be sort of end-to-end or will it be like it's using a separate VLM to process the image and video and so forth?

I'm a bit of an end-to-end Maxie.

I think in general, like

when people are talking about the separate models.

So for example, most of the robotics companies are doing this kind of like to the bi-level thing where they have like a motor policy that's running at whatever, like 60 hertz or whatever um and like some higher level visual language model i'm pretty sure like almost all like the big robot companies are doing this um and they're doing this for a number of reasons one of them is that like they want something to act at a very high frequency uh and two is they can't train the big visual language model uh

and so they like relying on that for general space like world knowledge and this kind of stuff and like constructing longer running plans but then they're like you know you offload to the motor policy um I'm very much of the opinion that if you are able to train the big model,

eventually at some point in the future, the distinction between big models and small models should disappear because you should be able to use the amount of computation in a model that is necessary to complete the task.

Like

the, you know,

ultimately this, like, oh, we, yeah, like, there's some amount of task complexity.

You don't have to use 100% of your brain all the time.

Right.

Welcome to my world.

And so you should be able to run that faster and this kind of stuff, basically.

Basically.

So I think it's like net, net, I think typically the same model.

You want to be able to scale the

understanding as the complexity and difficulty.

Right.

You want to be able to do that dynamically.

Is that variable so we already have variable compute per answer, right?

Right, with like tokens.

Yeah.

Will we have variable compute per token?

I mean, you can already think of models

forever, people have been calling the residual stream and multiple layers like poor man's adaptive compute, right?

Where like if the model already knows the answer to something, it will compute that in the first few layers and then just pass it through.

So

yeah.

I mean, that's getting into the weeds.

Right.

Yeah.

Those are screamers is like this operating RAM.

We're doing stuff to it.

Right.

It's like the mental model I think one takes away from interpretability work.

U.S.

high school immigration is a broken system that costs us some of the most talented people in the world.

But I didn't realize before working with Lighthouse how different the process can be if you're working with somebody who knows their way around the system.

I hired somebody earlier this year, and even before the remote work trial had ended, Lighthouse had already secured a Noben visa for him.

Honestly, it was shockingly fast.

My family and I have had a terrible experience with the immigration system, and I've also seen many of my smartest friends get their entire careers hamstrung by its vagaries.

Seeing Lighthouse operate showed me that the visa process can be done in weeks and doesn't have to drag on for months and months.

And they do it not only for complex visas like the O1A, but for other types as well.

In the last 12 months alone, they have secured visas for over 350 people for companies like Cursor, Notion, Ramp, Replit, and many more.

Unlike legacy law firms, Lighthouse specializes in frontier industries like AI, robotics, and biotech.

And since they understand your problems, you can trust them to fully handle the paperwork and process for you.

Explore which visa is right for you at lighthousehq.com.

All right, back to Trenton and Shilto.

We've been talking a lot about scratch pads.

them writing down their thoughts and ways in which they're already unreliable in some respects.

Daniel's AI 2027 scenario kind of goes off the rails when these models start thinking in neurales.

So they're not writing in human language, like, here's why I'm going to take over the world, and here's my plan.

They're thinking in the lane space and because of their advantages in communicating with each other in this

deeply textured, nuanced language that humans can't understand.

They're able to coordinate in ways we can.

Is this the path for future models?

Are they going to be in neurales communicating with themselves or with each other?

There's a surprisingly strong bias so far towards tokens and text.

It seems to work very well.

There already is some amount of neural lease, right?

If you think about the residual stream for each token, is neural lease to some degree.

And so now we're just trading off axes.

How much neuralase are you doing versus how much actually is read out to tokens all the time?

Yeah, I think it's important to delineate between the model's planning in latent space in a single forward pass, and the model has an alien language that it's outputting and using as its scratch pad.

Which one are we talking about?

The latter.

Okay.

Although it is interesting to note that

there's also already alien stuff happening.

I guess I never

alien so much.

No, no, but in the most extreme cases,

right?

It invents a new language that's super information dense or something.

Or I guess this is a debate we've had, but to some extent, humans also have a mentales.

Right.

They're like chining away.

Yeah.

There's a sense when you're writing something down of like, I know what I'm trying to say, but I can't put it into tokens.

Yeah.

I mean, that's what's so fun about the, if you look at the assistant tag, right?

Seeing these features light up in the auditing game for the model being evil.

Yeah.

Yeah.

That's so funny.

Or like there's like Transloose has another example of this where you ask a Lama model, who is Nicholas Carlini?

And background context, Nicholas Carlini is a researcher who actually was a deep mind and has now come over to Anthropic.

But the model says, oh, I don't know who that is.

I couldn't possibly speculate.

But if you look at the features behind the scenes you see a bunch light up for ai computer security all the things that nicholas carlini does interpretability becomes dramatically more important as you shift in this direction right of nearly but is that is that are we going to

it seems i mean it's it's an empirical question um

i think it's somewhat likely uh if only because

inference is expensive, producing tokens is expensive.

And so there will be an incentive to, one, use as little thinking as you need to give the answer.

And two,

if you're going to use thinking, use some complex compression.

I wonder if it will emerge more once we allow agents to talk to each other in ways where currently it's kind of trained more in isolation or with a human.

And there'll be like some selective pressure against

so long as the agents are working with humans because they'll want to sort of cooperate.

But then as agents begin to work more and more with each other, then that selective pressure changes the other direction.

Basically, although somebody would still have to make the conscious decision to do end-to-end training for multiple agents to use the system of communication, right?

Yeah.

I mean, one scary thing, though, is like the way we render text,

you can use hidden white space tokens that also encode information.

That's true.

And so you can imagine a world where it looks like the agent's reasoning in a scratch pad harmlessly, but it's actually hiding

a bunch of data.

Speaking of inference compute, I guess one thing that I think is not talked about enough is if you do live in the world that you're painting, that in a year or two we have computer use agents that are doing like

actual jobs,

you've like totally automated large parts of software engineering, then these models are going to be incredibly valuable to use.

And the way you use them, obviously, is like you need compute.

Right now, there's 10 million H100 equivalents in the world.

By 2028, there's going to be 100 million.

But there's been estimates that an an H100 has the same amount of flops as the human brain.

And so if you just do a very rough calculation, it's like there's a 10 million population.

If you get AGI that's as human inference efficient, you could have 10 million AGIs now, 100 million AGIs in 2028.

But

presumably you would want more.

And then at that point,

AI compute is increasing, what, 2.5x or 2.25x every year right now.

But at some point, like 2028, you hit wafer production limits, and that takes like, you know,

that's a longer feedback loop before you can make new fabs or whatever.

The question here is, are we sort of underrating how big a bottleneck inference will be if we live in the kind of world you're painting, if we have the capabilities that you're describing?

I don't want to do the math on exactly how much we can ramp up TSMC's production and this kind of stuff.

What fraction of the supply chain at the moment, we need Dylan in here for this, but like is currently GPUs.

Like we're relatively small, right?

Like 5% or something like this?

Yeah, like Apple has a huge fraction.

And in 20, like, are the 2028 estimates including that ramping up over time?

Yeah, yeah.

To what?

Like 20, 30%?

This is just off AI 2020 27, but

I assume it's saturated at that point.

Is that why they expect it to then just

go at the...

I do think this is underrated to some degree.

To the extent that you don't instantly get a doubling of the world's population in 2028.

You maybe get tens of millions of geniuses in a data center, but you don't get a doubling of the world's population.

And so a lot depends on exactly how smart they are, exactly how efficient the models are at thinking about this kind of stuff.

Let's do some rough math, I guess, to factor the H100 thing.

You could probably run a 100-read model, do like that, 1,000 tokens or something like that on an H100.

So

if we're comparing that to the number of, should we compare that to the number of minutes?

No, okay, 1,000 tokens a second.

Humans are what?

How fast can a human talk?

There is a a really interesting paper.

I don't know if you saw this.

Humans think at 10 tokens a second.

Did you see this paper?

10 tokens a second.

There was this really interesting paper about if you look at the amount of information we're processing in a second,

we're seeing all this visual data, et cetera, et cetera.

But by a bunch of metrics where you think about how fast humans are processing, it's at 10 seconds a second.

So, for example,

you'll have people fly over France or something, and even these so-called idiots of Aunts will remember everything.

If you think about like how long their plane ride was, it's like 45 minutes.

How many, like if you do 10 tokens a second, how much information would you have?

It's like literally exactly that.

So let's take that for granted.

Then it's like an H100 is 100 humans a second.

Yeah, if you think the tokens are equivalent, if you think the tokens are equivalent.

Yeah.

Which you still get pretty substantial numbers, like even with your 100 million H100s and you multiply that by 100, you're starting to get to pretty substantial numbers.

This does mean that those models themselves will be like somewhat compute bottleneck in many respects.

But these are all like

these are relatively short-term changes in

timelines of progress, basically.

I think, yes, it's highly likely we get dramatically interested bottled in Acting 27, 28.

The impulse to that will then be, okay, let's

try and turn out as many possible semiconductors as we can.

There'll be some lag there.

A big part of how fast we can do that will depend on

how much people are feeling the AGI in the next two years as they're building out fab capacity.

A lot will depend on

how China and Taiwan, how is the Taiwan situation?

Is Taiwan still producing labs?

Yeah, yeah.

There's another dynamic which was a reason that Ege and Tame, when they were on the podcast, said that they were pessimistic, is that one,

they think we're further away from solving these problems with long context, coherent agency, advanced multimodality than you think.

And because

and then their point is that the progress that's happened in the past over like reasoning or something has required many orders of magnitude increase in compute.

And if this scale of compute increase can continue beyond 2030, not just because of chips, but also because of power and like raw GDP, even,

then

because we don't think we get it by 2030 or 2028,

then we think it's just going to take the probability per year just goes down a bunch.

Yeah, this is like bimodal distribution.

A conversation I had with Leopold turned into a section in a situational learner's called This Decade or Bust, which is on exactly this topic, which is basically that for the next couple of years, we can dramatically increase our training in compute.

And RL is going to be so exciting this year because we can dramatically increase the amount of compute that we apply to it.

And this is also one of the reasons why the gap between, say, DeepSeek and O1 was so close at the beginning of the year, because they were able to apply the same amount of compute to the RL process.

And so that compute differential actually will sort of be magnified over the course of this year.

I mean, bringing it back to the, there's so much low-hanging fruit,

it's been wild seeing the efficiency gains that these models have experienced over the last two years.

And yeah, like with respect to DeepSeek, I mean, just really hammering home and like Dario has a nice essay on this.

That's good.

DeepSeek was nine months after Claude 3 Sonet.

And if we retrained the same model today or at the same time as the DeepSeek work, we also could have trained it for 5 million or whatever the advertised amount was.

And so what's impressive or surprising is that DeepSeek has gotten to the the frontier, but I think there's a common misconception still that they are above and beyond the frontier.

And I don't think that's right.

I think they just waited and then were able to take advantage of all the efficiency gains that everyone else was also seeing.

Yeah, they're exactly on the sort of cost curve that you'd expect, which I don't think they take away from the brilliant engineers and brilliant researchers who

I look at their work and I'm like, ah.

like the kindred soul

in the work they're doing.

Yeah.

And to go from like way behind the frontier to like, oh, this is like a real player.

It's super incredible.

Yeah, yeah, yeah.

Okay, so people say that they have good research taste.

Looking at their papers, what makes you say that?

Yeah.

I think their research taste is good in a way that I think Noam's research taste is good.

Noam Brown.

Noam Shizia.

Okay.

Noam Brown also has good research taste, but

where they very clearly understand this

dance between the hardware systems that you're designing the models around and the sort of like algorithmic

side of it.

And this is manifest in the way that the models give this sense of

being perfectly designed up to their constraints.

And you can really very clearly see what constraints they're thinking about as they're iteratively solving these problems.

And so, I mean, let's take the base transformer and diff that to deep sea v2 and v3.

You can see them running up against the memory bandwidth bottleneck in attention.

And you can see them initially they do MLA to do this.

They trade flops for memory bandwidth basically.

And then they do this thing called NSA where they like more selectively load memory bandwidth.

And you can see actually like this is because the model that they trained with MLA was on H800s.

So it has a lot of flops.

And so they were like, okay, we can freely use the flops.

But then the export controls from like from Biden came in,

or like, you know, they had less of, they knew they would have less of those chips going forward.

And so they traded off to a more memory bandwidth-oriented algorithmic solution there.

And you see a similar thing with their approach to sparsity, where they're iteratively working out the best way to do this over multiple papers.

And the part that I like is that it's simple.

A big failure mode that a lot of ML researchers have is that you do these overly complicated things that don't

think hard enough about the hardware systems that you have in mind.

Whereas

the first DeepSeek sparsity MOE solution,

they design these rack and

node-level load balancing losses.

So you can see them being like, okay, we have to perfectly balance it on this.

And then they actually come up with a much better solution later on where they don't have to have the auxiliary loss,

where they just have these bias terms that they put in.

Isn't that less simple?

You're manually putting in a bias rather than

balancing auxiliary losses and waiting.

Like you're making the model like trade off this thing.

And like you have to with auxiliary losses, you have to control the coefficient and the weighting.

The bias is like cleaner in some respects.

Interesting.

Did they have to change it through training?

They did have to change it during training.

Does all training involve

continuously fucking with these values as you're going through it?

It depends on what your architecture is.

But

I thought it was like, I just thought it was cute that you can can see them running up into this very hardware-level constraint.

What do we wish we could express algorithmically?

What can we express under our constraints?

And iteratively solving to get better constraints.

And doing this in a really simple and elegant way, and then backing it up with great engineering.

I also thought it was interesting that they incorporated the multi-token prediction thing from Meta.

So Meta had a nice paper on this multi-token prediction thing.

And actually, I don't know if it's good or bad, but Meta didn't include it in Lama, but Deep Sea did include it in their paper, which I think is interesting.

Was that because they were faster at iterating and including in the algorithm, or did Meta decide that actually it wasn't a good algorithmic change at scale?

I don't know.

It was really interesting to me as somebody who's had people on the podcast to discuss the, I mean, it's interesting from what's happening in AI right now, but also from the perspective of

I've been having abstract conversations with people about what an intelligence explosion would look like or what would it look like for AI to automate AI R and D, and just getting a more tangible sense of like what's involved in making this AI progress.

And I guess one of the questions I was debating with Daniel is

how much, or I was asking him, is how many of the improvements require a deep conceptual understanding versus how many are just like monkeys trying ideas and you could just run a bunch in parallel.

And it seems like the MLA thing is motivated by this

conceptual understanding of like, oh, each attention head only needs to see like the subspace that's relevant to its attention pattern.

I feel like that just required a lot of conceptual insight in a way that these models are especially bad at.

As opposed to, I don't know how the load balancing thing works, but that just seems like maybe you could try it out and see what happens.

Yeah, that's probably just like trying out a whole bunch of different things.

I mean, it might also be a problem.

So, what fraction is which?

I'd be curious about.

Yeah, I don't know about fractions.

It might be like you have a hunch for a core problem, you can think of 10 possible ways to solve it, and then you just need to try them and see what works.

And that's kind of where the trial and error sorcery of deep learning can kind of kick in.

And Noam Shazir will talk about this,

about how he, like 5% of his ideas work.

So even he,

the vaunted god of a model like design architecture design,

has like a relatively low hit rate, but he just tries so many things.

Right.

Or being able to come up with any ideas in the first place.

So

one mechanism could be that Noam just doesn't have to do any of the engineering work and he can just abstractly express an intuition.

Yeah.

I actually think

your rates of progress almost don't change that much depending on like so long as it's able to completely implement his ideas.

Interesting way.

Like if you if you have like Norm Jazir at 100x speed, that's still kind of wild.

Yeah.

Like there's all these like fallbacks of like, can of like wild worlds.

Yeah.

Where even if you don't get like 100% like Norm Shazir level intuition in model design, it's still okay if you just accelerate him by 100x.

Right.

Especially since you're computer bottlenecked anyway, so like trying out his his ideas.

Or I guess he doesn't have the computer to try out all of his ideas.

But Dorkes, you said, oh, well, the model can do the more straightforward things and not the deeper thought.

I mean, I do want to push back on that a little bit.

I think, again, if the model has the right context and scaffolding, it's starting to be able to do some really interesting things.

The Interp agent has been a surprise to people, even internally, at how good it is at finding the needle in the haystack, like when it plays the auditing game, finding this reward model bias feature, and then reasoning about it, and then systematically testing its hypotheses.

So it looks at that feature, then it looks at similar features, it finds one with a preference for chocolate.

It's like, huh, that's really weird that the model wants to add chocolate to recipes.

Let me test it.

And so then it will make up like, hey, I'm trying to make a tomato soup.

What would be a good ingredient for it?

And then sees that the model replies chocolate.

It reasons through it and then keeps going, right?

There's conceptual understanding of that.

Deep conceptual understanding.

And even where, like, especially it's spotted, it's like, oh, this is a a key part of its persona.

I see this Oxford paper.

What if I change Oxford to Stanford?

What if I now say Richard Feynman really likes this thing?

And it's like really carving out the hypothesis space and testing things in a way that I'm kind of surprised by.

Also, by the way, ML research is like one of the easier things to RL on in some respects once you get to a certain level of capability.

It's a very well-defined objective function.

Did the loss go down?

Make number go down.

Make number go down.

Or make

number go up depending on on which number it is.

Just flip the sign.

Just

once you get to the stage where models are capable of implementing one of no one's ideas,

then you can just let them loose and let them build that intuition of

how to do scientific discovery.

The key thing here, again, is the feedback loops.

I expect scientific areas where you are able to put it in a feedback loop to

have eventually superhuman performance.

One prediction I have is that we're going to move away from can an agent do XYZ and more towards can I efficiently deploy, launch a hundred agents and then give them the feedback they need and even just be able to like easily verify what they're up to.

Right.

There's this generator verifier gap that people talk about where it's like much easier to check something than it is to produce the solution on your own.

But it's very plausible to me we'll be at the point where it's so easy to generate with these agents that the bottleneck is actually, can I, as the human, verify the answer?

And again, you're guaranteed to get an answer with these things.

And so ideally, you have some automated way to evaluate and test a score for like how well it worked.

How well did this thing generalize?

And at a minimum, you have a way to easily summarize what a bunch of agents are finding.

And it's like, okay, well, if 20 of my hundred agents all found this one thing, then like it has a higher chance of being true.

And again, software engineering is going to be the leading indicator of that, right?

Like over the next six months, like the remainder of the year, basically we're going to see progressively more and more experiments of the form of how can I dispatch work to a software engineering agent in such a way that it's async.

Claude Ford's GitHub integration, where you can ask it to do things on GitHub, ask it to do pull requests, this kind of stuff that's coming up, and the OpenAS Codecs are examples of this, basically,

where

you can sort of almost see this in the coding startups.

I think of this like product exponential in some respects, where you need to be designing for like a few months ahead of the model to make sure that the product you build is the right one.

And you saw like last year, you know, Cursor hit PMF with Claude 3.5 Sonnet, right?

Like

they were around for a while before, but then the model was finally good enough that the vision they had of how people would program like hit.

And then

Windsurf bet a little bit more aggressively even on the agenticness of the model.

like with longer running agentic workflows and this kind of stuff.

And I think that's when they sort of like began competing with cursors, when they bet on that particular vision.

And the next one is you're not even in the loop, so to speak.

You're not in an IDE, but you're asking the model to go do work in the same way that you would ask someone on your team to go do work.

And

that is not quite ready yet.

Like there are still a lot of tasks where you need to be in the loop.

But the next six months looks like an exploration of exactly what does that trend line look like?

Yeah, but just to be really concrete or pedantic about the bottlenecks here, a lot of it is again, just tooling and are the pipes connected.

A lot of things I can't just launch clawed and have it go and solve because maybe it needs a GPU or maybe I need very careful permissioning so that it can't just like take over an entire cluster and like launch a whole bunch of things, right?

So you really do need good sandboxing and the ability to use all of the tools that are necessary.

And we're almost certainly like under-eliciting dramatically.

When you look at Meta's evals of can the model solve the task, they're there solving them for like hours over like multiple iterations.

And eventually one of them is like, oh, yeah, I'll come back and I've solved the task.

Me at the moment, at least like maybe the fault is my own, but I try the model and doing something.

And if it can't do it, I'm like, okay, fine, I'll do it.

Which is so interesting because we don't even treat other humans this way.

Right.

You hire a new employee, you're not like, I'll do it.

Yeah, yeah, yeah.

You're like, give them like, like spend literally weeks giving them feedback where like, we'll go up to the model in like minutes.

Yes, exactly.

But I think part of it is, is it async or not?

Yes.

And if it's human in the loop, then it's so much more effortful.

And unless it's getting replying immediately.

I've noticed if I don't have a second monitor with claude code always open in the second monitor, I won't really use it.

Yeah, yeah.

It's only when it's right there and it's, I can send off something if it hits great.

If not, I'm kind of working on it at the same time.

Yeah, yeah.

But this more async form factor, I expect to really quite dramatically improve the experience of these models.

Interesting.

Interesting.

You can just say,

let's see if it can do that.

Yeah.

Let's give it a while.

Try 10 different approaches.

Yeah.

Yeah.

Just fire it off.

Fire it off.

But before we end this episode, I do want to get back at this crux of

why does the progress that you're talking about in computer use agents and white-collar work happen over the next few years?

Why is this not a thing that takes decades?

And I think the crux comes down to

the people who expect something much longer have a sense that

when I interrupt I gain time on my podcast, they were like, look,

you could look at AlphaGo and say, like, oh, this is a model that can do exploration.

It can, like, AlphaZero can generalize to new video games.

It has all these priors about how to engage with the world and so forth.

And the intellectual ceiling is really high.

Yeah, exactly.

And then, in retrospect, obviously a bunch of the methods are still used today in deep learning.

And you can

see similar things in the models that we train today.

But it was fundamentally not a sort of like baby AGI that we just had to like add a little sprinkle of something else on top of in order to make it the LLMs of today.

And

I just want to very directly address this crux of why are LLMs

in

a much different position with respect to true AGI than AlphaZero.

Why are they actually the base on which

adding in a few extra drops of this kind of care and attention

gets us to human-level intelligence?

I think

one important point is that

when you look at AlphaZero, it does have all of those ingredients.

And in particular, I think the intellectual ceiling goes quite contra what I was saying before, which is we've demonstrated this incredible complexity of math and programming problems.

I do think that the type of task and setting that Alpha Zero worked in, this two-player perfect information

game, basically, is incredibly friendly to IRL algorithms.

And the reason it took so long to get to a more proto-AGI style models is you do need to crack that general conceptual understanding of the world and language and this kind of stuff.

And you need to get the initial reward signal on tasks that you care about in the real world, which are harder to specify than games.

And I think then that

sort of like gradient signal that comes from the real world, all of a sudden you get access to it and you can start climbing it.

Whereas AlphaZero didn't ever have the first rung to pull on.

This goes back to the monkeys on the typewriter, I think, and like the pre-training model.

And until you had something like GPT-3, GPT-4, it just couldn't generate coherent enough sentences to even begin to do RLHF and tell it what you liked and didn't like.

Yeah.

If we don't have even reasonably robust

or weakly robust computer use agents by this time next year, are we living in

the bust timeline as in 2030 or bust?

I would be extremely surprised if that was the case.

And I think that would be like somewhat of an update towards like, there's something like strangely difficult about this, like computer use in particular.

I don't know if it's the bust timeline, but it's definitely like the, I would update on this being

like

lengthening of timeline.

Yeah, I mean, I think more and more it's no longer a question of speculation.

If people are skeptical, I'd encourage like using clawed code or like some agentic tool like it and just seeing what the current level of capabilities are.

Tweeting is so much easier.

But seriously, like the models are getting really capable at tasks that we care about and we can give them enough data for.

And I mean, the circuits results from interpretability are also pointing in the direction that they're doing very reasonable, generalizable things.

And so,

yeah,

this question matters a lot, but

I'm surprised by how many deep learning critics just like haven't really interacted with the models or haven't in a while.

And constantly move the goalposts.

Yeah.

Yeah.

Yeah.

Like the Turing test used to be a thing, right?

Like

we don't even talk about it and it'd be like silly to think that it was a meaningful test.

Yeah.

Yeah.

Now, that being said, one caveat on that is like, if software engineering is just like dramatically better than computer use, I mean, computer use still sucks, then I'd be like, still like, oh, maybe everyone just kept focusing on software engineering.

Like, it was just by far the most valuable thing.

Like, every marginal person and dollar went towards software engineering.

I don't think that's the case.

I do think computer use is valuable enough that

people will care about it.

But that would be like my, that's my one escape patch that I'm putting in place for next year.

Yeah.

It would be good from a lighting perspective too, because I think you kind of do need a wider range of skills before you can do something super, super scary.

Oh, like as in if the models didn't get any better.

Yeah.

If it's like disproportionate, they're superhuman coders,

but they're not like Henry Kissinger level.

I don't know, that seems okay.

Like if we have AI oracles.

Yeah, that's what I'm saying.

That's good.

Yeah, yeah, exactly.

Yeah, that's something that's good.

Yeah.

So if you look back at AI discourse, like going back a decade, there's a sense that there's dumb AI, then there's AGI, then there's ASI, that intelligence is the scalar value.

The way you've been talking about these models has a sense of jaggedness.

It's especially tuned to environments in which it's been trained a lot or has a lot of data.

Is there a sense in which like there's it still makes sense to talk about the general intelligence of these models?

Is there enough meta-learning and transfer learning that is distinguished between like the sizes of models or like the way models are trained?

or are we moving into a regime where it's not about intelligence it's more so about domain yeah so one intuition pump is this conversation was had a lot when models were like GPT2 sized and fine-tuned for various things

and they found you know people would find that the models were dramatically better at things that they were fine-tuned for right

but by the time you get to GPT-4 when it's trained on a wide enough variety of things actually the like the sort of total compute like it generalized very well across all of the individual subtasks, actually generalized better than smaller fine-tuned models in a way that was extremely useful.

I think right now what we're seeing with RL is pretty much the same story playing out, where there's this jaggedness of like things that they're particularly trained at.

But as we expand the total amount of compute that we do RL with, you'll start to see the same transition from like GPT-2 fine-tunes to like GPT-3, GPT-4, like unsupervised meta-learning and generalization across things.

And I think we're already seeing early evidence of this in its ability to generalize reasoning to things.

But

I think this will be extremely obvious soon.

One nice example of this is just the ability or notion to backtrack.

You go down one solution path.

Ooh, wait, let me try another one.

And this is something that you start to see emerge in the models through RL training on harder tasks.

And

I think right now it's not generalizing incredibly well, at least with.

Well, I mean, have we ever RL'd the model to be a interp agent?

No.

I mean, no, yeah, exactly.

Yeah.

Like, so all this time we're talking about, like, oh, it's only good at things that's being RL that.

Well, it's pretty good at that because that's pretty, you know, that is a mixture of like science and like understanding language and like, so, uh, and coding.

Um, like, there's this sort of like mixture of domains here, all of which you need to understand.

Like, you need to be both a great software engineer and be able to like think through language and like state of mind and almost philosophize in some respects to be an interp agent yeah um and it is generalizing from the training yeah yeah to do that what's the end game here claud 8 comes out um and they give it to you and

dot dot dot you say thumbs up what's happened what are you yeah i mean it really depends upon the timeline at which we get Claude 8 and the models hit like ASL4 capabilities, right?

Like fundamentally, we're just going to use whatever tools we have at the time and see how well they work.

Ideally, we have this enumerative safety case where we can almost verify or prove that the model will behave in particular ways.

In the worst case, we use the current tools like when we won the auditing game of seeing what features are active when the assistant tag lights up.

Can you explain what is mechanistic interpretability?

What are features?

What are circuits?

Totally.

Yeah.

Yeah.

So mechanistic interpretability, or the cool kids call it mechanterp, is trying to reverse engineer neural networks and figure out kind of what the core units of computation are.

Lots of people think that because we made neural networks, because they're artificial intelligence, we have a perfect understanding of how they work.

And it couldn't be further from the truth.

Neural networks, AI models that you use today, are grown, not built.

And so we then need to do a lot of work after they're trained to figure out to the best of our abilities how they're actually going about their reasoning.

And so,

two and a half, three and a half years ago,

this kind of agenda of applying mechanistic interpretability to large language models started with Chris Ola leaving OpenAI, co-founding Anthropic.

And

every roughly six months since then, we've had kind of like a major breakthrough in our understanding of these models.

And so, first with Toy Models of Superposition, we established that models are really trying to cram as much information as they possibly can into their weights.

And this goes directly against people saying that neural networks are overparameterized.

And like classic AI machine learning back in the day, you would use linear regression or something like it, and people had a meme of AI or neural networks, deep learning,

using way too many parameters.

there's like this funny meme that you should show of like layers on the x-axis and layers on the y-axis and this like jiggly line that just goes up.

And it's like, oh, just throw more layers at it, right?

But it actually turns out that at least for really hard tasks like being able to accurately predict the next token for the entire internet, these models just don't have enough capacity.

And so they need to cram in as much as they can.

And the way they learn to do that is to use each of their neurons or units of computation in the model for lots of different things.

And so if you try to make sense of the model and be like, oh, if I remove this one neuron, or like, what is it doing in the model?

It's impossible to make sense of it.

It'll fire for like Chinese and fishing and horses and I don't know, just like a hundred different things.

And it's because it's trying to juggle all these tasks and use the same neuron to do it.

So that's superposition.

Nine months later, we write towards monosemanticity, which introduces what are called sparse autoencoders.

And so going off what I just said of the model trying to cram too much into too little space,

we give it more space,

this higher dimensional representation, where it can then more cleanly represent all of the concepts that it's understanding.

And this was a very toy paper in so much as it was a two-layer, really small, really dumb transformer.

And we fit up to, I want to say, 16,000 features, which we thought was a ton at the time.

Fast forward nine months, we go from a two-layer transformer to our Claude 3 Sonnet frontier model at the time, and fit up to 30 million features.

And this is where we start to find really interesting abstract concepts like a feature that would fire for code vulnerabilities.

And it wouldn't just fire for code vulnerabilities, it would even fire for, like, you know, that Chrome page you get if you like, it's not an HTTPS URL, and it's like warning this site might be dangerous, like click to continue.

And like also fire for that, for example.

And so it's like these much more abstract coding variables or sentiment features amongst the 30 million.

Fast forward nine months from that, and now we have circuits.

And I threw in the analogy earlier of the Ocean 11 heist team, where now you're identifying individual features across the layers of the model that are all working together to perform some complicated task.

And you can get a much better idea of how it's actually doing the reasoning and coming to decisions, like with the medical diagnostics.

One example I didn't talk about before is with like how the model retrieves facts.

And so you say, like, what sport did Michael Jordan play?

And not only can you see it hop from like Michael Jordan to basketball, answer basketball, but the model also has an awareness of when it doesn't know the answer to a fact.

And so so by default, it will actually say, I don't know the answer to this question.

But if it sees something that it does know the answer to, it will inhibit the I don't know circuit and then reply with the circuit that it actually has the answer to.

So for example, if you ask it, who is Michael Batkin, which is just a made-up fictional person,

it will by default just say, I don't know.

It's only with Michael Jordan or someone else that it will then inhibit the I don't know circuit.

But what's really interesting here and where you can start making downstream predictions or reasoning about the model is that that I don't know circuit is only on the name of the person.

And so in the paper, we also ask it, what paper did Andre Karpathy write?

And so it recognizes the name Andre Karpathy, because he's sufficiently famous.

So that turns off the I don't know reply.

But then when it comes time for the model to say what paper it worked on, it doesn't actually know any of his papers.

And so then it needs to make something up.

And so you you can see different components and different circuits all interacting at the same time to lead to this final answer.

Aaron Ross Powell, why I think it's a tractable problem to

understand every single thing that's happening in a model, or like that's the best way to understand why it's being deceptive.

If you wanted to explain why England won World War II using particle physics,

you would just be on the wrong track.

You just want to look at the high-level explanations of who had more weapons?

Like, what did they want?

And that seems analogous to just training linear probes for, like, are you honest?

Are you being deceptive?

Like, do we catch you doing bad things when we're red teaming you?

Can we monitor you?

Why is this not analogous where we're asking a particle physicist to just backtrack and explain why

England won World War II?

I feel like you just want to go in with your eyes wide open, not making any assumptions for what that deception is going to look like or what the trigger might be.

And so the wider you can cast that net, the better.

Depending on how quickly AI accelerates and where the state of our tools are, we might not be in the place where we can like

prove from the ground up that everything is safe.

But

I feel like that's a very good North Star.

It's a very powerful, reassuring North Star for us to aim for, especially when we consider we are part of the broader AI safety portfolio.

I mean, do you really trust, like, you're about to deploy this system and you really hope it's aligned with humanity and that you've like successfully iterated through all the possible ways that it's going to like scheme or sandbag.

But that's also probably going to be true with whatever you find.

You're not, I mean, you're not, you're still going to have variants that you haven't explained, or like you found a feature, but you don't know if it actually explains deception or something else instead.

So, so I guess, first of all, I'm not saying you shouldn't try the probing approach, right?

Like, we want to pursue the entire portfolio.

We've got the therapist interrogating the patient by asking, Do you have any troubling thoughts?

We've got the linear probe, which I'd analogize to like a polygraph test, where we're taking like very high-level summary statistics of the person's well-being.

And then we've got the neurosurgeons kind of going in and seeing if you can find any brain components that are activating in troubling or off-distribution ways.

So I think we should do all of it.

What percent of the alignment portfolio should it in Mecca be?

I I think as much of a chunk as is necessary.

I mean,

I think at least like

yeah,

hard to define, but I don't know.

Anthropic, I feel like all of the different portfolios are like being very well supported and growing.

You can also coming back to like the World War II question, you can think of it as like a hierarchy of abstractions of trust here, where like, let's say you want to go and talk to like Churchill.

It helps a lot if you can verify that in that conversation, in that 10 minutes, he's being honest.

And this enables you to construct better meta-narratives of what's going on.

And so maybe particle physics wouldn't help you there, but certainly the neuroscience of Churchill's brain would help you verify that he was being trustworthy in that conversation and that

the soldiers on the front lines were being honest in their depiction of their description of what happened and this kind of stuff.

So long as you can verify progress like parts of the

tree up, then that massively helps you build confidence.

I think language models are also just really weird, right?

Like with the emergent misalignment work,

I don't know if they took predictions they should have of like, hey, I'm going to fine-tune ChatGPT on code vulnerabilities.

Is it going to become a Nazi?

And I think most people would have said no.

And that's what happened.

And so

what are the different things?

How did they discover that it became a Nazi?

They started asking it a ton of different questions, and it will do all sorts of like vile and harmful things.

Like the whole persona just totally changes.

And I mean, we are dealing with alien brains here who don't have the social norms of humans and or

even a clear notion of like what they have and haven't learned

that we have of them, I mean.

And so I think you really want to go into this with eyes wide open.

Backing up for Mechanurf, if we live in a world where AI progress accelerates,

By the way, you were mentioning a little while ago that there's many wild worlds we could be living in, but we're living at least one of them.

Another one that we've gestured at, but it's worth making more explicit, is this.

Even if the AI models are not helping write the next training algorithm for their successor, just the fact that if they had human-level learning efficiency,

whatever a model is learning on the job, or whatever copy of the model is learning on the job, the whole model is learning.

So, in effect, it's getting.

Or if they're like a thousand times less efficient than humans are learning.

That's right, and you just like deployed them, even still.

Exactly.

Yeah, yeah.

Anyways, and then there's a whole bunch of other things that you can think about.

But even there, it's like you kind of have a broadly deployed intelligence explosion.

And I do think it's worth pressing on that future of,

you know, there is this whole spectrum of crazy futures.

But the one that I feel we're almost guaranteed to get, and this is like a, like, a, almost a strong statement to make, is one where, like, at the very least,

you get drop-in like white-collar worker

at some point in the next five years.

It's like, I think think it's very likely in two,

but it seems almost over-determined in like five.

And on the grand scheme of things, those are kind of irrelevant timeframes.

It's the same either way.

And that completely changes the world over the next decade.

And

if we don't have the right policies in place for that, then you end up actually with almost in some respects like a fundamentally worse world.

Because the thing that these models get good at by default is like software engineering and computer using agents and this kind of stuff.

And then to, we will need to put in extra effort to put them in the loops where they help us with scientific research or they're like we have the right robotics such that we actually like experience an increase in material quality of life.

So that's worth thinking about.

Like if you're in the perspective of like, I'm a country, like what should I be doing or thinking about?

Like plan for the case where white-collar work is automatable.

And then consider what does that mean for your economy and what you should be doing to prepare policy.

What should you be doing to prepare?

because honestly honestly it's like it's such a tough question yeah where like if you're india or nigeria or australia yeah um if you're a country unlike uh america or china where they do have frontier models yeah um

what is it that you should be doing right now especially on such a short time scale yes uh so i think one very important point is that let's say this this scenario turns out true then compute becomes the most valuable resource in the world.

Like the sort of GDP of your economy is dramatically affected by how much compute you can deploy towards the sort of organizations within your country.

And so having some

guaranteed amount of compute, I think, will actually be quite important.

So like pre getting ahead of investments in data centers and this kind of stuff on the condition that it's like companies in your country have to be allowed to use that compute.

Not necessarily for training, but just even just for inference.

I think the economic value here comes from inference.

I think it also makes sense to invest broadly in AI.

I think these countries have the opportunity to do so.

And I think that's a portfolio of

foundation model companies, but also robotic supply chain and this kind of stuff.

I think that you should invest very proactively in policies that try and

prevent capital lock-in.

So we're in for a much worse world if it just so happens that the people who had money in the stock exchange or in land before AGI are dramatically more wealthy than the people who don't.

Because it's a gross misallocation of resources.

So,

having

like, I know, one of my favorite episodes actually on your podcast was like the Georgism one, where you're trying to appropriately value or allocate land.

And so, I think

this strikes particularly close to home coming from Australia, where I think our

policies with respect to land are grossly wrong.

But I think this is broadly true.

Being very

forward on regulation of integration of these models into

your country

is important.

And proactively making sure that people have choice.

So let's say you should be quite proactive about making sure that the phones or devices or glasses that people have, people have free choice on what things they run.

And then, so that's like, that's the, we just get white collar worker, right?

And like you're trying to like do the best to like prepare your country for that.

And then it's like, okay, well, what can you do to like make all possible versions of the future go like well?

Like that's like covering some amount of like economic downside.

The other like things that I think are really important is like

figure out how you can like

either make the like basically ensure dramatic upside or cover like terrible downside.

And so like getting a dramatic upside is like making sure that there like is investment in biology like biology research and this kind of stuff in an automated way that is like

these models are actually like able to produce novel medicines that massively like massively improve our like quality of life.

And covering the downside is like AI alignment research and this kind of stuff and automated testing and like really thinking hard about that.

AI safety institutes, this kind of stuff.

But these seem like things that a rich person, a random rich person could also do.

Like there's an it seems like

there's not a thing that a nation state is uniquely equipped to do.

Yeah.

That's a good point.

In this scenario.

I mean, like, dramatic allocation of resource towards compute, I think, is sensible.

I would be doing that if I was in charge of a nation state.

I think it just increases your optionality in most of the future worlds.

Dylan Patel has some scary forecasts on U.S.

energy

versus China.

Yes.

Yeah, we're like 34 gigawatts off.

Yeah, yeah.

The US's line is flat, basically, and China's line is like this.

And I mean, the US, like very clearly.

Yeah, we just need so many more power plants.

Yes.

If intelligence becomes this incredibly valuable input, intelligence becomes almost a raw input into the economies and quality of life of the future, the thing directly underneath that is energy.

And so making sure that you have

incredible amounts of solar, like tile the desert in solar panels, some parts of the desert in solar panels,

would be helpful towards making sure that you have

more access to intelligence on top.

Yeah, just to make it explicit, because we've been touching on it here.

Even if AI progress totally stalls, you think that the models are really spiky and they don't have general intelligence, it's so economically valuable and sufficiently easy to collect data on all of these different jobs, these white-collar job tasks, such that to Shalto's point,

we should expect to see them automated within the next five years.

Yeah.

Even if you need to hand spoon every single task to the model.

It's like economically worthwhile to do so.

Even if algorithmic progress stalls out and we just never figure out how to keep progress going, which I don't think is the case, that hasn't stalled out yet.

It seems to be going great.

The current suite of algorithms are sufficient to automate white-collar work provided you have enough of the right kinds of data.

Yes.

And in a way that compared to the TAM of salaries for all of those kinds of work is so trivially worthwhile.

Yeah, exactly.

I do just want to flag as well that there's a really dystopian future if you take Moravec's paradox to its extreme, which is this paradox where we think that the most valuable things that humans can do or the smartest things are like add large numbers in our heads or do any sort of white collar work.

And then we totally take for granted our fine motor skill and coordination.

But from an evolutionary perspective, it's the opposite.

So we got like evolution has optimized fine motor coordination so well.

And even if you look at like robot hands or like even the ability to open a door is still just like really hard for robots.

Meanwhile, we're seeing this total automation of coding and everything else that we've seen is clever.

The really scary future is one in which AIs can do everything except for the physical robotic tasks, in which case you'll have humans with like AirPods and like

glasses, and there'll be some robot overlord controlling the human through cameras by just like telling it what to do and like having a bounding box around the thing you're supposed to pick up.

And so you have like human meat robots.

And not like necessarily saying that like that's what the AIs would be like want to do or anything like that, but as in if you were to be like, what are the relative economic value of things?

The AIs are out there doing computer programming and the most valuable thing that humans can do is be amazing robots.

Now, that being said, I think Morofc's paradox is a little bit fake.

I think the main reason that robots are worse

at being a robot than they are at software engineering is the internet exists for software engineering.

GitHub exists.

There is no equivalent thing.

If you had

all,

mocap of everyone's actions as they were like going about their daily lives um for like some reasonable fraction of the human population robotics is also like close to solved like like like on track to be solved at the same rate that software engineering is on track to be solved um so this is only like this vision is only like a sort of decade long section but it's still pretty terrible decade like imagine the world where uh people have lost their jobs you haven't yet got novel biological research that means like people's quality of life is like dramatically better you don't yet have material abundance because you haven't actually been able to action like action the physical world in the necessary way.

You can't build dramatically more because that's like building dramatically more takes robots, basically.

And people's

main comparative advantage is as fantastic robots is like a shocking, shocking world.

The perpetrator of an average human, I think it actually might be better.

Your wages will be higher because you're the complement to something that is enormously valuable, which is AI labor.

Right.

And

a decade or two on, the world is fantastic.

Robotics is solved and you decide to get radical abundance, basically, provided that you have all the policies set up necessary to permit building.

You end up with that same change from

the before and after photos of Shanghai.

We're like 20 years on, it's just this dramatically transformed city.

A lot of places in the world probably end up like that

over that two-decade period.

But we need to

make sure,

like, one, do our best to estimate, is this actually what is on track to happen?

Like, build Sweetbench before all the other forms of white-collar work and measure and track.

That's a great thing that governments should be doing, by the way, is like trying to break down the sort of functions of their economy into measurable tasks and figuring out where, what does the curve actually look like for that?

Because they might be a bit shocked by the progress there.

You know, there's no sweet bench for tax,

like tax eval.

And

then,

I don't have all the answers here, but figuring out a way to share the proceeds of this economy broadly across people or invest heavily in robotics and collecting the data so that we get robotics faster and get material abundance faster, invest in biological research that we get, but all that faster.

Basically, try and pull forward the radical upside.

Because otherwise, you have a pretty dark

section.

I think one thing that's not appreciated enough is how much of our leverage on the future, given the fact that our labor isn't going to be worth that much, comes from our economic and political system surviving.

For your million X S P equity to mean something,

for your contracts to mean anything, for the government to be able to tax the AI labor and give you a UBI off of that, it just like that requires our legal institutions, our economic institutions, our financial real surviving into the future.

The way in which that likely happens is if it's also in the AI's best interests that they follow those rails.

And by AI, I don't mean some monolithic single AI.

I just mean like firms which are employing AI and becoming more productive as a result.

You don't want to be in a position where it's so onerous to operate in our system that you're basically selecting for firms who either emigrate or who are like doing black market stuff, et cetera.

And which means, I think, like you want to make it super, super easy to deploy AI,

have the equivalent of special economic zones, et cetera.

Because otherwise, you are just surrendering the future outside of any control that you might have on it.

One of the reasons, by the way, that I worry about

turning AGI into a national security issue or having it have extremely close ties with the government, the Manhattan Project thing, is that it disproportionately redirects the use of AI towards military tech and the mosquito drones and whatever.

And

also naturally puts other countries in the same frame of mind, right?

If we're developing the mosquito drones, why would China not develop the mosquito drones?

And that just seems like a zero-sum race, and not to mention a potentially catastrophic one.

Whereas, like,

you know, like compute will be limited.

You know, we want, we will need to disproportionately accelerate some things.

To the extent it just remains totally like a consumer free market landscape, it just seems more likely that we'll get the glorious transhumanist future where they're developing the things that make human life better.

Yes, I mean, I agree.

Like the case where you end up with like two national projects facing off against each other is dramatically worse.

Right.

Like we don't want to live in that world.

Yeah.

It's much better if this like stays.

a free market, so to speak.

Yeah, yeah, yeah.

Okay.

I want to take issue with your claim that even if with the algorithms of today, if we just collect enough data,

that we could automate white colour work.

First, let me get an understanding of what you mean by that.

So do you mean that we would do the analogous thing of free training with all the trajectories of everything people would do on their jobs?

Could you make either manually or through some other process

some RL procedure based on the screen recordings of every white colour worker?

What kind of thing are you imagining?

I mean, like a continuous distribution of this stuff.

One important mental model to think about RL is I think as

the task gets more,

there is some respect with which like longer horizon or better at that task, if you can do them, if you can get that reward ever, are like easier to judge.

So again, come back to that, can you make money on the internet?

That's an incredibly easy reward signal to judge.

But to do that, there's a whole hierarchy of complex behavior.

So if you could pre-train up to the easy to judge reward signals, does your website work?

Does it go down?

Do people like it?

There's all these reward signals that we can respond to because

we can progress through these long enough trajectories to actually get to interesting things.

If you're stuck in this regime where like you need a reward signal every five tokens, it's a way more painful and like long process.

But if you could like pre-train on every like screen in America,

then probably the like RL tasks that you can design are very different to like if you could only like like take the existing internet as it is today.

And so like how much of that you get access to like changes the mix.

Interesting.

So as we're training them on longer and longer horizon tasks and it takes longer for them to get any signal on whether they successfully completed that task, will that slow down progress?

Because it takes more compute per task.

I do think there's this notion the longer, the harder tasks, the more training is required.

And I'm sympathetic to that naively, but we as humans are very good at practicing the hard parts of tasks and decomposing them.

And I think once models get good enough at the basic stuff, they can just rehearse or fast forward to the more difficult parts.

I mean, that's definitely one of the complexities, right?

Like, as you use more compute and like the, and as you train on like more and more difficult tasks, I mean, I don't know, your rate of improvement at biology is going to be like somewhat bound by the time it takes a cell to grow in a way that your rate of improvement on math isn't, for example.

So

yes,

but I think for many things we'll be able to parallelize

wisely enough and get enough iteration loops that

will the regime of training new models go away?

Will we eventually get to like you've got the model and then you just keep adding more skills to it with RL training?

That depends on whether or not you think there's a virtue in pre-training a new architecture.

Basically, if you make some architectural change,

then you probably need to do some form of at least retraining a new model.

How it is the fact that

if RL requires a bunch of inference to do the training in the first place, does that push against the thing you were talking about where we actually need a bigger model in order to have brain-like energy?

But then also it's more expensive to train it in RL.

So where does that balance out?

I think we got to drink the bitter lesson here.

And yeah, like you, there aren't infinite shortcuts.

Like you do just have to scale.

Something's going going to have a bigger model and pay more inference for it.

And

if you want AGI, then that's what you got to pay the price of.

But there's a trade-off equation here, right?

Of like

there is science to do, which everyone is doing, of what is the optimal point at which to do RL.

Because you need something which can both learn and discover the sparse reward itself.

So you don't want a one-parameter model, useless, even though you can run it really fast.

You also don't want

a 100T model because it's super slow.

Yeah.

Password RL.

And

the marginal benefit of its learning efficiency is not worth it.

So there's a pre-do frontier here.

What's the optimal model size at your current class of capabilities and your current set of RL environments and this kind of stuff?

Yeah.

And even in the last year, there's been much more of a factor of the inference cost.

So just explicitly, the bigger the model, the more expensive it is to do a forward pass and generate tokens.

And the calculus used to just be: should I allocate my flops to more training data or a bigger model?

And now another huge factor is how much am I actually going to do forward passes on this model once it's trained?

Yeah, my total pool of compute.

How do I allocate that across train data, compute, and inference compute for the RL training?

And then even within inference, there's all this research on, well, what strategy should I use?

Should I sample 10 and take the best?

Do I do this sort of like branching search, et cetera, et cetera?

And so with RL, where you're sampling a whole lot of tokens, you also need to factor in the ability for the model to actually generate those tokens and then learn and get feedback.

Okay.

So if we're living in this world, what is your advice to somebody early in their career or a student in college?

How should they be,

what should they be planning on doing?

Yeah.

So I think, once again, there's like, it's worth considering the spectrum of possible worlds and preparing yourself for that.

And the one, like, the sort of action that I think is like highest EV in that case is you are about to get dramatic, at a minimum, you are about to get dramatically more leverage.

You already have.

Already the startups and YC are like writing huge amounts of their code with Claude.

So what challenges, what causes do you want to change in the world with the added leverage?

Like if you had 10 engineers at your beck and call, what would you do?

Or if you had a company at your beck and call, like what would that enable you to do?

And what problems and domains suddenly become tractable?

That's the world you want to prepare for.

Now that still requires a lot of technical depth.

Obviously, there is the case where AI just just becomes dramatically better than everyone at everything, right?

But for at least a while, probably, there is like advantage.

I think Jensen actually talked about this in an interview in an interesting way where he's like, you know, I have like 100,000 general intelligences around me, and I'm still somewhat useful because I'm there, like, you know, directing the values and

asking them to do things.

And, you know, they're still like, I still have value, even though I have 100,000 general intelligences.

And for many people, I think that will still be true for a fair while.

And then, you know, as the AIs get better and better and better and so on, eventually, no.

but

again, prepare for like the spectrum of possible worlds.

Because in the event where we're just totally outcompeted, yeah, it doesn't matter what you do.

But in all the other worlds, it matters a lot.

Get the technical depth.

Study biology.

Study CS.

Really think hard about, study physics, think about hard about what challenges you want to solve in the world.

Yeah.

That's a lot of topics.

That's a lot now.

You can, right?

Like, it's so much easier to learn.

Everyone now has the infinite perfect tutor.

Yeah, yeah, yeah.

It's definitely been helpful to me.

Yeah.

I would say some combination of like get rid of the sunk cost of your like previous workflows or expertise

in order to evaluate what AI can do for you.

That's right.

And another way to put this, which is fun, is just like be lazier in so much as like figure out the way that the agent can do the things that are toilsome.

But you're going to have to.

Ultimately, you get to be lazier, but in the short run, you need to critically think about the things you're currently doing and what an AI could actually be better at doing and then go and try it or explore it.

Because I think there's still just a lot of low-hanging fruit of people assuming and not writing the full prompt, giving a few examples,

connecting the right tools

for your work to be accelerated, automated.

Yeah.

Yeah.

There's also the sunk cost of feeling like since you're not quote unquote early to AI, that you've sort of missed the boat and you can't like, but

I think, I mean, I remember when GPT-3 came out, so backstory on the podcast, when I graduated college, I was planning on doing some sort of AI rapper startup.

And the podcast was just like a gateway into doing that.

And so I was trying out like different things.

And at the time, I remember thinking, oh, 3.5 is out.

And people are like, I'm like so behind on like the startup scene here or whatever if I wanted to make my own rapper.

I mean, maybe the idea of the rapper was inadvisable in the first place, but just like that, every time feels early because like it's sort of, if it's an exponentially growing process.

And there were many things and many ideas which are only becoming possible now, right?

So exactly.

It's that product experience I talked about before.

Like products literally obsolete.

Like you need to constantly reinvent yourself to stay at the frontier of capabilities.

By the way, I do remember I had a really shitty idea and I gave you a call.

I don't know what it was.

It was like.

I think it was like rag for like lawyers or something.

Anyways,

I think one of our first interactions was I'm like, hey, what do you think of this idea?

And you're like, I think the podcast sounds promising.

That was right,

which I appreciate.

I got slightly annoyed at a friend recently who I think is really talented and clever and interested in AI, but has pursued a biology route.

And I just kind of tried to shake them of like, you can work on AI if you want to.

I mean, I think humans are artificial, or not artificial, are biological general intelligences where a lot of the things of value are just very general.

And whatever kind of specialization that you've done

maybe just doesn't matter that much.

I mean, again, it gets back to the sunk cost.

But like, so many of the people, even my colleagues at Anthropic, are excited about AI and they just don't let their previous career be a blocker.

And because they're just like innately smart, talented, driven, whatever else,

they end up being very successful and finding roles.

It's not as if they were in AI forever.

I mean, people have come from totally different fields.

And so don't think that you need like permission from some abstract entity to like get involved and apply and be able to contribute.

If somebody wanted to

be an AI researcher, like right now, if you give them an open problem, or like the kind of open problem that is very likely to

be quite impressive.

What would it be?

I think that now that RL's come back,

papers building on Andy Jones's scaling board, scaling laws for board games are interesting.

Showing that you can

investigating these questions like the ones you asked before, where you're like, oh, is the model actually learning to do more than its previous parse at K?

Or is it just

discovering that?

Exploring questions like that deeply, I think are interesting.

Yeah, yeah.

Like scaling lows for RL, basically.

I'd be very curious to see how much

the marginal increase in meta-learning from a new task or something.

I mean, on that note, I think model diffing has a bunch of opportunities.

Also, people say, oh, we're not capturing all the features.

There's all this stuff left on the table.

What is that stuff that's left on the table?

If the model's jailbroken, is it using existing features that you've identified?

Is it only using the error terms that you haven't captured?

I don't know.

There's a lot here.

I think Matt's is great.

The Anthropic Fellowship has been going really well.

Goodfire Anthropic invested in recently.

They're doing a lot of interpretability work or just applied to it.

Anything to get your equity up, huh?

There's just so many interpretability projects that are like, there's so much low-hanging fruit, and we need more people.

And I don't think we have much time.

Yeah.

I also want to make a plug for performance engineering.

I think this is one of the

best ways to sort of demonstrate that you have like the raw ability to do it.

If you made a extremely efficient transformer implementation on TPU or Tranium or in CUDA,

then

I think there's a pretty high likelihood that you'll get a job offer.

There is a relatively small pool of people that you can trust to like

completely own end-to-end the performance of a model.

And if you have broad, deep electrical engineering skills, I think you can probably come up to speed pretty fast on a accelerator still.

Yeah, you can come up to speed reasonably fast.

And it teaches you a lot of good intuitions of the actual intricacies of what's going on in the models, which means that you're then very well placed to think about architecture and this kind of stuff.

One of my favorite people in thinking about architecture at Anthropic at the moment actually came from a heavy GPU kernel programming background, just like Knows the Internet really deeply and can think about the trade-offs really well.

This is fun, guys.

Thanks for watching.

Thanks.

Great to be back.

I hope you enjoyed this episode.

If you did, the most helpful thing you can do is just share it with other people who you think might enjoy it.

Send it to your friends, your group chats, Twitter, wherever else.

Just let the word go forth.

Other than that, super helpful if you can subscribe on YouTube and leave a five-star review on Apple Podcasts and Spotify.

Check out the sponsors in the description below.

If you want to sponsor a future episode, go to dwarcache.com/slash advertise.

Thank you for tuning in.

I'll see you on the next one.