Hi, listeners.
Welcome back to No Priors.
Today, we're here with Eric Zellkman, previously of Stanford and XAI.
We're going to talk about the contributions he's made to research, reasoning and scaling up RL, as well as his new company, Humansend.
Eric, thank you so much for doing this.
Thank you.
You have had an amazing impact as a researcher, including starting from just your time at Stanford.
I want to hear about that.
But first background of how you got interested in machine learning at all.
I guess, going back like really far, I've been motivated by this question of like you have.
You know, all of these people out there have like all of these things that they're really talented in, all of these things that people are really passionate about, that you have like so much.
Like you know, there's just so much talent out there.
And I've always been like a little bit disappointed that, like you know, like so much of that talent doesn't get used, just because everyone has like circumstances and like has like these.
You know situations where you know they can't actually pursue those things.
And so for me, AI has... All of humanity is not living up to their full potential.
I mean... And then you've got to do AI.
The thing I've always been excited about is, like how do you actually build this technology that frees people up to kind of do the things that they are passionate about?
Like, how do you basically, you know, allow people to actually focus on those things?
Originally, I thought of automation as the most natural way of doing that.
You automate away the parts that people don't want to do, and that frees up people to do the things that they do want to do.
But I guess I realized increasingly that it's actually pretty complex.
You actually have to understand.
If you want to empower people to do what they want to do, you have to really understand what people actually want to do.
And building systems that understand kind of people's goals and outcomes is actually really hard.
Yeah.
Did you have this human-centric perspective when you were choosing research problems to work on originally?
I guess at the very beginning.
When I was choosing research problems, I was just interested in how do you actually make these things half decent?
So it's more increased capability.
I think for me, when I looked at AI or language models back in 2021 or whatever, I was like these things aren't very smart.
They can't do that much.
And there was some early work around there that showed that, for example, you could use a chain of thought to get models to answer more smartly.
But it was still like only like a small step improvement at that time, like there was still the you know.
The benefit of that was, you know as much as you can really get with just prompting.
And so back then I was like thinking about looking, how do you actually make them like half decent and actually solving these harder problems?
Can you give a broad like?
We have everything from researcher audience to business person audience here.
Can you give a broad intuition for STAR?
I guess the intuition is, if you have a model and it's able to solve these like basic, like these, like slightly harder questions by thinking about them, then what if you actually teach it like hey, this solution that you came up with, that got you to the right answer, good job.
Or, you know, if you, or if the model didn't, then you basically like uh, don't reward it.
I guess the original version of star actually had like or yeah uh, there were like no, there wasn't a baseline at the time.
Uh, we compared it to uh, reinforce.
Which is this like popular algorithm in i guess, reinforcement learning, like very simple, like policy gradient thing?
But yeah, I guess you know at the time it was like a very simple algorithm.
Just, you know, you iteratively generate solutions.
If the solutions get you to the right answer, you learn from them.
If they don't, you don't.
And then you just kind of keep doing this, as the model solves harder and harder problems and then learns from harder and harder problems.
Did you um?
At what point in the research, if at all?
Were you surprised by how well it worked, or did you have some intuition for this being like something scalable?
There was one experiment that I remember doing, though this was quite a while ago at this point.
Um, but we looked at the, I think it was like n-digit like addition or multiplication.
Sorry, it's been a second.
And one thing that was really interesting was that back then this was a task that was considered hard for language models.
Yeah, of course.
It was considered one of the examples of why they were still so stupid.
Yeah.
Exactly.
And I was like, okay.
And one of the really interesting things for me was that, as you actually trained for more and more iterations, the number of digits that it was actually able to do kept increasing.
And I think that this was like one of those big surprises for me.
Like, oh wow, there's no obvious plateau here.
And did you go directly from that to generally this should scale?
I think I was generally interested in like, yeah.
I think there were a few things though.
There was one part of it that we introduced to kind of.
We observed that there was a bunch of the data that the mall wasn't learning from.
And so we proposed another variant of this, where we actually were like what if you actually take the ones where it fails and you basically ask it to reason about why it should have gotten it right, and then you train as if it got it right?
And this version, was kind of a way of extending beyond the parts of the data that it couldn't see.
So if you only train on the positive examples, then you end up in this kind of potential minimum where there's just no more data that it can actually solve.
And so back then we were like what if we just show it the problems that it didn't solve and try to teach it from those?
But I guess another thing that other work has done since then is, oh, what if you just sample a lot?
And that also seems to work in those works.
STAR has become a broadly used part of the reasoning paradigm since you published.
Can you also describe, I think this is like sort of your last published work, like QSTAR?
Okay, so Quiet Star was kind of the last thing that I did back at Stanford, and it was really fun.
I guess we showed a few things that were kind of cool.
One of the main goals of that paper was to show that you could actually scale this up to pre-training skill by using basically pre-training style data.
I guess now there's like a bunch of these works that have come out recently around, like you know RL pre-training and stuff like that.
And that's, you know, I guess in some ways similar to some of what we showed in the Clyde Star work.
Instead of having question answer.
If you actually just have, like these arbitrary kind of chunks of text, for example, and he tries to predict what's going to come next, which is the standard language modeling objective, can you actually get models that more generally learn to reason?
One of the cooler things that I think is overlooked about the original Quiet Star paper is we showed a bunch of key improvements to the Star paper that were necessary to actually do this kind of thing.
So that was, for example, showing that it's really valuable for this algorithm to be online, showing that it's really valuable for to have a baseline where you, like you know the harder For harder problems, you learn more.
For easier problems, you don't learn quite as much.
And I think that there were a bunch of nuggets in there that even at the time I don't think I fully thought of as like oh wow, that's actually a cool improvement over the original thing.
So you ended up going to Grok for several years and you sorry, XAI for several years and you worked on a bunch of different paradigms.
So pre-training data for Grok 2 and then overall the reasoning recipe for Grok 3.
I'm sure I'm missing things, but tool use and agentic infrastructure for Grok 4.
I guess if you level set us today, how smart are models?
They can obviously do n-digit arithmetic at this point.
I guess in terms of IQ stuff.
I'd say there's a lot of, And if you're able to pose the problem very well like some very advanced physics problem or math problem, I would say they're reasonably smart.
I think...
A lot of the failures that people see... Give me a human comparison.
What is reasonably smart?
I think it's hard to compare directly because it's very jagged.
It's true that some of these, for example, some of the HLE questions that these malls are able to solve, are genuinely things that are like non-trivial for like actually, PhD researchers.
I'm not saying they're like open problems or anything, but they are like pretty non-trivial.
Also a lot of them are, like you know.
One interesting category of like these I spend a lot of time looking at kind of the HLE questions.
One interesting category of them.
Sorry, humanities last exam for anybody who isn't looking at these evals.
No, great.
Yeah, so looking at these humanities last exam questions, I kind of...
One kind of category that is like actually quite big.
Are these like trick questions that require you know?
Basically, people like if you're familiar with it, you'll be like oh, they're trying to get you to like assume something.
But actually, like if you think more carefully about this problem, that assumption doesn't hold.
And this turns out to be a bunch of those kinds of problems.
So I think it's like they're pretty smart, but also they're more, I think, tripped up by some of these tricky things.
But also they don't really.
I think one of the core things is that they're not smart like emotionally, or like they're not smart on the level of like actually understanding kind of what people care about or kind of like how to actually like help people accomplish the things that they care about.
I want to talk about this and your next mission.
But just on this topic, if even jagged intelligence within, like the IQ domain, which I think almost everybody in the industry has been focused on until now, what would you recommend for people who are not researchers to develop some sort of intuition for that surface?
Because that seems very important to making them useful.
Yeah, i guess one thing that's that i think is like really important to keep in mind is that, like the more kind of context you can give the current generation of models, the better you kind of are uh, the better off you are.
Like their answers are uh super sensitive to, like you know, whatever additional information you can give them.
Yeah, i think this is like a really important thing.
I would generally say, like existing models are particularly good at handling questions that are like easy to answer in kind of like a closed form like um, If there's a simple numerical answer to what you're asking or a simple way of choosing from a set of things, this is something that these models actually Obviously it's all dependent, but this is something that makes it easier for them all.
If you can imagine it being easy to check your answer, that actually, I think, makes it easier for the models.
What do you think is the most dominant explanation for attempts to use models and more verifiable domains like code, still failing at sophisticated tasks?
Is it just like the wrong context has been fed to them?
Is it?
Context window is simply not large enough to support the like scratch pad and continual testing.
Like why in those domains, what is the biggest challenge?
Part of it is there's, I think, a balance.
When people want to give users these models, it's actually important that they're not annoyingly slow.
And so I think there's actually a number of problems where, if you gave the models more time, they would actually be able to answer better.
But, for example, in the coding context, you kind of have to be reasonably responsive.
At least it depends.
It depends on the kind of setup right, like if you look at products, like you know, opening eyes codex, which you know is kind of this longer running background thing, uh versus like uh cursor, which is, like you know, more interactive, You have a bit more luxury with those more background approaches to tackle harder problems, I'd say.
Yeah, I think it's a tricky question.
A lot of things depend on how far the distribution of what you're asking is from the distribution that the models were actually trained with.
So if you happen to be asking a problem that's very similar to the kind of problems that it's seen before, then it'll do great.
And if you're asking a problem that's very out of domain.
So to some extent, this question is kind of hard to answer concretely unless you know basically what the URL data for a lot of these specific tasks is right.
And today um, obviously none of the model or code agent, code interface companies are going to release like a capability map for you of what their rl data looks like, which would be very useful because i mean intuitively unless if you just look outside of the pre-training internet data sets right, there are types of problems and types of code bases that are much further out of distribution.
And so when engineers try in those scenarios, obviously they get a dumb agent yeah, back right.
And and you know also, like the another thing that matters a lot is just like how verifiable are the things that you're trying to get them all to do?
I mean, obviously there's been, you know, a ton of work out there on making models like less dependent on verifiable rewards, Lots of cool published papers.
I believe most people would say that there's still a gap between how well these models perform on verifiable tasks versus not verifiable tasks.
Yeah, absolutely.
What is last real question on IQ?
But because it is where 90 plus percent of industry energy literally energy and compute is focused.
How would you characterize where we are in scaling and the obvious opportunity to improve from here?
There's still meaningful dimensions of scaling that haven't been fully explored in terms of IQ.
I think there's a lot of cool efforts out there.
There's a lot of cool stuff that can still be done on the capabilities access.
I do think that as you start thinking about some of these new kind of axes of scaling, it's actually very natural to realize that there are ways to do them in ways that incorporate people and there's ways to do them in ways that kind of leave people out more and more.
And being very mindful of.
Oh hey, I'm designing this new algorithm and it's going to scale IQ you know of this model by X amount.
To effectively keep people in the loop is actually like a very active decision.
And so, you know, I think in general, if you're thinking about these things, that's important.
Wouldn't it be fair to claim that the instinct of many labs is to try to get people out of the loop as much as possible from a scaling perspective?
Because that's very messy, right?
If I want to recruit people to, for example, take complex reasoning, traces off them in tasks that are not in distribution for me yet.
That is not as simple to execute on for an organization as like, more rollouts, right?
And so why is that important at all from a capabilities perspective?
I mean, that's a good transition to like, what are you doing?
Yeah, I'd say that the main thing is just that like, As you kind of have these models that you know expand in terms of like the horizon that they're automating.
You know you have these models.
The recent like or recent-ish IMO results are like a kind of a good example of this.
You have these models that go on for, like you know, hours of reasoning without any kind of human intervention.
And this has kind of been an increasing measure of success, I would say, for these labs.
So, for example, there's this METR benchmark that everyone likes to share whenever there's a new model.
And it's like oh, we went from being able to have these models work for two like complete two hour tasks autonomously, without human intervention, to 25 hour tasks without human intervention.
And obviously there's like, questions of like, what do those numbers actually mean?
And how should we take them like kind of at face value?
But regardless, this kind of in like the metric that you know people are looking at more and more to measure progress.
But as we get these models that increasingly remove people from the interaction, you end up with basically people having less say in the things that get built.
I think if you have a model that goes off and does its own thing for eight hours and comes back to you with something that is somewhat there,
I think this is a weird regime where people probably feel less real agency over the things that they're building.
I think also, I anticipate, that people will feel like they don't really understand the things that are being built.
That's already true.
I think it's already true.
20,000 lines of generated code looks good to me.
Yeah.
It's just like you make these PRs and they're like 100,000 lines.
I think in general, this is going to be part of the trend.
So do you think that it's important to have humans in the loop of, you know, producing the output or the reasoning, because the ceiling is higher with humans are in the loop, because it is more efficient, because we can error correct when models are off path, or philosophically because people want that, or like some combination of all three?
Yeah, I think it's probably some combination.
I think another thing that I kind of think about is the most natural thing to do as you kind of automate away the existing set of tasks is, you know, you kind of look at the world gdp, you like carve out the parts that are, like you know, most easy to replace with these models, and you know that that's kind of the things that you target like, oh wow, you know, coding is like a x billion dollar market, let's automate all of that.
Or, like you know, this other segment is like x billion dollar market, let's automate all of that.
But I actually think if you kind of empower people, if you have models that really understand what people are trying to accomplish and really support them in accomplishing those things, you have the potential to actually grow that pie instead of basically replacing all of those segments.
And I think in general, if the purpose of these models is to replace the person for this chunk of work, you end up with a lot less real innovation on what's possible.
Yeah, I think if you actually have malls that really understand what people's goals are and really empower them more, you end up in a very different situation.
Because we're going to push those capabilities into areas that are out of distribution for them.
Okay, yeah.
I think... Is that accurate?
I'm just... Yeah, no, I'd say so.
I think it's like when I say that you know, I'd like to work on models that like, empower people instead of replacing them.
People are like Oh yeah, sure.
Like, but I'd rather like you know, work on curing cancer or something.
Obviously that's a really important goal, right?
Building models that are able to kind of solve, you know, humanity's most difficult and most fundamental problems is like, incredibly important.
But I also think that like And I'm sure that many of the researchers in the field disagree.
I guess in the long run, we'll see kind of what plays out.
But I personally strongly believe that we're much more likely to solve a lot of these fundamental human problems by building models that are really good at collaborating with large groups of people, that are really good at understanding different people's goals, different people's ambitions, different people's values, understanding different people's weaknesses and how to kind of coordinate with these large groups of people to make everyone more effective.
And I think, the vision of this AI that goes off on its own for 20 hours, does its own thing and comes back with the answers to life, the universe and everything.
I think that this is...
Less likely.
I think it's, you know, this is like a, I guess we'll have to see, but I think it's less likely.
So that goes to you are starting a new company.
Humans and I remember being like actually quite fundamentally surprised, given all of your work on IQ and reasoning and coding and scale that you were interested in, essentially EQ.
And you also thought of EQ.
And tell me if this is a wrong characterization as, like the, The emotional or the interactive capabilities of models today have really shown up in things like character or like companionship tools only.
And you thought of it as also like enablement from a productivity perspective, right?
So tell me about like where this thread came from.
Yeah.
I guess I've been thinking about this kind of stuff for some time now.
Even back in my PhD, I think one of my I guess less well-known works was actually about we showed that you can train language models to simulate different kinds of students.
Right.
For tests.
Yeah, yeah.
And by simulating students, you can actually design better tests for those students.
And that was a really cool finding.
Like hey, if you have models that are really good at modeling people, you can actually design systems that are better for people.
And this was something that I found really cool and interesting.
And as we move towards the current capabilities frontier, it became more and more obvious that we have these incredibly smart models that are capable of so much, but they're not used for anywhere near what they're capable of.
Like the role that they play in people's lives is a lot less deep, a lot less positive than it could be.
And I spent a lot of time thinking about like, okay, why is that?
Like.
Why are these models not like more, like I said, deeply positively integrating people's lives?
And it seemed like a really big part of it is like, that fundamentally these models don't really understand people.
They don't understand people's goals.
I would say part of it is like the general kind of training paradigm that the field is in.
It's very, I would say, single task focused or task centric.
It's ludicrous that all the benchmarks are still oriented this way.
Or most of them.
There's very few benchmarks out there that actually try to consider what if you actually have a person that's interacting with this model.
At best you have some multi-turn benchmarks that try to simulate what an environment would respond to different inputs.
But even that is still far from considering.
Hey, if you actually have this model that interacts with the person for some amount of time, how does it actually affect that person's life?
It's really remarkable that the field is kind of like so stuck in this kind of task-centric regime.
But it makes a lot of sense.
One thing that I was told by some folks at Google is that one of the reasons is that it's actually very useful for credit assignment.
So being able to have these benchmarks that are very easy to quantify and very easy to relate to some immediate thing means that you can say, oh yeah, this team did 2 better than this team, so they deserve all of the resources.
Or, you know, this team improved the benchmark by like 10%, while this team improved it by 5%.
So, you know, let's allocate accordingly.
And I think in general, that's part of it.
I think another part of it is kind of more aligned with the easiest ways to train these models.
It's not easy to have these RL environments and stuff.
You have lots of these companies popping up obviously, that are trying to sell environments to different people.
And the most popular are, of course, in coding and computer use.
Yeah.
Rather than anything that requires simulating people.
Yeah, it's not that surprising that we're kind of in this current regime.
So what do models need to know about people, or like?
What capabilities are they either missing or have not been elicited from them?
The most fundamental thing is that the models don't understand the long-term implications of the things that they do and say.
When you treat every turn of a conversation as its own game And you basically think of it as like okay, you had this interaction, you're done.
You need to make sure that this one response has all of the possible answers, has all of the possible content.
You don't ever ask questions.
You don't ever try to clarify things.
You don't really tend to express uncertainty.
You don't tend to be proactive.
You don't tend to think about the long-term consequences like.
You see a lot of like even single turn side effects of this kind of regime like and most of them are treated as kind of their own problems to solve.
You see issues around like that people highlight around like sycophancy.
You see issues that you know there was recent news around, like you know the psychosis stuff.
There's a lot of these like harmful effects that you get if you think about things in this very single task or task-centric way.
But if you have models that actually consider the long-term implications of Oh hey, if I tell this person to start a company that sells gloves for catching ice cream, if I tell them that that sounds like a good business idea, they might actually go and they might actually build that business and they might realize that it was not actually a good business idea.
Having a model that can kind of roll out the long-term implications of the things that I said.
And then they won't trust me anymore.
And then they won't pay for my compute.
Exactly.
And then it's all over.
Exactly.
No, I'm kidding.
I think that's really interesting.
One of the very core principles we have at Conviction for how we make decisions is well, what is the very long-term thing we want, right?
And like if that is the customer, the founder in this case, or an LP, or even for us, like it actually simplifies things quite a bit.
If you say like, we're optimizing for like a decade plus versus like this interaction.
And so being single term versus multi-turn seems like a very different way to make decisions.
It seems very hard to collect data about multi-turn human interactions, especially when you get to times you know.
It's actually like analogous to a problem in biology of how do you study diseases that just take time to progress.
I think it's a really fundamental question.
I think there is actually like some good academic work that has started to explore some of this.
Yeah, there's some work recently around like, you know, RL from human interaction.
There's a cool paper called the CoLabLM, you know, that trains against like, you know, simulation.
There's a lot of very cool work kind of starting to explore this in academia.
But in general I would say there's a lot less attention being paid to this kind of stuff in industry.
Because I would say for most labs and maybe this is a strong statement I'd say for most labs, the human is kind of the intermediate, until you have this fully automated system.
And so spending a lot of time optimizing things for being really good at understanding and really good at interacting and really good at collaborating with things is kind of almost like an intermediate thing you have to do until you get to this fully automated point.
Can you paint a picture of like if we have models that better understand human objectives over different timescales and are good at interacting with humans?
How is that more integrated into like your life five years from now?
Yeah, I think you don't need to go that far out.
But yeah, I think you get a lot of behaviors that you currently don't really see in these models.
I think you have models that are much better at understanding how the things that you say and ask fit in to the overall context of the stuff that you're doing.
For example, if the model knows that you're going to some wedding, for example, And then you ask it about like booking, you know hotels in Paris, it might, you know consider oh hey, like around the time of this event.
You know, I know that this user like has all these things that are true about them, like a model that's generally able to kind of think about how every thing that you say fits into your understanding of that person uh, would just be like, i think, a very fundamentally different interaction.
Uh, because right now, if you want to ask a question like that, you kind of have to dump all of this context in.
You have to tell like oh, you know i, can you help me find a hotel in paris?
This is because you know i'm going to like wedding.
I have like, you know, these constraints.
I, you know, I have like, these people who need to be with me.
I have like, you know it needs to do this, it needs to be.
You know you have, you need basically just dump all of the context that's relevant to yourself into the model and which is also an expensive interaction and something that most people won't do.
Imagine if you had a friend where you had to re-explain everything about yourself to them every time you spoke.
Exactly.
Like can you imagine if every time you interacted with someone, you basically like they remember, like your name and like you know, maybe what you do and like just like the really high-level sketch of your life?
Like, it would be... that friendship probably would not last very long.
Yeah, I think that's kind of what the current models are.
So you'd argue that the like any investment in memory that today's models have is not it's not that interesting or that core to their capabilities today.
I would say that memory is definitely a feature that has been underinvested in by the field.
But I would say that it is kind of difficult to invest in memory in this very task-centric regime.
Because if you have a bunch of these independent tasks, the amount of information that each of those needs from other things that you've discussed is not all that high.
Because of the current paradigm, memory doesn't end up being super useful in the training.
And so these models are not particularly good at doing it.
So one other thing I said to you I think out of like fear instinct than anything else, but I feel like other people will have this reaction as well is I'm a unique snowflake.
You can't possibly simulate.
You know me and all of my self-consistency issues between like, I want to learn this today, but I don't actually want to do the work.
I want to eat cake, but I want to be in shape as well.
Like, you know, we have different time skills and change our minds.
I'm just a constant distribution shift.
And then you can't possibly bring all of us under distribution.
How do you react to that?
I think to a certain extent, it's probably a little bit true.
It's not easy to build these really good models of people.
But I do think that the task for the model needs to be that it should be trying to do that.
The model needs to actually be trying to learn about you, trying to learn about the things that you care about.
The actual objective of the model needs to be to kind of understand you, and it probably won't be perfect.
But boy, you can be a lot better than the current models.
That seems totally reasonable, actually.
Yeah, I think, and you know, it's something that I think as a field, we will probably get better at.
I'm not going to pretend that, you know, I'm going to one shot this problem.
But I think even like any serious effort, that gets you quite a long way.
So there is like a cult sci-fi series about the culture where you have these super intelligent minds and essentially all of the human and human-like races live in a society where the minds make most of the decisions.
And there's like, I forget the total humanoid population, but let's say there are 30 or 40 minds that are still relevant as people in terms of perhaps being out of distribution or providing reasoning that the minds cannot and everybody else just lives in a world of abundance where they're like rock climbing and hanging out or whatever, and they do not produce.
How is your view of abundance different?
Everyone kind of has things that they're passionate about.
And given the opportunity, I think People can do really cool things.
I think the role of the model should be to allow people to do those really cool things that everyone wants to do and accomplish those things that everyone wants to accomplish.
And I think we shouldn't outsource all of the thinking and all of the everything to these AI overlords or whatever.
I think what we really want are models that are able to empower us.
Amazing.
Okay, super unique mission, amazing research work.
You're hiring an early team, getting a lot of compute.
Who are you looking for on the recruiting side?
One thing that I think is actually probably a good thing that my previous company did is thinking of everyone kind of, to some extent, as like engineers.
I think I'm looking for really strong info folks who can build stuff.
I'm looking for really strong researchers who can build stuff.
I'm looking for really strong product folks who can build stuff.
I'm looking for people who like, have thought a lot about like users, who've thought a lot about like memory, you know, on the research side.
I'm looking for, you know, on the infra side, for people who've thought a lot about like distributed systems, really fast inference people who've, you know, been there to scale really big projects up.
On the product side, I think people who are really creative about new modes of interaction, people who really deeply care about building beautiful, tasteful products.
Awesome.
Thanks so much, Eric.
Thank you so much.
Congrats on the new company.
Thank you so much.
Find us on Twitter at NoPriorsPod.
Subscribe to our YouTube channel if you want to see our faces.
Follow the show on Apple Podcasts, Spotify, or wherever you listen.
That way you get a new episode every week.
And sign up for emails or find transcripts for every episode at no-priors.com.
