Humans&: Bridging IQ and EQ in Machine Learning with Eric Zelikman

0:05Hi, listeners. Welcome back to KnowPriors. Today, we're here with Eric Zeltman, 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 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 all of humanity is not living up to their full potential i mean and then you've got to do it i mean it's the 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 yeah allow people to actually focus on those things you know originally i thought of automation as kind of like the most natural way of doing that like Like you automate away the parts that like people kind of don't want to do.

1:38And that, you know, frees up people to do the things that they do want to do. But I guess I realized like increasingly that that's like, it's actually like 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. Did you have like this human centric perspective when you were choosing research problems to work on originally? I guess like at the very beginning, I was just like when I was choosing research problems, I was just interested in like, how do you actually make these things half decent?

2:21So it's more increased capability at all first. Yeah, I think I think for me, like, you know, when I looked at like AI, like, or, you know, language models back in like 2021 or whatever, you know, I was like, these things aren't very smart. They can't do that much. And there was some like early work around there, like that show that like, for example, you could use like chain of thought to like, you know, 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 The benefit of that was, you know, as much as you can really get with just prompting.

2:57And so back then I was like thinking about, okay, 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 a 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 start actually had like or yeah uh there were like no there wasn't a baseline at the time we compared it to uh reinforce which is this like popular algorithm in i guess reinforcement learning, like very simple, like policy gradient thing.

3:54But 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, 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.

4:28But we looked at the, I think it was like n-digit addition or multiplication. Sorry, it's been a second. And one thing that was really interesting was that this, back then, this was like a task that was considered like hard for language models. It was considered like one of the examples of why they were still so stupid. Yeah, exactly. And I was like, OK. 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.

5:01Like, 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 like 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, oh, what if you actually take the ones where it fails and you basically like ask it to reason about like why it should have gotten it right. And then you train as if it got it right.

5:35And 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.

6:14Can you also describe, I think this is like sort of your last published work, like QSTAR? So QuietStar was kind of the, yeah, 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 like pre-training skill by using like 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 the what we showed in the QuietStar work.

6:51Instead of having question and answer, if you actually just have like, you know, these arbitrary kind of like chunks of text, for example, and you try to predict what's going to come next, which is like the standard language modeling objective, can you actually get models that more generally learn to reason? One of the kind of cooler things that I think is kind of overlooked about the original QuietStar paper is we showed a bunch of like kind 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 to have a baseline where for harder problems you learn more, for easier problems you don't learn quite as much.

7:42And I think that there were a bunch of like nuggets in there that even at the time, I don't think I fully, you know, thought of as like, oh, that's actually like 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 GROC4. I guess if you level set us today, how smart are models? They can obviously do n-digit arithmetic at this point.

8:24I 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 like 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. Yeah. Like it's true that like some of these, for example, some of the HLE questions that these models 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.

9:07but they are like pretty non-trivial yeah 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 sorry humanities last exam sorry anybody who isn't looking at these evals no great um yeah so but yeah looking at these humanities last exam questions i kind of um 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.

9:51And this turns out to be like a bunch of those kinds of problems. So I think it's like it's they're pretty smart, but also they're more, I think, tripped up by some of these like 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?

10:38Because that seems very important to making them useful. Yeah. I guess one thing that I think is really important to keep in mind is that the more kind of context you can give the current generation of models, the better you kind of are, the better off you are. Their answers are super sensitive to whatever additional information you can give them. Yeah, I think this is 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 if there's like a, you know, a simple numerical answer to what you're asking or like a simple way of choosing from a set of things, this is something that these models actually like.

11:29Obviously, it's like all dependent, but this is something that makes it easier for the model. 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 in 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.

12:12When people kind of want to give users these models, it's actually important that they're not annoyingly slow. And so I think there's actually like a number of problems where like if you gave the models more time, you know, they would actually be able to answer better. but for example in the kind of coding context you kind of have to be reasonably responsive at least it depends on the kind of setup, right? Like if you look at products like OpenAI's Codex which is kind of this longer running background thing versus like Cursor which is more interactive you have a bit more luxury with those more background approaches to tackle harder problems, I'd say.

12:59Yeah, 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, you know, if you happen to be asking a problem that's very similar to the kind of problems that it's seen before, then, you know, it'll do great. And if you're asking a problem that's like very, yeah, out of domain, So like, to some extent, this question is kind of hard to answer concretely without, unless you know, like, basically what the RL data for a lot of these, you know, specific tasks is. Right. And today, 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.

14:11and so when engineers try in those scenarios obviously they get a dumb dumb agent 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 words lots of cool published papers i i 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 this last real question on IQ, but because it is where 90 plus percent of industry energy, literally energy and compute is focused.

14:59How 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, I think, 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 axis. I do think that as you start thinking about some of these new axes of scaling, it's actually very natural to realize that like 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 if you effectively like keep people to effectively keep people in the loop it's actually like a very active decision.

16:04And 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 like 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 um that is not as simple to execute on for an organization uh as like more rollouts right and so why is that important at all from a capabilities perspective yeah that's a good transition to like what are you doing yeah i'd say that it 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.

17:00You know, we 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, you know, 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, you know, there's this METR, like meter, like 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 2.5 hour tasks without human intervention.

17:41and obviously there's like questions of like what do those numbers actually mean um and how like should we take them like kind of at face value but regardless this kind of been like the metric that you know people are looking at more and more uh to measure progress but you know as we kind of uh get these models that increasingly you know remove people from the interaction you You end up with basically people having less say in kind of the things that get built. You end up with like, you know, I think if you have a model that goes off and does its own thing for like eight hours and comes back to you with something that is somewhat there.

18:24I think this is like a weird regime where like people probably feel less like real agency over the things that they're building. And I think also, I kind of anticipate that people 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 of like, you know, like. And I think in general, this is kind of 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?

19:07Because 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 like, you know, 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 like if you kind of empower people if you 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.

20:08And I think in general, if the purpose of these models is to kind of replace the person for this chunk of work, you end up with a lot less, I think, real innovation on kind of what's possible. Yeah. I think if you actually have models 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.

20:50People 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 you know, 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 working together, 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.

21:44And I think the vision of this AI that goes off on its own for like 20 hours, does its own thing and kind of like, you know, comes back with like, you know, the answers to life, the universe and everything. I think that this is like less likely. I think it's, you know, this is like, I guess we'll have to see, but I think it's less likely. So that goes to you are starting a new company, HumansAnd. 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 emotional or the interactive capabilities of models today have really shown up in things like character or like companionship tools only.

22:40And 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 show that you can train language models to simulate different kinds of students. Right. For tests. Yeah. 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.

23:21And this was something that I found really cool. and kind of as we move towards the current kind of capabilities frontier, it became more and more obvious that like the, you know, we have these incredibly smart models that are like 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.

24:07They don't understand people's goals. They're trained. 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. Yeah. Yeah. I mean, like, like. Or most of them. You know, I mean, even the ones that are like, like there's very few benchmarks out there that actually try to consider like, oh, what if you actually have like a person that's interacting with this model? Like, you know, at best you have like some, you know, multi-turn benchmarks that like try to simulate what an environment would respond in different, you know, to different inputs.

24:44But even that is like still like far from, you know, considering, hey, if you actually have this model that interacts with the person for like, you know, some amount of time, like, 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. And I think it but it makes a lot of sense. One thing that I was told by some folks that, you know, at Google, is that it one of the reasons is that like, it's actually very useful for like credit assignment. So like being able to have like these benchmarks that are very easy to quantify find very easy to like relate to some like immediate thing means that you can kind of say like oh yeah this like you know this this team did like two percent better than this team so they deserve like all of the resources or you know this team like improved the benchmark by like 10 % while this team improved it by five percent so you know let's let's allocate accordingly and I think in general like that's that's part of it I think another part of it is like kind of more aligned with the easiest ways to train these models.

25:52It's not easy to have these aural environments and stuff. You have lots of these companies popping up, obviously, that are trying to sell environments to different people. But the most popular are, of course, encoding 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 kind of don't understand the long term implications of the things that they do and say.

26:34When you treat every turn of a conversation as kind of its own game and you, you know, you basically think of it as like, OK, 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. You see a lot of even single term side effects of this kind of regime. And most of them are treated as kind of their own problems to solve.

27:11You see issues that people highlight around sycophancy. You see issues that there was recent news around the psychosis stuff. There's a lot of these 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 like you know a company that you know sells gloves for catching ice cream if I like 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 them all that can kind of rule out the long-term implications of the things and they won't trust me anymore and then they won't pay for my compute exactly exactly Exactly.

28:05No, I'm kidding. I think that's that's really interesting. Like 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-term seems like a very different way to make decisions. It seems very hard to collect data about multi-term human interactions, especially when you get to times.

28:45You 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 some, there's a cool paper called a Collab LLM, 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.

29:23because I would say for most labs, and maybe this is a strong statement, but I would 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?

30:03How is that more integrated into like your life five years from now? Yeah, I think you don't you don't need to go that far out. Two years. But yeah, I think 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. Like, for example, if the model knows that you're going to some wedding, for example, and then you ask it about booking 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 of these things that are true about them.

30:50Like a model that's generally able to kind of think about how every thing that you say fits into your understanding of that person would just be like, I think, a very fundamentally different interaction. 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 a 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 basically dump all of the context that's relevant to yourself into the model.