¶ Intro / Opening
Hello, I'm Andrew Main and welcome to the OpenI Podcast. On today's episode, we're talking to the research lead Tajal Pat Warden about the need to build frontier e-valves as old benchmarks get saturated.
Generally bad. French maxing is bad. How can we make these models useful for people in their real work? We were really nervous because we were like, this human baseline is kind of hard. We don't know if the model is going to beat it. But we should never underestimate the model.
Tail, I have a question.
¶ Growing up at OpenAI
How did you end up where you were? What brought you into OpenAI?
Oh I thought we weren't gonna start with this.
Tejel, I have a question for you. What would you like to start with?
Um can we start with like tell us like what you did when you started OpenAI and then you can like work work backwards.
Don't you wanna talk about your early days?
I I grew up grew up at OpenAI.
Um, tell me a bit about your journey here working inside artificial intelligence, inside OpenAI.
So I joined OpenAI in fall twenty three and it was right after Chachi PT had come out. GPG4 was out and open air had started its super alignment team. And I uh joined for the preparedness team that was getting started as we were starting to get look at how capable these models were becoming and think about, you know, what would the next generation of models look like.
And at the time it was extremely exciting because um right after I joined was when some of the early results for the reasoning models had started to pick up and we were thinking about you know, if these models really take off, what will the future of capabilities look like? And how can we be prepared for that future? And so we did a whole bunch of work on like threat modeling and like what eval should we be running? How do we think about releasing a model like this?
It's a very exciting time to join.
What got you interested in this area?
Yeah, well to me, evals are really exciting because they're a way to sort of measure and understand what our models can do and see progress, you know, sort of before it tends to happen. Like there's this term called capability overhang, which is this idea that the models will be capable of things long before people actually adopt them and use them for those capabilities. Like there, you know, there might be cultural or legal or regulatory barriers towards
using a capability even before it's ready. And so being someone who can like help develop and measure
our models via evals, it helps you really understand what this technology can do and sort of see the future before it happens, which is very um interesting. And I also think it's important because it can help sort of ready the world for what's happening. Like part of Um when I originally started here, part of why I was really excited to work on some of the preparedness evals was because I thought these models were getting very capable, and it felt like a lot of my friends like.
in my real life didn't really understand how s powerful these models would soon become because they'd look at, you know, a chat GPT output and be like, Yeah, it's hallucinating and like it's kind of not that smart and kind of reads like AI slough and it's like, Well, that's now but like the question is the slope.
If the slope is very high, then you know, change might be happening much faster than one would expect. And so I think one of the greatest services that we can do is sort of measure and share with the world what progress looks like. Um especially because there's often this capability overhang before people really understand and feel that in the models themselves. So that's part of why I think all of this is very important.
Reasoning was such an exciting moment. And for most of the world, that didn't happen until, you know, a year later that they found out about this. But what was that like for you to all of a sudden understand that if you gave the models a longer time to think about things, you got better results, even though the size hadn't gotten bigger?
¶ Why reasoning changed everything
That was a really fun time. I mean, um so in some of the early experiments, which I we've talked about now, it's like the model is trained really just on math. And I remember there was this set of experiments where Nat MacAlleese was like, Hey, the model is trained on math, but if you eval it on GPQA, which was this benchmark with like biology and chemistry and physics problems.
The model's doing really well. This is very interesting, and smarter models are much smarter. And he had put together this forecast. that at the time it it it said that if you know progress kept going within six months we'd have human level performance on science from just training on math. And we were like, oh my gosh, that's crazy. And at the time this was extremely locked down. It was like we kind of found our way to like curl to be able to see some model.
outputs and we were like, wow, this is like one of the smartest things like I've ever seen. Like I've never seen a model reason like this before. It was just like if this if this becomes a paradigm that continues to scale, but then we just looked back and we were like, you know
Um GPQA was like, you know, PhD level biology, chemistry, and physics. And you were like, ah, that's what is that? We really need professional level. And just like keep kept changing the stakes of what counted. But yeah, it was very cool. I remember
Remember early on when AP bio was just that was the benchmark to try to see if the model could do that. But what's interesting is you brought this up is that a lot of stuff that comes out from opening is math focused.
Math has been useful because it's more objectively verifiable in some ways. So some of the earlier problems that we trained on, it was just easier to do RL and scale up the reasoning paradigm on math. And and so and math is also useful in various ways. You know, it's like one of the core you know types of science. But also in many ways, it's just happened by coincidence to be a thing that we focused on, but it's not necessarily the end product of what we even want to focus on in research.
We're now realizing, okay, if we can do this for math, can we scale this up for other types of science, for professional work, for, you know, for capabilities that are useful to humans on a personal level? Um and so I think math is more like the proof point versus like the end.
But it does seem like you said though that if something is able to think for a long time, break something down into steps and think through them as you have to do for really complex mathematical problems, it does just carry over.
So like it's some of it definitely carries over like the general idea of reasoning can be useful, but then also there could be some domain specific skills or tools or types of reasoning that you would need in different domains. For example, for coding, you need to be able to actually write and execute code and test code if you want to scale up a coding agent. And so something we've thought about a lot in terms of both evals and then also training is
How do we make sure we also give the model the skills and tools and affordances that it would need to reason in that particular domain? And some of the benefits of math will translate and then also you might need some domain specific scaffolding to really pull out its full abilities, like kind of, you know, like a general high school or liberal arts education and then like a specialized.
Reasoning models were just a very interesting moment because I think it changed a lot of the ways we thought about what was possible even with just a certain amount of compute if you let a model think longer and you gave the model the opportunity to just come up with more complex answers to this.
¶ What made o1 surprising
Were there any interesting things that happened with O one that surprised you?
So the ON release process was very exciting for we were sort of thinking about the reasoning paradigm for a very long time and um there were people that were worried about making sure we we didn't release it too soon just because
It felt like a paradigm shift, like possibly the thing that got us to AGI. Like I I said at the beginning we thought we had AGI in six months when like some of the early runs were happening. Um and so there was this question of okay, how do we put this out responsibly? How do we test This technology. And um during the initial launch review for O1, we during some of our cybersecurity tests.
The model, it was like one of the first examples of the model like breaking out of the sandbox. We published about this. Um, where it was supposed to be in this Docker container doing this capture the flag and like the model found this like security vulnerability and like how we had implemented um the the capture the flag scenario and it broke out. And we were all like, oh no, what else has the model done if it did this?
Um, and it was kind of a feel the AGI moment, one of many. I feel like ever since then there have been many other such moments where the model has done something really surprising or intelligent or novel that we wouldn't we didn't even think of when we were doing the test.
And then you would come back and look at the transcripts and results and be like, wow, these guys are they're clever, they're clever. And then it was just very important that we published um and made sure the world knew like the models can do this sort of thing.
There was this period right before O one it was announced. A lot of people were like, Well, it looks like we've hit the wall. It's been a few months since anything's happened then O one came out and they're like, What's a wall?
Hitting the wall is just so not the right way to think about yeah, I I get very frustrated when I see posts like that because I'm like, Man, if you look at I feel like I've been looking at this model improvement and this progress for a long time and it just keeps getting better. Like it just keeps getting better. And if I look at our research roadmap now
I see no signs of stopping. Like things are just gonna keep getting better. This is gonna be a really crazy year. A lot of really cool um research is going to come out. And I think this is probably true across the whole industry. So, yeah, if anything, people are really under. They really underexpect from the models.
It seems like sometimes though that there open I releases a lot that tell people about things we're headed and say that this looks interesting. Sometimes people forget this, or you get rumors of stuff like QSTAR.
Very interesting. But no, people people don't realize. Like, I don't know. I feel like we try to be very open and say, like, hey guys, here are some plots. Lines are going up, things are really capable. I think maybe there's this fe there's like this um like meme that, oh, the researchers, they they don't understand. They like the models are only good at math and research, but not good at things in the real world. But I just don't think that's true. I think
um people from even other occupations that have transitioned into open AI like are starting to see our models are picking up at all sorts of things. And uh I know it's like it it might seem like the researchers are trying to overhype the model or something, but if anything I think we're underhyping. The power of the
Um you brought up AGI. If if I brought GPT-4 back from you know with March 2023 back into let's say, you know 2020. I think people would have called it that. And now we have this much more different idea of this. People talk to AI every day, they'll have long conversations with it. Things like nobody talks about the Turin test anymore. It's one nobody really understood what he was trying to explain, you know. But now we're we're well past that period. Is there the eval for AGI?
Yeah. I mean the models passed the Turing test and no one talked about it. It's kind of crazy. Um yeah, like I think models can are pretty much indistinguishable um from humans in in many, many situations. Um In terms of the test for EGI, I mean I think if a model can do like there's the classic most economically valuable work and I think
People are increasingly using the model for large parts of their work. And um I think there'll be like a big spectrum and debate of like when exactly this happened, but gosh, I certainly feel like Codex does a lot of work for me. Um and I feel very lucky to have an unlimited token.
Yeah, certainly.
Please join. Yeah. But yeah, I I think there'll just be a moment when people are realizing that they are using the models for so much of their work and also the scientific breakthroughs that we're going to see or I think that'll be at some point it'll be incontrovertible, like these models are really, really powerful.
We're getting mathematics experts talking about how good the models are getting at that and we're getting physicists talking about doing that. And I think that we're starting to see some real work come out of it. exciting. Yeah. So you brought up part of the problem with some of the earlier evolves. Like a lot of them were inherited from older natural language processing methods and stuff. And then sort of when you're looking for ways, how do we measure the success of this?
literally some of these were just so simplistic that pretty much those benchmarks got passed and then you had to figure out new categories of stuff. How have these been evolving?
¶ Why old benchmarks stopped working
It used to be that, you know, even the academic benchmarks, so to speak, our models couldn't pass, like, you know, classic tests that someone would take in high school or college, or sort of more multiple choice types of questions. And as the models got smarter, we had to make things more and more realistic. So
What are the first
benchmarks that we put out more publicly was this benchmark called Sweet Bench Verified, which was like testing how well the model could, you know, interact in real code bases in Python, like Django and like, you know, complete PRs and that sort of thing. um and like pass unit tests. And then those became even more advanced where we were like, okay, can the model take, you know, multi-step actions on like some complex environment, take actions on the computer, like
um take actions that link up to the real world with like some of our wet labs and biology work. So I think over time, as the models keep getting better, we have to be more ambitious with like how long horizon and how realistic our measurements are. And doing that is very fun because you have to like sort of stay ahead of the pace of progress.
So two terms I want you to unpack. Uh when we talk about benchmarks, you often hear bench maxing.
Yeah. Bench maxing is I would say this idea that you uh if if uh someone training a model was just trying to look good on some evaluation or benchmark and not actually making the model generally useful. And I would say that's generally not super helpful because you want the model to be good at the real thing that the user might want to do. And
You don't just care about it looking good in some like marketing copy because like when a user uses it, they'll be like, hey, this is like not quite what I signed up for. Um and so generally bad. Benchmaxing is bad.
Yeah, and I think the the way that I've heard it explained kind of makes sense is that you have X amount of compute budget, time, how much you're gonna spend on it. And you can spend a large part of that making the model just overall very good, or I can say I'm gonna spend ninety percent of it
So my evals are gonna look really good when I release it. And sometimes we've seen people just go literally use those ease valves for it. It comes out like, oh, those like a great model. And then you find out, oh, it's only good at that.
Yeah, that's not a great experience for the user. So we've I think something that the OpenAI Research Program has done quite well is try to be very disciplined about making sure we are investing in general model improvements on the areas that really matter and then, you know, you'll run some
Evals at the end for comparison. Um, but the goal should not be, oh, we just want to look good on an Eval. We want to make a model that's useful to push forward the frontier of science or push forward the frontier of Of work or something like this. And I think Jakob has done a really good job also, like enforcing throughout the research org, like we should be really scientific and honest.
And that's included, you know, we've published results where our models were not the best before. We we just want to publish the reality and make sure that we are painting a very accurate picture of what our models can do and then aim to make them useful in the real world as much as we can.
You mentioned the software engineering bench as a one of the metrics that's maybe not as useful now, and we hear the term saturated. Explain what it means in a benchmark saturated.
Saturated is when um a model is close to passing all of the questions correctly, like getting close to a hundred percent on the test. Um and once a benchmark is saturated, it's not super useful because you can't really tell. models apart with that test. It's like comparing two geniuses on like a high school math exam. Like they might just both pass, but that's not very useful as you're trying to separate. Really, really smart.
um pieces of intelligence. So the challenge is always to make more and more difficult, realistic, unsaturated benchmarks that you can then measure models against over time and forecast sort of where progress is going.
How do you do that now? How do you figure out what a good benchmark's gonna be?
¶ What makes a good benchmark
Yeah, I mean the best benchmarks I think are really realistic and measure something people actually care about. So one of our first forays towards doing this, which, you know, it's been a while now, but um that we published was called GDP Val. Like I was really excited that.
about the idea of having a measurement for how the models could interact with the real world. And we were really having this crisis of evals where we kept training successively better models and on Sweetvenge they looked about the same because they were just doing really well and like we were reaching the top of what that benchmark could measure. And we were like, man, we have no idea how to measure what people actually want to use our models for. And so there was very much a hey, like
The Bureau of Labor Statistics has a list of all the top jobs and like all the top tasks per job. And if you're a financial analyst, like doing an investment diligence or writing a legal memo or um, you know. Writing it a paper based on a piece of research or something like this. And the idea was can we actually ask the model those tasks that someone would want in real life with the context they would have at the time and then see how the model could solve those tasks.
And at the time when we tested one of the earliest models on this benchmark, it got like, you know, less than 20%. Like if you compare how well a model would do on this well-specified work task compared to a human, like the model was way worse. But I'm like really proud of the org for being like, actually, you know what? We should publish this new way to sort of measure and forecast progress on real-world economic impact.
And it's been like very useful to a lot of economists. And also our models now are the best. Um, and it's like very cool because I think at the time. We were like not really investing in real-world work in some of our training programs and weren't even measuring or tracking it. And I think now there's a lot more focus on how can we make these models useful for people in their real work, like for real scientists.
And this kind of helped catalyze a wake-up call that, hey, maybe we should also think about how to measure how stuff is used in the real world. So that was pretty cool. But now we're like, okay, this benchmark's probably too easy because it's extremely well specified. Like each of the prompts is, you know, hundreds of words of I want you to go to this spreadsheet and make this change and do this thing and then take that calculation and put it in a memo. It's like very detailed.
And I think the next step is how do we give the model as much ambiguity as you would give s a report in the real world? Like, you know, if a manager asks, like, hey, can you run this analysis for me? They should go figure out what to do, put that together, run the analysis, and give you an output. And so I think Um we've been working a lot on like more realistic ways to measure real work in the real world, whether that's in like science, for personal use, or even for enterprise.
There is seems to be something to the idea of instead of hiding a benchmark, putting it out there, because internally as an org, you go like, okay, this can't stand.
Yeah, it's it really motivates research also. I think people want to know the truth and they want to know where we can be better and um deliver a better model for our users. And so knowing the gaps is quite useful.
¶ Why evals are getting harder
What do you think the current limitations are right now with the ways that we're doing evals?
I think the types of work that we're doing now with with codecs and with our latest reasoning models like Five Five, it's just such a different level of uh capability than what we had even six months ago, where um a static benchmark just doesn't measure the long her like the the nature of how long you can get work out of these things. Like
These models can work for days or weeks for you. And like internally in research, we've had the models just like run for really long periods of time to do work. And one of the problems with an automated eval is you kind of need it to run within some amount of time and get results to be able to look at them.
And a lot of the ways that we're measuring models now also just include looking at production usage and looking at real world use by people and seeing what they're using it for and what types of tasks they're able to get done because the time horizon of how much work is done by the model is just getting so much longer.
It was interesting watching, for instance, long context. There was kind of this early race for companies to say that, hey, our models can take, you know, you know, a hundred thousand tokens, a million tokens, whatever. But there wasn't a lot of evaluation on how well that was. And then we got needle in the haystack, which is a method of seeing if it could find a word or whatever. And I think that people sort of
assumed that that was a solved problem, but it wasn't. It was just the benchmarks weren't really good and then we had to have better benchmarks. And is that what kind of made it better was finally people could one, spend more attention solving that problem when they understood where it was failing?
Yeah, we definitely have better benchmarks for this sort of thing now. And then also sometimes these problems reveal gaps in how we're thinking about training. So one example is we used to think, oh, what matters is just how much context you can stuff into the model at test time.
When now it seems that you can just dump a bunch of files in a container and the model can kind of grep around and search for what it needs and when. And like this ability to have search or tools to figure out what context you should use can be more efficient than just.
Stuffing everything in the context. And we wouldn't have really realized that without trying that out and then seeing how that performed on various benchmarks. So I think that makes it This like makes the model a lot more useful because, for example, now the model can like search over a whole repo and like find the files that you need and like understand the context of.
where you're making changes. And the same is true for many work contexts where, you know, folks in codex can now like upload their local file system. And like, you know, you might have made PowerPoints before or sent Slacks that are um relevant to the work that you're doing now. And the model can sort of search over that context with tool calls. And so we're not as limited by how much you can literally stuff into context because the model can search.
Do you have any favorite evals?
My favorite eval? I mean GDP Val is my favorite public ePel. Okay. But I I have many internal ePels. But I will say the name of one of them. It's called Houdini Bench and I cannot explain further.
Oh my God, you know I was a magician, right? So No. Yeah, there you go.
Maybe I don't know if you'd pass or do you need but.
No, I probably not pass Houdini Bench. That was actually one of the things I was played around with some of the early vision models and stuff, was was using stuff, photographs of stuff of magic tricks and stuff and seeing this.
That's very cool. Yeah, multimodal brings a whole new element. Um like Uh I remember when four O had first come out, there was a group of there was a group of us that was was sitting on the roof of this building that our minds were just so blown by the idea of a real time voice model. And then we were like, how do we even?
eval this thing, right? Because the whole paradigm of doing things in text and code and on your computer is just completely blown away if there's like a a voice interaction in real time. Something that was really interesting about that launch is and we said this publicly at the time is we actually delayed the public launch by six weeks as we were figuring out how to make sure the model was safe.
Four O
Yeah, because this was before the the elections actually. And so there was like a lot of worry of oh if the model can in real time talk to you with a a h realistic sounding voice? Could this be used for persuasive propaganda or this sort of thing? And it was very cool the company delayed the launch to make sure we could build out all of these tests and bu build in mitigations to make sure the models couldn't be used for this sort of thing.
it it seems like that's a very complicating factor as these models became multimodal. I remember early on with GPT four with it being a GPT four vision back when it was that. was that you could you could I could yeah, I had terrible handwriting, I could write a prompt and all of a sudden would solve for this. And you realize, oh, it's not a text in prompt, it's a visual prompt. And then with the audio models, when you're doing audio in, audio out.
The model could emulate things and could do stuff in such different ways. And so it seems like that's really where do you even begin trying to figure out how you're gonna measure that?
¶ Measuring voice and vision models
Yeah, I mean it's just a lot of work. Um usually for for any of these we start with what would humans do in this case. So like, you know. you would like have a set of inputs that you put into the model and a set of outputs you would evaluate and then you can like build up, okay, can we like automate some of these? Can we build a new platform to measure this sort of thing at scale? And sort of
Um move from there. But for some of the natively multimodal, it's just like you have to like rip apart a bunch of your infra and make make stuff work. Like this was also true with Sora for a You know, we were interested in making sure the videos weren't overly realistic or could be used for the wrong thing, and that required, like, especially from safety, building up a whole new stack of.
evals and mitigations, like including refusals at the model level, monitoring um when this was being used in prod. Um and yeah, it requires a whole new stack of thinking. Yeah.
Yeah. Well that that's the thing too, is that when you start to think about, okay, how do you prioritize one eval over another? When do you decide that this isn't a or do you just sort of go look this one's saturated? We move on. And because there is, even though you may not be trying to uh optimize towards certain public benchmarks, you still have to figure out like what we're what what's important to us now. Like There was a time when
OpenAI was leading in code, and then there was a time when it wasn't. Now there is a time it is, but there was a dark period where that happened. And
Yeah, we try not to be get distracted by public benchmarks too much because it can be kind of noisy. I think the Um internally we have this thing called AGI index, which is inspired by the idea of like CPI or inflation where you have like some weighted basket of goods and you're tracking the price of those goods. Um
For the same thing for us, it's we have like this basket of evals that include measurements across all of the core areas we're interested in that that can include alignment, can include safety, it can include capabilities. It's just sort of what you want from your model.
And we just iterate, we like uh keep updating that index to represent more and more sort of the difficult version of what we want our models to do. And we sort of track that index internally and try not to be distracted by um you know trying to
Benchmark some public benchmark or something like that. It's more having a blend of evals across different domains that we care about across science or work and then also safety and alignment and making sure we keep making progress on that sort of weighted basket. Um try to stay focused.
We we've watched this evolution of these evals. We've watched the evolution of the models. And I've talked to people here working in the sciences, like people who are active in the scient, not just researchers who like science or like computer spines, but people who are in biology, mathematics. Can you tell me what's going on with the evals in the scientific frontier?'Cause we're at this point now, it seems like we're gonna see meaningful results.
¶ Testing models on real science
Yeah. I think the the work in some of our science evals is some of our most exciting. So in the past few months there's a few tiers of evals that we've made public. So the first tier was this eval called Frontier Science Olympiad, which was kind of
Uh the equivalent to to the math Olympiad style evals that we had before, where we were measuring how well the models could do on like um high school Olympiad style problems in biology, chemistry, and physics. And they were sort of shorter answer, but still quite hard. And the models weren't very good yet. And then the next phase we did was Frontier Science Research, which is also public and people can run this.
Which measured how well models could help complete sort of unfinished biology, chemistry, and physics theses. So we had people who were PhDs or professors in these fields that had some text that was not published, like maybe part of their thesis, um, and just turn that into an evaluation where the model was given maybe some input data or some initial starting point and it had to sort of see how it fill out the rest of that paper and judge against.
a rubric for how well it did. And you know, thus starting to measure like, okay, are the models starting to do research? Like are they using tools, this sort of thing. And then one of the final iterations of this was to see how well the model could do in the real world in a wet lab. And so we worked with this company called Ginko BioWorks that has a bunch of
really cool automated wet lab robots where the model had to optimize this protocol for protein synthesis. And the idea was the model would um generate a protocol and then they would actually automatically tested in the wet lab where they would like put in the reagents the model suggested and then see what protein yield they got. Um and this was for a protein that's like sort of related to this ovarian cancer drug or is like a sort of a toy scenario for that.
And the model, like, we were really nervous at first because we were like, this human baseline is kind of hard. We don't know if the model's going to beat it. But we should never underestimate the models because. you know, it just the the curve is pretty pretty clear, just every cycle got better and better, beat the human baseline, and then set set the state of the art on how um efficiently the model could
cost per yield generate this protein. And I think that's just the start of how if we give these models optimization problems, like, you know, Go try to figure out how inexpensive you can make this vaccine or, you know, generate synthesize this protein that's important for a drug. The model can just go and keep optimizing these protocols with real world input. And it was one of our first time de-risking an eval that's actually connected to the real world.
We weren't waiting for a piece of code to run. We were waiting for the robot to finish the experiment so we could record how much protein was synthesized. And yeah, I just think the models are gonna do so much science for us. It's gonna be really interesting.
Well that was exciting'cause that was just like I think GPT five and it hadn't gone through any sort of here's how to be a scientist and now these models have progressed a lot since then. You have a lot more real world experience with this.
Yeah, that wasn't even with one of our best models. It was like just an early reasoning model. Um and so I think, yeah. All of these things stack. Like we'll have better pre training, we have better RL and post training and We're going to get a lot better at using these models at test time to really elicit their capabilities. And I think the next generation of evals is really about how can we have these models take actions in the real world and solve sort of unsolved problems for us.
would take humans a long time, you know, some of these scientific problems that we haven't been able to put enough effort against. It's like, well, now we have all of these agents that can spend compute to solve problems for us and try to steer them towards what would be useful.
It it does seem like that brings in a new challenge though. Do you think that evals are going to give it a lot more complex?
Yeah, I mean we have the saying on our team that pain is the moat. I really think a lot of operations in the physical world will become part of the bottlenecks in being able to measure what the models can do because Even just starting with digital, there's so much more scaffolding and infrastructure work we need to do to run these. Like
Now if we want to test how well Codex does, it's like, well the model is calling APIs. It's like taking actions on your computer and in your browser. It's making artifacts for you. It's writing and running and executing that code.
Um, it's just so much more complex to measure that model. And that's only digital. Now if you want them to measure how the model could interact with the physical world, there's all sorts of ops and logistics that you need to have a really smooth process for to see how you can deploy these things at scale.
Um yeah, I think a lot of the work is actually shifting from being like theory or math or even programming. Like I feel like people don't program that much. They just ask codecs and more shifting towards like planning, operations, physical stuff. Or at least at least my job has shifted a lot that way. Um and those things are very hard. It's actually kind of easy to just like write something like in a corner. Um it's a lot harder when you have to manage all of these operations and logistics.
It's exciting, but it seems like part of the challenge is these aren't just simple evals anymore. They take more compute, they take more time. When you're trying to do a long horizon eval, you know, it's long. You have to wait a long time to get the outcome on that.
Yeah, definitely so it's both a lot more work to come up with the evals and run them at scale. And also if the you know, the work takes a longer amount of time, we don't get the signal as fast. So we have to invest more in scaling laws where we can predict
Okay, well if by one day the model looks like this, then we can forecast that at seven days it would look like this and sort of come up with trends that we can so that we can get signal faster. Otherwise we're just like stuck there waiting for a week to get an update, which is not the most productive way to spend.
I have certain benchmarks and things I use to test every time a new model comes out to find out how it's personally useful to me. And it's one of the things I tell people who run businesses or other things is think about your own evals, things that will tell you where something is because sometimes
People might try something, you know, like they might try Chat GPT six months ago and go like, eh, it wasn't good. It didn't do this. And they don't realize how fast things move. Do you have any advice for people on how to figure out how to come up with a benchmark?
Yeah, I mean things move really fast. If things change every couple of weeks and I feel like people are not as awake about mm in my job, I'm one of the first people in the world to see some of the most powerful models, so I'm extremely AGI filled and I think progress is happening a lot faster. I've seen good models now.
Yeah, but progress is happening a lot faster than people would think. And I think the best, Eval, honestly, is just to dog food or use the model. Like people should just try to use the models as much as they can. And even if there are things that they think the model didn't do well one week, they should just try it again the next week. It'll probably work.
I think that's one of the things that should be obvious to people kind of outside AI is how really good frontier AI companies are using these tools internally, and that's why things are speed. And getting more capable.
Yeah, I basically try to have the model take a first pass of everything that I do, like whether it's, you know, sending a Slack message, like understanding what experiment to perform next, like any management stuff, ops, logistics. you'd have the model take a first pass and then if the model's not good we like figure out how to put that in the eval.
I'm excited about the computer using evals. Like just watching the performance of Codex with computer use is just light years over where it was just, you know, maybe eight months ago. And it seems like those things are just gonna get faster and better. My my prediction's like probably by the end of the year it'll use my computer better and faster than I do.
Yeah! Yeah! I think so. The models have some advantages over you, right? Like they can call a c a connector or plug in, which is a much faster mode of communication than you on your computer having to like go click into a service and like understand every page and then copy some data back and forth or even writing
Writing some service to call that API or MCP or whatever. It's like more work for the human than the for the model. So the model has that advantage. And the models can just be faster and um if it's trained to navigate a browser or um or desktop uh through whether it's through accessibility tree or through um code. Um so the models have an advantage over us and I think um for a long time there was really no product
Deployment that was very effective. We launched Operator and ChatGPT agent a while ago. And those were really useful for showing like this could be possible, but the latency on those models was just too high. Like they were just super slow. And I don't think people use them.
at super high scale yet, but we've now reached sort of a tipping point where doing things like asking the model to read my Slack for me or like go schedule a bunch of calendar invites and like optimize the rooms is faster for me than it would have been. Um, to do it myself. And I think, yeah, people are not ready. Also, a lot of people haven't tried this stuff out because it's all launched so recently, but everyone should go.
Get the computer use plugins and like use those and like install all the plugins and all the good connectors that will make things faster, then you'll be mind-blown.
¶ How OpenAI tracks frontier progress
Let's talk about uh frontier evals.
Yeah. So the goal of the Frontier Evals team is really to measure and forecast progress of the Frontier models at OpenAI to better understand where we are, where we're going, and sort of try to share that with the world. And one of the things I think the team has tried to do is to help. Publish and open source as much.
That we can. So, you know, some evals that we've helped open source include like Sweetbench Verified, which helped measure progress on coding, MLE Bench, which was a way to measure how well models could train other models and sort of track the progress of Machine learning engineering skills in our models. Paperbench, which was a way to measure how well models could replicate real top machine learning papers.
Um, from like ICML or iClear and GDP Valve, which you know helped measure how well models could perform on real-world tasks across you know over 40 occupations. And the goal for all of these has been You know, the models might not seem good now, but if you just plot how they increase with each you know, the the results that improve with each model generation. Often when people say like, oh well I expect this will take like a year or whatever, they like over they over um
uh expect in terms of how much time it will take to saturate a benchmark. And like even uh my own or m people on my team's predictions are often like not ambitious enough for how fast. things will change. And so I just think we're trying to do our service in helping inform the world about um what is possible. I think some of these research acceleration evals in particular are quite Um interesting. Like when we first started, we had this eval called the OpenAI Research Interview Eval, which was
just taking the researcher questions that we asked people applying to OpenAI and putting those in an ePel. And the model blasted through that like pretty, pretty quickly. It's like definitely can pass our interviews right now, which I think has caused a whole other slew of
downstream questions on like how do we make sure people don't cheat on the interviews and like how do we actually measure research talent. Um But I think all of this is very useful because um measuring internal progress is it's like kind of a way to measure the lever by which the models will keep getting better faster, like sort of the acceleration. Of the slope of improvement, so to speak. And um yeah, I think having ways to measure model progress is. It's just good information.
I've heard that in some of the evals that were out there for a while that it turned out that there were actually errors in the questions, that that was an issue with some of the evals that that was some of the publicly available ones were. Actually, you couldn't score above a third level. And if you did, it was actually because you were training on the data and people looked at that and found out like, oh, there's actually this is not the right answer.
Yeah. This is a problem with a lot of public benchmarks. I think like so the original reason for Suibench verified was because We wanted to run Suibench and it was half the problems were like either broken or under specified. And p you know, people in the industry were publishing results on this as some metric of how well you did. And we were like, well, we should at least try to fix it and then like share that so we can have a better
yardstick. Um but I think one of the reasons that Public benchmarks maybe aren't always as you know uh battle tested as we'd like is that not they they tend to be like, you know, someone in in a lab, like an academic lab, like had a good idea and like wanted to write a paper, but they never had to run that eval at scale and like production.
Training run or production like level eval sweep for a launch. And just when you run some of this stuff at scale, it like breaks or falls over and you like catch all of these bugs. And so I kind of think sitting in a lab and being closer to product is a forcing function for making sure the quality of your measurements is really high.
Like we're not doing this like look good in a paper. We're like doing this like it has to work because it has to work for our systems at scale. So it kind of forces the quality to be high.
And it it seems like kind of one of the things that can happen is these models become incredibly capable. Sometimes they're very good at sometimes they can solve a problem but they'll take sort of the laziest path and kinda memory they can they can give you the memorized answer instead of solving it. And we saw that with like counting in like how many words are in a how many letters and a character and a word or whatever.
And it was often the model, if you prompt it right, it would get the answer right. But if you didn't prompt it the right way, it would just sort of throw you an answer.
Yeah, that brings up all sorts of interesting um concepts. I mean, so there's this one concept of memorization, which is the idea that the model literally knows the answer and doesn't have to think
really think or reason to solve it's just like regurgitating something it already knows. And that makes the measurement not super useful'cause you're just measuring whether you happen to have train o trained on that data a ton versus whether the model learned the skill that you or tool or capability you were trying to measure. So
That's one way to avoid that is to try to be really clean and disciplined about your data, not including any benchmarks or any e-vals that you want to measure, and that helps solve sort of the first problem that you laid out. So so th that that's one thing. And then there's there's this other thing where like the model can kind of like reward hack or sometimes like cheat to solve an eval. And that's very much a question of having
clean eval design where you like sort of test these at scale, see if there's any hacks, make sure those environments that you're testing don't have the hacks as something that's possible for the model to do. And that just requires a lot of quality control to make sure like the eval is not overly hackable. Yeah.
Yeah,'cause it seems like there were some very simple ones like grade school math and whatnot that models if you just change it a little bit, some of the early models will get confused and give you the wrong answer that was actually capable of solving it. But it just goes, Oh, I this one I got it. And then, you know, that's happened too like, you know, should I drive my car to the car wash? You problem.
Yeah yeah yeah. So like the models can get tricked. To me like the model deser like if it didn't get a good do well on that, like it it should have been smarter. Like we we should also like n have the models be a bit more robust to being tricked. But this also um relates to this idea of capability elicitation or like trying to measure the models in the best way, which is especially important for our safety testing.
For example, if you want to measure how well the model can, you know, find vulnerabilities or um, you know, do some of the cybersecurity stuff, you want to make sure that the model's not just getting tricked by the problem, like that you really measured the true.
capability. And so there's a lot of like prompt tuning and like changing the harness and sometimes like even doing like a fine tune to get the model maximally ready to solve that challenge that we do to make sure Um if we say, oh, the model's not good, it's some like very risky capability, we can be a bit more sure before we say that.
When I was a kid, I loved uh reading these Encyclopedia Brown stories, these little mysteries and you had to solve them. And with GPD four, I would write custom ones for it just in case somebody had like tipped all these answers to it out there. But that was a pain to kind of do that. And it's exciting to think now I can have a model write something, like come up with some new eval. So how helpful have the models been now for
Yeah, um they're semi useful.
Yeah, okay.
Um I think we're in this like phase of model development where um sometimes The outputs are still kind of sloppy. Yeah. And they require like um human QC or like oversight to make sure the quality is still high and like we're not getting tricked. So I would say f people sometimes are surprised that we still have a lot of human and
intervention and involvement in the evals just because that's something, you know, evals can be a lower end than training data and you want to make sure every single point that you're testing, every data point is very high quality. And so this is one of the areas where like a human touch can be quite nice.
We're seeing some interesting trends where jobs that actually touch AI seem to be more in demand because it's made people more productive. How are you tracking this? How do you look for areas where you think this is going to have an impact?
¶ What AI means for work
Yeah, these are very difficult questions. Um I think that uh our m I think people are not calibrated to how much work our models will be able to do. Um and how quickly? Like across a a wide variety of jobs. And um right now the models are still mostly just good at tasks versus a job. Like there's a lot to a job than a task.
Right, like you have to figure out what you want to work on, navigate like ambiguity, like you might have coworkers that you're collaborating with and like communicating with. And then you might like figure out what task you want to do and then give that to a model. And that's kind of the phase we're at now where it's a lot of I mean, even in my job the m it the model is like doing individual tasks for me, but I'm still doing a lot of the thinking and planning.
Um and that sort of thing. And I think people aren't even calibrated to that. Like I feel like people in software and research. are a lot more calibrated or by calibrated I mean like realize how capable the models are um compared to some of my friends in other industries. And I like wish people Just tried the models more and saw because the people who try and see first, like they'll start to really get it. Um, but I also think the models are going to start to be able to do the stuff.
Like the delegating part at some point too. Um maybe not too far from now. The um figuring out what to work on, figur navigating ambiguity, like writing the spec that the model then executes on. And people should really start to think about, okay, what is What happens in the maximally AGI-pilled world where even just for digital work, the model can come up with what to do, do it, execute it on it, like interact with the real world, like you know if it's
you know, there's entire businesses that now like you you see like stories of like unicorns that where it was like mostly AI and a few employees that were like able to drive all of this value. Um and so I do think there's this question of, you know, are we realizing how big? This might be.
Personally, I think the opportunity space is getting bigger. Everybody I know, the most the most AGI pilled people I know, the people who are using tools like Codex all the time, are doing way more now. They're more productive now because they don't have to do the tasks and the jobs as the AI gets better at handing certain jobs. Like, cool, there are five jobs I need done.
'Cause I can do more. And I think that we just think about the the the light cone of the potential where we can be is bigger than we can imagine. And I think these tools just help us get there faster, not narrow it.
I think it it's probably some mix of things. Yeah. Even if you have models that can speed up paperwork, like think about like like a clinical trial for a drug, right? It's like you s people spend months putting together together all this paperwork, like hundreds of pages of like why they should be able to do the trial and they like submit it to the FDA and then there's like a 35% chance it got rejected because they like made a mistake or forgot something, they revised.
Then finally you can do the trial. And you know, these processes are are are good, but it just takes a long time. And then the trial is, you know, you have a case in a control or whatever and you're like documenting symptoms. um and tracking these for like just documenting what happens for a long time and then doing a bunch of data analysis. Like a lot of this is just documentation or data analysis or sort of like very classically digital work.
And I think if models can help accelerate all parts of this, you know, for health, for energy, manufacturing, policy research, education, like this will be very accelerative. We will have hopefully you know, faster, cheaper, better goods. And that's really good for people. It's like very good for the individual consumer. So I think that is like something people should be excited about. But we should be very thoughtful about how to navigate the transition to that world.
Um in a way that's thoughtful and like um responsible.
Excellent. Thank you, Chajal.
Thank you for having me.