OpenAI研究员：永远不要低估模型能力

Andrew Mayne: Hello, I'm Andrew Mayne, and welcome to the OpenAI podcast. Andrew Mayne: On today's episode, we're talking to the research lead, Andrew Mayne: Tejal Patwardhan, about the need to build frontier evals Andrew Mayne: as old benchmarks get saturated. Tejal Patwardhan: Generally bad. Benchmarking is bad. Tejal Patwardhan: How can we make these models useful for people in their real work? Tejal Patwardhan: We were really nervous because we were like, Tejal Patwardhan: this human baseline is kind of hard.

Tejal Patwardhan: We don't know if the model is going to beat it. Tejal Patwardhan: But we should never underestimate the model. Andrew Mayne: Tejal, I have a question. Andrew Mayne: How did you end up where you were? Andrew Mayne: What brought you into OpenAI? Tejal Patwardhan: Oh, I thought we weren't going to start with this. Andrew Mayne: Tejal, I have a question for you. Andrew Mayne: What would you like to start with?

Tejal Patwardhan: Can we start with, like, tell us, like, what you did when you started OpenAI, and then you can, like, work backwards. Andrew Mayne: Don't you want to talk about your early days? Tejal Patwardhan: No. Tejal Patwardhan: I grew up at OpenAI. Tejal Patwardhan: Okay. Andrew Mayne: Tell me a bit about your journey here working inside artificial intelligence, inside OpenAI. Tejal Patwardhan: So I joined OpenAI in fall 23, and it was right after ChatGPT had come out, GPT-4 was out, and OpenAI had started.

Tejal Patwardhan: its Superalignment team. Tejal Patwardhan: And I joined for the preparedness team that was getting started as we were starting to get a look at how capable these models were becoming and think about, you know, what would the next generation of models look like? Tejal Patwardhan: And at the time, it was extremely exciting because right after I joined was when some of the early results for the reasoning models had started to pick up.

Tejal Patwardhan: 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? Tejal Patwardhan: And so we did a whole bunch of work on like threat modeling and like what eval should we be running? Tejal Patwardhan: How do we think about releasing a model like this? Tejal Patwardhan: It's a very exciting time to join.

Andrew Mayne: What got you interested in this area? Tejal Patwardhan: 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. Tejal Patwardhan: 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.

Tejal Patwardhan: There might be cultural or legal or regulatory barriers towards using a capability even before it's ready. Tejal Patwardhan: And so being someone who can help develop and measure our models via evals, Tejal Patwardhan: it helps you really understand what this technology can do and sort of see the future before it happens, Tejal Patwardhan: which is very interesting. Tejal Patwardhan: And I also think it's important because it can help sort of ready the world for what's happening.

Tejal Patwardhan: When I originally started here, part of why I was really excited to work on some of the preparedness evals Tejal Patwardhan: was because I thought these models were getting very capable. Tejal Patwardhan: And it felt like a lot of my friends in my real life Tejal Patwardhan: didn't really understand how powerful these models would soon become Tejal Patwardhan: because they'd look at a ChatGPT output and be like, Tejal Patwardhan: yeah, it's hallucinating and it's kind of not that smart Tejal Patwardhan: and kind of reads like AI slop.

Tejal Patwardhan: And it's like, well, that's now. Tejal Patwardhan: But the question is the slope. Tejal Patwardhan: If the slope is very high, then change might be happening much faster Tejal Patwardhan: than one would expect. Tejal Patwardhan: And so I think one of the greatest services that we can do Tejal Patwardhan: is sort of measure and share with the world what progress looks like, Tejal Patwardhan: especially because there's often this capability overhang Tejal Patwardhan: before people really understand and feel that in the models themselves.

Tejal Patwardhan: So that's part of why I think all of this is very important. Andrew Mayne: Reasoning was such an exciting moment. Andrew Mayne: And for most of the world, that didn't happen until a year later Andrew Mayne: that they found out about this. Andrew Mayne: But what was that like for you to all of a sudden understand Andrew Mayne: that if you gave the models a longer time to think about things, Andrew Mayne: you got better results, even though the size hadn't gotten bigger.

That was a really fun time. Tejal Patwardhan: I mean, so in some of the early experiments, which we've talked about now, it's like the model is Tejal Patwardhan: trained really just on math. And I remember there was this set of experiments where Nat McAleese was Tejal Patwardhan: like, hey, the model is trained on math. But if you eval it on GPQA, which was this benchmark with Tejal Patwardhan: biology and chemistry and physics problems, the model is doing really well.

This is very interesting. Tejal Patwardhan: smarter models are much smarter. And he had put together this forecast that at the time it said Tejal Patwardhan: that if, you know, progress kept going within six months, we'd have human level performance on Tejal Patwardhan: science from just training on math. And we were like, oh my gosh, that's crazy. And at the time, Tejal Patwardhan: this was extremely locked down. It was like, we kind of found our way to like curl to be able to Tejal Patwardhan: see some model outputs.

And we were like, wow, this is like one of the smartest things like I've Tejal Patwardhan: ever seen. Like I've never seen a model reason like this before. It was just like, if this, Tejal Patwardhan: if this becomes a paradigm that continues to scale. Tejal Patwardhan: But then we just looked back and we were like, Tejal Patwardhan: you know, GPQA was like, you know, Tejal Patwardhan: PhD level biology, chemistry, and physics.

Tejal Patwardhan: And we were like, ah, that's, what is that? Tejal Patwardhan: We really need professional level. Tejal Patwardhan: And we just like kept changing the stakes of what counted. Tejal Patwardhan: But yeah, it was very cool. Andrew Mayne: I remember early on when AP Bio was just, Andrew Mayne: that was the benchmark to try to see Andrew Mayne: if the model could do that. Andrew Mayne: But what's interesting as you brought this up Andrew Mayne: is that a lot of stuff that comes out from OpenAI Andrew Mayne: is math focused.

Tejal Patwardhan: Math has been useful because it's more objectively verifiable in some ways. Tejal Patwardhan: 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. Tejal Patwardhan: And math is also useful in various ways. Tejal Patwardhan: You know, it's like one of the core types of science. Tejal Patwardhan: But also in many ways, it's just happened by coincidence to be a thing that we focused on.

Tejal Patwardhan: But it's not necessarily the end product of what we even want to focus on in research. Tejal Patwardhan: Like we're now realizing, OK, if we can do this for math, can we scale this up for other types of science, Tejal Patwardhan: for professional work, for, you know, for capabilities that are useful to humans on a personal level. Tejal Patwardhan: And so I think math is more like the proof point versus like the end goal.

Andrew Mayne: But it does seem like you said, though, that if something is able to think for a long time, Andrew Mayne: break something down into steps and think through them as you have

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。