大规模测试时计算与模型评估

听众朋友们好，我是 Sarah Goa，欢迎回到 No Priors。今天请到的是 Noam Brown，AI 推理领域的奠基人之一。我们聊了评估体系的缺陷、超大规模测试时算力、他对递归自我改进的看法，以及前沿竞争的下一个方向。欢迎回来，Noam，特别高兴再次邀请你。

Hi listeners, I'm Sarah Goa and welcome back to No Priors. Today I'm here with Noam Brown, one of our godfathers of AI reasoning. We talk about the broken state of evaluations, very large-scale test time compute, how he thinks about recursive self-improvement, and what's next on the horizon for competition at the frontier. Welcome, Noam. I'm so excited to have you back.

很高兴回来。

That's great to be back.

你是我们的第一位嘉宾。我为自己的选人眼光感到骄傲。考虑到推理时 Scaling 对整个行业的重要性，你也应该感到自豪，毕竟你真正开创了这条路。

You are our first guest. I'm very proud of my taste in friends and researchers for the pod. Given how important inference time scaling has become to the industry, you should be proud too, having actually pioneered it.

算是出了一份力吧，当然还有很多其他人。

Played a part, yeah, among many others.

你最近那篇关于大规模测试时算力的文章引起了很大共鸣，讲的是行业为什么没有足够严谨地评估这些模型。是什么促使你写这篇文章的？

You just wrote this essay that really resonated about large scale test time compute and why the industry is not evaluating these models as robustly as it should be. What was the motivation for it?

是的，动机是我们发布了 5.5，最初的反应是怀疑它是否真的更好。公平地说，这种怀疑只持续了几个小时，人们花时间试用后，发现它确实好得多。但我认为很多怀疑来自发布的基准测试表格。基本上，每当新模型发布时，都会有一个基准测试表格，横轴是各种基准，纵轴是不同模型的性能，你可以直接比较。每个模型在每个基准上只有一个数字。如果只看纸面数据，5.5 和 5.4 或其他模型的差异并不大，在某些基准上只提升了几个百分点。所以人们看了之后怀疑它是否真的更好。但一旦他们试用过，情况就不同了。我认为基准测试没有显示出更大优势的原因，是基准测试的呈现方式有问题——它们没有控制用于回答基准问题的测试时算力。结果发现，5.5 的思考效率更高。如果你以最大设置运行，5.4 思考的时间更长，返回响应的时间比 5.5 更长。一旦你控制了思考时间，就能看到 5.5 相比 5.4 有显著提升。我认为这就是人们日常使用时的感受。当我跟人提起这一点时，典型的反应是：那为什么不干脆让 5.5 也思考那么长时间呢？问题在于：应该让它思考多久？通常的回答是：直到性能饱和——基准性能会达到一个平台期，然后评估到那个点。但问题是，现在这个平台期点已经非常遥远了。在 2022 年的 GPT-3 时代，模型确实无法长时间高效思考，所以你可以一直运行直到饱和，那个点并不远。但如今，像 5.5 这样的现代模型，如果搭建得当，可以在某些基准上思考数周甚至更久才达到性能饱和。因此，饱和点已经远到无法合理测试了。我们现在必须引入一个耐心限制或 token 预算限制，这在几年前是不需要的。

Yeah, the motivation was we released 5.5 and the initial reaction was kind of skepticism that it was a substantially better model. To be fair, that only lasted for a few hours before people had some time to play around with it and try it out themselves and they saw that it was actually substantially better. But I think a lot of the skepticism came from the benchmark grid that was published. Basically, whenever a new model is released, there's this benchmark grid where they show all these different benchmarks on the x-axis and then the performance of different models on the y-axis and you can just compare different models. It's like a single number for a model on a single benchmark. And if you look on paper at the difference between like 5.5 and 5.4 or other models, it wasn't a huge improvement. It was only a few percentage points in some benchmarks. So, people looked at that and they were skeptical that it was actually a better model. Once they played around with it, the story changed. I think the reason why it doesn't show up as so much better on the benchmarks is because the benchmarks are being presented in the wrong way. They're not controlling for the amount of test time compute that is being used on that benchmark question. It turned out that 5.5 is just much more efficient with its thinking. If you run it at max settings, 5.4 is thinking for a lot longer. It takes longer to get back a response than 5.5. And once you control for the amount of thinking time, actually you can see that 5.5 is a substantial jump over 5.4. That is, I think, people's day-to-day experience with it. And then when I mention this to people, the typical question I get is like okay well why not just have 5.5 think for as long as 5.4 and the question is like well how long should they think for? Typically the response I get is well until the performance plateaus right there's at some point where the performance on the benchmark is going to plateau and you evaluate to that point. The thing is the point at which it plateaus is actually really far out these days. I mean, it was true in GPT-3 land back in 2022, the models couldn't really think productively for that long. And so, you could just run them until they plateau. It's not that far away. But what we're seeing today with the modern models is that 5.5 and other models can think for, if you scaffold them reasonably well, can think for weeks even before having performance plateau on some of these benchmarks. And so, the point at which they plateau is simply too far out to reasonably test. We all need to actually reinforce either a patience limit or a budget limit from a token perspective now, and that wasn't true a few years ago.

完全正确。所以我认为现在评估模型的正确方式是，要么为基准设定某种预算——无论是 token、成本还是时间——要么将性能绘制成测试时算力的函数，这样比较不同模型的性能就会清晰得多。

Exactly. And so I think the proper way to evaluate the models now is you either have some kind of budget for the benchmark whether it's tokens or cost or time or whatever, or you plot the performance as a function of the amount of test time compute that's going into the model and then it becomes much more clear how to compare the performance between these different models.

考虑到模型评估周期，以及许多任务的性能在相当长的时间内都不会趋于饱和，你如何应对这个问题？有些你想运行的评估，无论是预算还是时间，都超出了当前模型发布周期所能承受的范围。

Given the model evaluation cycle and the fact that performance does not asymptote for many tasks over quite a long period of time, what do you do about that issue? The fact that some of the evals that you would want to run are both beyond the scope of budget or time that's reasonable given the current model release cycle.

我认为，比如在网络安全方面，我们看到——实际上 AIS 的评估也显示——模型在 1 亿 token 时仍在改进。如果你运行 1 亿 token，它们还在继续提升，而且这需要很长时间。但你也看到性能并不是不连续的跳跃，而是可以看到在 1 亿 token 上的改进斜率。所以你可以先评估到某个预算，然后说：这是我们预测的性能。这方面研究还不多。我觉得如果哪位学者想找课题，这可以写一篇很好的论文：能否仅用 10 美元或 100 美元的推理预算，预测出 10,000 美元推理预算下的性能？

I mean, I think for things like cyber, we've seen and actually the AIS in their evaluations has shown that the models continue to improve at 100 million tokens. You know, if you run them for 100 million tokens, they're still improving beyond that point. And that can take a very long time to run. But you also do see that the performance is not just a discontinuous jump. It's actually like you can see the slope of improvement over those 100 million tokens. And so you could probably do some kind of evaluation up to a certain budget and then just say okay well this is what we project the performance to look like. And I think there hasn't been a lot of research on this yet. I actually think this would be a great paper to publish if there's any academics out there looking for something to research. Can you predict what the performance looks like at an inference budget of let's say $10,000 only using inference budgets up to $10 or $100?

那再问一个不太相关的问题：你觉得用户是不是系统性地让模型思考得不够久？

So maybe an orthogonal question for you. Do you think users are systematically not thinking long enough with their models about problems?

你说的“思考不够久”是什么意思？如果你能构建一个智能体或控制测试时算力的使用量，模型自身能做一部分，用户也能做一部分。你觉得行业对测试时算力的使用是最优的吗？是远远不足，还是模型本身的问题——它们需要更快地思考？

What do you mean by not thinking long enough? If you can build an agent or control the amount of test time compute being used, there's what is done by the model itself and there's what the user can do. Do you think that the industry is using test time compute an optimal amount, way undershooting it, or it's a problem in the models where they just need to be able to do that thinking faster?

我觉得这取决于具体问题。

I think it depends on the problem.

嗯，我觉得那种让模型思考一周再回应的想法，听起来不错，基准测试也确实好看，但实际用起来不太现实。因为你想，你问模型一个问题，然后得等上一周才能得到回复。

Uh I think this idea that the models you just let them think for a week or whatever and then they respond. It sounds nice and yes the benchmarks look great but it's not very practical when working because like okay you ask the model a question and then you sit there for a week waiting for it to come back to you.

我认为人们发现最有效的方式是与模型快速迭代，所以思考时间需要灵活。当需要快速响应用户时，就应该快速回应。而当需要长时间思考且用户也希望如此时，长时间思考就是合理的。我觉得在当前情况下，人们已经找到了合适的平衡点。

I think what people have found most effective is to kind of iterate quickly with the models and so the thinking time I think needs to be flexible. When it makes sense to respond quickly to the user, it should respond quickly. And then when it makes sense to think for a long time and the user wants it to think for a long time, then it makes sense to think for a long time. I think people have been striking the right balance given what they have to deal with right now.

你怎么看，现在有很多关于刷基准测试和操纵不同基准测试能力的讨论？你如何描述当前基准测试的格局？另外，你有没有特别偏好的、认为更能体现能力的基准测试？

How would you characterize, you know, there's a lot of talk about benchmark maxing and the ability to game different benchmarks? What would you characterize the landscape of benchmarks as today? And then do you have like favorites that you think are more indicative of capability than others?

刷基准测试也是我写那篇文章的动机之一。我认为很容易就能在基准测试上表现得比之前更好，比如通过把多个模型拼凑在一起。比方说，我们不是只运行一次模型，而是运行五次，然后取五次中最好的回答，或者让一个评判模型选出它认为最好的，这样就能得到比原模型高得多的分数。所以很容易在纸面上看起来好很多，但一旦控制测试时算力，实际上并没有更好。这是我对刷基准测试担心的一点。我的唯一顾虑是它有点误导性。至于基准测试本身，我认为总是存在过度优化基准测试的风险。我当然鼓励我的团队，而且我认为在 OpenAI 我们做得不错，不会刻意去优化某个特定基准。但一旦你发布一个基准测试，它就有被过度优化的风险。我认为解决这个问题的一个方法是保留一个不公开的私有测试集。

So the benchmark maxing thing is also motivation for writing the essay that I think it's really easy to show you can do much better than previous benchmarks or previous models on benchmarks by just for example scaffolding a bunch of models together. Um so if you say okay well we're going to instead of just running this model once we're going to run it five times and take the best of the five responses or like ask a judge which one it thinks is best then you can get much higher scores than that model. And so it's really easy to make something that looks a lot better on paper but is actually not better once you control for the amount of test time compute. That is one thing that I'm worried about when it comes to benchmark maxing. I mean it's a little misleading is the only concern that I have. And then as far as the benchmarks themselves, I think there is always a risk of just optimizing for the benchmark. And I've certainly encouraged my team and I think at OpenAI we're pretty good about not trying to optimize for specific benchmarks. But once you put out a benchmark, it's always at risk of just being optimized for. And I think one way to address that is to keep a held out private set that isn't publicly available.

判断一个模型是否明显更好的最常用建议就是亲自用一段时间。你有没有更复杂的方法推荐？比如除了 OpenAI 的私有测试集，你每次还会自己创建一套新的评估吗？

The most popular fallback advice for figuring out if a model is significantly better or not is to just play with it for a while. Do you have anything more sophisticated than that that you suggest people do? Like do you create your own set of new evaluations each time besides private holdback at OpenAI?

我觉得每个人都有自己的问题集，每当新模型出来时就会去问。对我来说，最近我用它们来制作扑克机器人，看看它们能做出多好的扑克机器人。我认为这是一个很好的评估，因为制作扑克机器人的开源代码很少，虽然有很多发表的论文，但你真的需要从头推理，需要大量的推理和迭代，还有很多小陷阱，我自己已经经历过，所以我能看出模型在哪些地方失败。现在它们已经非常擅长这个了。

I think everybody has their own set of questions that they like to ask the model whenever it comes out. Um for me lately it's been I use them to make poker bots and see how good they can make a poker bot. I think it's a nice eval because there is very little open source code for making poker bots and there's a lot of published papers on it but you really have to reason through everything and it's like requires a lot of just reasoning and iteration and like a lot of small gotchas that I can kind of I've already worked through myself so I can see where the models fail along the way. They've gotten really good at it now.

你能描述一下，在你们几次模型发布中，扑克机器人制作方面的推理能力是如何进步的吗？

Can you describe perhaps with your poker bot creation like how reasoning might have progressed in model releases for you guys over a few releases?

是的，早期模型在这方面非常糟糕，基本上什么都做不了。到了 5.2 版本，我能够和它一起制作一个河牌求解器。那是扑克的最后阶段。这本身我觉得非常令人印象深刻。我确实需要和它一起工作一点，但我真的很惊讶，因为我制作河牌求解器的速度可能比我单独做快五倍。有几件事让我卡住了。阻隔牌一直是个大问题。但总的来说，稍微引导一下，感觉就像和一个研究生合作：他们会遇到问题，但至少我知道问题是什么以及如何解决，我可以提建议，然后它就去执行，很快就能拿出很好的结果。嗯。特别是优化方面，我觉得非常出色。它能让代码比我快十倍，因为它优化得非常好。5.2 的缺点是它经常误导我，我总得非常小心地检查它，确保它真的做了它说的事情。有没有它没意识到或假装不是问题的明显错误？我记得有一次，我在玩一个模型（不是 5.2），作为一个单元测试，我告诉它：假设底池里有 100 美元，我弃牌了，我损失了多少？模型说 92 美元。我说，这太疯狂了。底池里有 100 美元，我弃牌了，怎么可能不是损失 100 美元？它说，哦，是 92，接近 100，没关系，没什么大不了的。我说，这显然是个问题，对吧？所以模型确实有这种经常误导你的问题。但到了 5.5 版本，我觉得好多了。它基本上可以零样本完成。实际上，我一直在做一个完整的扑克求解器。它基本上可以在我的轻微引导下完成整个工作。如果六个月或一年后，模型能够零样本完成整个扑克求解器，基本上一次性完成我的整个博士论文，我也不会感到惊讶。

Yeah, when the early models were really bad at it, like they could not basically do anything. And then 5.2 I was able to work with it to make a river solver. So that's like the final stage of poker. And that itself was I thought really impressive. I mean I had to work with it a little bit, but I was actually really impressed because I was able to make the river solver probably about five times faster than I would have alone. There were a couple things that I got tripped up on. Uh blockers was always a big big issue. But overall, like, you know, with a bit of gentle steering, it just kind of felt like a grad student where, okay, they would run into issues, but at least like I would know what those issues were and know how to fix it, and I could just make suggestions and it would go off and then do it and then pretty quickly it would actually come back with something really good. Mhm. And then especially the optimization I thought was very impressive. It was able to make it like 10 times faster than what I was able to do because it was just able to optimize the code so well. The downsides with 5.2 is I felt like it was gaslighting me a lot and I always had to be very careful checking it and making sure like, okay, is it actually doing what it said it did? Um, are there any things that are like glaring issues that it's not recognizing or it's just pretending aren't issues? Um, I remember there was like one point where for one of the models I was playing around with it, not 5.2, I kind of like as a unit test, I told it, okay, well, let's say I have $100 in the pot and I fold. How much am I losing? And the model said $92. And I was like, that's crazy. I have $100 in the pot and I just fold it. How do I not lose $100? And it said, oh, you know, it's 92. It's close to 100. It's fine. It's no big deal. And I was like, clearly this is a problem, right? So the models did have this problem where they would gaslight you a lot. But once we got to 5.5, I actually thought it was way better. It was able to basically do it zero shot. And in fact, I've been working on just doing a full scale poker solver. And it's basically able to do the whole thing with some gentle steering from me. And I wouldn't be surprised if you know 6 months or a year from now, the model is able to do zero shot an entire poker solver, basically my entire PhD thesis in one go.

我们来谈谈更大的影响，即需要根据推理速度、效率与 token 量或预算等标量来评估这些模型。你能描述一下你文章中包括安全评估在内的一些更大影响吗？

Let's talk about the larger implications of needing to evaluate these models relative to like speed of their reasoning or efficiency versus token volume or dollar budget or whatever your scalar is. Can you describe some of the larger implications in your essay including around safety evaluations?

是的，安全评估这件事。这有点像令人不快的事实。背景是，很多实验室都有所谓的负责任的扩展策略或准备框架，名称各异，但理念是每当模型发布时，都要经过一系列评估来衡量是否存在危险能力。这些模型能否做出我们不想让恶意行为者做的事情？如果模型能力不强，那就没什么大不了的。

Yeah, the safety evaluations thing. It's a bit of an inconvenient truth thing where okay, so I guess for background, a lot of the labs have these things called either responsible scaling policies, preparedness frameworks. They go by various names, but the idea is that whenever a model is released, they go through a series of evaluations to measure are there dangerous capabilities? Could these models do things that we wouldn't want a bad actor to do. And if the model isn't very capable, then it's no big deal.

但如果它非常强大，比如可能被用来制造生物武器，那么你就需要对此采取缓解措施。

But if it is very capable, if it could be used for example to make bioweapons, then you want to put in mitigations against that.

但问题是，你如何评估模型是否具备这种能力？他们有一些评估协议，但很多框架是在 ChatGPT 时代前后开发的，那时测试时计算扩展还不算什么。对于 GPT-3，你无法扩展测试时计算，即使给它 1000 万美元的预算，它也不会比花 10 美元或 1 美元做得更多。这些准备框架和负责任的扩展政策并没有真正考虑测试时计算量。他们只是说：“好吧，模型的能力是什么？”问题在于，我们现在所处的世界，模型的能力取决于你投入多少钱。基本上，如果你给它 1 万美元的预算，它能做的比 10 美元预算多得多。如果给它 1000 万美元，它能做得更多。那么，应该在什么预算下评估这些模型？现有的政策并没有真正解决这个问题。

But the question is okay, well, how do you evaluate whether the model is capable of that? And they have like various protocols about like how they do these valuations. But a lot of these frameworks were developed around the era of ChatGPT before after when test-time compute scaling was not really uh as much of a thing and it made sense like with GPT-3 you couldn't scale test-time compute like if you gave it a budget of $10 million and said okay well let's see what GPT-3 can do it really can't do that much more than what you could do with like $10 or $1. The preparedness frameworks and responsible scaling policies, they don't really account for the amount of test-time compute. They just say, "Okay, well, what's the capability of the model?" The problem is we're in a world now where the capability of the model is a function of how much money you put into it. Basically, if you give it a budget of $10,000, it can do a lot more than what it can do with a budget of $10. If you give it a budget of $10 million, it can do even more. And so, at what budget should you evaluate these models? The policies that exist today don't really address that question.

嗯。

Mhm.

有些做得好一些，但大多数情况下，这并非一个被重点考虑的因素。至于是否应该发布，我不想介入这个问题。我认为双方都有道理，但重要的是要认识到，这个问题被忽视了，我们只是在假装它不存在。我认为无论如何，都必须正视它。

Some do some do better than others, but for the most part, this is not really a factor that's being heavily considered. Now, whether it should be released anyway, I don't want to weigh into this question. I think there's arguments on both sides, but I think the important thing to recognize is that this is a question that is not being we're just kind of pretending that this issue doesn't exist. And I think it's important to just, one way or the other, account for it.

是的，这是能力问题的镜像：如果模型可以在非常大的预算下持续做更多事情，而没有在某些任务上达到瓶颈，那么它们也应该能够做我们社会不希望它们做的事情，对吧？因此，需要测试这一点以及分配多少预算。这似乎也与模型发布周期不同步，对吧？现在模型发布加速了，有时几天或几周就出一个新模型，而以前是 6 个月。你在文章中提到，真正评估一个智能体在长期任务上的表现，唯一的方法可能是让它运行一年。这对有用和有害的任务都是如此，对吧？那么你如何看待这一点与模型发布周期的关系？

Yeah, it's the mirror image of the capability question of if the models can continue to do more and more without asymptoting on some tasks at very large budgets, then they should also be able to do so for tasks we don't want them to do as a society, right? And so testing for that and what budget is allocated. It also seems out of sync from the model release cycle itself, right? There's been this acceleration of, you know, you get a new model every sometimes few days and weeks at this point versus 6 months. And you have a line in the essay where you say like the only way to truly evaluate an agent on some very long-running task might be to run it for a year. And that's going to be true of both like useful and negative tasks, right? And so how do you think about that versus the model release cycle?

是的，这也是一个有趣的动态：随着模型变得更强，它们能够更好地在更长的时间跨度内运作。以 GPT-3 为例，如果你想让它运行一周，你几乎无法将其搭建为能实际运行一周的有用工具。但我们现在看到，最新的模型，比如 o1，可以搭建起来进行一系列持续数周甚至数月的实验。

Yeah, this is also an interesting dynamic where basically as the models have become stronger, they're better able to operate over longer horizons. So again, with GPT-3, if you wanted to run it for a week, there's really not much you could do to scaffold it into something useful that could actually run for a week. But we're seeing now with the most recent models that you can actually scaffold, for example, o1 into doing a series of experiments that can run for weeks, for months.

你有没有给你的扑克求解器任务无限预算？

Have you given your poker solver task like infinite budget yet?

我还没有真正搭建一个系统，只是告诉它“好吧，运行几周”。我想我可以给它一个目标，然后让它放手去做。但我认为目前它 100%能完成河牌求解器。如果我只给它一个目标，我不认为它能达到完成完整扑克求解器的水平，如果我只给它一个目标并让它运行一个月。但我们很快就会达到那个点，我可能只需告诉它：“好，花一个月时间做这个，然后给我一个最先进的完整扑克求解器。”问题在于，如果你想评估模型运行一个月后的能力，唯一完全确定的方法就是让它实际运行一个月。如果你想了解六个月后的情况，唯一的方法就是运行六个月。稍后我们可以讨论一些应对措施。但重要的是要认识到，模型发布周期是每两三个月就发布一个新模型。一个模型出来，需要两三个月才能推向极限，然后下一个模型又出来了。所以没有人真正知道这些模型的能力上限，因为没有人让它们运行足够长的时间来真正了解。例如，当 o1 发布时，人们开始运行一些需要一周多才能完成的任务。因此，直到发布一周后，人们才意识到这有多重要。

I haven't really scaffolded something together where I just tell it like, "Okay, just run this for weeks." I think I could give it a goal and just tell it to go nuts. But I think at this point it could 100% do the river solver. If I just give it a goal, I don't think it's at the level yet where it could do like the full poker solver if I gave it just a goal and told it to go run for a month. But we're going to pretty soon be at that point where I probably could just tell it like, yeah, go work on this for a month and then come back to me with a full complete poker solver that's state-of-the-art. And the problem is if you want to evaluate the capabilities of a model, what it can do after running for a month, the only way to be fully sure is to actually run it for a month. And if you want to know after 6 months, the only way to know fully is to run it for six months. Now there are things we could do to address that a little bit later. But it's important to recognize the model release cycle is look, we're releasing new models like every two or three months at this point. And so a model comes out, it takes two or three months to push it to its limits and then you have another model come out. And so nobody actually knows what the ceiling of capabilities are for these models because nobody's actually run them for long enough to really tell. When o1 came out, for example, people started running things that took over a week for it to finish. And so, people actually didn't realize that this was a big deal until a week after it was released.

嗯。

Mhm.

我认为这种情况会越来越普遍。其影响相当有趣，因为实验室在发布前如何全面评估模型？实际上非常困难，因为唯一真正进行评估的方法就是延迟模型发布周期。而目前竞争压力很大，他们不会这么做。

I think that's going to be more and more true. The implications of that are, I think, pretty interesting because what do the labs do to like fully evaluate their models before they're released? It's actually very difficult because, yeah, you would have to the only way to really do the evaluations is then delay the model release cycle. And there's a lot of competitive pressure right now to not do that.

你认为已经发布的模型中是否存在令人兴奋的潜在能力，由于时间限制人们尚未充分探索？

Do you think there's like exciting latent capability in the models that are already released that people have not fully explored given timeline?

我绝对这么认为。实际上，一个很好的例子是埃尔德什单位距离问题。对于不了解的观众，几周前我们使用 OpenAI 的内部模型否定了单位埃尔德什单位距离猜想。我不是数学家，但这在数学界似乎是一件大事。这是许多数学家花费大量时间研究的第一个问题，而模型能够做到他们做不到的事情，并且以一种对数学家来说有趣且有用的方式完成。老实说，它的预算非常低。我们并没有投入太多精力。我们只是训练了一个新模型，好奇它能做什么。我们让它运行了一些问题。这个问题的预算很低，它就说：“哦，是的。”

I think absolutely. I think actually a really great example is the Erdős unit distance problem. So for the viewers that don't know, we used an internal model at OpenAI a few weeks ago to disprove the unit Erdős unit distance conjecture. Now, I'm not a mathematician, but this seems like it was a pretty big deal in the math community. It was like the first problem that a lot of mathematicians had really spent a lot of time on, and the model was able to do something that they weren't able to do and do it in a way that was actually interesting and useful for mathematicians. Honestly, it did it at a budget that was dirt cheap. I mean, we didn't put a lot of effort into this. We just trained a new model and we were just curious what it could do. And we ran it through some problems. And this one at a pretty low budget, it was like, "Oh, yeah."

我觉得我找到了一个反例。然后我们验证了，是的，这个证明是正确的。我们公布结果后，很多人发现你也可以从 5.5 那里得到答案。不过，这不像直接问 5.5 '嘿，这是单位距离猜想，反例是什么？'那么简单。你需要稍微搭建一下框架，稍微引导一下。所以有人发现：好吧，你让 5.5 列出一些解决这个问题的方法，然后它列出了其中一条有希望得到反例的路径，然后你告诉它'好的，再深入探索一下'。如果你这样重复足够多次，它最终会得出反例。这意味着，原则上你可以让 5.5 作为一个通用框架，列出许多不同的策略，然后对每个策略让它去调查，这样它很可能就能用通用框架得出反例。当然，这个框架会非常昂贵。我的意思是，粗略估计，大概要花一千到十万美元。但这是可能的，而且在我们之前，有人本可以用通用模型推翻厄尔多斯单位距离猜想。但没有人充分探索过：如果我把价值十万美元的算力投入 5.5，它能做什么？答案是，是的，你很可能从中得到类似的东西。所以人们应该更多地用当前这一代模型进行实验。

I think I have a disproof. And then we were able to verify that yeah, this proof is correct. After we announced the results, a bunch of people found that you could get the answer out of 5.5 as well. Now it's not as simple as just asking 5.5, 'Hey, here's the unit distance conjecture, what's the disproof?' You had to scaffold it a bit, you had to steer it a bit. So somebody found: okay, you ask 5.5 to list a bunch of ways that you could tackle this problem, and then for it, it lists one of the paths that are actually promising to get to the disproof, and then you tell it like, 'Okay, explore this some more.' And if you do this enough times, it actually ends up arriving at the disproof. Now what this means is you could in principle ask 5.5, as a general-purpose scaffold, to list a bunch of different strategies, and then for each strategy tell it to investigate that strategy, and then it would probably be able to arrive at the disproof with a general-purpose scaffold. Now that scaffold would be very expensive. I mean, it would probably cost, I just ballpark, like a thousand to $100,000. But it would be possible, and it would have been possible for somebody to disprove the Erdos unit distance conjecture before we did using a general-purpose model. And nobody had explored sufficiently what happens if I put $100,000 worth of compute into 5.5, what could it do? And the answer is like, yeah, you probably could get stuff like that out of it. So people should be experimenting more with the current generation.

嗯，这是一个有趣的问题：是否值得去实验？因为模型发布周期是每几个月一次。我们推出一个更强大的新模型，推翻厄尔多斯单位距离猜想的成本每次发布周期就会下降 10 倍或 100 倍，有些情况下甚至更多。你见过那个梗图：'哦，我为什么要费劲做工程工作？我应该等下一个模型，去度假，两个月后回来，它便宜了一千倍。'那么，你同意吗？你现在在 OpenAI 就是这么做的吗？就等着下一个模型发布？

Well, this is an interesting question of whether it's worth it to experiment, because again, the model release cycle is every couple months. We put out a new model that's even more powerful, and so the cost of disproving the Erdos unit distance conjecture drops by like 10 or 100x with every model release cycle, probably in some cases more. You've seen the meme that's like, 'Oh, why bother doing any engineering work when I should just wait for the next model, go on vacation, come back two months later, and it's a thousand times cheaper.' So, do you agree with that? Is that what you're doing right now at OpenAI, just waiting for the next model release?

我想，我是说，我们正处于一个进步非常快的时期，是的，模型正变得越来越强大。我可以说，在 OpenAI，我们积极避免的一件事是——你看，我们有很多数学家、物理学家，他们对这些模型现在能做的事情非常兴奋，尤其是内部模型。我们试图鼓励人们不要把时间都花在遍历所有数学开放问题、物理问题上，只是把模型推到极限，看看它们能证明或推翻什么。因为我们真的认为重点应该是如何制造更强大的模型？如何尽快将它们安全地推向世界，让全世界的科学家都能用这些模型自己解决问题。所以，是的，在某种意义上，我们正在考虑：是的，把所有精力都放在扩大这些模型规模、看看它们现在在极限状态下能做什么，这确实很诱人，但真正的重点应该是如何利用这些模型来制造更强大、更有能力的模型，这些模型可以更经济高效地完成所有事情。

I think, I mean, I will say we're in a period where progress is very fast, and yeah, the models are becoming more capable. I can say at OpenAI, one of the things we're actively not doing—and look, we have a lot of mathematicians, we have a lot of physicists, people are very excited about what these models can do right now, especially the internal models. We are trying to encourage people to not spend all their time just going through all the mathematical open problems, physics problems, and just seeing pushing the models to their limits to see what they can prove or disprove. Because we really think the focus should be on how do we make even more capable models? How can we get them out safely to the world as quickly as possible so that all the scientists in the world can use these models to solve the problems themselves. So yeah, in some sense we are thinking about this: yes, it's really tempting to just put all of our efforts into scaling up these models and see what they can do at their limits right now, but really the focus should be on how do we use these models to make even more powerful models, even more capable models that can do everything much more cost-effectively.

考虑到你对大规模测试时算力巨大影响的看法，你认为研究的方向或资源分配正在发生什么变化？这与递归自我改进的想法如何相互作用？例如，递归自我改进是任何实验室获得最佳能力模型的主导思路。

What is changing about the direction or allocation of resources for research in your mind, given your beliefs about this very large-scale impact of very large-scale test-time compute? How does this interact with the idea of recursive self-improvement, for example, where it's a dominant idea for how any lab gets to the best capability model?

所以，有一件事我应该澄清：我不认为我们已经到了那个地步，即你给它任意极高的推理预算，它就在所有方面都超级智能。那么，我们和那个目标之间还有什么差距呢？我认为，通过摆弄模型，首先，有一些基准测试中，即使给模型更多的推理预算，它们也不会改进。所以，我认为很多事实检索类问题就属于这一类：如果你问一个人亚伯拉罕·林肯是什么时候出生的，而他们不知道日期，他们可以坐在那里想一个星期，但如果他们无法访问维基百科之类的资源，他们思考一个星期也不会比思考五秒钟回答得更好。模型也一样。有趣的是，如果你给模型这类事实检索问题，并给它们一点思考时间，它们确实会做得更好。但如果你给它们一个星期，它们不会突然在记忆日期方面做得更好。所以有些基准测试中，它们显然会随着更多测试时算力而改进，有些则不会。我认为在另一个极端，有些基准测试中，它们显然会随着更多测试时算力而无限制地改进。我喜欢举的例子是数独。解决数独有一个非常简单的策略：尝试一堆不同的随机数字，然后看它是否符合条件，是否满足所有约束，如果不满足，就再试一组不同的随机数字组合。显然，只要有足够的时间，你就能用这个策略解决任何谜题。你可以很简单地看到，任何模型只要得到更多的测试时算力，就能不断做得更好。所以所有基准测试都存在于这两个极端之间。模型还没有达到那种水平，即只要给它们足够的测试时算力，它们就能完成我们所有的工作，因为有些基准测试它们不会改进。有些事情它们不会改进。我特别注意到，在研究方面，它们目前还没有很好的研究品味。所以我认为它们实际上是研究人员的很好补充，尤其是，我发现使用这些模型后我效率更高了，但它们还不能完全取代整个研究周期。那么，这会随着时间改变吗？可能吧。我的意思是，我认为模型正在全面进步。有些事情进步得比其他事情快。但它们还没有达到仅仅通过足够的测试时算力就能完全取代研究人员的程度。

So, one thing I should clarify: I don't think we're at the point where, okay, you just give it an arbitrary extremely high inference budget and it's just super intelligent across the board. What's between us and there then? I think having played around with the model, so okay, first of all, there are some benchmarks where the models will just not improve if they have more inference budget. So, I think a lot of factual retrieval kind of questions fall into this category: if you ask a person when Abraham Lincoln was born and they don't know the date, they could sit there, they could think about it for a week, if they don't have access to Wikipedia or something, they're not going to be able to do better answering that question if they thought about it for a week compared to 5 seconds. Same with the model. If you actually, interestingly enough, if you give the model these kinds of factual retrieval questions and you give them a little bit of time to think, they do actually do better. But if you give them a week, they're not suddenly going to do better at remembering dates. So there are some benchmarks where they clearly improve with more test-time compute and some where they don't. I think on the other extreme, there are benchmarks where they kind of obviously will keep improving without limit with more test-time compute. So the example I like to point to is Sudoku. There's a really simple strategy to solving Sudoku: just try a bunch of different random numbers and then see if it fits the criteria, if it matches all the constraints, and if it doesn't, just try a different random combination of numbers. Clearly, with enough time you will be able to solve any puzzle with this strategy. You can kind of trivially see that any model could keep doing better and better if it was just given more test-time compute. So you have all the benchmarks kind of existing somewhere between these two extremes. The models are not at the level where if you just give them enough test-time compute, they will be able to do all of our jobs, just because there are some benchmarks where they will not improve. There are some things where they will not improve. One thing I see for research in particular is they don't have very good research taste right now. So I think they're actually a very good complement to researchers, especially, I've found I'm much more effective by using these models, but they're not able to fully replace the whole research cycle. Now, does that change with time? Probably. I mean, I think the models are getting better across the board. Some things are getting better faster than others. But they're not at the point where they're fully replacing researchers with just enough test-time compute.

你能举一两个例子吗，比如让模型做研究任务，就像说‘这是个糟糕的主意’？

Can you give an example or two of asking the model to do a research task, like this is a terrible idea?

回到我的扑克求解器例子，模型优化我博士期间开发的算法的能力让我印象深刻。回想起来，我惊讶于自己当时效率有多低，它们能把速度提升 10 到 100 倍。然后我问：你能提出一个比我或任何人想出的更好的算法吗，查看所有已发表的工作，综合起来，尝试提出新颖的东西？它做不到。即使给很多时间，它还是做不到。如果我用某种框架约束它更多，也许最终能想出更好的东西，但那需要很多功夫。这不像说‘请提出一个更好的算法’那么简单。

Going back to my poker solver example, I was really impressed with the model's ability to optimize the algorithms I had developed in my PhD. It was shocking to see how inefficient I was in retrospect, and they were able to make it 10 to 100 times faster. Then I asked, can you come up with an algorithm better than what I or anyone else came up with, look at all published work, synthesize it, and try to come up with something novel? It's not able to do it. Even with a lot of time, it still can't. Now, if I scaffolded something and constrained it more, maybe it could eventually come up with something better, but it would take a lot. It's not as simple as saying, 'Please come up with a better algorithm.'

你认为这如何改进？

And how do you think that gets improved?

我观察到的是，随着每个模型发布周期，它在这些事情上确实变得更好。在我看来仍然很差，但不像以前那么差了。如果有一天，就像编程和数学那样，出现一个拐点，突然变得足够好、足够有用，我不会惊讶。如果我们在研究品味上也遇到那个点，我也不会惊讶。

What I've seen is with every model release cycle, it does get better at this sort of thing. It's still bad in my opinion, but it's not as bad as it used to be. I wouldn't be surprised if at some point, like with coding and math, there's an inflection point where suddenly it's actually good enough to be useful. I wouldn't be surprised if we encounter that point for research taste as well.

鉴于此，你如何看待今天的 RSI（递归自我改进）？我们应该怎么理解它？

Given that, what is your framing of RSI today? How should we think about it?

模型确实在加速实验室研究人员的工作。但我认为它们加速了一些事情，而不是全部。目前，如果某件事快了 100 倍，你就会受到那些没有快 100 倍的事情的瓶颈制约。随着时间的推移，我们受瓶颈制约的事情会减少，在这方面会有一个渐进的起飞。但这更多是转变研究人员的工作，而不是完全取代他们。

The models are definitely accelerating what researchers can do inside the labs. But I think they are accelerating some things and not others. Currently, we're at the point where if something goes 100 times faster, you get bottlenecked by the things that don't go 100 times faster. Over time, the things we're bottlenecked on will shrink, and there will be a kind of gradual takeoff in that respect. But it's more about transforming what researchers do rather than fully replacing them.

所以这实际上意味着你认为我们现在离非常快的起飞还很远。

So that actually implies you don't think we're close to a very fast takeoff right now.

我认为快速起飞是相对的；事情进展得非常快。但有一种假设是隔夜智能爆炸，模型发现突破让自己变得更聪明，立即导致更多突破，瞬间模型在各方面变得超级人类。我不认为我们会走向那个世界，很大程度上是因为模型依赖大规模测试时算力来实现其最大智能。如果需要那么多测试时算力来解锁全部能力，那么你就受到时间的瓶颈制约。事情只能进展得那么快，因为模型需要运行足够长的时间才能做真正强大的事情。时间本身成了瓶颈。我认为目前很多实验室就是这种情况。最终，我们所有人最大的瓶颈是时间，这就是为什么研究人员如此拼命工作。我们都看到了积压的能力，我们只是受限于做事情的速度。

I think fast takeoff is relative; things are moving very fast. But there is this hypothesis of an overnight intelligence explosion where models discover a breakthrough to make themselves smarter, leading to more breakthroughs immediately, and in moments models become superhuman across the board. I don't think we're headed to that world, largely because models rely so much on large-scale test-time compute to achieve their greatest intelligence. If it requires so much test-time compute to unlock full capabilities, then you're bottlenecked by time. Things can only go so fast because models need to run long enough to do something really powerful. Time itself becomes a bottleneck. I think that's the case right now for a lot of labs. Ultimately, the biggest bottleneck for all of us is time, and that's why researchers are working so intensely. We all see the overhang, the capabilities, and we're just bottlenecked by how quickly we can do things.

你认为前沿有哪些现在探索得较少？我们之前谈过多智能体。

What do you think is on the frontier that is less explored now? We've talked about multi-agent before.

我认为多智能体已经被探索得相当多了。在足够规模下，我认为还有很多可以做的。但这也是那些难以在小规模进行的事情之一。特别是多智能体，需要前沿模型才能完全解锁能力。我们看到了一些相当有趣的多智能体框架；它们能做很多事，但真的只是触及表面。我的一种思考方式：如果你看人类文明，人类在过去 5 万年里并没有变得更聪明，但今天能做更多，因为数十亿人类长时间思考并彼此积累的知识上构建。与 5 万年前相比，我们有非常好的检索和框架。这甚至不是框架；这是人类能够积累知识、分享并在此基础上构建的有机涌现特性。我们今天在 AI 模型上看不到这一点。它们出生在一个世界，存在很短的上文窗口，然后消失。你可以做一些事情来延续它们，但非常有限。我确实认为最终我们会进入一个它们能够大规模协调的世界。Multibook 和 OpenClaw 刚出来时有点被过度炒作，但它们指示了未来可能的方向。我确实认为最终我们会达到那种某种协调复合状态的世界，模型能够在更全球化的层面上分享知识并富有成效地在此基础上构建。

I think multi-agent is quite explored. At sufficient scale, I think there's a lot more that could be done. But it's also one of those things that's hard to do at small scale. Multi-agent in particular really requires frontier models to fully unlock capabilities. We've seen some pretty interesting multi-agent scaffolds; they can do a lot, but it's really just scratching the surface. One way I think about it: if you look at human civilization, humans haven't become smarter over the past 50,000 years, but they can do a lot more today because billions of humans have been thinking for a long time and building off each other's accumulated knowledge. We have very good retrieval and scaffolding compared to 50,000 years ago. It's not even a scaffold; it's an organic emergent property of humans being able to accumulate knowledge, share it, and build off it. We're not seeing that with AI models today. They are born into a world, exist for a very short context window, and then disappear. There are things you can do to continue them, but it's very limited. I do think eventually we will enter a world where they can coordinate on a large scale. Multibook and OpenClaw when they first came out were a bit overhyped, but they were an indication of where things could go. I do think eventually we get to that kind of world of some sort of coordinated compounding state, the ability of models to share knowledge on a more global level and build on that knowledge productively.

鉴于这些信念和你的工作，如果没有隔夜起飞，你如何描述前沿三大阵营之间的竞争？就是研究人员埋头苦干，做出好的高品味算法和投资决策，决定方向、算力分配、政策决策和评估决策。这感觉比冲向某个没人能追上的立即硬起飞更接地气。

Given this set of beliefs and your work, how would you characterize competition at the frontier between the three kingdoms if there is no overnight takeoff? It's just researchers grinding away, making good high-taste algorithmic and investment decisions about where to go, compute allocation, policy decisions, and eval decisions. That feels slightly more grounded than racing towards some immediate hard takeoff that nobody can catch you on.

我认为现在的竞争非常激烈。我确实认为今天的模型正在加速前沿实验室研究人员的工作。目前这有局限性，但利用模型改进模型研究的能力是真实存在的，它是一种放大力量。我认为这将继续成立。

I think the competition is very intense right now. I do think the models that exist today are accelerating what researchers at the frontier labs can do. There are limits to that right now, as I said, but the ability to use models to improve model research is a real thing, and it is an amplifying force. I think that will continue to be true.

我认为它们会随着时间的推移变得更加真实。让我感到安慰的是，我认为所有前沿实验室的研究人员都认识到利害攸关的是什么，以及这些模型的风险是什么。这让我感到欣慰，因为我觉得每个人都真正理解这是一件相当严肃的事情，它可以带来非常好的结果，也可以带来非常糟糕的结果。是的，实验室之间存在竞争动态，但我们也可以尝试找出如何共同实现积极结果，而不是非常消极的结果。

I think they'll become more true over time. One thing that I am comforted by is I think all the researchers at the frontier labs like all the frontier labs I think recognize what is at stake and what these models like what what the uh what the risks are and that's something that I I find comforting that I think everybody really understands like okay this is a pretty serious thing and it can lead to really great things or it can lead to really bad things and yes there's a competitive dynamic between the labs but like we can also try to figure out how we all get to the positive outcomes rather than the very negative outcomes.

我想你知道，我如果不问这个问题就失职了，因为你很早就正确认识到了测试时计算和推理作为框架的重要性。那么，你使用模型的方式中，有哪些是你应该鼓励其他人也去做的？是不是就是设定目标，然后一切交给模型？

I think you know I'd be remiss to ask just because you have been right very early for a long time um on the importance of test time compute and reasoning as a framework like are there ways in which you use the models that you should you would encourage others to right is it just goal everything

我认为对很多人来说，他们——我的意思是，这对你的听众来说可能不一定成立，但有很多人在 2012、2013 年左右尝试过 AI，觉得不能信任输出，因此不会将其用于真正高风险的决策。但实际上，我认为模型已经进步到非常适合这类事情的程度。比如，我向它咨询税务建议，或者最近我买了一套公寓，我问它需要填写哪些文件，以及这些文件是什么意思。它对于这类问题真的很好。嗯。所以我每天都在用它做很多这类事情。我认为它们现在——实际上已经有一段时间了——我感觉我可以信任输出，甚至可以说比信任人类专家更信任。是的。

I think for a lot of people they worked I mean this is probably not even true for your audience necessarily but there's a lot of people that experimented with AI back in like 2012 23 and felt like they couldn't trust the outputs and then don't use it for really high stakes decisions. And actually, I think the models have progressed to a point where they are very good for these kinds of things. I mean, I asked them tax advice or I bought a condo recently and I was asking it for advice on like, okay, well, what's all the paperwork that I have to fill out and like how to how do I what does it all mean? It's actually really good for these kinds of questions. Mhm. So, I use it dayto-day for for a lot of this kind of stuff. And I think they're at a point now where they've actually been at a point for a while now where I feel like I can just trust the outputs arguably more than I could trust the output from from a human, an expert human. Yeah.

好的。我有两个最后的问题。嗯，一个是，你认为研究界的其他人有什么不同意你的地方，或者还没有完全理解其重要性？哦，这些问题真好。我希望我有时间提前思考一下。你可以就和我一起待着，想想。是的。现在是不是有点奇怪，变成了共识？三年前你还有点不满，你说为什么人们不理解这有多重要？

Okay. I have two two final questions for you. Um, one is uh is there something you think that the rest of the research community doesn't agree with you on or doesn't understand the importance of quite yet? Oh, these are such good questions. I wish I had time to think about this ahead of time. You can you can just hang out with me and think about it. Yeah. Is it weird to be like consensus now? You're a bit salty three years ago when you're like, why don't people understand how important this is?

我仍然觉得这不是共识，因为你知道人们仍然不以这种方式发布基准测试。哦，确实如此。是的。这实际上就是为什么——我认为这是惯性。是的，但这正是我写那篇文章的原因。我只是想说，我的意思是我们可以讨论这个，但动机的一部分是，我和研究人员交谈时提到，用 x 轴展示基准测试是有意义的，无论是 token 数、成本还是时间，应该有一个 x 轴。每个人都会说，是的，有道理，我们应该这样做。但每个人都没有以应有的重视程度去行动，比如我们必须衡量正确的东西。实际上，他们的回应是，人们期望我们发布那个网格。然后，为什么人们期望发布网格？因为每个人都发布网格。所以你最终陷入了一个糟糕的均衡，每个人都知道这是一个糟糕的均衡，但没有人想打破它。我觉得，好吧，如果我站出来说，看，伙计们，让我们都认识到我们处于一个糟糕的均衡，然后转向一个不同的均衡，我们用 x 轴来绘制东西，希望下次有模型发布时，公司可以放心地不发布那个网格，至少不在最前面、最显眼的位置发布。这样我们就可以对这些模型进行更有成效的评估。

I still I still feel like it's not consensus though because like you know people still don't publish the benchmarks this way. Oh, that's true. Yeah. Like that's actually why I think that's like inertia. That's kind of Yeah, but that's kind of why I wrote the essay. I was just like look I mean we can talk about this but like yeah the part of the motivation is like I would talk to researchers about um we it makes sense to show the benchmarks with an x-axis whether it's tokens or cost or time there should be an x-axis and everybody would say like yeah that makes sense we should do that but everybody they're not acting with the importance of like good heart like this is we have to measure the correct thing well really their response is people expect us to to publish the grid and then okay why do people expect the grid to be published because everybody publishes the grid. And so you kind of end up in this this bad equilibrium where everybody kind of knows that it's a bad equilibrium, but like nobody wants to break out. And I I felt like, okay, well, if I just hopefully come out and say like, look guys, let's all recognize that we're in a bad equilibrium and let's move to this different equilibrium where we're we're plotting things with an X-axis that hopefully that can, you know, next time there's a model release, a company can feel comfortable not publishing the grid, at least not at the very front uh the top line. and um we can have a more productive evaluation of these models.

那么最后一个问题。你如何看待那些在专业领域的公司，它们认为自己的价值本质上就是路由层、选择层，你知道，我的目标由一系列离散任务组成。有些任务需要更多智能，有些需要更少。作为供应商，我的工作是在考虑预算约束的情况下解决问题或实现最优结果。所以我会管理并行化，以及从哪个模型上花费多少推理算力。因为我认为前沿实验室的观点是，这种路由既发生在 API 背后、应用程序背后，也有一部分发生在模型本身。嗯，这些部分显然正在被所有这些应用程序外部化。

Then a last question for you. How do you think about companies across all of these specialized domains who feel the value that they have is essentially like the routing layer, the choice layer of you know my goal is composed of a bunch of discrete tasks. Some require more intelligence and less and within my my job as a vendor is to um solve that problem or achieve the optimal outcome with uh taking into account the budget constraints. And so I will manage like the parallelization and how how much inference do you spend on it from what model because I I think the the frontier lab point of view is that that routing happens both within the you know behind the API behind the application and then some of it in the model itself. Um and that's pieces of that are clearly being externalized in all these applications.

是的，我确实认为这与基准测试应该用 token 数或成本的 x 轴来评估有关。嗯，我最近看到一些评估显示，通过路由层，你可以通过模型之间的共识获得更好的性能。是的。我绝对相信，如果你在模型之间做共识，你会获得比任何单个模型更好的性能，但重要的是要问，你是否会比让那个模型思考更长时间做得更好？嗯，一旦你控制了测试时计算量，它实际上是否仍然更好？这是你想要弄清楚的问题。好的，这是一个非常有原则的观点，即路由没问题，但都受制于同样的预算问题。是的，没错。嗯，如果你把它放在同一个尺度上，那么你可以做出最优决策，我想也许我会赢。嗯。我甚至不一定相信路由会更好，但还有一个问题：它是否会显著更好？它是否非常脆弱？它是否反映了现实世界的用例，与基准测试相比？因为你可能遇到的一个问题是，你可以通过路由针对某些基准测试进行优化，然后显示，哦，是的，我们在这些基准测试上看到了很大的改进。但在现实世界的用例中，它实际上并没有显著的改进。所以我会说，至少，你要控制测试时计算，然后你还要对基准测试保持通常的怀疑态度。

Yeah, I do I do think this is related to the fact that like benchmarks should be evaluated with an x-axis of tokens or cost. Um I I have seen some evals recently that show like okay well with with a routing layer you can achieve much better performance um by basically doing consensus among the models. Yeah. And like I definitely believe that if you do consensus among the models that you're going to achieve better performance than any individual model but it's important to ask like are you going to do better than having that model basically think for longer. Um like once you control for the amount of test time compute is it is it actually still doing better? That's that's the question that you want to figure out. Okay, that's very principle of view which is like yes routing is fine but it's all subject to the same budget question. Yeah, right. Um if you put it on the same scaler then you can make an optimal decision and I think maybe I win. Mhm. I I I don't even know necessarily that I I would believe that the routing does better but then there's still a question of um is it going to do significantly better? Is it very fragile? Is it um reflective of real world use cases compared to benchmarks? because like one issue you could run into is that you could optimize for certain benchmarks with the routing and then show like oh yeah we see this big improvement on these benchmarks. Um but in real world use cases it actually ends up not being a significant improvement. So I I would say at the very least like I would say you want to control for testime compute and then you also want to have all the same um skepticism about benchmarks that you would normally have.

太棒了。非常感谢你，也感谢你致力于打破这种虚假均衡的使命。是的，很高兴回来。在 Twitter 上找到我们，账号是 no prior pod。如果你想看到我们的脸，请订阅我们的 YouTube 频道。在 Apple Podcasts、Spotify 或任何你收听的地方关注这个节目。这样你每周都能收到新一期。

Awesome. N thanks so much and and for being on the mission for uh breaking us out of this false equilibrium. Yeah, it's great to be back. Find us on Twitter at no prior pod. 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.

请访问 no-briers.com 注册邮件或查找每期节目的文字稿。

And sign up for emails or find transcripts for every episode at no-briers.com.