进化式 AI 与科学发现：对话 Robert Lange

Robert Lange，非常高兴你能来 MLST。

Robert Lange, it's amazing to have you on MLST.

谢谢 Tim，很高兴再次做客。

Thank you, Tim. It's a pleasure to be back.

你现在在 Sakana 工作？跟我们聊聊吧。

So, you're working for Sakana? Tell us about that.

Sakana 是一家日本 AI 初创公司，主要做面向日本的 AI，同时在研究方面探索一些新颖或宏大的想法。

Sakana, yeah, is a Japanese AI startup working mostly on AI for Japan and at the same time exploring novel or ambitious ideas on the research side.

公司成立已经一年多了，你是创始研究员之一，对吧？

It's been around for over a year now. You're one of the founding researchers, right?

没错。Sakana 到现在已经快两年了，大概一年九个月吧。回顾早期，公司在组织上变化很大，但精神上我们一直在践行 Ken Stanley 的开放结局理念，探索许多目前在机器学习社区可能得不到资源支持的想法。

Exactly. So, Sakana has been around for now almost 2 years, like 1 and 3 quarters, I would say. And it's pretty fascinating to look back at the early days and how much the company organizationally has changed, but in spirit we're trying to embrace Ken Stanley's open-endedness idea and explore many different ideas which might not get the resources right now in the ML community more generally.

我们还有几期在日本的 Sakana 访谈即将上线。我就不剧透了，CEO 是 David Ha。你知道，像 Clune 和 Stanley 这样的巨擘，David Ha 也是其中之一。

And we've got a few interviews coming out with Sakana that we filmed here in Japan. So, I won't spoil the surprise, but the CEO is David Ha. And David, you know, there are these epic giants out there like Clune and Stanley. David Ha is one of these people.

David 的工作对我个人的博士研究影响很大。他在超网络和神经网络调制方面做了很多有趣的工作，还有进化计算和进化优化，这些也塑造了我博士期间的研究方向。

David's work has had a lot of influence on my personal PhD. He did a lot of fascinating work on hypernetworks and modulation in neural networks, but also on evolutionary computation and evolutionary optimization, and that sort of also painted my path during the PhD.

你们发表了一篇名为《Ashinka Evolved》的论文，我们刚说到这名字有点“进化进化”的意思，因为日语里 shinka 就是进化，这在日语里挺常见的，喜欢用多语双关命名。在聊这个之前，我们采访过 Alpha Evolved 团队，几周前也采访了 Jeremy Berman。你们的论文更像是这些工作的进阶版——用语言模型生成程序，采用进化方法：生成程序、改进程序、用评估器评估，迭代多步。你们的方法做了很多其他方法没做的事。给我讲讲这篇论文吧。

You've released a paper called Ashinka Evolved and we were just saying that kind of means evolve evolve because in Japanese shinka is evolve, but that's quite common. It's common to have these multilingual double namings in Japanese. Just before we get there, we interviewed the Alpha Evolved team and I also interviewed Jeremy Berman a few weeks ago. And your paper is very much like a more sophisticated version of those in the sense that it's using language models to generate programs and it's doing an evolutionary approach where we generate the program, we refine the generated program and we have an evaluator and we do this over several steps. And your approach does many things that the other ones don't do. Tell me about the paper.

首先，这篇论文部分受到了 Alpha Evolve 的启发。我认为那是很棒的工作，我认识 Alex 和 Matei，他们在做非常了不起的科学。使用这些进化式 LLM 方法，一个关键点是样本效率。很多系统对给定任务采样上千个程序，而我们 Shinka Evolve 试图通过引入一系列技术创新来降低成本和计算评估时间。我们证明，只需极少的程序评估，就能改进他们论文中展示的圆填充等经典结果。更广泛地说，我认为我们正处在一个转折点，这些进化驱动的 LLM 系统能够真正革新科学发现。我们希望朝更民主化的方向迈出了一步。代码是开源的，由于样本效率高，我们希望很多人能使用这个系统，做出自己的科学发现。

First off, of course, this was partially inspired by Alpha Evolve. I think it's great work. I know Alex and Matei and I think they were doing incredible science. One thing that is important about using all of these evolutionary LLM-driven methods is sample efficiency. Many of these systems sample like a thousand programs for a given task, and what we tried to do with Shinka Evolve was to essentially cut down costs as well as computation evaluation time by introducing a set of technical innovations to this evolutionary search. We showed that it's possible with very few program evaluations to basically improve upon, for example, the circle packing canonical result that they showed in their paper. More generally speaking, I think we're right now at an inflection point where these evolutionary-driven LLM systems can really revolutionize scientific discovery. We hope to have made a step forward to making this more democratically accessible. The code is open source available, and by its sample efficient nature, we hope that many people can interact with the system and make their own scientific discoveries as well.

这确实是个重要观点。我们能用这些基础模型，而且想想看，像 GPT-5 和 Grok-4 这样强大的模型，让它们多步改进解决方案时效果会好得多，这难道不迷人吗？为什么？一个天真的问题是，为什么它们不是开箱即用就很好？

Yeah, that's actually a really important point because I suppose we can use these foundation models. And first of all, isn't it just fascinating to reflect that we have these amazing models out there that we can access, like GPT-5 and Grok-4, and they are so much better when you get them to refine their solution in several steps. Why is that? I mean, a naive question would be why aren't they just good out of the box?

理论上，如果有足够多的随机样本，就像猴子敲键盘，它们最终也可能做到。但原则上，这又回到了进化原理：你需要先收集一堆垫脚石，然后在此基础上构建，才能真正找到创新或逐步优化创新。我认为，语言模型配上合适的进化框架，在规模化发现方面极其强大。

Potentially, with enough random samples, it's sort of this monkey typing on the keyboard, they would potentially be able to get there. But in principle, it's coming back to the principles of evolution: you need to collect a bunch of stepping stones first and then build on top of them to really find innovations or to tune innovations down the line. I think language models with the right sort of evolutionary harness are extremely powerful in terms of scaling up to make discoveries.

我觉得进化论中的很多类比都适用于科学研究，对吧？我们遍历一棵由不同想法或实验组成的树，然后在论文中只报告其中一条路径。

I think a lot of analogies from evolution transfer to scientific research, right? In the sense that we traverse a tree of different ideas or different experiments and then in the paper we report one path through that tree.

当我们自主运行 LLM 时，它们往往什么有趣的事也不会发生。

When we run LLMs autonomously, they tend to just kind of like nothing interesting happens.

但很多时候，针对特定问题的创新可能需要先发明另一个问题，对吧？自动提出这种归约或问题求解的递归性质，这些系统目前并不一定内在地具备。很多时候，生成大量解决方案比严格验证它们要容易得多。

But oftentimes innovation for a specific problem might require first inventing a different problem, right? Sort of automatically coming up with this reduction or this recursive nature of problem solving is something these systems right now don't necessarily have built in intrinsically. Often times it's easier to generate a lot of solutions than to actually hard verify them.

我之所以还不那么担心劳动力市场被颠覆，是因为我仍然深信人类是世界上深刻理解和创造力的源泉。如果我不相信这一点，我会非常担心。所以我认为它会是人类擅长那些潜在维度的放大器。

The reason why I'm not that worried yet about labor market disruption is I still believe deeply that humans are the source of deep understanding and creativity in the world. If I didn't believe that, I would be very worried. So I think it's going to be an amplifier of these latent dimensions humans are great at.

我认为一个关键转折点是当新的 Transformer 架构或类似重大突破被 AI 发现，而我们都在使用它的时候。

And I think one of the Rubicon moments is when the new Transformers architecture or something massive is discovered by AI and we're all using it.

Nvidia GTC 周一在圣何塞开幕，线上免费参加。本周已经泄露了一个叫 Nemo Claw 的东西，它是一个开源智能体平台，如果属实，可能是今年最大的公告之一。仅凭这一点就值得看 Jensen 的主题演讲。我要送出一台 DGX Spock。Nvidia 刚涨价 700 美元，你可能听说了内存短缺的事，对吧？所以现在售价 4700 美元，非常非常贵。顺便说一句，Hugging Face 的 Merve 过生日时收到了一台，她说她真的哭了。所以这真是一套很酷的设备。如果你通过描述中的链接注册，并至少参加一场会议，就有机会抽奖。这是一个大型会议。物理 AI 和机器人将是突破性主题，Jensen 的主题演讲在周一太平洋时间上午 11 点。链接在描述中，别错过。

Nvidia GTC starts Monday in San Jose and it's free to attend virtually online. There's already been a leak this week of something called Nemo Claw, which is an open-source agent platform and if it's real, it could be one of the bigger announcements this year. So it's definitely worth watching Jensen's keynote for that alone. I'm giving away a DGX Spock. Nvidia just hiked the price $700. You probably heard about these memory shortages, right? So yeah, it's now $4700, which is very very expensive. And Merve from Hugging Face, by the way, she got one for her birthday and she said she literally cried. So it's a really cool bit of kit. If you register through my link in the description and you attend at least one session, then you are in the draw. This is a massive conference. Physical AI and robotics are going to be the breakout theme and Jensen does the keynote Monday at 11:00 a.m. Pacific. The link is in the description. Don't miss it.

嗯，是的，我认为 Jeremy，以及 Alpha Evolve 论文，还有我们在 DeepMind Go 机器上做的工作，都表明这种“垫脚石”积累加上迭代验证，以及从现实世界或合成评估器中收集信息和证据，对此非常重要。

And um yeah, I think Jeremy, as well as the Alpha Evolve paper, as well as sort of work we've done on like the Deep Mind Go machine, for example, shows that this um sort of stepping stone accumulation plus iterative verification and um collecting sort of information and evidence from the real world real synthetic evaluator is really important for that.

非常酷。而“垫脚石”收集，这来自 Kenneth Stanley。那篇精彩的论文《为什么伟大不能被计划》提到，他主张系统最好不要收敛。在自然进化中，我们只是尝试各种不同的事物，而伟大往往遵循一条多样化的路径，这意味着你必须做一些起初看起来很愚蠢，但后来却变得极其有用的事情。

Very cool. And stepping stone collection, so this is came from Kenneth Stanley. It's a wonderful paper, Why Greatness Cannot Be Planned. And he said that it's it's better to have systems that don't converge. So, in natural evolution, we are just trying all of these different things, and greatness quite often follows a diverse path, which means you have to do things which initially seem quite stupid, and then later on they turn out to be incredibly useful.

是的。

Yeah.

我们试图设计算法，允许存在一群有点奇怪的东西，然后我们锁定并稍微收敛。但我们仍然在收敛。我们仍在构建不会永远发散的系统。我们失去了什么？

We're trying to design algorithms that can kind of allow for a population of slightly weird things, and and then we kind of lock in and and converge a little bit. So, we we're still converging though. So, we're still building systems that don't diverge forever. What are we losing?

在完成 Shinka Evolve 之后，我发现一个极其重要的问题，就是“问题”本身的问题。到目前为止，所有这些系统（也许除了 AI Scientist，我们也可以谈谈它）都是给定问题的。你有一个评估器，一个正确性检查器，你只针对那个单一问题采样程序。但通常，针对特定问题的创新可能需要先发明一个不同的问题。例如，我认为在 Alpha Evolve 展示的矩阵乘法结果中，你可以递归地将算法应用于更大的矩阵。这实际上是一个重要的结果。但自动提出这种归约或问题求解的递归性质，这些系统目前并不一定内在地具备。因此，我认为未来不仅要做开放式的解决方案优化，还要做问题与解决方案的共同进化，以收集更多样化的垫脚石，真正启动这个开放式过程。因为对我来说，一个重要的生活目标或成就是希望看到一个过程不仅能运行一周或几周，甚至可能运行数年，收集更多样化、有趣的垫脚石。

One one thing I find extremely important after having done the Shinka Evolve is um sort of this problem problem, right? So, um with all of these systems so far, maybe except for the AI scientist, which we can also talk about, the problem is given, right? So, you have an evaluator, you have a correctness checker, and you sample programs only on that single problem, right? But, oftentimes innovation for a specific problem might require first inventing a different problem, right? So, for example, I think in the matrix multiplication result that the Alpha Evolve people show, you can recursively apply sort of the algorithm to larger matrices. So, it's actually an important result, right? But, sort of automatically coming up with this reduction or like this, let's say, recursive nature of problem solving is something these systems right now not necessarily have built in intrinsically, right? So, I think going forward it's going to be really important to not only sort of do open-ended, let's say, optimization of solutions, but sort of do the co-evolution of problem and solution together in order to collect even more diverse stepping stones, and um to really kick off this this open-ended process. Cuz also to me, like one of the the big life goals or achievements I would want to see is really having a process that can run not only for, let's say, a week or many weeks, but like for years even potentially, right? Collecting even more diverse, interesting stepping stones.

是的，我和 Joel Lehman 聊过，他提到了奈特不确定性，即机器学习算法非常擅长处理“未知的未知”。从某种意义上说，“未知的未知”指的就是那些以后可能有用的垫脚石。当我们目前运行这些算法时，LLM 和推理系统也是如此：当我们给它们一个具体任务时，它们表现得非常好。而你指出的是，我们可能需要发明新的、不相关的问题，并找到解决方案，这些方案可能与我们试图做的事情相关。这感觉有点像第 22 条军规，对吧？我们说的是，比如圆形填充。这是我的评估函数，我希望你多样化，然后收敛到解决方案。顺便说一句，我对 Genie 也有同样的想法：它只给你你要求的东西。你输入一个提示，比如瑞士湖，水上有船，山在岸边，然后我想，鸟在哪里？哦，我忘了在提示里加鸟。那么，我们如何有意义地构建系统，真正引入其他可能有用的未知事物呢？

Yeah, I spoke to Joel Lehman and he was talking about the knightian uncertainty, which is that machine learning algorithms are very good with unknown unknowns. And and in a sense, the unknown unknown is talking about these these stepping stones that might be useful later. And when we run these algorithms at the moment, it's the same with LLMs and reasoning systems, is that they're very, very good when we give them a specific thing. And what you're pointing to is we might need to invent new unrelated problems and find the solutions which might then be related to what we're trying to do. So, that feels like a bit of a catch-22 situation, right? So, we're saying, you know, circle packing. Here's my evaluation function and I want you to sort of diversify and then you know kind of and then converge towards the solution. It's just I had the same thought with Genie by the way that it it gives you exactly what what you ask for. So you put put a prompt in and you know like a Swiss lake with you know with boats on the water and mountains on the side and I was thinking where are the birds? Oh, I forgot to put birds in the prompt. Right, so how can we meaningfully build systems that actually kind of bring in other unknown things that might be useful?

我认为一个灵感或我个人想研究的东西，是像 Jeff Clune 等人提出的 PowerPlay 或 Poet 那样的系统。其中有一组任务和一个解决方案生成器，两者以类似自动课程学习的方式共同进化。Poet 最初的自然应用是强化学习，但我认为现在可以将其扩展到更广泛的科学领域，至少在有模拟器可用于运行这些评估时。通过这种共同进化，你总是试图最大化生成器的能力，同时增加这个凸包或潜在更多样化的问题。

I think one inspiration or thing I would personally want to sort of research are systems like outlined in in PowerPlay or Poet by by Jeff Clune and others. So where there is essentially like a a set of tasks and a solution generator and both of them sort of co-evolve in this almost like auto curriculum play like style, right? And I think sort of the in Poet the natural first application was sort of reinforcement learning, but I think this can now be broadened up to to science more generally, right? At least when there's a simulator available to for for running these evaluations. And by doing such a co-evolution, you always try to to max out the capabilities of that generator while sort of uh increasing this this convex hull or potentially even more diverse problems while doing so.

我知道总有一种主导思想，即使对于 Poet，它有一个种群，很多环境和智能体，环境是复杂化的。所以智能体有一个有效的课程，以递增的复杂度学习事物。但即便如此，系统中是否存在一种设计偏差，即某处有代码逐步复杂化环境，而这不也是由人类设计的吗？所以它也只给你你要求的东西。

I know that there's always the leading thought that even with Poet which was this thing where you had like a population you had like a lot of environments and agents and the environments were in complexified. So the agents would have a kind of effective curriculum to to learn things in increasing complexity. But even then isn't there a kind of design bias in the system where there's some code somewhere which complexifies the environment step by step and wouldn't that also just be designed by the human? So, it would also just give you exactly what you ask for.

最终，这归结为语言模型可能进行外推或内插的假设。即使这些东西最终由人类设计，但存在许多未知的未知，我们在设计时没有考虑到。因此，LLM 有可能仅仅因为我们之前没有想过，就做出新颖的发现。

Ultimately, this comes down to like the hypothesis that language models can potentially do extrapolation or interpolation, right? In the sense that even though these things might be in the end designed by humans, there are many unknown unknowns, right? That we humans didn't think of while designing them, right? So, potentially, it is possible for an LLM to Yeah, find a novel discovery simply by us not having thought about it before, right?

当我们自主运行 LLM 时，它们往往不会发生什么有趣的事情。根据你给它们的提示，它们会朝那个方向走几步，然后就没有新的有趣新奇事物出现了。我认为即使你以智能体方式将它们与环境反馈连接起来，它们似乎仍然非常依赖于初始条件。有了 LLM，我们能构建一个真正适应新奇事物、能够发现新事物的系统吗？

When we run LLMs autonomously, they they tend to just kind of like nothing interesting happens. So, depending on the prompt you give them, they'll kind of go a few steps in that direction, and then no new interesting novelty emerges. And I think even if you wire them agentically with environmental feedback, they they still seem quite parasitic on their starting conditions. With an LLM, could we build a system which actually adapted to novelty? That could actually discover new things.

我认为这很大程度上取决于你给 LLM 什么作为起点。例如，在 Shinka Evolve 中，我们经常看到，如果你给一个已经针对当前问题优化得很好的初始解决方案程序，你仍然会陷入局部最优，不会引入太多新奇性。而如果你从一个贫乏的解决方案开始，就有更多的多样性空间。我认为这又回到了我之前的研究，即元学习。

I think it really kind of also depends on what do you give the LLM as a starting point, right? So, for example, in Shinka Evolve, we from time on time saw that if you give an initial solution program, which is already pretty optimized on the problem at hand, you still kind of get stuck in in local optima, right? Where not a lot of novelty is introduced, right? While if you start off from like an impoverished solution, there's much more room for diversity. And I think this is sort of coming back to um sort of what I did before in my research, namely meta-learning.

这是一种经典的权衡：要么从一个非常不受约束的起点出发，比如一个非常简单的解决方案，为优化留出更多空间，但这可能需要开放性和很长时间才能找到好的解决方案；要么从一个已经受到归纳偏置强烈约束的起点出发，这样在收敛方面可能更高效，但你就得不到那种开放式的创新收益。

It's this classical trade-off: you can either start from something very unconstrained, like a very simple solution, giving much more room for optimization, but this might require open-endedness and a long time to find a good solution. Or you start from something already very constrained by inductive biases, which might be much more efficient in terms of convergence, but you don't get the open-ended novelty benefit.

是的，我想我们的目标是构建不是由人类设计的系统。例如，如果我利用自己的深刻理解，LLM 在你深入理解某件事时确实很擅长。同样，我们可以启动 Archinka Evolve，并放入一个利用我理解的初始解决方案。我们希望 AI 系统能让任何人都能使用。所以一个非专家可以说，“我想解决这个问题”，然后它就会解决。我们应该谈谈进化方法。为了保持多样性，你有一个程序种群，它们被分成岛屿。给我讲讲这个。

Yes, now I suppose where we want to get to is building systems which are not designed by humans. For example, if I'm leveraging my deep understanding, LLMs are really good if you understand something deeply. Similarly, we could kick off Archinka Evolve and put a starting solution in there which leverages my understanding. We want AI systems that anyone could use. So a non-expert could say, 'I want to solve this problem,' and it will solve it. We should talk about the evolutionary approach. To maintain diversity, you had a population of programs separated into islands. Tell me about that.

Archinka Evolve 的工作方式类似于 Alpha Evolve：你维护一个程序档案（类似数据库），然后用一组灵感程序采样父程序。接着让 LLM 对该程序进行改进，提供代码来编辑或重写整个程序，甚至可能交叉不同的程序。然后你查询 LLM，得到一个程序，并在当前问题上评估它。例如，增加正方形内一组圆的半径之和。你运行这个过程，每次从评估器收集证据，添加到数据库，然后重复。你不是顺序执行，而是并行处理许多不同的程序。每次添加一个程序时，你都会尝试将该程序收集的知识扩散到整个数据库。一种理解方式是，你有一棵树，每个节点代表一个程序，你根据父节点分支。有趣的是，这些方法确实倾向于 Scaling，但理想情况下我们可以让 Scaling 以更快的速度发生。我们在 Shinka Evolve 中尝试通过一系列创新来实现这一点，包括模型集成，即我们不仅使用 Gemini，而是使用所有前沿模型提供商，并找出一种智能方式来为给定的父程序使用每个模型。对于某个程序，在某些情况下使用 GPT 模型可能更好，在其他情况下使用 Gemini 模型。我们引入了一种自适应优先级方案，可以在运行算法时动态调整进化算法。这也回到了命名上：Shinka Evolve 意味着我们使用 LLM 应用的这种进化算法在优化程序的同时也在共同进化。

The way Archinka Evolve, similar to Alpha Evolve, works is you keep an archive, like a database, of programs, and then you sample parent programs with a set of inspiration programs. Then you ask an LLM to make an improvement to that program, to provide code to edit or rewrite an entire program, or potentially even cross over different programs. Then you query the LLM, get a program out, and evaluate it on the problem at hand. For example, increasing the sum of the radii of a bunch of circles in a square. You run this, each time collecting evidence from the evaluator, adding it to the database, and repeating this process. You don't do this sequentially, but in parallel for many different programs. Each time a program is added, you try to diffuse the knowledge collected by that program across the entire database. One way to think about this is you have a tree where each node represents a program, and you branch off based on the parent nodes. Interestingly, these approaches do tend to scale, but ideally we can make the scaling happen at a faster rate. This is something we tried in Shinka Evolve by doing a bunch of innovations, including model ensembling, so we're not using just Gemini but all frontier model providers, figuring out a smart way to use each model for a given parent. If you have a certain program, in some situations it might be better to use a GPT model, in others a Gemini model. We introduced an adaptive prioritization scheme that can adapt the evolutionary algorithm on the fly while running the algorithm. This comes back to the naming: Shinka Evolve means that this evolutionary algorithm we apply using LLMs also co-evolves at the same time while we optimize the programs.

关于圆填充问题，你有一张图显示了它如何快速收敛。我们会把图放在屏幕上。性能迅速上升，然后缓慢收敛。你在论文中说它使用了三个核心创新。我在想：如果你运行这个 50 次，每次结果会一样吗？它在多大程度上是跳出框框思考？Sebastian Boubeck 总是在 Twitter 上发帖说 GPT-5 发现了新东西，但总有一个问题：它只是在搜索互联网，找到以前发现过的东西，然后用新方式组合它们吗？它真的能跳出框框思考吗？

On the circle packing problem, you had a plot showing how it converged quickly. We'll show it on screen. Performance jumped up and then slowly converged. You said in the paper it used three core innovations. My thinking was: if you ran this 50 times, would it be the same every time? And to what extent is it thinking outside the box? Sebastian Boubeck always posts on Twitter about how GPT-5 has discovered new things, but there's always the question: is it just searching the internet, finding things that have been found before, combining them in a new way? Could it really think outside the box?

是的，这几乎是一个主观问题。首先，我不知道互联网上所有尝试圆填充的问题。但我在我们描绘的树中看到的是，例如，两个程序之间发生交叉操作，不同的概念被组合在一起。一个重要部分是圆的初始化。另一个是优化，基本上是一个约束优化程序。最后一部分是再加热阶段，添加噪声并挤出更多策略。对我来说，这种通过树传播信息的方式非常迷人，其中踏脚石以互补的方式被使用。关于多次重新运行程序，存在一些随机性。我们使用语言模型，由于它们服务器端的排队设备调度，我们无法消除所有噪声。我们看到，对于解决方案的总体质量，是可以重新获得的，但有时是通过不同的程序，这主要是由于随机性。对于许多问题，并不只有一个解决方案能达到那个分数，而是程序空间中的一个谱或区域，它们相似。关于圆填充问题的一个有趣之处，回到我最初讨论的问题问题，是原本我们使用了一个公式，其中正确性检查有非常小的容差。圆可以稍微重叠，然后我们减小半径，解决方案就是精确的。这并没有改变分数太多，所以它仍然是 state-of-the-art，但本质上是一个代理问题。然后我们在精确设置上重新运行了 Shinka Evolve，发现获得相同质量需要更长的时间。这指向了我一开始讨论的内容：有时代理问题对于发现可能非常有价值，而拥有一种自动化的方式来高效设计这些代理问题，未来可能非常重要。

Yeah, this is almost a subjective question. First off, I don't know all problems on the internet that try circle packing. But what I can see in the tree we depict is, for example, a crossover operation between two programs happening where different concepts are combined. One important part is the initialization of the circles. Another is the optimization, basically a constraint optimization program. The final part is a reheating stage where noise is added and more strategy is squeezed out. To me, this propagation of information through the tree is really fascinating, where stepping stones are used in a complementary fashion. Regarding rerunning the program multiple times, there is some stochasticity. We're using language models, and due to queuing device scheduling on their server side, we can't get rid of all the noise. We've seen that for the general quality of the solution, it is possible to reobtain it, but sometimes with a different program, mostly due to stochasticity. For many problems, there isn't one solution that achieves that score, but a spectrum or region in program space that resembles the same. One interesting thing about the circle packing problem, coming back to the problem problem I discussed initially, was that originally we used a formulation where correctness is checked with a very tiny amount of slack. The circles could overlap a tiny bit, and then we reduced the radii and the solution was exact. This didn't change the score by too much, so it was still state-of-the-art, but it was essentially a proxy problem. We then re-ran Shinka Evolve on the exact setting and found it took a little longer to obtain the same quality. This points to what I discussed in the beginning: sometimes surrogate problems might be extremely valuable for making discoveries, and having an automated way to design these surrogate problems efficiently might be really important going forward.

是的，这绝对令人着迷。这让我想起了支持向量机，我们通过引入松弛变量使优化变得可处理，你可以将其视为一种代理问题。但我在想，Shinkai Evolve 或 Alpha Evolve 会知道引入代理问题吗？因为作为理解的设计者，我们可以跳出框框思考，做这样的事情。因为如果适应度函数有禁止圆相交的约束，那么算法可能不会想到提出代理问题。

Yeah, that's absolutely fascinating. It reminds me of support vector machines where we make the optimization tractable by introducing slack variables and you can think of that as a kind of surrogate problem. But then I'm thinking, well, would Shinkai evolve or Alpha evolve, would it know to introduce a surrogate problem? Because you know, as designers who understand, we can think outside the box and we can do stuff like that. Because presumably if the fitness function had the constraint that there were no circle intersections, then it wouldn't occur to the algorithm to come up with the surrogate problem.

完全正确。是的，这是目前的一个大限制，对吧？目前，我们认为问题是固定的，并针对该问题进行优化。但想想人类，我们非常擅长发明自己的问题，或者重新表述问题以便实际处理它。我认为数学中的许多创新都来自于对问题采取非常不同的视角，比如将数论应用于线性代数，反之亦然。我认为目前这些系统还没有达到这种迁移水平。

Exactly. Yeah, this is a big limitation right now, right? So, at this current point in time, we take the problem to be fixed and we optimize for that problem. But when you think about humans, we're really really good at sort of inventing our own problems, right? Or reformulating the problem so that we can actually sort of work with it, right? So, I think a lot of sort of the innovations in let's say mathematics come from taking a very different perspective on a problem, right? So, taking sort of number theory and applying it to linear algebra or the other way around and I think right now these systems are not yet at the point of achieving such level of, let's say, transfer.

是的，这让我想起我和 Leon 聊过这个。你们有那个 Sudoku Bench。很多人看 Cracking the Cryptic 的 YouTube 频道，他们正是这样做的。他们基于捕捉问题本质或方面的抽象来发明新问题。然后他们做一些类似于 Shinkai Evolve 的事情。他们进行这种进化，将不同的解决方案结合起来，融合两者的最佳方面，并开辟一条通向新解决方案的分叉路径。

Yes, and it reminded me I spoke to Leon about this. You've got this Sudoku bench. And a lot of folks watch Cracking the Cryptic YouTube channel, and that's exactly what they do. They invent new problems based on abstractions that capture the essence or aspects of the problem you're solving. And then they do something which is similar to Shinkai evolves. They do this kind of evolution where they take these different solutions and they kind of combine the best aspects of both of them, and they forge a divergent path to a new solution.

是的。

Yeah.

这似乎正是我们需要做的本质。

And that seems to be the essence of what we need to do.

是的，当然。Jeff Clune、Shangwen Yu 和 Song Lu 也有一些关于自动能力发现的工作。他们研究生成任务的语言模型，但这是以一种非结构化的方式进行的，并不是为了促成某个目标问题的解决方案。我认为建立这些联系在未来会非常富有成效。

Yeah, for sure. I mean, there is some work also by Jeff Clune, Shangwen Yu, and Song Lu on automatic automated capability discovery. So, there they look at language models that generate tasks, right? But it's in a let's say unstructured way in the sense that it's not done in order to enable the solution to one target problem, right? And I think sort of doing these connections is going to be very fruitful down the line.

非常酷。另一件事，我们会在屏幕上展示进化图。对于圆填充问题，我看了它，首先它看起来非常简洁，这很好。它似乎很快找到了一个最优路径。我在想，也许有一些自然模式，我们可以抽象地利用它来指导未来的进化。但我在想的另一件事是，目前机器学习的问题是我们没有真正内嵌语义。所以我们有一个验证器，观察奖励，进行模式化探索，朝着目标前进。我喜欢机械形式的推理，我们实际上知道程序组件的含义。这之所以重要，是因为当我们合并来自两个不同岛屿的最佳程序时，这是一种一阶交互。合并它们可能没有意义。LLM 很棒，你可以给它们任何程序对，它们会找到合并的方法。但更原则性的方式难道不是存在某种语义原语，我们知道它们如何组合吗？所以有一个乐高类比，我们基于原则构建，而不是基于性能开辟路径。

Very cool. Now, the other thing, we'll show the graph on the screen, the evolutionary graph. So, for the circle packing problem, I was looking at that and first of all, it looked incredibly parsimonious, which is good. It looked like it had found an optimal path to the solution very quickly. And I was thinking in my mind, well, maybe there's some natural pattern that there's something about that that we could use in the abstract to guide the evolution in the future. But the other thing I'm thinking about is right now, the problem with machine learning is that we don't really have semantics baked in. So, what we're doing is we have a verifier, we're looking at the reward, and we're sort of like doing patterned exploration, and we're taking steps towards the target. And I love mechanistic forms of reasoning where we actually know something about what the program components mean. And the reason this is important is when we're merging together the best performing programs from two different islands, that's a kind of first-order interaction. And it may not make sense to merge them together. It's wonderful that LLMs you can give them any pairs of programs and it will find a way to merge them together. But wouldn't a more principled way be that there's some kind of semantic primitives here and we know they fit together. So there's this Lego analogy that we're kind of building up based on principles rather than forging a path based on the performance.

是的。嗯，这是个好观点。我们在 Shinkai Evolve 中做的一件事是保留一个草稿本。每个程序都被总结，然后从程序总结中我们保留一组全局见解，这些见解是从这些程序中提取或共享的。然后基于草稿本，我们构建元建议，这些建议成为系统提示的一部分。这样你可以尝试在语义上把握一些发现，但一个普遍的问题（同样依赖于任务）是，你因此将知识扩散到整个树中。但有时你希望事物更加隔离。这总是一种权衡，你必须为你的问题找到知识扩散程度和程序硬岛之间的正确位置。我们正尝试朝自动以最优方式调整的方向迈进，但这非常依赖于程序。另一个你已经涉及的点是 Jeremy 对 Arc AGI 的解决方案，即在指令空间中进行解决方案进化，而不是程序空间。我认为这很重要，正如我们通过构建元草稿本试图同时做这两件事。再次强调，这取决于问题。我玩过 Arc AGI 1 和 Arc AGI 2。我认为在 AGI 1 上，程序方向实际上相当有效，就像 Jeremy 说的，它是确定性的，更容易在进化过程中获得清晰的改进信号。而在 AGI 2 上，语义进化似乎更高效。所以理想情况下，我们可以得到一个系统，能够自动决定在可行且易于引导的情况下采用程序化方法，还是采用进化指令或 LLM 驱动的输入输出映射的语义方法。

Yeah. Um that's a good point. So um one thing we do in Shinkai evolve as well as we keep essentially a scratchpad. So each program is being summarized and then from the program summaries we keep sort of a set of global insights, let's say, that were shared or like extracted from these programs. And then based off of the scratchpad we construct um sort of meta recommendations that then become part of the system prompt, right? So um that way you can try to sort of semantically grasp some of the discoveries, but a general problem which is again sort of task dependent is um thereby you sort of diffuse that knowledge across the tree, right? But sometimes you want things to be much more isolated, right? It's always like um a trade-off where you somehow have to find for your problem the right uh position on the spectrum of how much knowledge diffusion do you want to have and how much sort of uh let's say hard islands of programs do you want to have, right? And um yeah, we're trying to make steps in the direction of sort of automatically adjusting this in an optimal way, but again it's very program sensitive. And then sort of I think another point where you're already sort of going into um is sort of Jeremy Jeremy's solution to Arc AGI, right? And sort of doing um solution evolution in the instruction space, right? Instead of the program space. I do think that this is uh something important and we're like I said with like the construction of this meta scratchpad trying to do sort of both at the same time. Uh again it's problem dependent. Like I played around a little bit with Arc AGI 1 AGI 1 and Arc AGI 2. And I think on AGI 1 actually the transform sort of program direction is actually quite effective, right? It's like Jeremy said, it's deterministic and it's easier to sort of get clear signal to improve on during your evolution process. While on others like AGI 2, like this whole sort of semantic evolution seems to be more efficient. So, I think ideally we we can get a system that can automatically in some sense decide whether or not it wants to take like a programmatic approach in settings where it's actually feasible and easier to to bootstrap off or it takes the semantic approach of evolving instructions or like LLM driven input output mappings.

是的，这很有趣。因为符号 AI 的人会说，我不喜欢连接主义，因为它不理解，连接主义中唯一的语义是相似性概念，它并不真正理解事物。所以他们会说，从实体关系图开始，然后通过组合和第一原理构建。但这行不通。所以我们使用神经网络，因为它们非常灵活，理解世界的许多事物，但它们没有我们想要的那种约束。所以我们使用这些技巧。Jeremy 进化了程序描述。

Yeah, it's so interesting because you know, like a symbolic AI person would say, oh, I don't like connectionism because it doesn't understand you know, the only semantics in connectionism is this notion of similarity. It doesn't really understand things. So, they would say, well, just start with an entity relationship graph and then just kind of build up using composition and first principles. That doesn't work, right? So, we're using neural networks because they're incredibly flexible and they understand a lot of things about the world, but they don't have the kind of constraints that we want. So, what we do is we use these tricks. So, Jeremy evolved program descriptions.

在你的程序选择中，你使用了基于嵌入的相似性进行语义新颖性检测。你有一个基于余弦的自我相似性矩阵。你还有这个元草稿本。所以，我们看到了一系列迷人的可能性，仍然使用神经网络，你可以通过所有这些不同的技巧注入语义，但它们都有权衡。

On your program selection, you had a semantic novelty detection, using embedding-based similarity. You had a kind of self-similarity matrix based on cosines. And you've got this meta scratchpad. So, we're seeing this fascinating spectrum of possibilities where still using neural networks, you can imbue semantics using all of these different tricks, but they all come with trade-offs.

是的，当然。我觉得有趣的是，计算机科学经历了很长一段由人类设计算法的时期。然后我们有了 Andrej Karpathy 的软件 2.0 范式，使用神经网络执行特定功能。现在我们到了使用 LLM 更普遍地设计算法或解决方案的阶段。尽管大型前沿语言模型是极端的黑箱，但输出——程序、指令——可以是透明的。所以，这为做研究或基本上任何事情开辟了一个新范式。但我们才刚刚开始摸索合适的用户界面。

Yeah, for sure. I think it's interesting we've had a long period of computer science where algorithms were designed by humans. Then we had Andrej Karpathy's Software 2.0 paradigm where we used neural networks that performed a certain function. And now we're at this point where we're using LLMs to design algorithms or solutions more generally. Even though large frontier language models are extreme black boxes, the outputs—programs, instructions—can be. So, it opens up a new paradigm for doing research or basically anything. But we're just at the starting point of figuring out the right user interface for that.

论文中的另一项创新是使用 UCB（上置信界），来自多臂老虎机文献。这是一个拉杆问题，最初不知道拉哪个，随着时间的推移减少不确定性，拉那些有效的，存在探索-利用困境。你实现了这一点，用于决定使用哪个 LLM——Gemini、Grok-1 等。

The other innovation in the paper was using UCB, upper confidence bound, from the multi-armed bandit literature. It's this problem where you pull levers, initially not knowing which to pull, and over time you reduce uncertainty and pull the ones that work, with an exploration-exploitation dilemma. You've implemented that for figuring out which LLM—Gemini, Grok-1, etc.—to use.

我们使用模型集成来提出程序变异。直觉上，SWE-bench 上最好的前沿模型总是最好的变异提议模型，但实际情况并非总是如此。在这个进化环境中，很难将明确的功劳分配给单个模型。例如，一个改进由 GPT-5 实现，下一个由 Sonnet 4.5 实现，不清楚第二个突变的性能提升是来自 GPT-5 收集了第一个垫脚石，还是来自 Sonnet 4.5。因此，我们没有均匀采样模型，而是实现了一种基于老虎机的方法，每个模型是一个臂。我们观察这个模型通过创建变异改进父节点的频率。我们调整后验概率，首先探索所有臂一次，然后随着时间的推移偏好那些之前对类似节点产生过改进的模型。

We're using a model ensemble to propose program mutations. Intuitively, the best frontier model on SWE-bench is always the best mutation proposal model, but that's not always the case. In this evolutionary setting, it's extremely hard to assign clear credit to a single model. For example, one improvement is implemented by GPT-5, the next by Sonnet 4.5, and it's unclear if the performance gain from the second mutation originated from GPT-5 collecting the first stepping stone or from Sonnet 4.5. So instead of uniformly sampling models, we implement a bandit-based approach where each model is one arm. We look at how often this model improved performance of a parent node by creating a mutation. We adjust the posterior probability to first explore all arms once, then over time prefer models that yielded improvements before for similar nodes.

UCB 的优点是它具有理论遗憾，仅比最优切换路径差对数级别。但如果我理解正确，UCB 基于全局评分，每个 LLM 的平均分。我们想要的是上下文相关的切换决策，知道对于这个特定程序 Gemini 更好。它是否会收敛到单个前沿模型，在细微的情况下我们可能仍然得到错误的模型？

The great thing about UCB is it has a theoretical regret, only log worse than the optimal switching path. But if I understand correctly, UCB is based on a global rating, a mean score of every LLM. What we want is a contextual switching decision, knowing for this particular program Gemini is better. Does it converge to a single frontier model, and in nuanced situations might we still get the wrong model?

一般来说，所有模型都分配了一定的概率，所以它不会只选一个模型而停止使用其他模型。仍然有机会实现开放式的偶然发现。对于我们考虑的问题，我们没有看到任何一个模型明显主导所有其他模型。这确实取决于进化过程的进程，哪个模型更好，而 UCB 或老虎机方法有效地动态调整这一点。

In general, there is some probability allocated to all models, so it's not like it can pick one model and stop using others. There's still a chance for open-endedness serendipity. For the problems we considered, we haven't seen one model clearly dominate all others. It really depends on the course of the evolutionary process which model is better, and UCB or the bandit approach dynamically adjusts this efficiently.

未来是否有可能使用 LLM 来做这个判断？

Would it be possible in the future to use an LLM to make this judgment?

有可能。在这种情况下，你将 LLM 视为代理模型。已有工作表明语言模型可以充当代理模型。真正的问题是如何向 LLM 表示信息。如果你使用原始程序及其适应度评估，你会很快耗尽上下文。所以你需要压缩，以便以正确的方式呈现信息来进行模型优先级排序。

Potentially. In that case, you think of the LLM as a surrogate model. There has been work showing language models can act as surrogate models. The real question is how to represent the information to the LLM. If you use raw programs and their fitness evaluations, you quickly run out of context. So you need compression to present the information in the right way for prioritization of models.

我之前没有意识到上下文有多长。我们能否使用一个 80 亿参数的 Llama 模型进行主动微调？我们在 Grok 上运行一个程序，得到一个分数，随着时间的推移，它会知道 Grok 擅长这些问题。

I hadn't appreciated how long the context is. Could we use an 8 billion Llama model with active fine-tuning? We run a program on Grok, get a score, and over time it would know Grok is good at these problems.

有可能。我不确定如果我们只评估 150 个程序，微调效率如何，但原则上可以想象。这更多是工程方面的问题，不一定是最优雅的做法。它确实可能发生。但对于所有这些事情，我们从最直观的算法组件开始，而 UCB 在这里确实发挥了作用。非常感谢 Eduardo Santana 向我们介绍了 Shinka。

Potentially. I'm not sure how efficient fine-tuning is if we're only evaluating 150 programs, but in principle one could imagine. It's more on the engineering side, not necessarily the prettiest to do. It could happen. But for all these things, we started with the most intuitive algorithmic component, and UCB really did the job here. Much credit to Eduardo Santana who introduced us to Shinka.

让我们谈谈差异和变异。你生成程序，有点受 Alpha Evolve 启发。他们有门控，你可以门控代码中可变的部分。给我讲讲这个。

Let's talk about the diffs and mutations. You generate programs, inspired a bit by Alpha Evolve. They had gating where you gate part of the code which is mutable. Tell me about that.

程序是一个长字符串。为了确保某些对评估至关重要的部分（如导入）不会被 LLM 变异删除，我们使用标记来声明哪些部分是可变异和可进化的。当我们得到差异提议时，很容易通过编程使它们不可变，因此只有代码片段的其余部分被更改。

A program is a long string. To ensure certain parts essential for evaluation, like imports, are not deleted by LLM mutations, we use markers that state which parts are mutable and evolvable. It's easy to programmatically make them immutable when we get a diff proposal, so only the rest of the code snippet is changed.

我们实现了一种带反射的拒绝采样方法：如果 LLM 偶然试图变异某个部分，该提议会被拒绝，然后重新采样一个新提议。这样可以在一定程度上缓解某些安全或安保问题，并获得稳健的变异。一个更大的问题是如何将这种单文件变异设置扩展到多文件变异，即处理整个代码库。原则上，你可以将多个代码库表示在一个文件中，但层次结构可能实际上很有用。Aider 这个编码工具有一些想法，比如构建仓库地图并具有一定的抽象层次，但它们也有正反两方面的权衡。

We sort of implement a type rejection sampling with reflection approach where if an LLM by chance for example tries to mutate this part, it's going to be rejected and you resample a new proposal. And thereby you can somewhat mitigate certain security or safety problems and get a robust sort of mutation. One of the bigger questions is how can you turn this from a single file mutation setup to a multi-file mutation setup, so working on entire codebases. In principle, you can represent many codebases in a single file, but the hierarchical structure might be actually useful. And there are some ideas from Aider, this coding tool, where you construct a repository map and have some level of abstraction, but they also come with positive and negative trade-offs.

顺便说一句，我很喜欢 Aider。感觉未来代码生成系统实际上会像 Shinkurou 一样。想想看，它会使用某种 Git 仓库。也许 Cursor 已经做到了，因为在 Cursor 中你可以恢复之前的检查点。但它可以探索不同的分支，并将检查点合并在一起。显然，你只需用自然语言说出你想做什么。但我们还没讨论变异。刚才我们只说了差异。还有一种选择是完整文件重写。但还有交叉的概念。那么这是如何工作的呢？

I love Aider by the way. It feels that in the future, code generation systems will actually resemble Shinkurou of all. And if you think about it, it'll be using some kind of Git repo. Maybe Cursor already does this because in Cursor you can restore previous checkpoints. But it can be exploring different branches and merging checkpoints together. Obviously you just say in natural language what you want to do. But we didn't talk about mutation by the way. So we just spoke about diffs. And there's also an option to do a full file rewrite. But there's also this notion of crossover. So how does that work?

在 Alpha Evolve 之上的一项小创新——我认为他们只使用了基于变异的操作——是我们希望有更大的灵活性来完全重写程序，以产生一个完全不同的垫脚石。同样，你可以让部分代码可变，但与其提议一个补丁来更改某些部分，我们实际上重写了整个程序。这有时是有帮助的，但并不总是有明显的好处，但它允许你在搜索中获得更多多样性。所以这是除了基于差异补丁的方法之外的另一种变异类型。另一种是交叉变异，我们不仅采样一个父程序，而是两个不同的程序，并要求系统进行互补改进。同样，在某些问题上这非常有用，而在其他问题上则不然。但总的来说，我们发现操作符的多样性也有助于发现新事物。

A small innovation on top of Alpha Evolve where I believe they only use this based mutations is that here we wanted to have more flexibility to entirely rewrite the program, to come up with a completely different stepping stone if you will. So again, you can make part of the code mutable, but instead of proposing a patch to change certain parts, we essentially rewrite the entire program. And this sometimes is helpful. It's not always a clear benefit, but it allows you to get more diversity into the search. So this is one type of mutation next to this diff patch based approach. And the other one is a crossover mutation where we sample not only a single parent program, but two different ones and we ask the system to make a complementary improvement. Again, on some problems this is really helpful and on others it's not. But in general we found that having a diversity in terms of operators is also helpful in discovering new things.

我想跟进你之前提到的观点，即这是一种新范式。我也这么认为。我深信不疑。我认为现在我们正处于起步阶段，我们仍然主要将聊天助手界面视为与 LLM 交互的方式，但大多数时候它本质上是单线程的。我们坐在电脑前，与聊天交互，看到编辑器中发生的变化，接受它们等等。但我认为这也只是迈向更分布式的研究、优化等思考方式的垫脚石。所以我喜欢想象氛围编码、氛围聊天。另一方面，我们有氛围优化和氛围研究，我理想的未来场景是：作为一名研究人员，白天你与 Shinka 或 AI 科学家这样的系统一起工作，在某种意义上像牧羊人一样掌舵；到了晚上，你按下播放键然后上床睡觉，后台有多个实验在运行，LLM 自动提出新实验，积累证据；早上你回来时，有一个多线程系统在并行运行，你更像是船的牧羊人，而不是实际执行实验和分析的人。哦，是的，你仍然在分析，但你没有在执行。这一切都由系统本身完成。

I wanted to follow up on the point you made before about this being a new paradigm. I think so too. I'm really convinced. I think right now we're at the beginning where we still think a lot about this chat assistant interface as the way how we interact with LLMs, but it's most of the times inherently single threaded. So we're sitting in front of the computer, interacting with the chat, seeing changes as they occur in the editor, accepting them and so on. But I think this is also just a stepping stone towards a more distributed way of thinking about research, optimization and so on. So I like to think of vibe coding, vibe chatting. And on the other hand we have vibe optimization and vibe researching where my ideal future scenario is one in which you as a researcher during the day work with a system like Shinka or the AI scientist, you steer the ship like a shepherd in some sense, and then during the night you press play and go to bed, and in the background you have multiple experiments running and automatically new ones being proposed by LLMs, evidence being accumulated, and then in the morning you come back and you have a multi-threaded system running in parallel and you're more like the shepherd of the ship than the person actually executing experiments and analyzing. Oh yeah, you're still analyzing but you're not executing. This is happening by the system itself.

是的，而且这可能会越来越半监督甚至主动。我的意思是，OpenAI 有一个新产品，它知道你对什么感兴趣，当你睡觉时它会去探索并找到你的脉搏。没错。我们现在的情况是，我们是相当有技术能力的人。所以 MATLAB 和 Mathematica 非常强大，但你需要知道如何精确地表达问题。而我可以想象一个未来，我们只用自然语言表达问题，或者仅仅基于我们与语言模型的交互，平台就知道我们感兴趣的是什么，它可以代表我们去寻找东西，因为这关乎将这项技术民主化，让那些可能并不确切知道自己要找什么的人也能使用。

Yes, and increasingly this might be semi-supervised or even proactive. I mean, there's that new product from OpenAI where it knows what you're interested in and while you sleep it's going off and find your pulse. That's right. And we're in the situation now where we're reasonably technical people. So MATLAB and Mathematica are supremely powerful but you need to know how to express problems precisely. Whereas I can imagine a future where we express problems just in natural language or maybe just based on our interactions with language models the platform knows what we're interested in and it can just go and find things on our behalf because this is about democratizing this technology to people who perhaps don't know exactly what they're looking for.

我认为其中一个更大的问题是验证方面。通常，生成大量解决方案比严格验证它们更容易。语言模型能够进行软验证，查看代码并隐式运行执行堆栈跟踪，但这并不精确。我认为这些奖励黑客的概念，以及不做真正发现而是走捷径，是我们需要投入更多时间和精力去解决的问题，以确保这确实朝着正确的方向前进。我希望语言模型在某个时候能够自己高效地做到这一点，无论是通过代码实现还是隐式进行。但这也是问题的一部分：不仅要提出解决方案，还要同时进行自动验证。

I think one of the bigger problems there is this verification aspect to it. In the sense that oftentimes it's easier to generate a lot of solutions than to actually hard verify them. Language models are capable of doing soft verification, looking at code and latently running a stack trace of execution, but it's not exact. And I think these notions of reward hacking and not doing real discoveries but shortcutting them is one where we need to put more time and effort into figuring out how to make sure that this actually moves in the right direction. And I would hope that language models at some point can do this efficiently themselves, either implementing in code or latently doing it. But this is also part of the problem: it's not only coming up with the problem, but also with the automatic verification at the same point.

是啊，这是一个诱人的想法：世界上存在自然模式，构建新颖解决方案的积木已经存在。也许它们存在是有原因的。也许它们只是反映了宇宙中的自然规律，因为总有一个问题：智能在于适应新奇。世界总是在变化。明天的世界会有我们今天无法解释的事物。但我们确实拥有可以轻松重组以解释未来的抽象知识。而 LLM 可能已经拥有了这些积木。

Yeah, isn't it a tantalizing idea that there are natural patterns in the world, and the building blocks to construct novel solutions are already there? And maybe they're there for a reason. Maybe they just reflect natural regularities in the universe because there's always this question of, intelligence is about adapting to novelty. So the world is always changing. And the world tomorrow will have things that we can't explain with our knowledge today. But we do have abstract knowledge that could be easily recombined to explain the future. And LLMs might already have those building blocks.

是的，当然。

Yeah, for sure.

我认为从某种意义上说，你越思考奥卡姆剃刀如何适用于我们世界中的一切——无论是语言还是科学——就越有趣，因为这些人工产物现在进入了我们的语言模型，并且可能在一定程度上被捕捉到了。不过，我认为这也可能是一种归纳偏置，最终会导致局部最优，对吧？你需要更多的复杂性。但我确实认为，采用这种进化突变风格方法的系统，最终可能仍会将系统推出这些局部最优。

I think in some sense, the more you think about Occam's razor applying to everything in our world, whether it's language or science, it's pretty interesting because these artifacts now go into our language models of today, and potentially there is some amount of this being captured. I think though, it might also be an inductive bias that leads to a local optimum at some point, right? And you need more complexity. But I do think with systems that do this evolutionary mutation style approach, you might still push the system out of these local optima eventually.

是的，还有适应性的重要性这一概念。这正是 Chollet 所说的智能。自从我们有了这些在推理时真正实现适应性的模型——比如测试时主动微调和推理模型等——它们开始在 ARC 上取得不俗的表现。

Yes, and then there's also the notion of the importance of adaptivity. So, this is what Chollet says intelligence is. And since we've had these models that actually do adaptivity at inference time, so things like test time active fine-tuning, and the reasoning models, and so on. They started getting non-trivial performance on ARC.

嗯。

Mhm.

现在，适配巨大的基础模型非常非常昂贵。你知道，这只是一个实际顾虑，我们还没有做到。但我们可以构建像 Shinka Evolve 这样的系统，它兼顾了两者的优点。它利用冻结的基础模型，但提供了适应性。适应性的目的是应对新异事物，创建新的构建块，在这种有原则的树状结构中合成新的构建块，使我们能够适应新异事物。

Now, it's very, very expensive to adapt huge foundation models. You know, it's just a practical concern where we haven't done that yet. But, what we can do is build systems like Shinka Evolve that leverage the best of both worlds. So, they leverage frozen foundation models, but they give you adaptivity. And the purpose of adaptivity is to respond to novelties, to create new building blocks, synthesize new building blocks in this principled tree-like structure that allow us to adapt to novelty.

是的。

Yeah.

所以我们鱼与熊掌兼得。

So, we are having our cake and eating it.

我觉得很有趣的是，Jeremy 在你的播客中，当你问他关于 Shinka 时，他说他不相信使用像 Shinka 这样的系统能获得很多百分点的提升，但你可以让它更高效，对吧？这大致就是他的回答要点。而对我来说，一旦你让它更高效，你就可以再次扩展它，对吧？所以，如果你有一个更便宜的系统，可以生成更多的指令，我预计由于开放性的本质，你可能会从中获得一些改进。目前我没有任何证据，我很想收集这些证据。这又是开放性的魔力在起作用：只要这些 ARC AGI 的训练示例为你的最终测试提交提供了良好的信号，你就应该能够进步。

I found it very interesting that Jeremy basically in your podcast when you asked him about Shinka was saying like he doesn't believe that there are a lot of percentage points to be gained by using a system like Shinka, but you can make it much more efficient, right? That was sort of the gist of his answer. And to me it's like once you have made it much more efficient, you can scale it up again, right? So, if you essentially have a cheaper system that can generate many more instructions, I would expect that by the nature of open-endedness, you might get some amount of improvement out of it. Right now, I don't have any evidence for it. I would love to collect that evidence. It's again like the magic of open-endedness that comes into play that as long as these training examples of ARC AGI give you a good signal for your final test submission, you should be able to progress.

是的，这是一个很好的过渡，因为在圆填充问题上，它的样本效率如此之高，以至于在与 LLM 的不到 200 次交互中就收敛到了解决方案。但我在想，这很好，但它仍然非常依赖初始条件。你知道，我们谈论这种设计偏差等等。所以，我们投入的东西非常重要。但现在我们可以做的是横向扩展。我们可以运行这个一千次，并让另一个过程来提示、生成、培育初始条件，因为每次运行 Shinka Evolve 时，它都在搜索认知树的部分。如果我们大规模地扩展它，会发生什么？

Yes, and that is a great segue because certainly on the circle packing problem, it was so sample efficient that in less than 200 interactions with an LLM, you converged on the solution. But, I was thinking that great, but it's still quite dependent on the starting conditions. You know, we talk about this design bias and so on. So, what we put in is very important. But, now what we could do is scale out. So, we could run this a thousand times, and we could have another process which prompts, generates, breeds the starting conditions because every time we run a Shinka Evolve, what it's doing is it's searching parts of the epistemic tree. And what would happen if we just scaled that out massively?

我们还没试过，但你甚至可以从一个空程序开始，对吧？那基本上是一样的，对吧？然后你会从那个空程序分支出去。我期望如此。是的，我们没这么做只是因为成本和时间原因。但我确实认为，在很多方面，这个问题将推动我们走向真正的开放性愿景，即运行一个系统大约一个月左右，对吧？真正试图榨出成果。是的，我不确定我们是否完全达到了，但我会尽最大努力实现它。

We haven't tried but you could even start with an empty program, right? Which would be basically the same, right? And then you would branch off of that empty program. I would expect. Yeah, we haven't done this simply out of cost and time reasons. But I do think in many ways this is the question that will push us towards this true open-ended vision of running a system for like a month or so, right? Really trying to squeeze this out. Yeah, I'm not sure if we're entirely there yet but I will do my best that we will.

这之所以有趣，是因为我们实际上知道不能从零开始。如果我们从最原始的构建块开始，搜索空间会巨大，而且没有学习信号。所以我们知道需要从堆栈稍高的地方开始，但我们可以大规模并行化。假设我们有一千个不同的 Shinka Evolve 实例。它不一定是尴尬并行的。我们仍然可以有一些共享。所以在它们执行期间，我们仍然可以进行一点交叉，然后也许我们可以在一个类似的元进化循环中运行所有 Shinka Evolve 实例。我的怀疑是，与 Jeremy 相反，我同意你的看法。我们知道存在多样化的垫脚石，可以显著改进许多这些解决方案。我们只是还没有扩展它。

And the reason this is interesting is we know as a practical matter that we can't start with nothing. If we were just starting from the most primitive building blocks, the search space would just be huge and there'd be no learning signal. So we know we need to start a little way up the stack but we can massively parallelize that. So that, let's say we have a thousand different instantiations of Shinka Evolve. It doesn't have to be embarrassingly parallel. We could still have some sharing. So during their execution we could still have a little bit of crossover and maybe then we could run all the Shinka Evolve instantiations in a similar kind of meta evolution loop. And my suspicion is, contra Jeremy, I agree with you. We know there are diverse stepping stones out there that could dramatically improve many of these solutions. We simply haven't scaled it up yet.

是的。我还相信，使用像 Shinka Evolve 这样的系统能够自动检测对于给定问题，基于指令的优化方法还是基于变换的方法才是正确的做法。有时甚至可能是混合的，对吧？有些事情你可能用 Python 比用语言更容易表达，对吧？所以我真的很有兴趣探索这一点。

Yeah. I also believe that using a system like Shinka Evolve could be able to automatically detect whether an instruction-based optimization approach for a given problem or a transform-based approach is actually the right thing to do. And sometimes potentially it's even the mixture, right? There are some things you can probably easier articulate in Python than you can articulate in language, right? So I would be really interested in exploring that.

是的，你之前提到 Jeff Clune。Jeff Clune 的论文是什么？那个生成问题的东西。

Yeah, I mean you said earlier about Jeff Clune. What was Jeff Clune's paper? The thing that generates problems.

是能力发现。

It's capability discovery.

我在 NIPS 上确实和他谈过这个，但类似的东西也可能很迷人。你知道，我们也在生成问题和解决方案，然后移动后端。但我认为这在商业上的落地方式将是一种新型的 GPT，每个人都在解决不同类型的问题，系统将像一种 Tinker Revolve 的大规模分布式版本，数学家在这里使用平台解决这个问题，它会发现共性，并将它们联系起来。因为我认为在这个过程中也需要利用人类的创造力。

I did speak to him about this at NIPS, but something like that could be fascinating as well. You know, where we're also generating the problems and solutions and then kind of moving the back end. But I think the way this will land commercially is there'll be a new type of GPT where everyone is solving different types of problems and the system will be like a kind of Tinker Revolve with a massively distributed version where mathematicians are using the platform over here to solve this problem and it will see commonalities and it will kind of link them together. Because you need to leverage human creativity in this process as well, I think.

未来的一个大挑战将是如何改变我们的激励机制，使其真正规模化，对吧？我认为，例如，需要一定量的经济或机制设计，以确保每个人都仍然乐于参与其中，对吧？所以，也许我们会有更多的排行榜，用于任何可以数值评分的东西。我认为，看到算力、这些自动化智能体、人类的引导和操控最终如何改变和革新科学，乃至更广泛的社会，将会非常有趣。

A big challenge going forward is going to be how do we change our incentive system for this to actually scale, right? I think for example, some amount of economy will be needed or some amount of mechanism design in order to make sure that everyone is still happy to engage in it, right? So, maybe we're going to have many more leaderboards for whatever is numerically scorable. And I think this will be really interesting to see how compute, these automated agents, human shepherding and steering will ultimately change and revolutionize science and I guess society more generally.

Rob，展望未来，旧金山有很多人想要扩展语言模型，他们正在加入隐式的自适应和组合形式，构建控制器并进行带有可验证反馈的强化学习等。我认为你持有略微不同的观点，即我们需要更加开放，需要使用进化算法等。你认为他们是在走向死胡同吗？你认为他们可能会改变方向吗？我的意思是，这会走向何方？

And Rob, looking at the future, we've got a lot of people in San Francisco that want to scale language models and they are adding in implicit forms of adaptivity and composition so that they're building controllers and they're doing reinforcement learning with verifiable feedback and so on. I think that you subscribe to the slightly different idea that we need to be far more open-ended and we need to be using evolutionary algorithms and so on, but do you think that they are on a path to nowhere? Do you think they might change tack? I mean, where is this going?

实际上，我认为这些东西是可以互补的。比如说，假设你微调自己的模型，使其成为圆形填充专家。我确实相信，将不同类型的强化学习微调模型混合到模型集合中，然后有一种好的方法自适应地选择使用哪个模型，这并不是一个坏主意。所以我完全信奉这种开放性的哲学，阅读 Ken 和 Joel 的书确实是我生命中的一个重要时刻，我想看看我们能将这一点推到多远。我认为我们还没有达到收敛，无论是模型的能力、我们围绕它们搭建的框架，还是我们人类与它们交互的方式。所以对我来说，这三点是：模型能力、模型框架和用户界面，我认为在这三个角度上我们还有很多可以推进的。

So, I actually think that these things can be complementary, right? In the sense like let's say you fine-tune your own model to be a circle packing expert, right? So I do believe that mixing in different sorts of RL fine-tuned models into the ensemble of models and then having a good way to adaptively select which model to use is not a bad idea, right? So to me I just fully subscribe to this philosophy of open-endedness and reading Ken's and Joel's book was really a fundamental moment in my life and I want to see how far we can push this and I think we're not yet at convergence where either the capabilities of the models have converged or the way we scaffold around them or the way we humans interface with them. So to me there are really these three points: model capability, model scaffolding, and the user interface, and I think we have a lot still to push on all three angles.

很好。我们唯一没谈到的是，我们讨论了圆形填充问题，但你还把它应用到了其他一些事情上。能跟我们讲讲吗？

Beautiful. The only thing we didn't talk about was we spoke about the circle packing problem but you also applied it to a few other things. Can you tell us about that?

我们做的一件事是使用了一个名为 ADAS 的框架，即智能体系统的自动设计。基本上，不是手动编写智能体框架，而是使用 LLM 为特定任务编写智能体框架。我们研究了数学任务 AMIE，并使用 Chinchilla 来进化一个智能体，即用智能体来进化智能体。我们发现，这可以显著提高像 GPT-4.1 Nano 这样非常便宜的模型的性能，而且智能体框架实际上能够泛化到其他语言模型或不同年份的 AMIE。这是一个应用。另一个重要的应用是 ALE bench。ALE bench 基本上是 Sakana 的其他同事（包括 Yuki，他也是论文的作者之一）所做的工作，涉及之前由 AtCoder（日本著名的竞技编程组织）完成和执行的启发式编程竞赛问题。我们展示了 Shinka 也可以很好地作为“酷科学家”工作。基本上，我们获取了之前设计的 ALE 智能体得到的初始解决方案，然后使用 Shinka 在这些初始解决方案之上进行优化，并表明在其中一个编程任务中，如果这个智能体和 Shinka 的组合参加了挑战，它基本上会排名第二。所以我认为有证据表明 Shinka 可以作为“酷科学家”工作，不仅适用于 LLM 智能体，甚至可能适用于人类，就像我们之前讨论的那样。最后，我们研究的最后一个应用是设计混合专家负载均衡损失函数。在 Sakana，我们之前做过一些工作，叫做 DiscoPop。我想我们在上次播客中讨论过，我们使用 LLM 来设计目标函数，当时我们做的是偏好优化和后训练，而这里我们做的是混合专家的负载均衡。同样，我们发现，即使在仅仅 20 代内，我们也能探索到不仅是一个单一目标函数，而是一个凸包，其中存在性能和负载均衡之间的不同权衡。所以我认为这是 Shinka 的另一个应用，它不仅仅是找到最佳解决方案，而是本质上照亮了一个程序空间，其中总是存在潜在的权衡，例如运行时间和圆形填充质量之间的权衡。拥有一个能够探索所有这些的系统也很重要。

So one thing we did was we used a framework called ADAS, automatic design of agentic systems, where basically instead of manually writing an agent scaffold you use an LLM to write agent scaffolds for a specific task, right? So what we did is we looked at mathematics tasks, AMIE, and we used Chinchilla to evolve an agent, right? So using an agent to evolve an agent and we found that we could dramatically improve the performance of very cheap models like GPT-4.1 Nano, but the agent scaffold was actually able to generalize to other language models or to different years of AMIE, right? That was one application. One important other application that we did was to ALE bench. ALE bench is basically work done by other folks at Sakana including Yuki who's also part of the paper which is considering heuristic programming contest problems previously done and executed by AtCoder, which is this famous Japanese competitive programming organization, and we showed that Shinka can also work very well as a cool scientist. So basically we took initial solutions obtained by an ALE agent that was previously designed and then we optimized on top of these initial solutions with Shinka and showed that on one of these programming tasks, if the combination of this agent and Shinka would have competed in the challenge, it would have ranked second place basically. So I think there's some evidence that Shinka can work as a cool scientist, not only for LLM agents but potentially even for humans like we discussed before. And then finally, the final application that we looked at was designing mixture of expert load balancing loss functions. So at Sakana we've done some previous work called DiscoPop. I think we discussed this during the last podcast we did where we're using LLMs to design objective functions and back then we did it for preference optimization and post-training, and here we did it for load balancing of mixtures of experts. Also there we found that within even only 20 generations we were able to explore not only a single objective function but a convex hull where there are different trade-offs between performance and load balancing and so on. So I think this is another application of Shinka where it's not only about finding the best solution but essentially illuminating a program space where there are always potential trade-offs between, for example, runtime and the quality of the circle packing, right? And having a system that can explore all of these is important as well.

我很兴奋看到你将这个应用到 ARC 挑战中。你对此有什么想法？

I'm very excited to see you apply this to the ARC challenge. What are your thoughts about that?

我还在收集结果。所以，在完成之前我不想做出任何硬性断言。但我希望有机会肯定能改善这些系统的成本，甚至可能提高性能，但拭目以待吧。

I still need to collect results. So yeah, I don't want to make any hard claims before having done this. But I would hope that there is some chance of for sure improving the cost of these systems and then potentially even performance, but yeah, to be seen.

哦，非常令人兴奋。所以你已经开始做一些实验了。可能很快会有激动人心的消息。

Oh, very exciting. So you've done some experiments. Exciting news is potentially coming.

我已经开始研究了。

I've started looking into it.

嗯，总的来说，你对 ARC 有什么看法？

Yeah, I mean, what are your thoughts in general about ARC though?

我认为它很好。它非常重要，填补了一个重要的空白。我确实非常尊重 Francois，并在论文刚出来时就读了它，当时没有人认为能取得超过 10% 的成绩。从社会层面来看，自那以来我们取得了多大的进展，这也相当令人着迷。有时当你深陷战斗模式或工作模式时，你不能忘记一年前你在哪里。然后回头看看，真是令人惊叹。还有自 01 年以来我们走了多远。

I think it's good. I think it's really important and I think it fills an important gap. And I do really deeply respect Francois and read the paper when it first came out and no one thought of actually being able to get numbers above 10%, right? And it's also pretty fascinating on a society level how far we've come since then. And sometimes while you're deep in battle mode or work mode, you can't forget where you were 1 year ago. And then just looking back, it's pretty amazing. Also how far we've come since 01.

是的，说到这个，我认为 Francois 没有得到足够的认可，因为这是一个非常好的基准。而且不一定是因为人们认为的原因，因为 Francois 总是说我们需要一个对人类容易、对 AI 困难的基准。从某种意义上说，情况并非如此。当 ARC V2 出来时，我说过它实际上对人类来说非常困难。你知道，有一个任务让 Dog 困惑了大约 15 分钟。我们三个人看着它，就是解不出来。这是那种取决于你的视角，你可能立刻得到答案，也可能得不到的事情。所以有这样的批评。人们说 ARC V3 甚至更难。但我觉得这有点偏离重点。我认为他的意思是，很多这类竞技编程问题，数据集被污染了。这些问题之前部分或全部被解决过，这意味着当你查看认知树时，解决它们的许多构建块在树中位置很高。他关注的是那些数据集污染很少的问题，需要从非常抽象的构建块开始解决。

Yes, and saying that, I think Francois doesn't get enough credit because it's such a good benchmark. And not necessarily for reasons people think because Francois is always saying that we need to have a benchmark which is easy for humans and hard for AIs. And in a sense, that's not quite the case. I said when ARC V2 came out that it's actually very difficult for humans. You know, there was one task where Dog was stumped for about 15 minutes. There were three of us looking at it and we just couldn't. And it's one of those things that depending on your perspective, you might get it straight away or you might not. So there's that criticism. And people have said that ARC V3 is even harder. You know, but I think that's rather missing the point. I think he's saying that with a lot of these competitive coding problems, the dataset is contaminated. These are problems that have been solved before in part or in whole, which means when you look at the epistemic tree, many of the building blocks for solving them are very high up in the tree. He's looking at these problems that there is very little dataset contamination and they need to be solved from very abstract building blocks.

所以你从树状结构更低处开始，通过组合非常抽象的构建块来合成模型，这本质上是智能的核心。正因如此，我认为 ARC 确实在推动我们构建自适应系统，这些系统可以说是智能的。

So you're starting much lower down the tree and you're synthesizing a model by composing together very abstract building blocks which is the essence of intelligence. And I think for that reason ARC is really kind of pushing us to build adaptive systems which we could say are intelligent.

是的，我同意。从很多方面来说，我真的很期待未来几年，看看我们能把这个推多远，以及之后能获得多少泛化能力。因为我相信，当你观察较新的模型时，它们在转换式代码演化或为 ARC 生成输出方面，比在基于指令的层面上要好得多。我认为这至少可能已经是 ARC AGI-1 上一定程度过拟合的小迹象，对吧？

Yeah. I agree. I mean in many ways I'm really looking forward to the next years and seeing how far we can push this and then also how much generalization we can get afterwards because I believe when you look at the more recent models, they're getting much better at the transform-style code evolution or outputting for ARC than they are on the instruction-based level and I think this might already be a small sign of some amount of overtraining on ARC AGI-1 at least, right?

我确实相信，在完全科学自动化和我所做的那种工作到来之前，工作的某些方面将被自动化。但我可以想象，我每天处理的某些维度肯定会被 AI 冲击。那么问题就是，是否会有新的维度被打开，由我们人类来填补，对吧？我认为我之前提到的“引导”之类的概念，我真的希望那是前进的方向，对吧？从某种意义上说，人类是掌舵者，而他们的生产力被大幅放大。

I do believe there are some aspects of work which will be automated before it comes to sort of fully science automation and the type of work I'm doing, but I could imagine that certain parts of the dimensions that I deal with every day are for sure going to be hit by AI. And then the question is are there going to be new dimensions opened up that we as humans will fill in, right? And I think what I said before about shepherding and so on. I really hope that that's the way forward, right? In the sense that humans are the ones steering the ship while just being massively amplified in their productivity.

目前我并没有看到那种被预测的就业市场颠覆。从个人经验来看，从某种意义上说，它让招聘变得非常困难。你知道，剧本写手用 ChatGPT，我一眼就能看出来。而作家和文字编辑实际上比以前更抢手，因为他们要修正 ChatGPT 生成的那些垃圾内容。还有云计算的类比。你知道，英国年薪 6 万英镑的 IT 系统管理员，他们重新包装成云和 DevOps 工程师，收入翻了一倍多。人们非常有适应能力。他们看到新趋势、新潮流，就适应并在此基础上增加价值。这个趋势已经持续了很长时间。你认为 AI 会如此具有变革性，以至于超越人们的适应能力吗？

Right now I am not really seeing the kind of job market disruption that was being predicted. I know from personal experience that in a sense it's made it very difficult to hire people. You know, script writers use ChatGPT. I can spot it instantly. And writers and copy editors are actually in more demand than they were before fixing all of the crap that has been generated with ChatGPT. And there's the cloud analogy as well. So, you know, IT system administrators who were earning 60,000 pounds a year in the UK, they rebranded as cloud and DevOps engineers and they more than doubled their pay. And people are very adaptive. They see new trends, new bandwagons, and they just adapt and they add value on top. And that has been the trend for a very long time. Do you think that AI is going to be so transformative that it will transcend people's ability to adapt?

我认为这只是速度的问题。我之前谈到文化进化和技术进化，似乎我们人类需要更多的适应和更多的时间来习惯技术，以开辟这些我们可以填补的互补性利基，对吧？首先，我认为我们还没有达到技术发展的天花板，对吧？所以，也许几年后，我们需要的垃圾编辑工作会减少，就像你说的，但我确实认为我们需要更多时间来适应与这些系统交互的不同模式，对吧？我想每个人都能与聊天助手交互，但我认为这是与 AI 智能体交互的最初级形式，对吧？所以，是的，我认为我们需要把握好这一切的节奏，我们需要在人机界面、UI/UX 设计以及如何确保人类在这种体验中感到满足方面进行更多的探索。

I think it's just a question of speed rate. So, I was talking about cultural evolution and technological evolution, and it seems like we humans we need more adaptation and more time to get used to the technology to carve out these niches where we can fill in and it's complementary, right? So, first off, I think we're still not at the ceiling of the technological progression, right? So, maybe in a couple of years we will need less of slop editing like you said, but I do think we need some more time to adapt to the different modalities of interacting with these systems, right? I think everyone can sort of interact with a chat assistant, but I think this is the most naive form of interacting with AI agents, for example, right? So, yeah, I think we need to get the pacing of all of this right, and we need to do much more exploration in human machine interfaces, UI UX design, and how to make sure that humans feel fulfilled during this experience.

这特别相关，因为你知道，你是 AI 科学家论文的幕后推手，现在有了第二版。请允许我稍微怀疑一下。你知道，我们之前谈到演化系统来做特定的事情。目前，尽管它们很好，但感觉它们仍然相当依赖人类监督者的指令和意图。所以，这很大程度上是人与系统之间的交流。因为其含义是，未来我们可能会有如此自主、如此开放的系统，能够找出有价值的研究方向，以至于不再需要人类。而我之所以还不那么担心劳动力市场颠覆，是因为我仍然深信人类是世界上深度理解和创造力的源泉。如果我不相信这一点，我会非常担心。

This is particularly relevant because, you know, you were behind the AI scientist paper and there's now a version two of that. Allow me to be a tiny bit skeptical. You know, we were talking about when we evolve systems to do a particular thing. And at the moment, it feels like as good as they are, they are still quite parasitic on the instructions and intentions of the human supervisor. So, it's very much an exchange between the humans and the system. Because the implication is that in the future, we might have systems that are so autonomous and so open-ended and can figure out valuable things to research that humans wouldn't be needed anymore. And the reason why I'm not that worried yet about labor market disruption is I still believe deeply that humans are the source of deep understanding and creativity in the world. If I didn't believe that, I would be very worried.

我同意。对我来说，AI 科学家 V1 和现在的 V2 是对潜在变革的一瞥，但我完全同意，要取得真正重大的科学突破，比如每天多个突破之类的，你仍然需要人类在循环中，要么播种，要么指导探索方向，要么验证、检查，并实际传递这些见解，对吧？所以，我认为不会所有博士都失业。这更像是人类与这项技术的核心进化，在理想的未来，它将使人类能够专注于他们真正擅长的事情，对吧？所以，我认为它将放大人类擅长的这些潜在维度，对吧？我认为关键的一点是，我们人类要尽早与这些系统互动，以便真正对这个发展过程拥有影响力和所有权，对吧？最终是集体智慧将共同塑造所有这些系统。

I agree. To me, the AI scientist V1 and now V2 are glimpses into a potential transformation, but I fully agree in order to make really big scientific breakthroughs like multiple of them every day or whatever, you still need humans in the loop to either seed or guide the direction in which to explore or to verify, check, and actually transfer these insights, right? So, I think it's not going to be like all PhDs will be unemployed. It's more going to be a core evolution of humans with this technology and potentially in an ideal future for me, it will allow humans to focus on what they're really great at, right? So, I think it's going to be an amplifier of these latent dimensions humans are great at, right? I think something that's critical is that we as humans try to interact with these systems as early as possible in order to actually have influence and ownership over this development process, right? It's ultimately collective intelligence that will shape all of these systems together.

你认为这些系统会变得极其复杂，以至于在某种程度上脱离人类吗？

And do you think these systems can become incredibly sophisticated such that they are somewhat detached from humans?

嗯，我的意思是，对于 AI 科学家 V2，我们发布了一篇提交给 ICLR 研讨会的论文，它能够在元评审之前通过接受阈值。所以，我确实认为至少对于研讨会级别的贡献，我们正在接近目标。虽然并非 AI 科学家论文的每次提交都能达到那个阈值，但我们已经到了可以讨论嘈杂评审过程的地步，而且只要你有大预算，你可能会从中得到一些东西。我认为未来对于更大的创新等，目前你仍然需要人类，但我们正处于实现这一目标的 GPT-1 时刻，可能在 10 年后，一旦基础设施建立起来，情况将大不相同，对吧？所以，有像 Periodic Labs 这样的地方，对吧？他们正在建造真正的物理实验室，配备机器人系统来自动执行实验。这需要一些时间，但可以想象，随着我们在这些系统上进行强化学习，并实际考虑负面结果和真正的假设检验，让这些系统成为带有验证器的优秀假设检验者，我们可能能够解锁更多能力。

Well, I mean with the AI scientist V2, we released that one paper that we submitted to an ICLR workshop was able to pass the acceptance threshold before meta review. So, I do think at least for workshop level contributions, we're getting there. While not every submission in AI scientist paper does or is reaching that threshold, we're at the point where we can even talk about noisy review processes and this actually being something that as long as you have a large budget, you might get something out of it. I think going forward for the bigger innovations and so on, for now you still need humans, but we're at the GPT-1 moment of making this a reality and potentially in 10 years this is going to look very different once the infrastructure for it has been built up, right? So, there are places like Periodic Labs, right? Which are now building real physical labs with robotic systems to automatically execute experiments. This will take some time, but it is imaginable for sure that as we do RL on these types of systems and we actually also account for negative results and for actual hypothesis testing. So, getting these systems to be real good hypothesis testers with verifiers in the loop that we might be able to unlock many more capabilities.

嗯，我想我不希望听起来像个勒德分子。所以，完全有可能这只是因为我没有想象未来的能力。因此，未来这些系统有可能理解得非常深刻并具有创造力。我认为现在的问题是，它们只在认知树中向下理解了几层。所以，它们能做表面层次的重新组合，在已经发现的事物范围内发现新东西，但我们对认知树的理解非常深入，这意味着我们的创造潜力锥体要宽得多。这个差距有可能被弥合。那之后会发生什么？

Yeah, I mean I suppose I don't want to sound like a Luddite. So, it's entirely possible that this is just, you know, I don't have the imagination to think about the future. So, it is possible that in the future these systems might understand very deeply and be creative. I think right now the problem is they only understand things a few levels down in the epistemic tree. So, they can do some surface-level recombination and discover new things in the basin of things that are already discovered, but we understand things very deep down in the epistemic tree, which means our cone of creative potential is much wider. It's possible that that gap might be closed. What would happen then?

我认为科学过程本质上就像树搜索，对吧？所以，我觉得很多进化论的类比可以迁移到科学研究中。我们遍历不同想法或实验的树，然后在论文中报告树中的一条路径。我之前提到过，我们需要更完整的树数据集来训练这些大语言模型系统，让它们真正学会如何进行这种探索和觅食。同时，我觉得文化层面也会发生进化。我们会更擅长掌舵，我可以想象未来世界的研究方式将完全不同。而且我确信，现在 99% 的机器学习研究已经借助 AI 完成了，对吧？想想 ChatGPT 头脑风暴、Cursor 编程、Claude Code 等等。长期来看，我们会从“与 AI 一起”逐渐转向“由 AI 执行”，人类则负责更高层次的编排和监督。

The way I kind of think about the scientific process is like a tree search ultimately, right? So, I think a lot of analogies from evolution transfer to scientific research. In the sense that we traverse a tree of different ideas or different experiments and then in the paper we report one path through that tree. And I think what I alluded to before, we need much more like full tree datasets for training these LLM systems to actually learn how to do this exploration and foraging, basically. At the same time, I feel like evolution will also take place on the cultural level for us. We will get better at steering the ship and I can imagine that in the future world the way we do research will be completely different. And I'm pretty sure that right now already 99% of machine learning research is done with AI assistance, right? Think about ChatGPT brainstorming, cursor coding, Claude Code, etc. In the long run, we're going to move on that spectrum from with AI closer to by AI and then more high-level orchestration and overseeing by humans.

还有一个概念是价值函数在多大程度上与人类内在耦合。一种观点认为，AI 会发展出自己的思想，基本上超越人类，拥有不依赖于我们的自主性。我个人不赞同这种观点。另一种观点是，比如 AI 科学家第 10 版，它会持续进行认知觅食，发现有用的新事物，而这些事物必须对我们有用。因为如果它发现的东西对我们没用，我们就不会使用它们，什么也不会发生。那么，你认为价值函数会始终与人类耦合吗？

There's also the notion of how intrinsically coupled to humans is the value function. So, one school of thought is that AI will develop a mind of its own and it will basically transcend humanity and it will just have agency which is not parasitic on ours. I personally don't subscribe to that view, but the other view is that it is like, let's say the AI scientist, version 10, it's going to be continually epistemic foraging. It's going to be finding new things that are useful and they kind of have to be useful to us. Because if it finds things that are not useful to us, then we just won't use them and nothing will happen. So, do you think there'll always be a kind of coupled value function to humans?

Jeff Clune 有关于 Omni 的工作，对吧？利用大语言模型作为人类兴趣的摊销概念。我认为最终我们训练这些系统的方式是与人类数据耦合的。未来，它还将与使用验证器收集的人类数据耦合。所以，我很难相信长期来看，当你用 AI 科学家智能体运行这种开放性范式时，它会完全偏离到完全不可解释或与人类关心的问题无关的方向。而且，人类可以在一定程度上引导搜索的方向。所以，你可以告诉系统，比如，尝试做癌症研究，解决我们关心的问题。最终，是我们控制着投入多少算力。

Jeff Clune had this work on Omni, right? And using LLMs as amortized notions of interestingness for humans. And I think ultimately the way we train these systems is coupled in human data, right? And going forward it will also be coupled with human data that is collected using verifiers. So, I have a hard time believing that in the long run when you run this open-endedness paradigm with AI scientist agents, it's going to completely divert to something that's either fully non-interpretable or unrelated to problems we as humans care about. And then again, humans can steer to a certain degree where the search happens. So, you can tell a system, okay, try to do cancer research, and work on problems that we care about. And ultimately, we are the ones who control how much flops are being pushed into this.

是的，因为作为一个思想实验，我可以想象，比如在数学领域，如果 AI 科学家能提出全新的问题表述并解决它们。这些是人类从未想到过的东西，也许人类对答案不那么感兴趣，因为人类没有花时间思考过。想想看，我们可以无限探索数学的系统发育。在某个点上，也许我们就不再关心了。也许我们可以永远地开拓那个空间。

Yeah, because as a thought experiment, I can imagine, let's say, in the world of mathematics, what if an AI scientist could come up with entirely new problem formulations and then solve them. And these are things that humans had never conceived of before and maybe they would be less interested in the answer because humans hadn't spent time thinking about it. And if you think about it, we could just explore the phylogeny of mathematics to the nth degree. And at some point, maybe we just wouldn't care anymore. Maybe we can just carve out that space just forever and ever.

是的，但也许在某个地方有一块垫脚石，能让我们真正关心的不同领域产生新创新，对吧？所以，很难先验地判断某件事是否有趣。

Yeah, but maybe down the road there is a stepping stone that enables a new innovation in a different field that we actually care about, right? So, it's very hard to say a priori whether or not something is interesting or not.

是的，还有一个概念——我喜欢这种多元智能和多元思维的想法。也许我们可以在一个完全陌生的空间里创造人工制品。我们甚至可能赋予它们道德价值，可能不想关闭电源，因为我们希望这些陌生的人工制品存活下去。

Yes, and there's also the notion of I love this idea of diverse intelligences and diverse minds. And maybe we could just create artifacts in a space which is completely alien to us. And we might even ascribe moral value to them, and we might not want to turn off the power because we want these alien artifacts to stay alive.

也许吧。我读了很多科幻小说，但我不太愿意猜测所有这些，但我非常确定一件事：我们进行研究和科学的方式将在未来 5 年、10 年和 20 年发生根本性变化。我希望我们能够借助 AI 解决一些目前看似遥不可及的重大问题。

Maybe. I read a lot of science fiction, but I would sort of shy away from speculating about all of this, but I do think one thing I'm extremely certain of is that the way we conduct research and science is going to fundamentally change in the next 5 years, 10 years, and 20 years. And I hope that we're going to be able to tackle some of the biggest problems which are still seemingly unreachable right now with and by AI.

陶哲轩发帖说他一直在用 GPT-5，这加快了他的速度，省去了很多苦差事。但悲观的观点是——Scott Aaronson 也发过类似的内容——也许懒惰在作祟，使用 AI 模型实际上阻碍了我们跳出框框思考。它鼓励我们在已知事物的附近搜索。这很有用，有一个知道所有实验、所有 20 年前人们做过的事情的人工制品非常有用，但现在没有人真正把他们的才华和天赋应用到全新的领域。

So, Terrence Tao has posted that he's been using GPT-5 and it's been speeding him up. It's taking away a lot of the drudgery. But the cynical take is that and Scott Aaronson posted something similar as well. The cynical take is that maybe laziness is stepping in and in some pernicious way, using AI models is actually stopping us from thinking outside the box. So, it's encouraging us to search in the neighborhood of things that are known. And that is very useful. It's very useful to have an artifact that knows all of the experiments, all of the things that were ever done by people 20 years ago, but now we don't have people really applying their brilliance, their talent in completely new areas.

所以，首先，这些专家已经在日常工作中使用这项技术，这很好，对吧？而且我认为，真正顶级的科学家尝试推动这些系统的能力极限，或者找出这些系统无法做到的盲点，也非常重要。

So, first off, it's great that these experts are already using the technology in their day-to-day work, right? And I think it's also important that really, really top-level scientists try to push what's capable with these systems or squeeze out where there might be black spots or stuff where these systems can't do.

嗯，其次，我认为这归根结底是自律以及我们如何培养下一代的问题，对吧？所以，在个人层面上的自律，比如你在多大程度上只是点击接受这些系统提出的所有内容，以及在教育下一代方面的责任，我们需要教给孩子们，这些系统输出的内容不一定总是真实的，事实在某种程度上可能是主观的，而且需要对所得到的信息进行更多研究。我认为这将是我们必须经历的文化演变，并尽力从中获益。

Um second off, I think it comes down sort of to discipline and how we raise sort of the next generation, right? So, discipline on a personal level, like how much do you just sort of tap accept everything that's being proposed by these systems and responsibility in terms of educating the next generation in the sense that we need to sort of teach our kids that ultimately what comes out of these systems might not always be true, that facts can be sort of subjective, if you will, and that there needs to be more research about what's being given to you. And I think this will be, like I said, this cultural evolution that we have to step through and try to make the best out of.

是的，自动驾驶这个类比很有意思，因为在使用 Cursor 时有一种倾向，你知道，模型生成得如此之快，你甚至来不及阅读输出的 token，然后你就一直按接受。这和汽车一样，一旦自动驾驶太强，你就完全放松了。然后你会看到一种分歧，因为思考必须基于你自己的路径。存在这种路径依赖。当你开始被另一种思路寄生时，你就停止思考自己的路径，不再处于驾驶位了。

Yeah, the autopilot thing is very interesting because there is a tendency using cursor just to, you know, at some point the models are getting so quickly that you can't even read the tokens coming at you and then you just press accept and you press accept. It's the same thing in cars that as soon as you have too strong of an autopilot, you just completely switch off. And then you see a divergence because there's something about thinking that it must be grounded on your path. There's this path dependence. And when you start kind of becoming parasitized by this other train of thought, then you stop thinking about your path and then you're not in the driver's seat anymore.

这听起来有点苛刻，但有时我想知道这些系统，比如这些编程助手，是否几乎像毒品一样，对吧？你会沉迷其中，用光所有预算，然后需要再次加载，一旦完全达到预算上限，你会觉得：“好吧，我现在该怎么办？”我认为一旦发生这种情况，你真的应该重新思考你的工作方式。对我来说，现在有些部分自动接受是可以接受的，有些部分则绝对不行，你需要深入其中。我认为我们现在处于一种奇怪的非平衡状态，一切都在不断变化，对吧？系统或模型在变，功能在变，系统擅长的部分也在变，我们人类需要不断适应。我认为这是一个巨大的认知挑战，我们都需要意识到有些问题和挑战我们必须适应。我认为最好的方法就是尽可能多地与这项技术互动，也许从这种体验中找到新的研究思路。

This is now like a bit of a harsh statement, but sometimes I wonder if these systems, like these coding assistants, are almost like drugs, right? In the sense that you become addicted, you use up all your sort of budget, and then you need to load up again, and once you fully reach sort of the budget limit, you feel like, "Okay, what am I going to do now?" And I think once that happens to you, you should really sort of rethink the way how you work, right? And to me, right now, there are certain parts where like sort of auto accepting is acceptable, and there are certain parts where it's definitely not, and you really need to go deep into it. And I think right now, we're sort of in this weird non-equilibrium state where things are moving constantly, right? So, the systems or the models are changing, the features are changing, the sort of parts where the systems are good is changing all the time, and we humans need to constantly adapt to that, right? And I think it's a big cognitive challenge, and I think we just all need to be aware that there are certain problems and certain challenges that we have to adapt to. I think the best way to do so is just interact with this technology as much as you can, and maybe find new research ideas out of that experience.

那么 AI Scientist V2 与 V1 有何不同？

And how is AI scientist V2 different to V1?

在 V1 中，我们采用了基于模板的方法。我们有一个基础实验，然后让 LLM 通过 Semantic Scholar 调用和文献搜索来生成想法，接着基于模板实现这些想法，基本上是代码差异。然后线性执行实验计划，最后写一篇论文。可能发生的情况是，有一个想法没有成功，但论文中的实验仍然线性执行，最终写出一篇论文。这已经很令人印象深刻，因为它看起来很像科学，但如果你考虑人类的科学和科学方法，它更像是研究，就像我之前说的，对吧？你会调整接下来要执行的内容，并根据积累的证据进行改进。这就是卡尔·波普尔的可证伪性概念，对吧？我们为假设收集证据并拒绝其他假设，循环进行，直到我们想要发表或有所发现。我们尝试将这个理念直接构建到 AI Scientist V2 的智能体式框架中。现在它基本上是一个可并行化的智能体式树搜索，不再需要模板实验，而是由 LLM 自己起草。因此，AI Scientist V2 可以应用于更多场景。核心是这种新的智能体式树搜索范式，然后我们进行了一些小的技术改进，比如使用 VLM 阅读器来判断论文标题是否与图表对齐，并将其扩展到更多的计算节点，最后再写一篇论文。

In V1, we used sort of a template-based approach. So, we had like a base experiment, and then for that base experiment, we asked sort of an LLM to generate ideas sort of with Semantic Scholar calls and sort of literature search, and then it implemented sort of these ideas based on the template, right? It did basically code diffs. And then it linearly executed like an experiment plan and wrote a paper in the end. And so, what could happen was that there was an idea, and that idea didn't work out, right? But then in the end, the paper like the experiments were still executed linearly, and you wrote a paper. And this was already impressive in the sense that it looked very much like science, but if you think about human sort of science and like the scientific method. It's much more like research, like I said before, right? You sort of adapt what you're going to execute next, and you sort of refine based on evidence that you accumulated, right? So, this is sort of the notion of falsificationism from Karl Popper, right? In the sense that we collect evidence for hypotheses and reject others, and we do so in a loop, basically, until we want to publish or we find something. And we try to take this notion and directly build it into the agentic scaffolding for the AI scientist V2. So, now it's basically like a parallelizable agentic tree search, where there's no longer a template experiment needed, but this is drafted up by the LLM itself. And thereby, the AI scientist V2 can be applied to many more sort of settings, if you will. So, at the core is sort of this new agentic tree search paradigm, and then we use sort of a couple of minor technical changes like using a VLM reader for sort of figuring out if captions of a paper are aligned with the figures, and we scale this up to many more sort of computational nodes, and then write a paper in the end again.

所以，我尽量以最礼貌的方式说，但批评者可能会说——我不想用“垃圾”这个词——但批评者可能会说，我们生产的论文看起来像论文。它们有图表、有结果、有特定风格的文字，但缺乏深层的认知根基，这意味着在知识树的顶端附近，我们看到了一些组合上的新颖性，但这并不反映深刻的理解。你如何回应这种指责？

So, I'm trying to say this in the most polite way possible, but a critic might say I don't want to use the word slop, but a critic might say we are producing papers which appear like papers. So, they have figures and they have results and they have things written in a certain way, but they're not grounded deep down the epistemic phylogeny, which means that they have, you know, near the top of the tree, we're seeing some novelty in composition happening, but it doesn't reflect a deep understanding. What would you say to that charge?

可以肯定的是，并非 AI Scientist V2 生成的每篇论文都达到《自然》级别的发表标准，对吧？这确实是事实。所以，肯定有一些所谓的“垃圾”或内容，并非科学上的重大发现。但最终我们证明了，获得 workshop 级别的论文是可能的。而且你认为这基本上是第一次，我们至少能够完全自主地投入算力和 API 调用，获得一定数量的科学见解。至少对我来说，目前这是一种很好的方式，用于原型设计想法或研究某个领域，获得起点和初步结果，然后在此基础上继续工作。但当然，还需要更多工作来使整个过程更稳健、更高效，并本质上产生更多真正的阳性结果。

It's for sure that not every paper that comes out of the AI scientist V2 is a nature-worthy publication, right? That's for sure the case. So, definitely there is some amount of, let's say, slop or content that is not like a scientific big discovery being written up by the AI scientist for you. But ultimately we showed that it was possible to obtain a workshop level paper. And you think this is sort of the first time basically where we can see that at least now we're able to fully autonomously spend compute, spend API calls to obtain some amount of scientific insights. And for me at least right now, it's a good way to sort of prototype ideas or to investigate a certain field, get like starting point, initial results, and then to work on top of it. But for sure more work needs to be done to make this entire process more robust, more efficient, and essentially produce many more sort of true positives as you will.

是的，而且这可能就像我们从 GPT-3 升级到 GPT-4 时，保真度大幅提升一样。因为对我来说，“垃圾”仅仅意味着缺乏深层的、有根基的理解。原则上没有理由认为这些东西不能拥有深层的、有根基的理解。它们只是还没有而已。所以这是可以随着时间改进的，但可能改进得很慢。然后在某个时刻，我们可能会想：“天哪，我们有了一个 AI 科学家。”

Yeah, and there might be one of these things, you know, like when we moved from GPT-3 to GPT-4, there was just a massive increase in fidelity. Because the thing is with slop, to me it simply means lack of deep grounded understanding. And there's no reason in principle why these things couldn't have a deep grounded understanding. They just don't have it yet. So it's something that could improve over time, but it's likely to improve quite slowly. And then at some point we might just think, "Oh my god, we've got an AI scientist."

是的，对我来说，这又回到了我们之前讨论的内容。

Yeah, I mean like to me this kind of comes back to what we were discussing about before.

首先，循环中有一个验证器，对吧？也就是说，实验实际上是在计算机上执行的，因此数值结果可以反馈到系统中，以提出下一个要探索的内容。但我们还没有做出像残差连接那样渗透到机器学习各个领域的发现。我认为我们真正需要的是让这些系统更好地整合多个实验的知识，并更好地根据先前的见解提出下一个假设。这可能需要对这些轨迹进行一定量的后训练，但我很乐观，我们也可以通过多样性和以高效但规模化的方式扩展这些系统来实现这一点。

So first off, there is a verifier in the loop, right? In the sense that experiments are actually executed on a computer, so the numerical results can be fed back into the system to come up with the next thing to explore. But we haven't made a discovery like a residual connection that has diffused into everything in machine learning. I think what we really need is to make these systems much better at integrating knowledge over multiple experiments and become better at formulating the next hypothesis based on previous insights. This might require some amount of post-training on these traces, but I'm pretty positive that we might also get there with just diversity and scaling these systems up in an efficient but scaled-up way.

我在想，第一个突破性发现，是会像 AI Scientist 论文，还是像 Chinchilla Evolve？例如，我们可以大规模扩展 Chinchilla Evolve，然后说：“我想发现一种新的架构设计。”

I'm just thinking that the first breakthrough discovery, would it resemble the AI scientist paper or would it resemble Chinchilla Evolve? So, for example, we could do a massively scaled-up Chinchilla Evolve, and we could say, 'I want to discover a new architectural design.'

是的。

Yeah.

然后会发生吗？接着我们会让 AI Scientist 论文来撰写并做消融实验等？也许这就是它的模式。

And would that happen, and then we would get the AI scientist paper to kind of write it up and do ablations and stuff? Maybe that would be the pattern of it.

在某种程度上，我一直在思考如何可能将这两种范式结合起来。AI Scientist 和 Chinchilla 或 Alpha Evolve 风格的优化算法。我确实认为在自动验证方面和问题表述方面还有一些工作要做。论文写作部分实际上是 AI Scientist 中最不重要的。它是一种我们人类习惯的形式因素，有助于锚定我们对科学发现的心智模型。但最终，我不确定论文是否会在 20 年后成为知识传播的媒介。我还在想，我们是否可以让论文更容易被智能体访问。现在它是一份 LaTeX 文档，但你可以想象为每篇论文配备几个模型上下文协议，这样每个图表都可重现，数据可访问，本质上让 LLM 智能体更容易复制工作或在此基础上进行后续工作，通过该接口对论文进行 epsilon 改进和消融实验。但老实说，我不确定这会不会发生，因为有很多改进科学制品格式的好主意，但人们似乎仍然喜欢已经存在了几百年的论文格式。所以我认为这又是一个激励问题，需要真正展示如果这样的东西存在，它将使 AI 智能体在科学发现方面取得更快的进展。

To a certain degree, I've been thinking a lot about how you can potentially even combine these two paradigms. The AI scientist and the Chinchilla or Alpha Evolve-style optimization algorithms. I do think there's some work to be done on the auto-verification aspect and the problem formulation aspect. The paper writing part is actually the least important about the AI scientist. It's a form factor that we humans are used to, and it helps anchor our mental model of a scientific discovery. But ultimately, I'm not sure if the paper is going to be the knowledge transmission medium in, say, 20 years. Something else I've been thinking about is whether we can make papers much easier agentically accessible. Right now it's a LaTeX document, but you could imagine equipping every paper with several model context protocols, so that every figure is reproducible, data is accessible, and essentially make it much easier for LLM agents to either replicate work or to work off of them afterwards, doing epsilon improvements, ablations yourself through that interface to your paper. But to be entirely honest, I'm not sure if it's going to happen because there have been many great ideas for improving the format of scientific artifacts and people still seem to like the paper format which has existed for hundreds of years. So I think it's a question of incentives again and really showing that if something like that would exist, it would enable much faster progress of AI agents for scientific discovery.

是的，论文是一个很好的人机界面。这和自动驾驶类似，对吧？我们可以彻底改造道路网络，安装传感器，从而大幅改进监控、可观测性和优化。但我对这个想法很着迷。所以你的意思不仅仅是实验的可重复性，还包括图表的设计方式、代码等等。因为这样我们就可以创建一个巨大的游乐场，智能体可以重新利用、重新组合、重新研究其他科学家已发表的工作。这也让我想到：拥有一个自动化科学家，是让同行评审变得更重要还是更不重要？

Yeah, paper is a great human interface. It's a similar thing with automated driving, right? That we could revolutionize the road network to have sensors and we could dramatically improve the monitoring and observability and optimization. But I'm fascinated by that idea. So you're saying not just reproducibility of the experiments, but also the way that the figures are designed and the code and so on. Because then we could create this huge playground where agents can repurpose, recombine, restudy work that has been published by other scientists. And it also made me think: does having an automated scientist make peer review more or less important?

我确实认为它实际上让同行评审变得更重要，至少目前是这样。因为我们现在有一种机制可以生成大量论文，这首先增加了人类评审员的工作量，我们需要一种有效的方法进行筛选，然后只将最精华的部分留给人类验证。所以我认为目前最终的验证仍然是人类以及结果在社区中的传播。我们需要更好的工具来进行这种自动筛选和验证。比如我们有 AI Scientist 附带的 AI 评审员，但你可能实际上需要某种形式的实验执行来真正验证一切。例如，OpenAI 在 PaperBench 上的工作，正试图利用 LLM 软件验证朝这个方向发展。所以我乐观地认为我们将在未来几年内解决这个问题。

I do think it actually makes it more important, at least for now. In the sense that we now have a mechanism that could generate many papers, and it first increases the workload on human reviewers, and we need some effective way for filtering and then essentially only taking the cream of the crop for human verification afterwards. So I think for now the ultimate verification is still the human and the diffusion of the result through the community. And we need better tools for doing this automatic filtering and verification. Like we have the AI reviewer that comes with the AI scientist, but you actually probably need some form of experiment execution for actually verifying everything. There is, for example, work by OpenAI on PaperBench and trying to go into that direction using LLM software verification. So I'm hopeful that we're going to figure this out in the next years.

是的，我认为一个关键转折点是当 AI 发现一种新的 Transformer 架构或类似重大成果，而我们都在使用它。我担心的是，像谷歌这样拥有足够算力的公司，他们会运行 AI Scientist，并拥有许多这样的发现，这就是为什么拥有能够高效发现科学新事物的研究如此重要。

Yeah, and I think one of the Rubicon moments is when a new transformer architecture or something massive is discovered by AI and we're all using it. My worry is that probably folks like Google who have enough compute power, they're going to be running AI scientists and they're going to own many of these discoveries, which is why it's so important to have work which can efficiently discover new things in science.

拥有公开可用的工作也很重要。我认为通过 AI Scientist 和 Shinka，我们真正试图确保我们可以运用我们所有人的集体智慧来塑造未来的可能面貌。

And it's important to have work that's openly available. I think with the AI scientist and Shinka we're really trying to make sure that we can apply the collective intelligence of all of us to shape how this might look in the future.

太棒了。Rob，非常高兴你能来参加节目。顺便说一句，Sakana 正在招聘优秀的工程师。所以如果这听起来像是一个绝佳的机会，请联系 Rob 和团队。我相信你正在研究一些令人兴奋的新东西。

Amazing. Well, Rob, this has been so fantastic to have you on the show. Sakana is hiring amazing engineers by the way. So if this sounds like an amazing opportunity, get in touch with Rob and the guys. And I trust you're working on some exciting new things that are coming up.

是的，我希望将来能再次和你讨论其中的一些内容。

Yes, and I hope to be able to talk to you in the future again about some of this.

当然。Rob，非常感谢你的到来。

Absolutely. Rob, thank you so much for coming on.

非常感谢，Tim。

Thank you so much, Tim.