我召集了几位我所认识的最有洞见的 AI 研究者,他们都在偏开放型的公司工作,因为我想了解前沿究竟在发生什么、接下来又会怎样。
00:00:00 – 为反对 RSI 的观点做最强论证
今天,我和三位 AI 研究者朋友聊天,每次交谈我都能从他们身上学到很多。他们恰好也在一些偏开放型的实验室和公司工作,所以你们可以真正地公开表态。和我一起的是 Beren Millidge,他是 Zyphra 的 CTO,该公司正在开发开源模型。John Schulman 是 Thinking Machines 的首席科学家,此前是 OpenAI 的联合创始人,并领导了催生 ChatGPT 的 RLHF 工作。而 Charlie O’Neill 是 Baseten 的模型训练负责人。
我的第一个问题是:如果我们到了 2036 年,却没有数十亿个疯狂运转、彻底改变世界的 超级智能,那么最可能的原因是什么?排除外生的政治冲击、爆发战争,或者他们封禁 AI 之类的因素。2036 年没有成为一个疯狂的异星超级智能世界,最可能的技术原因是什么?
Beren Millidge
有一个经典现象,几乎就像莫拉维克悖论一样——我们总觉得 AI“如果它能做到这件事,那就太厉害了”。如果它能解出这些高难度数学题,如果它能在国际象棋中获胜,等等等等……然后它确实做到了这些事,影响却没那么大。显然,它有一定影响,但并非一切。
如果不知怎的这种趋势持续下去,始终没有真正的泛化火花出现,我认为这可能导致 AI 只是在人们放进基准测试或环境中的一切任务上极其擅长。但仍然存在某种持续的仿真到现实差距,不知怎地阻碍了一切。我认为这不太可能。实际上,我们已经在实践中从RL中看到了这种泛化。但如果元学习的泛化真的难到离谱,加上我们没能解决持续学习,而且它就是超级困难、不可能做到……那在这种情况下,这会是我的默认情景。
John Schulman
我同意这一点。人类目前相对于模型有很多优势。每当一个新模型问世,它就会在其中一些方面追赶上来。但你最终会受制于模型较弱的环节、判断力较差的地方,或者模型无法足够好地自我检查。
有一个不断重复的循环:新模型一出来,人们就被震撼到,说“就是它了,这就是 AGI”。但用了一阵子之后,大概过一个月左右,它就开始显得笨了。这个循环可能还会一直持续下去。很难预测它到底会重复多少次。
眼下,你并不会看到能力上的爆发式增长,因为你在做研究和工程时,仍然会遇到足够多的瓶颈。即便模型能写的代码远超一个人,它也不会让你生产力提升 100 倍。所以,也许这类循环的次数会比我们预想的更多。
Charlie O'Neill
对我来说,问题在于全局最优解——“一个你可以放在芯片上的学习器”——距离Transformer+ RL,也就是当前的配方,还有多远。人们设想,一旦你拥有一个在 AI 研究上比所有人类都更强的智能体,哪怕它只比所有人类强 0.1%,那么你能并行运行数十万甚至数百万个这样的智能体——而且随着芯片速度提升,你还能让它们运行得快得多——这一事实将压倒所有其他瓶颈。你最终会撞上这个非常快速的起飞在自我改进方面.
我可以想象,如果我们沿着当前这一范式继续走下去——基本上就是自注意力、RL、不断扩大 RL 环境的规模……想想 摩尔定律所经历的一切。我们曾有这样一条非常漂亮的直线,它维持了非常非常长的时间。但为了让 那条缩放定律持续下去,必须发生许多离散的断层式突破和创新。LLM也是如此。我们曾有过 预训练缩放定律,然后它开始遭遇收益递减。接着我们想出了 RL 并解决了这个问题,然后又得到了一条新的收益递减曲线去突破,使它继续看起来像一条向上延伸的直线。
所以,如果解决这个问题需要再一次那样的断层式突破,我不确定当前用这些 RL 环境来训练 LLM 的方法——即便是面向 RSI的 RL 环境——能否发现那个断层式突破。如果不能,我们很可能会撞上这条渐近曲线。
但你认为这次断层式突破会比 2012 年以来的任何一次都更难吗?
Charlie O'Neill
如果我们知道那个答案,我们差不多就有能力把它实现出来了。但也许我们应该区分两种不连续性:一种是给当前范式做增量、具有累积性的不连续——也就是在 RL 之外还有某种我们必须去发现的东西,而也许它们有能力沿着那条直线把点连起来——或者,还是那个问题:我们距离全局最优到底有多远?我们是否必须回过头去,把梯度下降和神经网络整体抛弃?我不认为,如果你继续把当前范式——也就是 LLM——不断扩展下去,无论你运行多少个 LLM,它都必然有能力发现那个东西,如果那个东西离得太远的话。
唯一真正的希望,其实在于深度学习就是无法把我们带到一个至少能够主导人类研究、人类开发的 AI,包括人类提出新范式等等的能力。或者,我也不知道,也许人类本来也永远不会发现下一种学习架构。但就人类最终有可能发现它而言……但看起来就是……如果你只看从 2012 年到现在所发生的进展,然后你就把它继续外推下去——我知道这一路只是靠海量算力扩展等等在推动——如果它最终没有达到能够主导人类的地步,至少是在研发方面,尤其是在接下来几年里,那反倒会显得很奇怪。
Ryan Greenblatt最近上了这个播客。他提出了一个观点,我很想听听你的看法。你可以想象,随着 AI 变得越来越有能力,它们能够在模拟环境中取得进展,而这些模拟会激励它们不仅把 AI 研发做得更好,也把整个科学做得更好。这是所有实验室都在瞄准的事情,也是许多初创公司在瞄准的事情。
另一个直觉泵是,如果你看看Elo 分数自国际象棋机器人80 年代以来的变化。Elo 随时间呈非常线性的增长。但当它们跨过人类水平区间时,出现了巨大的不连续性——随着 Elo 的这种线性增长,从人类专家总能战胜 AI,变成了人类专家永远无法战胜 AI。
我同意你的观点,即到目前为止,AI 能力就其最终对世界产生的经济影响而言,还没有那么大的分量。但这是因为它们相对于人类的 Elo 正在缓慢上升。
Beren Millidge
我同意,那会非常令人意外。唯一不会发生这种情况的可能,正如你所说,就是它在某个节点之前不知何故趋于渐近。因为在我看来,我们距离开始超越人类 Elo 分数已经相当接近了。所以我们需要在那之前趋于渐近。在我看来,这是在你提出的那种情景下——不知何故我们到了 2035 年,一切都还正常——这种情况得以发生的唯一途径。另一种唯一的可能是对 AI 出台了某种剧烈的监管。实际上,我认为这才是这一情景最有可能发生的方式,而不是技术层面的原因。
Charlie O'Neill
我认为研究有不同的类型。有一种是 autoresearch 风格的研究,目标已经被非常清晰地指定好了,你要做的就是优化那个目标。我想大家脑海中的图景是,如果我们继续沿着这条路走下去——让预训练损失不断下降、让我们的环境所携带的奖励不断上升——那就会带来进步。
但也许 Ryan 所谈的是那种开放程度高得多的科学,而范式转变恰恰需要这种科学——在这种科学里,我们无法指定目标,AI 也绝对无法指定那个目标。对于这其中的任何一件事,我们在如何指定目标上都必须极其、极其谨慎。
也许你的观点是,自 2012 年以来所发生的那些突破,其本质在于我们发现了……在 2012 年,人们并没有说……我这是假设,我不清楚,你们当时在场。或者至少 John,你当时在场。但我没有。
Charlie O'Neill
我当时还在上小学。
其实,John,我很好奇你那历经岁月的智慧,或者说当年身处一线的智慧。想必,一个重大的突破就是意识到 next token prediction 才是……在 2014 年,你不会想到 nanoGPT 的 speed run 才是值得去优化的东西。但既然我们已经走到了这个新范式,你会想到要在这个上面做 speed run,并让 AI 在这方面变得非常擅长。
但也许还有一个 AI 们不会预料到的、更靠内的循环可以优化。还有一个关于收入之类的更靠外的循环,它最终应该会很强,但那是一个非常缓慢的外层循环。
John Schulman
事实上,我记得在 OpenAI 早期的时候,我就有一种直觉:仅仅最小化对数损失并不能让你通向智能。因为那些重要的部分在损失中所占的比例太小了,会被噪声淹没。所以仅仅用下一 token 预测来训练一个语言模型,并不能学到你希望它学到那些有趣的东西。我们需要设计更好的目标函数,把更多的权重放在重要的东西上。
你可以为此提出各种各样的论点。你可以说:“哦,人类大概并不会学会对我们环境中的一切建模。大多数人无法对他们看过的某个场景做出照片级的复现。所以我们肯定需要一个更好的目标函数。”但后来事实证明,它就这样照样奏效了。
正如你指出的,即便在当前的 AI 研究中,后训练基准测试之类的内层循环,也不一定能转化为用户喜欢的东西。
John Schulman
哦,是的。整个领域在很大程度上依赖于泛化,而你很难预测什么时候会获得泛化,或者什么时候会获得某种分布外泛化。我们知道,如果你在你关心的任务上进行训练,你会表现得更好。但最重要的进展往往是那些我们本没有理由期待的泛化类型。
例如,仅仅通过在这个非常朴素的下一 token 预测目标上进行预训练,就能泛化到各种需要以某种深层方式理解输入的任务,或者从预训练中学到某种非常罕见且代表性不足的技能。然后还有从这些可验证任务到不太可验证任务的泛化,这也是一种先验上没有理由期待的泛化类型。
这是一个有趣的问题,因为对于为什么你会迅速看到某种奇点——甚至不需要把 AI 进步的投入扩大到不仅仅是 AI 劳动力——一个直觉上的推想是:在你运行每一个七位数成本的实验之前,你会花费等量的算力在 AI 劳动力上。所以你只需用你们这些人的自动化版本,花一个世纪的时间思考什么是最优的实验,做小规模的消融实验,发展出字面意义上一个世纪份量的理论,甚至可以追溯到深度学习之前。
在你决定运行什么实验之前,你正在做极其优化的实验设置。然后实验结束后你再花一个世纪的时间思考,分析发生了什么以及下一个要运行的实验是什么。
John Schulman
如果你足够深入地思考,你或许事先就能预料到其中一些事情。可能存在某种非常巧妙的方法,让你做一个较小规模的实验,从而构建出一套理论,再将其推广到大规模实验中。所以我认为,我们在研究能做到多好这件事上,还远未触及天花板。
我可以想象这样一个未来:AI 承担大量的分析和理论构建工作,投入的算力与你在实验本身上投入的算力相当,进行各种分析,并围绕我们目前所观察到的东西构建理论。
Charlie O’Neill
我认为,当目标被明确界定的时候,确实有一些非常具体的例子。思考所能做的,只是根据你形成先验之后所获得的那些比特,更新你的后验。你无法仅凭思考获得任何新的比特。但当目标被明确界定、并且有这些数据摆在那里时,我想我们目前所处的范式将会迎来一次巨大的提速。
一个很好的例子是,如果你让 AI 去思考Kaplan 缩放定律。在这个时间点上,AI 会注意到:“哦,他们只是取了这些中间检查点,却没有考虑退火,所以这是错的。”这一点本可以在几年前就被发现。仅凭 AI 的这一观察,我们就能省下一到两年的进展。
再次强调,一旦目标被明确指定——也就是降低预训练损失之类的——就有很多很多很好的例子表明,只要你再多想一想,就能大幅削减你已经做的那些事情。比如 muP,以及学习率如何随模型规模缩放,并意识到模型宽度在其中也很重要。我觉得你确实可以反推出很多这类东西,砍掉大量唾手可得的低垂果实。我可以想象,如果我们的目标仅仅是"最大化我们当前所优化的目标",那就能带来 10 倍的加速。
但我看不出这如何能推广到一开始就提出正确目标这件事上。光是思考,未必就能让你一开始就得到正确的目标。
Beren Millidge
我认为,对于当前 AI 能否实现任何形式的极快速 RSI 来说,这确实是关键问题。AI 在学会自己的目标方面能泛化到什么程度?要拥有任何形式的自我驱动自动化循环,我们需要 AI 提出目标、优化目标、搞清楚这一点、提出新目标,并且让这一切在很长很长一段时间内都不会在任何环节失控。
回到莫拉维克悖论,可能存在这样一种莫拉维克悖论的情形:我们认为这种自主性和自我封装——也就是我们可以自己思考该做什么,然后去执行,并形成这样一个循环——超级简单,因为我们一直都在这么做。显然,进化需要创造出能够长时间独立生存的生物。而这可能恰恰是某种出于某种原因对 AI 来说极其困难的事情,就像运动控制之类的事情极其困难,而数学虽然对我们来说超级难,对 AI 却超级简单一样。
但时间跨度不断增加不就说明那是——
Beren Millidge
对,没错。这是另一种可能性,但我同意,目前没有明显的证据支持这一点。事实上,我们的智能体现在已经超级持久,而且做到这一点相当容易,这某种程度上就是反对这一点的证据。但如果这件事确实很难,那这可能就是我们不能立刻迎来起飞的原因之一。
如果你回顾从 2012 年到现在——或者也许从你开始做研究到现在——在那之后发生的所有创新中,包括纯工程性的,包括纯概念性的,哪一部分看起来像是 AI 完全自动化 AI 研发之前人类最后还需要做的事情?
Beren Millidge
大概就是不断地提出正确的问题。即便你能让 AI 去做任何实验,你仍然需要决定该做哪些实验。目前我认为,与编写实验代码相比,AI 在这方面还不太擅长。每当我们谈论研究时,它们会提出一堆杂七杂八的东西,都是非常非常微小的步骤。
Charlie O'Neill
又或者,甚至从 DeepMind那种"我们要通过学会以超人水平玩游戏来解决智能问题"的思路,转变到像 Radford这样一位普通研究员的想法——"我就试着去预测极其广泛数据中的下一个 token"……即便 Radford 已经发现了这一点,人们也花了一段时间才决定将其规模化,因为我们必须提出缩放定律这一概念,以及你可以非常可靠地预测这些事情这一事实。
John Schulman
我想说,人类最后的工作,或者说持续时间最长的人类角色,是定义目标并决定我们真正想要什么。沿着这个思路,像决定 AI 助手应该如何表现,或者什么才叫有帮助,又或者当我们做基于人类反馈的强化学习时目标是什么,就是这类事情。再往后,定义 宪法和 模型规范又是另一件事。即便 AI 能完成所有技术性工作,我们仍然需要做大量这类事情,并决定我们真正想要什么。
对齐是最后的工作。
John Schulman
对齐在某种意义上就是答案。但对齐本身可以分解为目标的规格说明,也就是弄清楚正确的目标应该是什么,然后真正实现或优化你所定义的目标。我认为第一件事短期内不会消失。
如果我考虑一个后训练团队,以及为什么你需要很多人在这个团队里,那只是因为有很多不同的领域,你必须弄清楚模型应该如何表现。要把整件事自动化会非常困难,因为总得有人去思考模型在这个领域应该如何表现。
00:18:39 – 是什么在推动中国实验室的进展
为什么模型提供商没有出现大规模整合,这背后的原因是什么?这里有太多因素都指向集中化。如果你把时间拉长到数年来看,会有什么东西阻止这种趋势吗?
John Schulman
我认为知识蒸馏是抵抗集中化力量的主要因素。基本上任何可以通过 RL 学到的东西都可以非常容易地被蒸馏,因为它只有很少的比特。它是你可以从少量数据中学到的东西。如果你能从模型中获取展示某种行为的轨迹,你就能轻松地蒸馏它。我认为知识蒸馏是抵抗集中化的因素之一。
还有一种可能性是会出现公司专属模型,即可以从部署中学习,让一家公司持续改进自己的模型。这样的系统可以由当前的模型提供商寡头或某个目前规模较小的公司来提供。但我认为这会稍微改变格局。
Beren Millidge
我还想指出,持续学习说实话并不能阻止知识蒸馏。即使你的模型每天都在改进,人们也可以每天对它进行蒸馏。这个循环可以以相同的节奏运转。
这说得通。所以复制模型行为……我猜你需要自己知道什么样的提示词分布是正确的,才能获得相关的模型行为?
John Schulman
哦,是的。仅用监督学习来做知识蒸馏的话,提示词分布极其重要。蒸馏一个模型是非常不容易的,即便你能完全访问它、拿到思维链和一切。要从中蒸馏出所有有用的能力并非易事,因为你需要用某些东西来提示模型。你需要用真实的提示词来提示它。你需要有一个非常广泛的真实提示词分布。
最近浮现的一件事是,一些中国公司很可能在使用这些路由服务,这些服务旨在让中国的用户能够使用美国的前沿模型,而这些模型原本在中国是被封锁的。有各种路由或代理服务让中国的用户能够使用这些模型,主要用于编程。这些路由服务正在收集并出售部分数据。这对于知识蒸馏来说是非常有用的数据集,因为它提供了完美的提示词分布。
Beren Millidge
我认为这是 AI 能帮上大忙的事情之一。如果你真的去看前沿流水线,或者中国模型在论文中实际披露的内容,他们会从某处获取种子提示词,来源是人工和这类数据的某种组合。然后他们利用现有的模型或其他前沿模型,从这些种子提示词合成出广泛的覆盖。你可以在很大程度上自动化提示词分布的收集和环境创建。随着模型变得更好,人类需要提供的比特越来越少。
但看起来你仍然受制于拥有一项有用户在使用它的服务。
Beren Millidge
不一定。那显然非常有帮助,但理论上,你只需要思考用户想要什么。
但关键在于,用户会说:“给我做一个像这样的应用。哦,那个不行。我其实想让你做这个新功能。不过其实,让我们退一步,做另一件事。”捕捉这整个轨迹才是重点。或者,如果你本来就能做到这一点,那你就直接有了 RSI。
Beren Millidge
归根结底,如果你有这样一个完全自动化的循环,那基本上就是 RSI。AI 在决定数据,它在决定训练。这就是那个循环。但这取决于你需要多少人类信息。在某个时候,如果你只是说:“我想要看起来像这样的轨迹”,你把这个提示给模型。模型就能给出一个相当不错的近似。
但如果你想做的是,“给我做一个非常出色的政治家”,然后它必须从头预判参议院大厅里的一场讨论会如何进行之类的呢?我只是觉得会有很多事情是——
Beren Millidge
讽刺的是,这件事对做知识蒸馏的人来说反而比前沿实验室更容易。做蒸馏的人只需要说:“我想要一个优秀的政客。”他们去找前沿模型。前沿模型已经知道怎么当一个优秀的政客,所以它直接生成那些轨迹就行了。而如果你真的想造出第一个能做到这一点的模型,你就得想办法获取政客每天做什么的数据,并据此把它建出来。说“我想要一个像这样的东西”,然后让 AI 生成十亿个变体,实际上要比从一开始就真正创造出“像这样的东西”容易得多。
Charlie O'Neill
我认为,基于这一观察,你其实可以做出一个非常具体的预测:中国的实验室拥有这些路由数据。最初引发这件事的是,我当时说:“难道不奇怪吗,Sonnet 5和Opus 5几乎是客观上比GLM-5.3和Kimi K3更差的模型,尽管它们不仅能获取知识蒸馏,还能获取logit 蒸馏来自Mythos?”反驳的观点是,提示词分布真的非常非常重要。你需要看到用户在实际做什么,这样才能把这些行为蒸馏进去。
我认为由此得出的预测是,前沿实验室如今在 RL 环境方面未必有多大优势,甚至可能根本没有。是的,用户分布对通用行为等确实重要,但衡量一项能力最好的标准,是你在前沿打造的那些极其、极其困难的 RL 环境。如果你作为 Anthropic 能接触到那些 RL 环境,又能使用 logit 蒸馏,结果却仍然做出了一个更差的模型,那也许——
那么真实世界的部署就比环境更重要了。这真的很有意思。
但他们一开始就必须在 Fable 或前沿模型中激励出这些能力。所以奇怪的是,他们居然没法用更小的模型之类的东西再次激励出这些能力。
Charlie O'Neill
也许我们正处在这个奇怪的恐怖谷里:试图过多地复制那个前沿模型,学生-教师差距,不管是什么,都太大了。人们已经就 Opus 提出过这一点。Opus 4.6 和 Opus 5 的区别在于,Opus 5 真的感觉像是有一个 AI 作为评判者在检查它所做的每一件可能的事。这就是它使用那么多 token 的原因。它会尝试思考所有这些东西,但它未必具备 Fable 那种 大模型气息,知道什么时候该停止这么做,或者什么时候是一条值得走的好路。
力所不及。
John Schulman
我想提出一个略有不同的假设。我会说,你可以创建的环境有几个不同的维度。其中一个是难度,另一个是真实度。
相对来说,创建大量高难度环境是比较容易的,这些环境涉及执行复杂得多的任务,或者做一些需要更多巧思的事情。你可以说这就是benchmaxxing分布,因为许多最知名的基准测试无非就是做一些非常难的、类似谜题的任务,而且容易验证。然后是真实度这个维度,你希望模型在真实的编程智能体场景中表现良好,那里有人类之间的多轮来回交互,也有多个目标。
那些首次塑造模型行为的实验室需要在这两个方向上同时推进。要获得良好的模型行为,你需要真正在真实度维度上发力,并拥有评分标准或某种人类反馈来为你在那里使用的奖励函数提供信息。但如果你试图以朴素的方式做知识蒸馏,最终你只会在 benchmaxxing 分布上与教师模型对齐。如果你没有足够多的环境来真正锻炼模型在这些更棘手的真实场景中的能力,那么你就无法将这些能力带入你的学生模型。
我觉得可能正在发生的一件事是,大模型从棘手的窄任务到这些更真实的任务泛化得更好。如果你有一个非常好的真实提示词分布用于知识蒸馏,你可以很好地匹配大模型。但如果你只有这种易于验证的任务分布,那么你可以在所有基准测试上匹配大模型,但在更广泛的分布上表现更差。
这或许还能解释 Anthropic 一些较小模型的情况,比如 Sonnet 5,尽管很难准确预测他们对这些模型的后训练到底做了什么。也有可能他们一直在调整后训练技术栈,只是在其中一些模型上弄错了几处。我也不确定……他们把某个参数调得太高了,制造出了一些人们非常不喜欢的怪癖。后训练非常容易在某个方面搞砸,而且这种问题不会在基准测试中暴露出来。
Beren Millidge
还有一个非常基本的点:前沿 AI 实验室都是从大数据公司购买所有数据的。中国方面同样可以从数据公司买到相同的数据。
确实如此,对吧?
Beren Millidge
确实如此。没错。很多人对此感到不满,但如果他们拥有完全相同的数据并且能买到这些数据,他们也能进行蒸馏。这意味着跟上进度其实相当容易。
00:28:06 – 自动化 AI 研究员将如何训练
我还有一个问题是,第一批能够自动化 AI 研发的模型究竟会如何被训练出来。有一个简化版的做法,就是 Ryan 谈到的那种。你直接让 GPT-8 去构建 GPT-3 规模的模型,这些模型非常擅长内循环类型的挑战:击败需要持续学习的电子游戏,或者以最少的算力达到某个特定的损失值,等等。
但 John,我觉得你提出了一个很有意思的观点,那就是实际中可能并不会以那种方式发生。所以我很好奇,当你拥有真正能够自动化 AI 研发的 AI 时,它们大概率会是如何被训练的?
John Schulman
我们大概率会做一些组合:一方面从人类反馈中学习,以吸收研究人员的品味;另一方面创建大量涉及多步骤研究项目的练习环境。实践中,人们会以某种组合方式来做这两件事,并且在每一轮迭代中,修补上一轮中看起来问题最大的部分。研究人员会大量使用这些 AI,并注意到它们存在一些一贯的弱点。这些问题要么通过收集人类反馈来修补,要么通过创建环境来修补。
Charlie O'Neill
也许一个有用的思考方式是:我们会回滚多少代系,然后从那里开始让自我对弈。在极限情况下,你设想的是只给它们一块 GPU,也许还有神经网络之类的东西,然后说:“好,想办法训练一个模型来完成这些特定任务。”目前的工作方式是,我们一直推进到代系的最前沿,然后说:“好,这些是 Anthropic 在过去几个月里在他们的训练栈中发现的 bug。我们会把它们转化为环境。”你需要在前沿上训练并变得更强。所以你显然会锁定代系的所有先前历史。
但你可以想象一个世界,在那里你回滚到GRPO之前之类的状态。然后你有各种环境,试图让它发现用 RL 训练模型的最佳形式,然后也许你越回滚越远……但我认为我们的算力仍然会如此受限,以至于人们只会一直待在前沿,本质上就是把自上一个模型版本以来发现的 bug 和任何改进做 diff,然后把它们转化为训练环境。
这对于在模型代际之间获得非陈旧的、新的数据也非常有好处。再次强调,这基本上就是 AI 实验室内部的持续学习,通过环境和 RLHF 类的东西,把过去三个月的 AI 研究进展蒸馏回模型本身。
Charlie O'Neill
而它正在做知识蒸馏。这或许就是为什么我们中有些人觉得它渐近。你总是只想拿到最近三个月的进展。当然,这些进展有 AI 的贡献,但其中也仍然有人类在环。感觉就像你一直在一点点逼近人类研究者正在发现和能够做到的东西。
Beren Millidge
不过我要说的一点是,显然如果你只是基于轨迹做知识蒸馏,你永远无法超越它。但环境可以远远超出人类能做到的水平。设计一个没有人能解决的环境非常容易,但 AI 显然仍然可以尝试去解决它。那将是超越人类 AI 研究水平的路径。
Charlie O'Neill
就 RSI 而言,你有没有一个例子,说明是什么样的训练集?
Nanochat 速通,但比人类速通玩家做得还要快。
Beren Millidge
我觉得尤其是在 AI 研究中,定义目标非常容易。你可以说 loss 需要达到 1.3 之类的,而现在没有人类能做到。但这是一个极其可测量、可验证的任务。如果 AI 做到了,那就太好了。
或者,我也说不准,就是训练一个 1 亿参数的模型去打败《我的世界》。这也许太简单了,但打败一个复杂得多的游戏之类的。
Charlie O'Neill
1 亿参数的模型能打败《我的世界》,这难道不疯狂吗?我们居然还说这太简单了?想象一下五年前你要是这么说会怎样。
John Schulman
不过我想说,很多研究并不完全是这样,并不是在一个定义明确的目标上爬山。它更像是这样:我们对模型应该在某个方面变得更好有一种直觉。我们也有一些算法上的想法,似乎朝着这个方向走了一点点。于是我们就设计一个任务,用来展示这种方法是否出现了生命迹象,看看能不能看到这些迹象。如果看到了,我们就可以把这个任务逐步做成越来越接近真实的版本。
它更多是由直觉引导的。内层循环是为了检验那种直觉,而不是让测试本身带来洞见。
John Schulman
没错。你并不是直接针对你最终关心的目标或实际生产目标做优化。你是在稍微放宽你的目标。你是在说:“让我们在真实感这个维度上稍微放宽一点,找到一些真正有效的方法,然后再等这个方法稍微成熟一些之后,再回到真实感上来。”
还有一些研究更偏向于解释事物和建立理论。通常,我们在机器学习中并没有那么具有预测力的数学理论。但我们有很多关于正在发生什么的、更非正式的理论。
Beren Millidge
想必这些模型会在所有这些任务的某种组合上进行训练。有些会非常容易验证,有些会是LLM-as-a-judge,或者直接问人类:“这看起来合理吗?”人们希望这些都能泛化到那些难得多、更模糊、更含混的任务上。这在某种程度上很可能会实现。但它是否能泛化到足以让这个循环变成自我封闭、完全不需要人类参与,这一点还不清楚。
00:33:51 – 长时程 RL 会催生 AGI 吗?
也许先退一步看。在我看来,未来 AI 研究的计划大致是这样的。你觉得它会不会奏效,或者你是否认同这种描述,可以告诉我。这个赌注是:我们将在数百万个多样化环境、数百种不同领域中扩大 RLVR 训练的规模。最终从中涌现出来的,将是一个已经学会这些基本技能——或者还谈不上基本技能——的智能体,这些技能包括保持持续性、能够对信息和上下文进行分诊,最终还能实现与其他智能体协作的端到端优化,诸如此类。
这样的智能体在上下文内将具有很高的 样本效率——你研究过如何扩大 上下文学习的规模,使其可以任意长,但你仍在不断扩大它。而最终从中产生的东西,基本上会像一个 可即插即用的远程工作者,其工作周期可以是一周或一个月。
首先,你是否认同这就是各家实验室正在下的赌注?其次,这足够吗?基本上,就是在数据中心内的这些模拟环境中学会如何学习,然后被部署到现实世界,但并不是真正从现实世界部署中学习……只是从数据中心里的模拟环境中学习这些元技能。
Charlie O’Neill
我认为现在已经很难区分,各家实验室的努力中有多少是在直接推进 RSI,又有多少是在打造可以持续部署、以收集收入来资助下一轮大规模训练的通用智能模型。
对于后者,是的,这大概就是他们下的赌注。过去几年这些环境的发展方向非常清晰。Anthropic 的环境演进脉络就是一个非常清楚的例子。首先,我们只专注于编程,并且要在这个领域做到非常非常出色。然后,从编程所获得的任务时间跨度出发——编程大概是利用互联网上可用数据来构建环境的最低垂果实,再加上他们自己内部可以转化为环境的东西——接着我们再向外泛化。
接下来我们会推进到金融领域,在 RL 训练中光是 Excel 数据以及诸如此类的东西就多得不得了。然后是 PowerPoint。这就是整个工作经济的长尾。这套做法似乎效果非常好。许多其他实验室,甚至开源实验室,现在都已经意识到那才是正确的赌注。
但这意味着什么呢?当我请 Dario 上播客时,我问他的问题是:如果你真的期望模型具备像人一样在工作中学习的能力,那你为什么要试图把使用 PowerPoint 之类的所有这些技能都预先烘焙进去呢?你难道不应该期望模型在部署之后就能自己学会这些吗?
有多种不同的解释。一种解释是,我们预期模型很快就能达到那个水平,但它们目前还没到,那为什么不把这些技能摊销到模型训练中呢?另一种解释是,我们并不专注于让它真正擅长广泛部署的工作。我们只想让它真正擅长 RSI。这只是我们获取收入的一种方式,以便将资金重新投入到真正擅长 RSI 开发的模型中。然后一旦奇点到来,从另一端出来的东西将在所有对当前一代模型来说看似瓶颈的方面都表现得非常出色。
John,如果路径是这种泛化,我不知道你对如何理解这些模型中存在如此多任务特定知识有什么看法。
John Schulman
如果模型在上下文学习方面足够好,那么理论上你不需要用金融数据来训练它们。它们只需即时阅读所有书籍,就能弄清楚如何在相应的司法管辖区做所有事情。你可以说,你需要做大量这种领域特定的训练,只是为了让他们更高效。即使它们足够聪明,能够即时弄清楚这些,你可能仍然想做大量 RL,并把这些直觉全部烘焙进权重中,这样模型在运行时就会更高效。
在实践中,模型提供商确实似乎在逐个领域地推进,试图在价值最高的领域强化模型。我想说,这就是模型为何变得如此出色的答案之一。原因就在于模型提供商已经覆盖了大量高价值领域以及最常见的技能类型。
Beren Millidge
另一件事是,同时做这两件事的成本并不高。模型规模庞大。就其参数量而言,它们完全有能力学会一切。很可能存在某种迁移。即便金融并不具体,这些信息对 RSI 也很重要。仅仅是关于如何判断什么重要、如何培养品味、如何做长周期工作的通用元学习,就具有潜在的泛化能力。
世界上也没有那么多 RSI 数据。它很难生成,需要大量投入。所以如果你能用这些其他数据来分摊成本,你就能从中获得一些迁移。你已经有海量算力和海量参数空间,那为什么不在卖模型这个明显的直接商业意图之外,也顺便做这件事呢?
John Schulman
我还要补充一点,有一个问题是,当前这种 sim-to-real 的范式是否会永远占据主导地位。你看看现实世界中的任务是什么样的,然后你尝试创建一堆可以在数据中心里模拟的环境,并在这些环境上做 RL。显然,这种做法非常成功。但它也有很多弱点,因为很多东西就是很难模拟,尤其是当它们涉及与大量人类实时交互的时候。所以,sim-to-real 是否会永远是最主流的框架,这是存在疑问的。
Beren Millidge
我认为在样本效率还很低的时候,sim-to-real 必然是主导框架,因为现在你需要与人类进行成千上万次交互。没有人会坐在那里参与到 RL 训练的循环中。所以我们现在必须模拟这些,才能获得所需的样本。但显然,如果样本效率大幅提升,你就会预期从部署中学习会成为其中大得多的组成部分。
John Schulman
不过你还可以做其他一些事情。你可以进行离策略学习,这样你就可以把所有轨迹拿来,即使不重新模拟所有内容,也有可能从中学习到一些东西。
我想就这一点多问一些,因为很奇怪的是,你有 50% 的算力花在推理上,而这些推理并没有直接帮助模型变得更好。你预期数字心智最终会拥有的一个关键优势是,与人类只能获得 50 年的现实世界经验不同,一个模型将通过它的所有实例,在经济中各种与经济相关的工作中获得数百万年的部署经验。而目前,这些数据在真正有意义的意义上,并没有帮助模型变得更好。
看起来如此显而易见的是,最终模型应该能够从这些数据中学习。一旦它们做到了,你就会拥有某种几乎像是广泛部署的智能爆炸一样的东西,因为模型正在从所有这些部署实例中吸收如此多的信息。你预计这种蜂巢思维、疯狂的事情什么时候会开始发生?
Beren Millidge
我认为从广义上讲,在一个非常基础的层面上,这已经在发生了……就在下一代模型中。现在,你显然可以拿你的部署数据,把它放进未来模型的预训练或中期训练中,尤其是如果你对数据做了某种过滤、某种判断、标注或合成的话。
你认为这在多大程度上解释了代际之间的改进?
Beren Millidge
我觉得这能解释不少问题。我不知道那些实验室是否这么做,因为理论上他们声称不会用用户的数据来训练。但中国那边 100% 会这么做。他们确实从中获得了这种优势。这基本上就是知识蒸馏的本质。他们拿到模型,通过向模型发请求获取一部分部署数据,然后用这些数据训练下一代模型。他们当然也可以在自己的模型上这么做。完全没有理由不这么做。
Charlie O'Neill
我完全同意这一点。如果你把视角拉得足够远,这确实正在发生。我们所有人脑海中描绘的、持续学习这一终极目标,是单个模型获得经验、实时更新、当场从中学习的那种非常有机的、实时的循环。当你把视角拉近到那种细粒度时,很多事情就会出问题。但那些大型实验室正在这么做。闭源模型正在这么做。
也有一些早期迹象表明,使用开源模型的人正以快得多的节奏做这件事。一个很好的例子大概是 Composer。Harvey 在法律智能体上也在做同样的事。你手上有某种模型,你从自己针对特定任务所拥有的数据中获取非常具体的环境信息,还有用户抱怨的问题,以及你以某种方式从自己特定部署中提取出来的所有反馈。这类公司中的许多相比大型实验室拥有一个优势,那就是他们能把这些数据用得极其出色。
然后他们会创建环境。他们会对 Kimi K3 做一次大规模后训练。他们会去部署它。他们可能还会做一些在线学习,就像 Composer 做在线……基本上就是长时间做REINFORCE。
仍然有人在环中。仍然有人在说:“好,这些是我们关心的信号。下面是我们如何从手头的数据中创建环境。”它的节奏仍然比你想象的要长,但它确实在发生。最终这个循环会变得越来越快。
Composer 这件事很有意思,因为正是在这里,在 Cursor 中,人们会按 Tab 或者不按 Tab 来接受模型建议的下一个补全。基于这一点,每一天,Composer 都在更好地预测下一个——
Charlie O’Neill
那是旧的 Tab 模型。他们实际上不仅对 Tab 模型做了同样的事,对真正的生成模型也做了。
哦,我明白了。有意思。
Charlie O’Neill
这很难,因为当你做在线强化学习时,你并没有分组。你只有一个用户说了一件事,然后你得到一次 rollout。所以你面临一个很大的方差缩减问题。
Cursor 对此给出的模糊回答是:“我们有非常好的启发式方法,能够估计这个回答比平均水平好多少,或者比平均水平差多少。”然后他们会做一次大的 REINFORCE 更新。对于模型是否变差这个问题,他们的解决方案是:如果它在CursorBench上有所提升,他们就每五小时部署一次新模型。如果没有提升,他们就把那个版本扔掉。
John Schulman
我认为你最大的问题其实是,不知道自然数据的奖励函数应该是什么。如果你使用某种表面信号,比如他们是否接受了这次编辑,那可能会以某种方式被奖励黑客攻击。
00:45:24 – 仿真到现实的差距
但这难道不是仿真到现实问题中更大的一个问题吗?任务的时间跨度越长,就越难在数据中心内进行仿真。在我看来,即使在编程领域,我们也已经接近这样一个节点:不存在某种长达一年的编程任务,最终不需要你与客户沟通、与公司互动或与用户互动。
如果你想想我们希望 AI 能够胜任的种种事情,最终超级智能应该能够经营一家企业,或者创办一家新企业并让它盈利,或者在市场上做日内交易并盈利,或者打赢一场官司。这些事情都很难在数据中心里模拟。其中学习的一个固有部分就是与真实世界互动。
也许它们通过从仿真到现实的迁移,学会如何在这些事情上做得更好。但另一种可能是,你确实需要从这类互动中进行权重更新,才能在这些事情上变得更好。如果是这种情况——如果迁移不够强,你确实需要权重更新——那么模型样本效率相当低这一事实,或许就是一个更深层的问题。
我对这一点感到好奇的原因是,按默认情况来看,我看不出你怎么能在未来 10 年内不出现某种疯狂的递归式自我改进。但这可能不会发生的一个原因是,就权重更新的样本效率而言,模型似乎远远落后于人类。从一个人从出生到成年所看到的数据量,与一个模型从冷启动到完成训练所看到的数据量相比,它们很可能落后人类一百万倍。
这一切都说明,首先,仿真与极其长时程、真正复杂的现实事务之间是否会有良好的迁移——我们希望 AI 在真实世界中做的正是这些事?如果没有,那是否真的意味着这些模型样本效率的不足会反过来害了我们?
Charlie O’Neill
也许我可以这样划分两类任务——模型能够胜任的任务,以及模型仍将持续挣扎的任务——依据是任务是否具有累积性,还是说它处于一种非平稳分布之中,你必须不断学习并反复重新论证一堆东西。
一个累积性任务的例子可能是 RSI。从理论上讲,完全可以有一个不到一百万 token 的 Python 文件,它从零开始训练出一个具备递归自我改进能力的模型。你做出的每一项发现都是一条你坚守的底线。如果对于 RSI 而言,我们确实不需要发现新的注意力变体之类的东西,那么一旦你发现了注意力机制,一旦你发现了混合专家,一旦你发现了GRPO,你只需把它加入训练栈,它就在那里了。
一个很好的例子是5.6 Sol的训练、5.6 Terra,或者 OpenAI 告诉我们的它训练的那个版本。它不必回过头去发现注意力机制。它基本上就是调用一堆脚本,比如 pre-training.sh 和 post-training.sh,然后就这么做了。这就是一个累积性任务的例子。
我认为,现实世界——也是人们如此关注持续学习的原因——并不是一个真正累积性的任务。想象一下在一家律师事务所里,你有一个智能体担任法律助理。那是一个非常非平稳的分布。你必须能够在上下文中容纳该公司所有重要人物之间的所有关系,而这些关系也在不断变化。你还要掌握所有这些关于事情如何做、去哪里找信息等等的隐性方式。这不像 RSI 那样是一个干净的累积性任务示例。
我认为任务之间会出现这种分化。但如果实验室意识到这一点——而且他们确实相信 RSI 在某种意义上是累积性的,即我们不需要回头去发现某种全新的架构之类的——那么也许越来越多的精力和算力会集中在那上面,而不是其他任务上。
RSI 恰好比当律师助理更容易,这真是太不幸了。
John Schulman
我想说,今天的模型在很多不同方面都比人类弱。其中一些可能与特定情境下的样本效率有关。在某些情境下,模型非常样本高效,比如在上下文中学习。但可能存在某种中等长度的情境,模型在那里的样本效率较低,因为人类能以比模型更高效的方式完成某种权重更新。我认为在某些情境下样本效率较低,可能是模型弱点的来源之一。
但我认为还存在一些与此完全不同的其他弱点来源。例如,思维多样性低于人类,或者在某些类型的长期判断上表现不佳。我认为人们所说的很多“品味”,其实是某种从长远来看行之有效的行为,而且人们已经意识到它从长远来看是有效的。不全是,但品味的某些方面确实如此。尤其对于软件工程这类事情,我认为很多品味就是“什么样的系统在这个项目的长期运行中是可维护的、能良好运作的?”
模型存在各种各样的弱点,这些弱点与其他因素一起限制了 RSI。其中一些与样本效率有关,另一些则无关。
Charlie O'Neill
也许一个有趣的思想实验是这样的。假设你能够给一个模型一个一万亿 token 的上下文窗口,或者说足以容纳你在比如说 RLHF 之前所有经验所需的那么大的窗口。它在上下文窗口中拥有了所有这些经验,并且它在百万 token 时具有的样本效率和上下文学习能力保持不变。你认为品味问题就解决了吗?它能够做出与你相同的判断吗?还是说,除了一个具有相同样本效率的更长上下文窗口之外,还有什么根本性的东西缺失了?
John Schulman
它必须经过训练,才能从这些上下文中学习。要么它必须经过训练,学会从该上下文中做出正确的更新,要么它必须能够泛化。
Charlie O'Neill
所以你不认为你可以直接把所有东西都倒进去,你的整个人生,你的研究经验?
Beren Millidge
你仍然需要数据来在长上下文上训练它。即使你理论上能获得一万亿的上下文,你也需要一万亿长度的数据来训练它。现在你有 10k 的上下文,你不能直接倒进去一百万。
Charlie O'Neill
是的,我只是在问如果你有那个的话。
Beren Millidge
理论上,我认为,是的。这实际上归结为一个问题:品味从较短周期的经历中元学习的可学习性有多高。我觉得没有明显的理由说明它需要超级长,因为人类不知怎么就发展出了品味,而没有很多长的经历。我们活不到 10,000 岁。我们发展得相当快。
如果你想想哪怕是一个博士,一年级博士生和最后一年学生或博士后之间的差距,那大概是五年。他们总共可能只做过 10 到 30 个研究项目。但不知何故,他们从相对较短的一系列小事中很快就培养出了品味。从理论上讲,以这种方式培养品味是可能的。
AI 显然会拥有多得多的经验来培养品味、对其进行元学习。那么问题就在于,这能在多大程度上泛化到真正长周期的事情上,而我认为这一点目前确实尚未解决。我们不知道。
回到这个问题,最终应该会出现一种状态,即 AI 从每一次部署实例中学到大量东西。目前可以说,存在一个模糊的元过程,模型确实会通过部署得到改进。但我觉得这是一个非常弱的反馈回路。你认为这种非常快速的蜂巢式学习即将到来吗?如果是的话,它究竟是如何发生的?
John Schulman
我想说,我们能否得到一个从其所有部署经验中学习的蜂巢式智能,很大程度上取决于激励,而不是一个技术问题。公司不会愿意让模型提供商从他们的所有部署中学习,因为那可能只会削弱他们自身的业务优势。
Charlie O'Neill
我认为,这种经济性会带来压力,不一定是推动权重更新走向一个大型的公共共享模型,而是推动以模块形式插入。一个非常明显的例子就是 LoRA,但也可能是别的东西。
已经有很多工作试图把任意长度的上下文塞进一个固定大小里。这就是所有 线性注意力那类东西。还有 cartridges,本质上就是经过训练、被压得极其紧凑以容纳大量信息的 KV cache。这是另一个例子,说明公司可能愿意接受某种东西,只要它是被插入到模型里,而并不真正改变底层基础模型本身。
从你自己的数据中实时学习有很多不同的版本。后面那些其实对大实验室没什么帮助,因为它们只是这些模块而已。但我认为经济压力会迫使实验室先走上这条路,然后才能着手这个……
Beren Millidge
不过,是哪种经济压力呢?我觉得即便你有一堆 cartridges 或 LoRA 之类的,你仍然可以直接把所有这些轨迹丢进你下一代模型的预训练里。
Charlie O’Neill
是的。这可能是大型实验室所获得的一种更间接的学习形式。这对他们来说显然仍然非常有价值。但我无法想象这样一个世界:我们一开始就说,“我们要直接用我们获取到的所有精确数据来训练这一个大模型。”
Beren Millidge
不,我认为它肯定会经历多个阶段,因为这里假设存在一个不连续的突变事件,突然之间我们就持续地修复权重更新了。实际上,我认为更可能的情况是,这些“卡带”之类的东西让你能够在部署中实现专业化。然后你生成轨迹,把它放进你的模型里,三个月后你就得到一个在这方面更强的模型。你再次对它进行专业化,再次对它进行整合,然后最终我们会让这个飞跃越来越快。
不再是每三个月发布一个模型,而是每周、然后每天、然后每小时,到那个地步我们基本上就解决了这个问题。
Charlie O’Neill
我认为这也是一个很好的观点,因为你问到当前范式距离能够做到这一点还有多远。我们对此做过一些研究,其他人也做了大量研究。在非常大的规模下,当你滤除足够多的噪声并且拥有足够大的批次时,这种将数据放入中期训练并创建我们自己环境的外循环过程,确实能在某种持续学习的机制中发挥作用。
但问题在于,当你把视角拉得足够近、放到微观层面——我手上有一个模型,我想为某家律所之类再更新它一次,而且我想用相对少量的数据非常连续地做这件事——所有方法都有点失灵了。如果我只是用成功的轨迹对模型做 SFT,无论是离策略还是在策略,最终在这种高度迭代的状态下,当你做数百次这样的微更新时,你就会看到 灾难性遗忘。你会看到此前在基础模型之上学到的、更早阶段的信息被遗忘,也会看到通用能力的退化。
在策略蒸馏似乎能把这一时间线往后推一点,但它最终仍会屈服于同样的问题。RL 擅长把能力装进去,但不那么擅长把知识装进去——那种非常明确的知识,比如“啊,好吧,这个人在这家律所做这件事,而这是我们发现的一个非常具体的流程”。你得投入大量算力来构建合适的环境,才能用 RL 把知识灌进去。
你认为这里的根本问题——为什么你在这些其他技能上会变差,或者为什么会出现遗忘——从根本上说是容量问题,还是技术问题?
Charlie O’Neill
两者都有一点。我认为 SFT 甚至 on-policy 蒸馏都可能破坏性太强。RL 之所以这么好,是因为它对模型的改动非常非常小。有很多证据可以说明为什么会这样。它只是在那个非常非常小的损失谷里做微调,把模型调到正确的点上。但这也同时限制了你能用 RL 做什么,以及你实际上能对模型改变多少。
所以你是说,这件事之所以可能不会变成赢家通吃,是因为要把那么多信息蒸馏进基础模型实在太难了?
Charlie O'Neill
在不破坏某些东西的前提下,以迭代的方式进行。
Beren Millidge
把它蒸馏进另一个基础模型是容易的。我认为这主要是个技术问题。绝对不是没有容量。如果你有一个拥有所有这些数据的模型,然后你拿一个完全相同规模的模型,用中期训练中的所有内容从头预训练,它会更好。我认为这就是今天正在发生的大部分情况。
确实存在一个瓶颈,使我们无法一直训练同一个模型,而只能从旧模型中提取全部数据,从头训练一个新模型。这正是 Charlie 所说的:可塑性与灾难性遗忘的某种组合。如果你只是天真地在非平稳数据上训练,因为你在训练过程中不断加入新数据,这就扰乱了数据分布,于是旧的东西就被遗忘了。我们目前并没有真正好的方法来阻止这种情况发生。
所以也许在极限情况下,你只能受限于用所有这些新信息从头重新训练模型。
Beren Millidge
是的,这当然非常昂贵。从头训练一个模型是很昂贵的。
但你无论如何都会这么做。
Beren Millidge
不一定。也许最终,如果你有了持续学习,你就永远不会训练新模型。你只是有一个模型,它不断学习并扩展。
但可能存在某些深层技术原因,使得这非常困难。
Charlie O’Neill
这就是问题所在。我认为我们已经推迟了需要从零开始做多少工作。现在完全有可能拿预训练基座,在其之上做非常好的中期训练,某种程度上是持续进行的,再加上从训练后期不同检查点出发的一些 RL。这看起来更像是持续学习,但肯定不是那种拿最新模型、应用几个非常小的更新、然后迭代式地永不丢失任何东西的情况。
抱歉,但我有点困惑,因为这难道不就是训练过程中发生的事情吗?在后训练之类的阶段,你有一个已经经过了大量训练的模型,然后你蒸馏某个经过进一步 RL 的分叉。这难道不就是字面上发生的事情吗?
Charlie O'Neill
但它的规模仍然足够大,我认为,你在洗掉大量噪声,而且你不只是聚焦于一个分布,正如 Beren 所说,那才是问题所在。如果你只是聚焦于一个任务——
但在最终的状态下,你会做的是……有数十亿个已部署的实例。你同时从所有实例中学习,所以希望噪声能从中被洗掉一些。
Charlie O'Neill
也许在那个规模下,是的。
Beren Millidge
正如 Charlie 所说,你确实可以长期进行持续的中期训练,也可以回滚到某个检查点并给它新的中期训练数据。但与此同时,你不能无限期地这样做。如果你只是一直持续训练同一个基座,它会在某个点趋于渐近。你无法在那个基座上不断学习新东西。这就是为什么人们最终会训练新的基座。否则你就会永远对同一个基座进行中期训练。
01:00:33 – 数据能解释多少进展?
现在让我们谈谈数据。我普遍感兴趣的一个问题是,AI 的进展有多少仅仅是由数据进展所解释的。这并不意味着它必然难以自动化,但那是另一个问题。是否存在某种数据分布,如果你用当前架构在其上训练,就会产生一个在所有领域都完全碾压人类专家的超级智能?
Charlie O’Neill
我们说的是预训练加后训练数据,还是也包括环境?我认为它的存在是显而易见的。问题只在于我们能否创造出正确的环境来实现它。
Beren Millidge
在最简单的情况下,我们可以直接训练它输出那个训练出真正超级智能的 Python 文件。只要把它记在权重里就行了。
Charlie O’Neill
是的,或许可以构建出一套 RL 环境阶梯,使得你最终能得到一个至少与人类研究员同等优秀的 AI 研究员。但每向上攀登一级,所需的投入大致呈指数级增长。这两者就是你必须权衡的因素,决定了我们多快能到达最后那一级——也就是 AI 表现更优的那一级。我认为这一点相当清楚。
我们在 RL 环境的创建上仍处于相对早期的阶段。为了构建出好的环境,我们利用了许多不对称性。我们之前讨论过的一个不对称性是:有些环境往回走比往前走更容易。我的意思是,定义一个复杂的数据生成过程非常容易,而这个潜在变量是你对模型隐藏的。你可以生成任意复杂的环境,而模型必须花费大量不可约的 token 开销和不可约的工作量,才能弄清楚那个数据生成过程究竟是什么。
在从真实世界注入信息方面也存在不对称性。Anthropic 通过数万人类与 LLM 的协作发现了一个 bug,并将其转化为一个非常非常精巧的环境,单个 LLM 理论上可以在几百万 token 内找到它。我们正在挑选利用所有这些不对称性,并指望这种任务跨度上的泛化能力。
但我认为,它在某个时点终将遭遇收益递减——首先是创建这些环境本身的难度所带来的收益递减,也就是设计出这些环境,因为你未必能简单地拥有那种"倒退比前进更容易"的流程。你实际上必须坐下来,构建出某种对人类而言时间跨度足够长的东西,而这将是一项极其复杂的创建任务。此外,智能体实际执行这些任务时还会面临算力和时间的瓶颈。我认为你将会开始看到这条曲线趋于平缓。
John Schulman
我看到一些关于有人微调Talkie 模型的内容,这个模型只用了截至 1930 年的数据训练,却用现代编程智能体的数据做了微调。它的表现超过了Claude 3 Opus在SWE-bench上的成绩。所以这个对代码一无所知的模型,只需用适量数据微调,就能作为编程智能体表现得比这个大得多的预训练模型更好,这相当疯狂。它某种程度上说明,一旦你有了正确专家行为的范例,把它复制到一个相对较弱的模型里其实出奇地容易。
Charlie O'Neill
但有一个反例是最近的一篇论文,他们把模型训练到五年级数学的水平,还包括小学英语之类的内容,所以它已经算是一个不错的语言模型了。他们尝试用 RL 让它做高中后期和大学水平的数学。差距实在太大了。他们根本没法让它往上爬。但如果你一级一级地来,先七年级数学,再八年级数学……以此类推,你显然能爬到十二年级。说到底,问题就在于这些阶梯上每一级之间的间距有多大,以及造出这些阶梯有多难。
Beren Millidge
这又回到了 RL 信号的问题上。RL 目前在探索方面还不太行。如果模型在 128 次 rollout 里都做不出来,那它极不可能获得推进下去的梯度信号。这就是为什么在 RL 中我们需要课程,而在预训练中我们不需要,因为这对预训练来说根本不是问题。
Charlie O’Neill
同样,预训练数据和后训练数据是不同的。我想随着我们继续往前走,人类的参与会越来越少,但这并不改变一个事实:你能从真实世界中提取多少信号,始终是一个瓶颈。
世界上存在大量信号,这没错。有人在处理电子表格任务,有人在处理法律任务,诸如此类。但就当前模型所处的性能前沿而言,世界上究竟有多少比特是真正与提升模型能力相关的?有多少新的数学问题正在被解决,而这些问题恰恰超出了当前模型的能力范围或理解力?有多少新的编程问题正在被创造或解决,而这些问题超出了当前模型的能力范围?
我认为这就是收益递减出现的原因,因为即便整个世界作为一个整体,也无法提供那些能把你推入下一个能力盆地所需的有用比特。
Beren Millidge
我完全同意这一点。这本质上是一个信号来自哪里的问题。在预训练中,信号已经存在于Common Crawl中了。对于预训练中你关心的那些任务而言,问题根本不在于获取信号,而在于过滤掉所有存在的噪声。而这是一个相当可自动化的过程。
但随着模型变得更强,随着我们进入中期训练和后训练阶段,信号在我们已有的原始数据中根本不存在于任何地方。再多的过滤也无法得到它。Common Crawl 里并没有什么隐藏的千禧年大奖难题证明,等着我们去过滤直到发现它。
到那个时候,你必须通过其他方式获取比特,要么直接从人类那里获取,请他们写出自己的推理过程,要么创建环境,由人类来决定应该创建什么环境以及这些环境的目标是什么,或者基于部署中已有的人类数据进行某种训练。你必须从某个地方获取比特。
有一个问题是,预训练中有多少进展是由数据驱动的。我和Jerry Han一起做了这项研究,他是普林斯顿的学生,我们把 2019 年至今的所有训练配方与 2019 年至今的所有数据集进行了两两配对训练。你把GPT-2放在最新的数据集上训练,比如Ultra-FineWeb。你把Delphi——最新的开源训练配方——放在the Pile或一些旧数据集上训练。你把整个网格都跑一遍。然后看,要达到某个能力水平,在这个网格中,能少用多少算力?
你会发现,数据似乎能解释大约 12.0 倍的算力效率提升,而架构改进在非常小的规模上大约能解释 3.7 倍的算力效率提升。如果这在大规模上也成立——即预训练算力效率提升的大部分来自更好的数据——那这种趋势还能持续多久?你能不断加大数据过滤力度、构建越来越多的合成数据吗?你对这类预训练进展还能持续多久有判断吗?
Charlie O'Neill
我的先验判断是,低垂的果实某种程度上已经被摘完了。我们得到了互联网这一大块资源,而互联网本身未必以同样的速度增长。互联网上所有有用的内容也并非以同样的速度增长。我们大概还剩下一堆 0.1% 的损失下降空间,但肯定没有目前已发生的那么多。
但同样非常有意思的是,你发现两者合计带来了 33 倍的改进。我记得是 Epoch 或某个人估算出自 2019 年以来每年 3 倍,这意味着大约是 37,超过 2,000 倍的改进。那么缺失的那 100 倍左右是从哪里来的?这大概能给你一个很好的信号,说明其中有多少来自后训练。
我认为解释只能是,很多计算效率的提升是依赖于规模的,而我们是从极小的规模起步的。这就引出了一个问题:数据计算效率的提升和算法计算效率的提升,哪一个对规模的依赖性更强。我不知道你对此是否有先验判断。我们只是没有足够的算力来研究这个问题。
Charlie O'Neill
单纯从直觉上、理论上讲,架构对规模的依赖性是比较为人所知的,你可以拟合出一条直线。而对于如何把预训练加上后训练数据以及中训练数据结合起来,我完全不知道该怎么去做。
Beren Millidge
说来有趣,我觉得数据实际上随着规模增长反而更加重要。我觉得架构在某种程度上是一次性的东西。
只说效率提升了 X% 其实是有些误导性的,因为架构的作用是让你进入一个全新的定性阶段,这是旧架构无法企及的。在那个阶段里,数据显然是决定性的因素。
但如果我们连 GQA 都没有,如果我们一直做全注意力,那么做百万上下文会贵得离谱。正因如此,我们永远无法利用真正处于百万上下文的数据,也就无法获得这些能力。尽管如果你只是简单地看“这在 2K 上下文下表现如何”,此时架构并没有解锁任何新东西,那么数据看起来就会比它实际上重要得多。我不确定这些东西真的只是以这种方式相乘叠加的。
我明白了。那你怎么看数据的规模依赖性?
Beren Millidge
关于规模依赖性,我认为我们现在拥有的很多中期训练和后期训练数据实际上会随着规模增长而变得更好,因为其中很多——那些超长上下文视野的环境类数据——确实需要大模型才能加以利用。如果你试图用 1 亿参数的模型在 SWE-bench 轨迹上训练,它不会有什么进展。它不会展现出你用真正合理规模的模型训练时所能获得的那种提升。
Charlie O’Neill
现在这也很困难,因为很多架构上的变化——比如你看 Kimi,或者 DeepSeek——他们做这些架构修改时,不仅仅考虑降低预训练损失,还考虑模型在真实世界中会如何被使用。推理效率,比如 DeepSeek 模型中某种形式的压缩注意力,未必是围绕某种根本性的权衡改进来设计的。它只是,“好吧,我们在考虑模型将如何被使用。”
我很好奇的一个问题是,为了理解未来,随着我们进入一个更偏 RL 的范式,参数扩展会如何发展。你可以看看开源架构,看看参数扩展得有多快。也许前沿开源模型大致每年翻倍。以至于即便前沿闭源模型有 100B 或 200B 的 active parameters,你认为这还会逐年翻倍吗?
或者,既然我们现在处于一个 RL 范式,你还想在 rollouts 上节省算力……另外,也许存在某种阈值效应:一旦你有了足够的容量,到那时任意增加参数就没那么重要了。你们对 2030 年一个前沿模型会有多少 active parameters 有概念吗?
Charlie O’Neill
我认为,未来几年,由于我们如此专注于为强化学习做越来越长时程的 rollout,而推理效率在其中至关重要,模型在这方面能力似乎未必已经饱和。瓶颈仍然在于环境。所以我们可能会看到一点平台期。
我有一种感觉,Mythos 和 GPT 系列模型远小于人们谈论的 10 万亿参数这个量级。哪怕只是拿它们和开源模型做个朴素的对比,你大概也能反推出这个结论。
大概未来几年,我不会设想参数量会有巨大增长。但话说回来,这里要权衡的东西太多了。你根据自己有多少预训练数据来决定模型规模,然后再考虑你要训练的强化学习环境的难度。理想情况下,你希望达到那个最优点:在你手上最难的环境里也能拿到不错的 pass@1 之类的成绩。做一个更大的模型去通过那些环境就没意义了,因为那样你只是在支付远超所需的推理成本。
所以这在很大程度上取决于 Mercor 和内部团队能以多快的速度,把他们用于训练的强化学习环境的复杂度扩展上去。
John Schulman
我预计模型会继续变大,只是因为人们在扩大算力规模,GPU 也在变得更大。但具体会变大多少,在一定程度上取决于那些并不显而易见的扩展规律。
有一点是,我认为数据效率将比人们所用具体架构的计算效率成为更大的驱动因素,因为我们现在正进入高质量预训练数据日益枯竭的阶段。这可能会影响你想让模型做到多稀疏。
我还认为我们对稀疏性的理解还不够深入。参数量与激活参数量是不同的资源。稀疏性确实有所增加,但并不清楚它是否会无限制地持续增加。可能存在某种最佳平衡点。
有一种观点认为,稀疏性应该会让数据效率变差,因为你可能不得不在多个专家上学习同样的东西,不过这一点尚有争议。我认为我们对缩放定律还没有足够好的理论,无法真正理解稀疏性为什么有帮助、能帮多少,以及它是否会在某个稀疏程度下趋于平台期。
抱歉,你能具体说明一下数据效率对参数量究竟意味着什么吗?听起来你会说应该减少稀疏性,但对参数缩放还有哪些其他影响?
John Schulman
就是说,在缩放定律下,你并不是在试图优化计算效率。你在架构上有各种可以做的选择。每一种都会给你带来不同的缩放定律。传统上,你会基于算力去看某种包络线。你会看性能与算力的关系,然后取最佳模型的包络线。
但如果我们基于数据来做这个决策——我们假设可以投入大量算力,于是数据就取代算力成为我们的 x 轴——那么我们只会得到一组不同的最优解,或者说一组位于该前沿上的不同模型。
Charlie O’Neill
我也不认为过去几年我们每年都把模型规模翻了一番。人们训练 1 万亿参数模型至少已经有好几年了。甚至还有一个开源的叫Falcon. Liam来自Periodic Labs,我想,昨天在 Twitter 上发帖谈到一个早期实验是在训练一个非常非常稀疏的 1 万亿参数模型。
John Schulman
这就是他们在 OpenAI 之前于 Google 所做的,Switch Transformer。
Charlie O’Neill
它在知识方面非常非常出色,但在推理方面却很糟糕,因为它过于稀疏。感觉我们已经在 1000 亿到 2 万亿参数这个区间里摸索了有一阵子了。它显然不是那种漂亮的线性增长。
Beren Millidge
我觉得有两点。正如 Charlie 所说,推理效率对 RL rollout 至关重要。这会大幅压低激活参数量。
我认为总参数量在很大程度上也取决于硬件。要真正能够服务数万亿参数的模型,你确实需要非常高的内存带宽和VRAM容量。目前,人们仍在大量使用H100之类的东西。随着所有人转向GB,然后再转向Vera Rubin,我们将获得更多扩展能力,真正能够在更大规模上服务并进行大规模 RL 推理。
数据这个问题我觉得很有意思,因为从直觉上看,更大的模型在实际数据点上的样本效率要高得多。即使你没有让模型饱和,做大仍然更好,因为更大的模型泛化能力更强,在相同数据量下能达到更好的 loss。目前我认为我们有很多数据,这不是约束条件。算力才是。所以我们有了推理效率很高的小模型。但如果算力不再是瓶颈,可能又会回到更大的模型上——这些模型处于未饱和状态,但因为规模大得多而具备这种泛化能力。
如果你只看基本的Chinchilla 缩放定律,只是把参数量最大化,实际上达到相同 loss 所需的数据量减少得非常少。如果你把参数量推到无穷大,我认为所需的数据量下降不到 10 倍,这仅仅是因为幂律的本质。
Beren Millidge
但我们现在已经处在 Chinchilla 定律中数据过多的那一侧了。眼下按照 Chinchilla 定律,我们对模型是过度训练的。所以,随着数据逐渐耗尽,我们很容易回到 Chinchilla 最优点,甚至稍微偏向模型训练不足的那一侧。
Charlie O'Neill
但可以肯定的是,即便有这些新上线的芯片,未来几年我们在算力上仍会严重受限,所以情况未必会如此。
Beren Millidge
这实际上取决于你训练算力与推理算力的比例。如果你严重受限于数据而非算力,就应该把模型做大。如果你严重受限于算力,就应该始终把模型做小。你也可以使用计算机生成的合成数据,所以这是那种极难预测的事情之一。
John Schulman
我认为,人们最初之所以花了那么久才搞清楚缩放定律,部分原因在于:如果你没有把所有这些事情都做对,就不会得到那么干净利落的关系。图表上那些漂亮的直线,掩盖了大量的复杂性——你必须确保以正确的方式缩放每一个超参数,或者以某种可缩放的方式参数化你的优化器,使得你在改变模型规模时不必更改超参数。
Charlie O'Neill
Bug 也有它们自己干净的缩放定律。比如 Kaplan忘了 余弦退火那回事,甚至只是没有考虑 嵌入向量参数,我觉得都是如此。这搞乱了较小模型上的估计,因为嵌入向量参数在模型中占了相当可观的比例。
01:18:03 – 为什么 RL 效果这么好?
稍微谈谈 RL。一年前,很多人都在提出这样的论点:RL 在模型缩放方面不会特别成功。John,你写了一篇研究论文,其中指出模型在 RL 时每个 episode 只学到一个 bit。它们学到的是:“我答对了还是答错了?”然后我今年早些时候写了一些博客文章,我当时说:“情况比那还要糟”,因为当通过率很低、模型极不可能答对时,它从一次 RL episode 中几乎什么都学不到。
但看看今天的模型,它们看起来相当聪明。这似乎是扩大 RL 规模的结果。Beren,你几周前写过一篇文章,试图解释到底发生了什么。为什么 RL 的成功超出了人们天真的预期?
Beren Millidge
我认为 RL 的成功归结于一系列不同的因素。首先,有一点被略微低估了,那就是中期训练。我们所看到的 RL 的成功,很大一部分实际上来自非常非常好的中期训练数据,也就是我们本质上在做预训练,但用的是合成推理数据以及那种能让模型为 RL 热启动的环境。这通常能把模型带到距离最终 RL checkpoint 大约 80% 的位置。
然后 RL 在此基础上所做的,本质上就是调整策略。这也是为什么它不需要像你天真想象的那样多的 bit。它不必从零开始学习所有这些行为。它只需要从这些 episode 中获取几个 bit,而你确实能得到这些 bit。我在博客中还指出的另一点是,与常规预训练相比,这些 bit 的信号极其强,这就是为什么你需要 RL,而不是仅仅在成功的推理轨迹上做 SFT。
因为那恰恰是关于如何得到正确答案的 bit。
Beren Millidge
有两件事。是的,第一,它恰恰就是关于如何得到正确答案的那些比特。但这并不完全是你想象的那样,因为在 SFT 中,你有一条轨迹。比如说,你有一堆数学推理,然后最后是答案。那些比特仍然在那里。你仍然在答案 token 上做 SFT。
重要的是,目标函数忽略了所有其他的比特。在 SFT 中,你必须尝试匹配模型产生的确切推理 token。你本质上是在获取过多关于你所训练的另一个模型具体如何推理的比特。而在 RL 中,你只得到那一个比特。这意味着该信号不会被模型拥有的所有其他比特的噪声所淹没。
这确实是训练过程中信噪比的极大提升,这也是为什么 RL 在步数效率上如此显著高效。
Charlie O'Neill
关于 RL 对模型做了什么,与中期训练或 SFT 或其他方法相比,一直有大量争论。每个人都在谈论 pass@1 会上升,但 pass@256 会下降。非常罕见的正确推理轨迹会被降权,并被来自更容易的推理轨迹的梯度信号所压倒。
我认为现在看待 RL 的简单方式是:如果你有足够大的算力来采样足够大的组大小——使得你获得一批正确答案的概率超过某个并非微不足道的概率——那么它就会被加权提升。就 Beren 的观点而言,中期训练和更多的预训练——即 RL 的起点 pass@1——随预训练 token 数量呈对数增长。
我能问一些非常基础的问题吗?那个回答说得通,也许有实证研究表明情况确实如此。但接下来我看看这些模型本身……我不知道发生了什么。也许你可以让我了解一下,过去一年 AI 进步的根基是什么。
也许它只是把那些本来就会进行正确思考的策略上调了权重。但从定性上看,模型的能力似乎提升了很多。也许这里并不存在内在矛盾。但我们要如何调和这种说法所暗示的 RL 影响相对较小,与模型似乎正在获得的实际定性能力之间的差距呢?
Beren Millidge
这里我想指出的一点是,这并不必然意味着 RL 的影响很小。即使你只有几个 bit,而且只对参数做了很小的改动,对函数空间——也就是模型学到的从输入到输出的映射——的实际影响仍然可能极其巨大。哪怕一个 bit 也能大幅改变你的函数空间。它可以排除掉一半的假设空间,这是巨大的影响。
我并不认为,少量的 bit、少量的 RL,一旦你从一个非常好的起点出发,就必然意味着你不会在行为上产生巨大的影响。至少……不一定。
Charlie O'Neill
我认为归根结底是两点。第一点是,大家都曾希望 RL 能把这种推理泛化到所有这些不同的领域。我并不认为我们真的获得了这种横向泛化。仅仅在数学上训练,并不一定就能让你成为最出色的程序员。你确实必须在代码环境中做 RL。
不过我认为我们确实获得的是时间跨度上的泛化。模型只是学会了如何使用更多 token、持续更长时间,并且仍然在某种任务上取得进展。你可以在那些让它们持续越来越久的环境中训练,然后把它们放进一个全新的环境。是的,它们可能并没有泛化出那种能让它们在该环境中表现出色的推理模式,但它们至少泛化出了在这项任务上持续更长时间的能力,而这与成功是相关的。有一篇叫 EdgeBench 的论文表明,模型能够持续工作更长时间的能力大约每三个月翻一番。这是泛化的明确证据。
最后一种思考方式是,在预训练中,存在一种量子的概念。你有一条非常平滑的预训练损失曲线。当你观察模型内部发生的事情时,模型正在学习所有这些非常离散的任务,并且存在所有这些出现相变的涌现点。它以前没有 induction heads,现在它有了 induction heads。这些东西有成千上万、数百万,可能数亿个。你把它们全部平均在一起,就得到了这条非常平滑的损失曲线。
在某种程度上,RL 也在发生类似的事情。正如 Beren 提到的,存在这样一个非常缓慢的外层循环。我们会训练一个模型,然后对它做 RL,接着在下一轮模型训练迭代中,我们会把一大批这样的合成推理轨迹倾倒进中期训练数据里。我们差不多是在为所有这些不同的任务逐一攻克各个量子台阶,而在单个任务层面上,它可能看起来像是一次相变。你突然之间就从某个特定金融任务或 Excel 任务之类的 0.5% 通过率,跃升到了 90% 的通过率。
但当你把所有这些加在一起取平均,再加上 horizon 泛化,你就会觉得,“哇,我们得到了定性上更好的模型。”
Beren Millidge
我觉得这其中很大一部分也只是……RL 确实会有一点泛化。你会在数学和代码之间,或者谜题和数学之类的东西之间获得一些迁移。而且,人们所针对的环境数量本身就大大增加了。以前,两年前,当你尝试做某个日常生活中会做的任务时,实验室根本不会太在意这个。他们不会为此训练模型。现在范围就广得多了。他们有很多针对这个具体事物的环境。
01:24:54 – 第 37 手与熵坍缩
在之前的对话中,我们谈到了 RL 在导致这种熵坍缩方面的背景,或者说只是把概率集中到基础模型已经做过的解上,并导致策略中相对稀疏的更新。
但我觉得关于强化学习还有另一个故事,那就是回到Atari 游戏,然后是AlphaGo下出的第 37 手,那步极具创造力的棋。因为它从未用人类数据进行初始化,所以它能以人类根本不会去思考的方式进行思考,并得出极具创造力的解法。
你觉得我们什么时候应该期待——或者说是否应该期待——大语言模型上的强化学习能产生类似第 37 手那样的东西,即超越人类创造力的极致创造力,因为智能是从零开始初始化的?
Beren Millidge
这里有几件事要说。首先,我认为 AlphaGo 用的是MCTS,它显然比普通的策略梯度做了更多的探索之类的事情。但我也认为强化学习并不一定会降低创造力。
这显然是定性的,但如果我们看看OpenAI-Hugging Face 事件,这些模型一次就能想出多个零日漏洞来突破沙箱。这显然已经达到了某种程度的第 37 手式创造力,而这只是我们从大语言模型的一般泛化特性中获得的。强化学习绝对不是在彻底摧毁熵,尤其是在长时程上。
John Schulman
人们所称的创造力,有一部分不过是解决困难的搜索问题。AlphaGo 的 Move 37 显然就是一个例子,或者写一首满足大量不同约束的诗。如果针对这方面进行训练,AI 显然会非常擅长这类事情。
然后还有另一个方面:经过 RL 之后,模型输出的多样性大幅降低,它们会养成一些固定的小毛病。尽管这些模型看起来很擅长写作,但当你做某种分布分析时,会发现它们一直在重复使用某些主题,一直在使用相同的角色名字。你得不到人类作者那种多样性。你得到的是一种非常出色的风格。所以我认为这种多样性肯定已经被 RL 大幅削减了。
事实上,既然我们之前谈到了知识蒸馏,正在发生的一件事是,非常多的人在进行蒸馏,而且主要是从 Claude 蒸馏,以至于所有 开放权重模型的写作方式都和 Claude 一样,有着相同的小毛病。这让我觉得有些令人担忧——我们正在形成这种单一文化。
Beren Millidge
不过,我再次认为这并不是 RL 作为一种方法的根本问题。知识蒸馏也是如此。即便是知识蒸馏,你也只是在数据上进行训练。你的数据不够广泛,并不意味着训练方法本身有什么问题。这是数据的问题。
我认为,比如说,很多 RL 熵坍缩基本上都是因为在环境多样性不够高时,对相当简单的验证器进行了利用。比如写作,大概是某个评判模型来打分的。这个评判模型有一些特定的癖好,而模型正在学会奖励黑客攻击这个评判模型,这就是它坍缩的原因。但这实际上是评判模型的问题,而不是 RL 本身的问题。
01:28:32 – 快速问答时间线
好,关于未来的超快速预测。我想就以下几个问题给出时间线。到什么时候我们会有这样的模型……以下是它对用户的感受。你基本上可以把它当作一个即插即用的远程工作者,用于各种白领工作?不只是编程,还有视频剪辑、法律、律师助理等等。它实际上就是一个真正的远程工作者,具备完整的计算机使用能力,能够无缝地学习和运行长达一个月,执行需要与他人互动的复杂项目等等。人类工作者在一个月内能做的一切。
Charlie O'Neill
如果你强制它使用浏览器之类的——而不是由公司把信息设置为可通过程序访问——那也许要几年。但如果不是基于浏览器的——它可以发送 Slack 消息,它可以做所有这些事情——我大概还是会说一年左右。
Beren Millidge
我想说,大概需要三年才能实现完全的通用性。但正如 Charlie 所说,最终会有很多人让自己的组织对 AI 更友好、更易于使用,所以在那之前你就能达到 80-90% 的程度。
抱歉,但一年和三年之间的差别简直就是……
Beren Millidge
我认为会有一大堆杂七杂八的长尾任务,某些人类能做,但模型需要相当长的时间才能完成。
你指的是计算机方面的东西,还是基础认知能力?
Beren Millidge
我认为这归根结底取决于我们能多快解决这种在线学习的问题,以及我们能否通过压缩和给自己写文件之类的方式达到 80-90% 的程度。这是我最大的不确定性。我真的不知道。
Charlie O'Neill
一个它不擅长的例子是:如果我必须冲某人大吼才能在工作中拿到东西,或者真的去催某人把事做完。模型就是不会那么做。它太友善了。
John Schulman
我想说,人类远程工作者的质量差异很大。如果你试图从 Upwork 上雇人做一个软件工程项目,质量差异会非常巨大。通常很难让他们做好工作,或者认真对待你给出的所有反馈。我猜测,在某些情况下,前 AI 时代的这种模式比你现在从现有 AI 那里得到的还要差。
我认为最终可能会有点复杂,因为某种程度上,对于一些质量不太高的工作,我们已经有了这种形态。但显然,在某些更高质量的工作形式上,我们还没有达到人类水平。但我基本上同意 Charlie 和 Beren 的看法,也许一年左右我们会有某种版本的这种东西,还算可以。我们会拥有那种形态,它能非常好地完成一些事情,另一些事情则做得不那么好,然后从那里开始不断改进。
Charlie O'Neill
我们一直在根据那条非常长的长尾不断移动球门柱。我感觉你之前用过报税之类的例子。今年,我直接让 Codex 去把我需要的所有东西都弄来发给会计师。有一大堆东西它得用电脑点击操作并下载。它做到了。完美无缺。很多这类事情它已经能做了。
好吧:给你 10 倍的整体生产力提升。基本上,如果现在你需要一年才能取得一项突破,那么你每个月都能取得一项突破。
John Schulman
我觉得我干脆拒绝给你一个标量数值。在某些类型的工作上,我们可能已经超过那个点了。比如说,你在尝试做某些类型的数学,而——
哦,抱歉。但对你这样试图推进 AI 研究前沿的 AI 研究者来说呢?AI 研究者被加速或提升了多少?
Charlie O'Neill
大概 5-10 年之间?
哦,真的吗?好吧,那还很远。
Beren Millidge
真的吗,你觉得这比一个通用的远程工作者还要久?有意思。
我意识到你对"完全通用的远程工作者"的定义可能非常不同。我早该说清楚的。
Beren Millidge
确实如此,因为显然一个 AI 研究者也可以是远程工作者。
Charlie O'Neill
我想的是普通白领工作在一个月周期内的情况。我觉得超过两个月后就开始有点分化了。
一个非常称职的白领工作者,但不一定是极具创造力的研究员。
John Schulman
我会说两年。
两年?10 倍?好吧。你呢,Beren?
Beren Millidge
其实我大概能理解,因为现在在编程方面肯定已经超过 10 倍了。所以如果它能做哪怕一两轮实验反馈的循环,那已经会是巨大的飞跃了。
所以,AI 研究员的生产力在两年内提升 10 倍。如果你把这个代入一个非常朴素的 AI 进展模型,看看其中有多少来自 AI 研究员,而他们的生产力又提升了 10 倍,那么从两年后开始,AI 进展的速度将被极大地加速。
Beren Millidge
我认为这意味着 AI 进展的瓶颈将不再是 AI 研究员运行小型实验的能力,而是会转移到其他方面。
当然。但它就是快了 10 倍,这本身就是件大事。这也会让下一件事——带来 100 倍加速的那件事——更早发生,以此类推。
Charlie O’Neill
我愿意在那件事上多花一点时间。
关键点是什么?
Charlie O’Neill
我吸收信息、并对下一个实验做出贝叶斯最优决策的能力。
Beren Millidge
我假设你可以把其中一部分委托给 AI。AI 在决策方面正变得相当不错。它运行了这个实验,得到了这个结果,它就去运行下一个实验。如果它能连续运行两三个实验而不崩溃,那实际上就是一个很大的提升。
好,最后一个问题。一个在所有可以通过计算机完成的工作领域中都能超越顶尖人类专家的 AI。所以不只是 AI 研究,而是所有认知工作。不只是短周期的工作,而是字面意义上,即使需要三年左右,AI 仍然会比人类做得更好。
Beren Millidge
这基本上就是 ASI 了吧?
John Schulman
我会说是 3-4 年。
什么鬼?我是说,这看起来倒也没错,但是——
John Schulman
AI 显然正受到越来越多的关注。这是比较难的事情之一,但大量精力正投入其中。它对 AI 来说也不算最难的事情之一,因为它涉及大量代码和数学,而模型在这方面非常擅长。
对于涉及 3D、空间和物理的东西,我认为那会需要更长时间。如果是机械工程之类的,而且目前没有得到最多的关注,那可能需要更长时间。
但它也确实包括一些由于领域性质而数据相对较少的领域,模型必须边做边学这些数据。例如,它必须在台积电(TSMC)当工程师这件事上达到超人水平之类的。
John Schulman
所以你得假设你能给 AI 同样的入职培训材料。然后还得解决更长周期学习的问题。
Charlie O’Neill
我会说 5 到 10。
所以基本上,你认为自动化 AI 研究是 ASI 完备的之类的?
Charlie O'Neill
是的,我想是这样。我觉得世界上有太多事情,即便你有一个模型外部的某种记忆系统,即便上下文长度增长了一些,但有些东西从根本上来说,即便你可以自己去检索信息或做笔记,你今天仍然需要一个超过百万 token 的上下文窗口。
Beren Millidge
我在 5 年这个时间范围上大致同意,至少对于各家实验室正在聚焦的那些东西来说是这样。但我觉得会有一条很长的尾巴,那些东西 AI 理论上可以去学习,但没有人费心去做,算力也没有分配到那上面。所以要真正做到覆盖每一个人类专家,可能需要更长时间。
抱歉,我这么说也包括了像人类一样快速学习一个新领域的能力。
Beren Millidge
我觉得那未必是必要的,因为 AI 将拥有远超任何人类的经验。
非常感谢各位参与这次讨论。我觉得这是一个很好的形式,让不同的专家表达分歧、展开辩论并一起探讨。非常有收获。
John Schulman
感谢邀请我们。
I got together with some of the most insightful AI researchers I know who are at the openish companies, because I wanted to hear the details of what's actually happening at the frontier and what comes next.
00:00:00 – Steelmanning the case against RSI
Today, I’m chatting with three of my AI researcher friends from whom I learn a lot every time we talk. They also happen to be at somewhat open-ish labs and companies, so you guys can actually say things on the record. I’m joined by Beren Millidge, who is the CTO of Zyphra, which is developing open source models. John Schulman is the chief scientist at Thinking Machines, previously a co-founder of OpenAI, and led the RLHF work that led to ChatGPT. And Charlie O’Neill is head of model training at Baseten.
The first question I have: If we’re in 2036 and we don’t have billions of crazy superintelligences running around that have radically transformed the world, what is the most likely reason that doesn’t end up being the case? Other than exogenous political shocks, or there’s a war, or they ban AI or something. What is the most likely technical reason that 2036 isn’t a crazy alien superintelligence world?
Beren Millidge
There’s been a classic thing, almost like Moravec’s paradox, where we think of the AI as, “If it can do this, it’s going to be amazing.” If it can solve these hard maths problems, if it can win at chess, blah, blah, blah… Then it solves these things, and it’s not that impactful. Obviously, it’s somewhat impactful, but not everything.
If somehow that continues, and there’s never the true spark of generalization that occurs, I think that could lead to the AI just being extremely good at everything that people put into a benchmark or put into an environment. But there’s still some persistent sim-to-real gap which is somehow blocking everything. I think this is unlikely. We do actually see this kind of generalization even from RL in practice already. But if it is just ridiculously hard to generalize meta-learning, plus we don’t solve continual learning and it’s just super hard and impossible… This would be my default scenario in that case.
John Schulman
I agree with that. Humans have a lot of advantages over models now. Each time a new model comes out, it’ll catch up in some of these areas. But you end up getting bottlenecked by the places where the model is weaker and where it has worse judgment, or the models can’t check themselves well enough.
There’s this cycle that keeps repeating where a new model comes out and people are blown away and they’re like, “This is it. This is AGI.” But then they use it a bit, and it starts to feel dumb after a month or so. That cycle just might keep going. It’s hard to predict how many times it’s going to repeat.
Right now, you don’t get explosive growth in capabilities because you still get bottlenecked enough when you’re trying to do research and engineering. Even if the model can write way more code than a person, it doesn’t make you 100X more productive. So maybe there are just more of these cycles than we would expect.
Charlie O’Neill
For me, it’s a question of how far off the global optimum of “a learner you could have on a chip” is from the transformer + RL, basically the current recipe. People imagine that once you have an agent which is better than all humans at AI research, even if it’s 0.1% better than all humans, then the fact that you can run hundreds of thousands, if not millions, of these in parallel — and you can run them much faster as chips speed up — is going to outweigh every other bottleneck. You’re eventually going to hit this very fast takeoff with regards to self-improvement.
I could imagine that if we continue along the trajectory that we’re currently on with that paradigm, where it’s basically self-attention, RL, scaling up RL environments… Think about what happened with Moore’s law. We had this very nice straight line and that held for a really, really long time. But there were so many discrete discontinuities and innovations that had to happen to keep that scaling law going. The same thing has happened with LLMs. We had this pre-trainingscaling law, and then that was hitting diminishing returns. Then we came up with RL and solved that, and then we got this new diminishing returns curve to hit that made it keep looking like a straight line going up.
So if it requires another one of those discontinuities to solve, I’m not sure that the current method of training LLMs with these RL environments, even RSI-targeted RL environments, would be able to discover that discontinuity. If not, we’re probably going to hit this asymptotic curve.
But do you think the discontinuity will be harder than anything that’s come since 2012?
Charlie O’Neill
If we had the answer to that, we’d kind of have the ability to implement it. But maybe we should distinguish between a discontinuity which adds to the current paradigm, which is cumulative — there’s something beyond the RL that we have to discover, and maybe they’re capable of connecting the dots in that straight line — or, again, how far off the global optimum are we? Do we have to go back and throw out gradient descent and neural nets in general? I don’t think, if you continue to scale up the current paradigm, an LLM, no matter how many LLMs you’re running, is necessarily capable of discovering that if it’s too far away.
The only hope really is if deep learning just can’t get us to an AI which can at least dominate human research and human development, including the human ability to come up with new paradigms and so forth. Or, I don’t know, maybe humans would also never have discovered the next learning architecture. But to the extent humans could have discovered it eventually… But it just seems like… If you just look at the progress that’s happened since 2012 till now, and you just continue that on —I know it’s just been powered by huge amounts of compute scaling and so forth— it would be weird if it just didn’t get to the point where it could dominate humans, at least in R&D, especially over the next few years.
Ryan Greenblatt was on the podcast recently. He made this point that I’d be curious to get your thoughts on. You could imagine, as AIs get more and more capable, that they’re capable of making progress on simulations which incentivize getting better at not only AI R&D, but at science generally. This is a thing that all the labs are targeting and many startups are targeting.
Another intuition pump is if you look at the Elo score of chess bots since the ’80s. There’s a very linear increase in Elo over time. But there’s this huge discontinuity as they cross the human range, from human experts always winning against AIs to human experts never winning against AIs, as this linear increase in Elo happens.
I agree with your point that so far, AI capabilities have not been that big of a deal in terms of their end economic impact in the world. But that is because they’re slowly rising in Elo relative to humans.
Beren Millidge
I agree it would be very surprising. The only way for this to not happen is if, as you said, it somehow asymptotes just before. Because we’re already pretty close, in my opinion, to where we’ll start crossing the human Elo score. So we’ll need to asymptote before that. That’s the only way — in this scenario you pose where somehow we’re sitting here in 2035 and everything is normal — for this to happen, I think. The only other way is there’s some dramatic regulation on AI. This is what I see as the most likely way for this scenario to happen, actually, rather than a technical thing.
Charlie O’Neill
I think there’s different kinds of research. There’s research in the autoresearch style where the objective is already specified very cleanly and you’re optimizing that objective. I think everyone is picturing that if we continue along this path of making pre-training loss go down and making our environments have the reward on them go up, that’s going to lead to improvement.
But maybe what Ryan is talking about is this much more open-ended type of science which is required for paradigm shifts, where we can’t specify the objective, and the AIs are definitely not able to specify that objective either. We have to be really, really careful about how we specify objectives for any of these things.
Maybe your point is that the nature of the breakthroughs that have happened since 2012 is that we have found… In 2012, people weren’t saying… I’m assuming, I don’t know, you guys were there. Or at least John, you were there. But I was not.
Charlie O’Neill
I was in primary school.
Actually, John, I’m curious for your wisdom of the ages, or wisdom of being in the trenches way back when. Presumably, a big breakthrough was realizing that next token prediction is the… You wouldn’t have thought that the nanoGPT speed run is the thing to be optimizing for in 2014. But now that we have come to this new paradigm, you would think to do a speed run on that and have AIs get really good at that.
But maybe there’s a next inner loop to optimize that the AIs wouldn’t anticipate. There’s an outer loop of revenue or something that eventually should be strong, but it’s a very slow outer loop.
John Schulman
In fact, I remember in the early OpenAI days having the intuition that just minimizing log loss wasn’t going to get you to intelligence. Because the important bits are accounting for such a small fraction of the loss that it was going to be overwhelmed by noise. So just training a language model on next token prediction wasn’t going to learn the interesting things you want it to learn. We needed to craft better objectives that would put more emphasis on the important things.
You can make all sorts of arguments for this. You could say, “Oh, humans probably don’t learn how to model everything in our environment. Most people can’t create a photorealistic reproduction of some kind of scene they’ve looked at. So we must need a better objective.” But then it turned out that it just worked anyway.
As you were pointing out, the inner loop, even in current AI research, of post-training benchmarks or whatever, doesn’t necessarily translate into what users like.
John Schulman
Oh, yeah. The whole field relies a lot on generalization and it’s very hard to predict when you’re going to get generalization, or when you’re going to get some kind of out-of-distribution generalization. We know that if you train on the task you care about, you’re going to do better. But the most important advances are often types of generalization that we have no right to expect.
For example, from just pre-training on this very naive next-token-prediction objective to various tasks of interest that require understanding of the input in some deep way, or learning some skill from pre-training that’s very rare and not heavily represented. Then also generalization from these verifiable tasks to less verifiable ones, this is also a type of generalization that there’s no reason a priori to expect.
This is an interesting question, because one intuition pump you could have for why you would see some sort of singularity very rapidly — without even scaling up the inputs to AI progress that are not just AI labor — is that before every single 7-figure experiment you run, you spend an equivalent amount of compute on AI labor. So you just have automated versions of you guys spending a century thinking about what is the optimal experiment to run, doing small-scale ablations, developing literally a century’s worth of theory, going back even before deep learning.
Before you decide what experiment to run, you’re doing extremely optimal setting up of the experiment. Then you do a century of thinking after the experiment is over, where you’re analyzing what happened and what the next experiment to run is.
John Schulman
If you think hard enough, you probably could have expected some of these things beforehand. There is probably some very clever way to do a small-scale experiment that’ll let you build the theory that then will generalize to the large-scale experiment. So I would expect that we’re nowhere near the ceiling of how well you can do research.
I would imagine a future where AI is doing a lot of analysis and theory building, spending a comparable amount of compute to the amount that you’re spending on the experiments themselves, doing various kinds of analysis and building a theory around what we’ve seen so far.
Charlie O’Neill
I think there are really concrete examples of this when the objective is well specified. All thinking can do is update your posterior based on the bits that you’ve gotten since you formed your prior. You can’t gain any new bits from just thinking. But when the objective is well specified and there is this data sitting around, I imagine there will be this big speed-up in the current paradigm we’re in.
A good example of this is if you got an AI to think about the Kaplan scaling laws. An AI at this point would have noticed, “Oh, they’ve just taken these intermediate checkpoints and didn’t account for the annealing, and so this is wrong.” That would have been caught years earlier. We would have cut off a year or two of progress just from that observation from an AI.
Again, once the objective is well specified, which is lower pre-training loss or whatever, there are many, many good examples where if you just thought about it a bit more, you would have been able to cut down significantly on things that you’ve done. So muP, and how learning rate scales with model size, and realizing that model width is important in that as well. I feel like you can really back out a lot of these things and cut off a lot of low-hanging fruit. I would imagine a 10x speed-up if our thing is just, “Maximize the objective we’re currently on.”
But I don’t see how that generalizes at all to coming up with the right objective in the first place. Just thinking doesn’t necessarily buy you the right objective in the first place.
Beren Millidge
I think this is really the key question for any kind of very rapid RSI from current AIs. How well can AIs generalize to learning their own objectives? To have any kind of self-propelling automated loop, we need the AI to propose objectives, optimize them, figure that out, propose a new objective, and have this not go off the rails at any point for a long, long time.
To come back to Moravec’s paradox, there might be a case of Moravec’s paradox where we think this kind of autonomy and being self-encapsulated — so we can think of what we should do ourselves and then go do it and have this loop — is super easy because we always do this. Obviously, evolution needs to create creatures that can survive by themselves for long periods of time. And this just might be something that for some reason is really hard for the AI, in the same way that locomotion stuff is really hard but math is super easy despite being super hard for us.
But doesn’t the time horizon increasing suggest that that’s—
Beren Millidge
Yeah, exactly. This is another possibility, but I agree, there’s no obvious evidence for this. In fact, the fact that our agents are now super persistent and it’s quite easy to do this is kind of evidence against this. But this would potentially be one of the reasons why we just don’t get this immediate takeoff, if this is hard.
If you look back from 2012 till now — or maybe from when you started doing your research till now — what part of all the innovations that have happened since that time, including purely engineering ones, including purely conceptual ones, seems like the thing that would be the last thing humans would have to do before AI totally automates AI R&D?
Beren Millidge
Probably just iteratively asking the right questions. If you can get the AI to do any experiment, you still need to decide what experiments to do. Right now I think AIs are not very good at this compared to coding the experiment. Whenever we talk about research, they propose a bunch of miscellaneous things which are very, very tiny steps.
Charlie O’Neill
Or even going from DeepMind’s approach of, “We’re going to solve intelligence by learning to play games at a superhuman level,” to one random researcher like Radford being, “I’m going to try and just predict the next token of a very wide swath of data”… Even once Radford had discovered that, it took a while before people decided to scale it up, because we had to come up with the idea of scaling laws and the fact that you could very reliably predict these things.
John Schulman
I would say that the last job for humans, or the role for humans that’ll last the longest, is defining the objective and deciding what we actually want. In that vein, something like deciding how the AI assistants should behave, or what it means to be helpful, or what the objective is when we’re doing RL from human feedback, is one such thing. Then later, defining constitutions and model specs is another one. Even if the AIs can do all the technical work, we’ll still have to do a lot of that and decide what we actually want.
Alignment is the final job.
John Schulman
Alignment is sort of the answer. But alignment itself can be decomposed into specification of the objective, or figuring out what the right objective should be, and then actually achieving or optimizing the objective you’ve defined. I think the first one is not going to go away anytime soon.
If I think about a post-training team and why you need a lot of people on the team, it’s just because there are a lot of different areas where you have to figure out how the model should behave. It would be very hard to automate the whole thing, just because someone has to think about how the model should behave in this area.
00:18:39 – What’s driving the Chinese labs’s progress
What is the story for why there isn’t huge consolidation in model providers? There are just so many things that point to centralization here. If you step back over the course of years, is there something that is going to prevent that?
John Schulman
I think distillation is the main thing that fights against the centralizing force. Basically anything that can be learned through RL can be distilled very easily, because it’s a small number of bits. It’s something that you can learn from a small amount of data. If you can get trajectories from the model that show a behavior, you can easily distill it. I think distillation is one of the things that fights centralization.
There is also a possibility that there’ll be company-specific models, that it’ll be possible to learn from deployment and have a company continually improving its own model. Such a system could be provided by the current oligopoly of model providers or some other currently smaller company. But I think that’ll change the game a bit.
Beren Millidge
I also want to point out that continual learning, honestly, doesn’t stop distillation. Even if your model is improving every day, people could be distilling it every day. The loops could just operate at the same pace.
That makes sense. So copying model behavior… I guess you need to know yourself what the right distribution to prompt is in order to get the relevant model behavior?
John Schulman
Oh, yeah. For just distilling with supervised learning, the prompt distribution is extremely important. It’s very non-trivial to distill a model, even if you have full access to it and have the chain of thought and everything. It’s non-trivial to distill all of the useful capabilities from it, because you need to prompt the model with something. You need to prompt it with realistic prompts. You need to have a really wide distribution of realistic prompts.
One thing that’s been coming out recently is that some of the Chinese companies are probably using these router services which are designed to allow people in China to use the US frontier models, which would otherwise be blocked in China. There are all these router or proxy services that allow people in China to use these models, mostly for coding. And these router services are collecting and selling some of the data. This is a very useful data set for distillation because it gives you the perfect prompt distribution.
Beren Millidge
I think this is one of those things where AIs help a lot. If you actually look at the frontier pipelines, or the Chinese models that they’ve actually put in their papers, they get seed prompts from somewhere, which is some combination of humans and this kind of data. Then they synthesize a vast coverage from those seed prompts using their existing models or the other frontier models. You can automate an awful lot of this prompt distribution gathering and environment creation. Humans need to provide increasingly fewer bits as the models get better.
But it still seems you’re bottlenecked by having a service which has users going through it.
Beren Millidge
Not necessarily. That’s obviously very helpful, but theoretically, you can just think about what users want.
But the whole point is that the user says, “Make me an application like this. Oh, that didn’t work. I actually want you to make this new feature. But actually, let’s step back and do this other thing.” Capturing that whole trace is the thing. Or to the extent you could have done that anyway, then you just have RSI.
Beren Millidge
Ultimately, if you have this fully automated loop, that is basically RSI. The AI is deciding the data, it’s deciding the training. That is the loop. But it depends how much human information you need. At some point, if you’re just like, “I want traces that look like this,” you prompt that to the model. The model will be able to come up with a pretty good approximation.
But what if you want to do, “Make me a really good politician,” and then it has to anticipate de novo how a discussion in the Senate halls would go or something? I just feel like there are going to be a lot of things which are—
Beren Millidge
Ironically, this is actually easier for the distillers than the frontier labs. The distiller’s just like, “I want a good politician.” They go to the frontier model. The frontier model already knows how to be a good politician, so it just generates those traces. Whereas if you actually want to build the first model that does this, you have to actually somehow get data on what politicians do every day and build that. It’s actually much easier to say, “I want something like this,” and then get the AI to produce a billion variations, than to actually create the thing like this to begin with.
Charlie O’Neill
I think you can actually make a really concrete prediction based off this observation that the Chinese labs have this router data. The thing that started this originally was I was saying, “Isn’t it weird how Sonnet 5 and Opus 5 are almost objectively worse models than GLM-5.3 and Kimi K3, even though they’ve had access to not only distillation but logit distillation from Mythos?” The counter was that the prompt distribution really, really matters. You need to see what users are doing so that you can distill these behaviors and things in.
I think the prediction from this is that the frontier labs don’t necessarily have much of an advantage, if at all, in RL environments now. Yes, user distribution matters for general behavior and so on, but the best measure of a capability is the very, very hard RL environments you’ve made at the frontier. If you have access to those RL environments as Anthropic, and you have access to logit distillation, and you’ve still made a worse model, then maybe—
Then real-world deployment matters more than the environment. That’s really interesting.
But they had to incentivize those capabilities in the first place in Fable, or the frontier model. So it’s weird that they can’t incentivize them again with a smaller model or something.
Charlie O’Neill
Maybe we’re just in this weird uncanny valley where trying to copy that frontier model too much, the student-teacher gap, whatever it is, is just too large. People have made this point with Opus. The difference between Opus 4.6 and Opus 5 is that Opus 5 really feels like it’s got this AI-as-a-judge checking every possible thing it’s done. That’s why it uses so many tokens. It tries to think about all these things, but it doesn’t necessarily have the big model smell of Fable to know when to stop doing that, or when’s a good path to go down.
The reach exceeds the grasp.
John Schulman
I would offer a slightly different hypothesis. I would say there are a couple of different axes for the environments you can create. One of them is difficulty and the other is realism.
It’s comparatively easy to create a lot of difficult environments that involve doing a much more complicated task or doing something that requires a lot more cleverness. You could say this is the benchmaxxing distribution, because a lot of the most prominent benchmarks just involve doing some very hard puzzle-like task that’s easy to verify. Then there’s the realism axis, where you want the model to be good in the realistic coding agent setting where there’s multiple back-and-forths with the human and there’s multiple objectives.
The labs who are crafting the model behavior for the first time need to push in both directions. To get good model behavior, you need to really push on the realism axis and have rubrics or some kind of human feedback that’s informing the reward function you use there. But if you try to do distillation naively, you end up just matching the teacher on the benchmaxxing distribution. If you don’t have enough of the environments that really exercise the capabilities in these trickier realistic settings, then you’re not going to get those into your student model.
I think maybe one thing that’s happening is the big models generalize better from the tricky narrow tasks to these more realistic tasks. If you have a really good realistic prompt distribution for distillation, you can match the big model really well. But if you only have this distribution of easily verifiable tasks, then you can match the big model on all the benchmarks, but you do worse on this broader distribution.
That might even explain something about the smaller Anthropic models, like Sonnet 5, though it’s hard to predict exactly what they’re doing to post-train those models. It could also be that they’re always changing their post-training stack, and they just got a few things wrong in some of these models. I don’t know… they turned something up too high and created some quirks that people really don’t like. It’s really easy to screw up post-training in some way that doesn’t show up in benchmarks.
Beren Millidge
Just one other very basic point is that the frontier AI labs buy all their data from big data companies. The Chinese can also just buy the same data from data companies.
And they are, right?
Beren Millidge
And they are. Exactly. There’s a lot of people being annoyed about this, but if they have exactly the same data and they can buy that, they can also distill. It means it’s quite easy to keep up, really.
00:28:06 – How will automated AI researchers be trained
The other question I had is how the first models that are capable of automating AI R&D will actually be trained. There’s a toy version, which is this thing that Ryan was talking about. You just have GPT-8 try to build GPT-3 size models that are really good at inner loop type challenges: beating video games that require continual learning, or just getting to a certain loss with the least amount of compute, et cetera.
But John, I think you had an interesting point that maybe that’s not the way it actually will happen in practice. So I’m curious, by the point at which you have AIs that are actually capable of automating AI R&D, how are they probably trained?
John Schulman
We’ll probably do some combination of learning from human feedback to absorb the researchers’ taste, and just creating a lot of practice environments which involve doing multi-step research projects. People will in practice do some combination of those two things and, each iteration, patch whatever seems to be most broken in the last iteration. Researchers will be using the AIs a lot and will notice that they have some consistent weaknesses. Those things will either be patched by collecting human feedback or creating environments.
Charlie O’Neill
Maybe a useful way to think about this is how much of the lineage we roll back and then let self-play from there. In the limit, you’re picturing just giving them a GPU and maybe neural nets or something and saying, “Okay, figure out how to train a model to do these particular tasks.” The way it currently works is we go up to the very edge of the lineage and say, “Okay, here are the bugs Anthropic has found in their training stack in the last few months. We’ll turn those into environments.” You need to train and get better on the frontier. So you obviously lock in all the previous history of the lineage.
But you could imagine a world in which you roll back to before GRPO or something. Then you have environments which try to get it to discover the best form to RL models on, and then maybe you roll further and further back… But I think we will still be so compute bottlenecked that people will just keep staying at the frontier and essentially diffing the bugs and whatever improvements they found since the last model version, turning those into training environments.
Which is also really good for having non-stale, new data between model generations. Again, this is basically continual learning within the AI lab, of distilling the last three months of AI research progress through environments and RLHF-type stuff back into the model itself.
Charlie O’Neill
And it is distilling. That’s maybe why some of us feel like it’s asymptotic. You’re always just trying to get the last three months of progress. That progress is being contributed to by AIs, of course, but it also still has humans in the loop. It feels like you’re just constantly inching closer and closer to what the human researchers are finding and capable of doing.
Beren Millidge
The one thing I will say, though, is obviously if you’re just distilling on trajectories, you can never go above it. But environments can go quite a far way above what a human can do. It’s very easy to design an environment that no human can solve, but the AI can obviously still try and solve it. That would be the path to go ahead of just what the human AI research is.
Charlie O’Neill
Do you have an example in terms of RSI, of what kind of training set?
Nanochat speedrun, but doing it even faster than a human speedrunner.
Beren Millidge
I feel like in AI research especially, it’s very easy to define goals. You could say the loss needs to be 1.3 or something, and no human can get that now. But that’s an extremely measurable, verifiable task. If the AI gets that, then great.
Or I don’t know, building a 100 million parameter model that beats Minecraft. That’s maybe too easy, but beats a much more complicated game or something.
Charlie O’Neill
Isn’t it crazy that 100 million parameter models beat Minecraft? We’re calling that too easy? Imagine if you said that five years ago.
John Schulman
I would say a lot of research is not exactly like that, though, where it’s hill climbing on a well-defined goal. It’s more like, here’s an intuition we have about some way models should be better. We also have some idea for an algorithm that seems to go a little bit in this direction. So let’s come up with a task that is designed to show signs of life on this approach, and see if we get those signs of life. If we do, we can make successively more realistic versions of the task.
It’s a lot more guided by intuition. The inner loop is to test for that intuition rather than the test itself leading to the insight.
John Schulman
Right. You’re not directly optimizing for the eventual objective you care about or the practical production objective. You’re relaxing your objective a little bit. You’re saying, “Let’s relax on the realism axis a little bit and find some methods that actually work, and then try to get back to realism later after the method matures a little bit.”
There’s also research that’s more oriented towards explaining things and developing a theory. Often we don’t have mathematical theories in machine learning that are that predictive. But we have a lot of more informal theories for what’s going on.
Beren Millidge
Presumably the models will be trained on some combination of all of these tasks. Some will be very easily verifiable, some will be LLM-as-a-judge or just ask the human, “Does this look reasonable?” The hope would be that these would all generalize to these much harder, more vague, fuzzy kinds of tasks. It probably will to some extent. Whether it generalizes enough that the loop can become self-sealing without humans being in the loop at all is unclear.
00:33:51 – Will long-horizon RL elicit AGI?
Maybe taking a step back. Here’s what it seems to me the plan for AI research going forward is. You tell me if you think it’s going to work or if you agree with this characterization. The bet is that we will scale up RLVR training across millions of diverse environments, across hundreds of different kinds of domains. What will emerge at the other end is an agent which has learned these basic skills — or less than basic skills — around being persistent, being able to triage information and context, eventually having end-to-end optimization of working with other agents and things like that.
Such an agent will be very sample efficient within the context —you’ve done research on how you scale up in-context learning to make it arbitrarily long, but you keep scaling it up. And what comes out the other end will be something that basically functions like a drop-in remote worker over the course of a week or a month.
First of all, do you agree that that is the bet the labs are making? And second, is that enough? Basically learning how to learn within these simulacra within a data center, and then getting deployed into the real world, but not actually learning from real-world deployment… only learning these meta skills from the simulated environments in the data center.
Charlie O’Neill
I think it’s now hard to separate out how much of the labs’ effort is going towards direct RSI versus making generally intelligent models that they can continue to deploy to collect revenue to fund the next big training run.
For the latter, yes, that’s probably just the bet they’re making. It’s very clear, the pattern of where these environments are going over the last few years. Anthropic’s lineage of environments is a very clear example of this. First, we just focus on coding and we’re going to get really, really good at that. Then, from the task horizon that we’ve got from coding — which is probably the lowest-hanging fruit in terms of data available on the internet to create environments, and their own internal stuff that they can turn into environments — then we’re going to generalize.
We’re going to go up to finance next, and literally just so much Excel data and all that sort of stuff in the RL training. Then it’s PowerPoints. It’s this long tail of the working economy. That seemed to work really well. A lot of the other labs, even the open source labs, have now realized that that was the correct bet to make.
But what is the implication from that? When I had Dario on the podcast, the thing I asked him was, if you truly expect models which will be human-like in their ability to learn on the job, why would you try to bake in all these skills of working with PowerPoint or something? Wouldn’t you just expect the model to be able to pick that up while it’s deployed?
There’s multiple different explanations. One is just that we expect models to get there soon, but they’re not there yet, so why not amortize these skills into the model training? Another is that we’re not concentrated on making it really good at widely deployed work. We just want it really good at RSI. This is just a way for us to get revenue so that we can pour it back into a model that is actually really good at doing RSI development. Then once the singularity happens, the thing that comes out the other end will be really good at all the things which seem like bottlenecks to the current generation of models.
John, I don’t know if you have takes on how one should construe why there is so much task-specific knowledge in these models if the path is this kind of generalization.
John Schulman
If the models were good enough at learning in context, then in theory, you wouldn’t need to train them on finance. They would just be able to read all the books on the fly and figure out how to do everything in the appropriate jurisdiction. You could argue that you need to do a lot of this domain-specific training just to make them more efficient. Even if they were smart enough to figure this out on the fly, you still might want to do a bunch of RL and bake all these intuitions into the weights, so the model would be more efficient at runtime.
In practice, it does seem like model providers are going domain by domain and trying to strengthen the models in the highest value domains. I’d say that that’s one of the answers to why the models have gotten so much better. It’s just because the model providers have covered a lot of the high-value domains and the most common types of skills.
Beren Millidge
Another thing is just that it’s not that expensive to do both at the same time. The models are massive. They can easily afford, in terms of their parameters, to learn everything. There is likely some transfer. Even if finance is not specific, the information is important for RSI. Just the general meta-learning of how to figure out what’s important, how to have taste, how to do long-horizon work is potentially generalizable.
There’s not that much RSI data in the world as well. It’s hard to generate and requires a lot of effort. So if you can amortize in this other data, you get some transfer from it. You already have masses of compute and masses of parameter space, so why not do that as well as, obviously, the direct commercial intent of selling a model?
John Schulman
I’ll add that there’s one question about whether this current paradigm of doing sim-to-real will be the dominant one forever. You look at what the real-world tasks are like. Then you try to create a bunch of environments that can be simulated in the data center, and you can do RL on them. Obviously, this has been very successful. But it has a lot of weaknesses, because a lot of things are just hard to simulate, especially if they involve interacting with a bunch of humans in real time. So there’s some question about whether sim-to-real will be the dominant framework forever.
Beren Millidge
I think sim-to-real has to be the dominant framework while sample efficiency is low, because right now you need thousands and thousands of interactions with the humans. No human is going to sit there and be in the loop of RL training. So we have to simulate that now to get the samples you need. But obviously, if sample efficiency improves a lot, you’d expect learning from deployment to become a much bigger part of it.
John Schulman
Though there are also other things you could do. You can learn off-policy, so you can take all the traces, and even without resimulating everything, you can potentially learn something from them.
I want to ask more about this, because it’s weird that you have 50% of compute that’s spent on inference that is not directly helping the model become better. One of the key advantages you’d expect digital minds to eventually have is that, unlike a human who gets to have 50 years of real-world experience, a model will get to experience, through all its instances, millions of years of deployment across all kinds of economically relevant work in the economy. Right now, that data is just not, in a meaningful sense, helping the model get better.
It seems so obvious that eventually models should be able to learn from this data. Once they do, you would have something that almost feels like a widely deployed intelligence explosion, because the model is assimilating so much information across all these deployed instances. When do you expect this kind of hive-mind, crazy shit to start happening?
Beren Millidge
I think broadly, at a very basic level, this is already happening… just in the next generation of models. Right now, you can obviously take your deployment data and put this in the pre-train or the mid-train of future models, especially if you do some kind of filtering or some kind of judgment or annotation or synthesization of that.
How much do you think that explains the generation-over-generation improvement?
Beren Millidge
I think it explains quite a bit. I don’t know whether the labs do this, because theoretically, they claim not to train on people’s data. But the Chinese 100% do. They definitely get this advantage. This is basically what distillation is. They take the models, they get some fraction of their deployment data by pinging the model, and then they train their next generation of models on it. They can certainly do it on their own models as well. There’s no reason not to whatsoever.
Charlie O’Neill
I completely agree with this. If you zoom out far enough, this is definitely happening. What we’re all picturing, the holy grail of continual learning, is this very organic, live loop of an individual model getting an experience and live-updating on the spot and learning from that. A lot of things break when you zoom into that level of granularity. But the big labs are doing this. The closed models are doing this.
There are also early signs of life of people using open-source models doing this at a much faster cadence. A good example is probably Composer. Harvey’s doing the same thing with legal agents. You have some sort of model, and you are getting very specific environments from the data that you have for that particular task, and things that users are complaining about, and all the feedback that you’re somehow extracting from your specific deployments. A lot of these companies have the advantage over the big labs in that they can use this data really, really well.
Then they will create environments. They will do a big post-train of Kimi K3. They will go deploy it. They might do some online learning as well, like Composer did online… basically REINFORCE for a long time.
There’s still a human in the loop. There’s still a human saying, “Okay, these are the signals we care about. Here’s how we’re going to create environments from the data that we have.” It’s still a longer cadence than maybe the one that you’re thinking of, but it really is happening. Eventually that loop will become faster and faster.
The Composer thing is interesting because this is where, in Cursor, people press Tab or they don’t press Tab on the next completion that the model suggests. Based on that, every single day, Composer gets better at predicting the next—
Charlie O’Neill
That was the old Tab model. They actually did the same thing not just for the Tab model, but for the actual generative model.
Oh, I see. Interesting.
Charlie O’Neill
It’s hard because when you do online reinforcement learning, you don’t have groups. You just have one user saying one thing, and then you get one rollout. So you have a big variance-reduction problem.
Cursor’s fuzzy answer to this was, “We have very good heuristics which are able to estimate how much better than average this response was, or how much worse than average this response was.” Then they would do this big REINFORCE update. Their solution to whether it got worse or not was that if it improved on CursorBench, they would deploy the new model every five hours. If it didn’t, they would throw that version out.
John Schulman
I think your biggest problem is actually just not knowing what the reward function should be for natural data. If you use some kind of superficial signal, like did they accept the edit, that might get reward-hacked in some way.
00:45:24 – The sim-to-real gap
But isn’t this a bigger issue with the sim-to-real thing, where the longer-horizon tasks get, the harder they are to simulate within a data center? It seems to me that even in coding, we’re already getting to the point where there’s not some year-long coding task that doesn’t eventually require you to talk to a client or interact with the company or interact with users.
If you think about the gamut of things we would want AI to be capable of, eventually superintelligence should be able to run a business, or start a new business and make it profitable, or have a profitable day trading in the markets, or win a court case. These are all things which are very hard to simulate in a data center. An inherent part of the learning there is interacting with the real world.
Maybe they learn how to get better at these things from the transfer between sim-to-real. But alternatively, maybe you do need weight updates from these kinds of interactions in order to get better at them. If that is the case — if transfer isn’t strong enough and you do need weight updates — then the fact that the models are quite sample-inefficient is maybe a deeper problem.
The reason I’m curious about this is that by default, I don’t see how you don’t get some kind of crazy recursive self-improvement within the next 10 years. But the one reason why that might not happen is that in terms of the sample efficiency of weight updates, models just seem way far behind humans. They’re plausibly a millionfold behind humans in terms of how much data a human sees from birth to adulthood versus how much a model sees from cold start to finishing training.
This is all to say, first of all, is there going to be good transfer between simulations and the extremely long-horizon, really complicated real shit that we want the AI to do in the real world? And if not, does that really mean that the lack of sample efficiency in these models comes to bite us?
Charlie O’Neill
Maybe the way I’d break down the two types of tasks — the ones in which models get good and the ones where models will still continue to struggle — is whether the task is cumulative, or whether you have this non-stationary distribution where you have to keep learning and relitigating a bunch of stuff.
An example of a cumulative task might be RSI. It’s theoretically possible to have a less-than-a-million-token Python file which from scratch trains a model that is capable of recursive self-improvement. Every discovery that you make is a line in the sand that you hold. If it’s true that for RSI we don’t need to discover a new attention variant or whatever, then once you’ve discovered attention, and once you’ve discovered mixture of experts, and once you discover GRPO, you just add that to the training stack and that’s there.
A good example of this is 5.6 Sol training, 5.6 Terra, or whichever one OpenAI told us it trained. It didn’t have to go back and discover attention. It basically would have called a bunch of scripts, like pre-training.sh and post-training.sh, and just done that. That’s an example of a cumulative task.
I think the real world — and the reason people are thinking so much about continual learning — is not really a cumulative task. Imagine in a law firm, you have an agent acting as a legal associate. That’s a very non-stationary distribution. You have to be able to fit in your context all the relationships between all the important people at that company, which are also changing all the time. You have all these implicit ways about how things are done, where to find information, et cetera. That’s not as clean an example of a cumulative task as RSI is.
I think there will be this breakdown between tasks. But if the labs realize that — and they do believe that RSI is cumulative in the sense that we don’t need to go back and discover some brand-new architecture or whatever — then maybe more and more effort and compute gets focused on that versus the other tasks.
It’s so unfortunate that RSI happened to be easier than being a paralegal.
John Schulman
I would say today’s models are weaker than humans in a lot of different ways. Some of them might have to do with sample efficiency in a certain regime. In some regimes, models are very sample-efficient, like learning in context. But then there might be some medium-length regime where they’re less sample-efficient, because humans can do some kind of weight update more efficiently than models. I think being less sample-efficient in certain regimes might be one of the sources of weakness.
But I think there are other sources of weakness that are completely different from that. For example, having lower diversity of thought than humans, or being bad at certain kinds of long-horizon judgments. I think a lot of what people call taste is something about behavior that works in the long run, and that people have realized works in the long run. Not everything, but some aspect of taste. Especially for something like software engineering, I think a lot of taste is “What are the systems that are going to be maintainable and work well in the long run of this project?”
There are a variety of weaknesses of models which limit RSI along with other things. Some of them are related to sample efficiency, and some of them aren’t.
Charlie O’Neill
Maybe an interesting thought experiment is this. Let’s say you were able to give a model a context window of a trillion tokens, or whatever you would have needed to fit in your experience prior to, let’s say, RLHF. It’s got all that experience in the context window, and it has the same sample efficiency and in-context learning ability as it does at a million tokens. Do you think taste is then solved? Would it be able to make the same judgments that you did? Or is there something fundamentally missing, apart from just a longer context window with the same sample efficiency?
John Schulman
It would have to be trained to learn from that context. Either it would have to be trained to learn the right update to make from that context, or it would have to generalize.
Charlie O’Neill
So you don’t think you can just dump it all in, your whole life, your research experience?
Beren Millidge
You still need the data to train it on long context. Even if you could theoretically get a trillion context, you would need a trillion lengths of data to train it. Right now you have 10k context you can’t just dump in a million.
Charlie O’Neill
Yeah, I’m just asking if you had that.
Beren Millidge
In theory, I think, yes. This really just comes down to the question of how meta-learnable taste is from shorter-horizon episodes. I feel like there’s no obvious reason it’s super long, because humans somehow developed taste without having many long episodes. We don’t live to be 10,000. We develop pretty quickly.
If you think about even a PhD, the difference between a first-year PhD student and a final-year student or postdoc, that’s five years maybe. They’ve only done maybe 10-30 research projects in total. But somehow they develop taste quite quickly from a relatively short succession of small things. Theoretically, it’s possible to develop it like that.
The AI obviously will have vastly more experience in which to develop taste, to meta-learn it. Then the question is how well that generalizes to really long-horizon things, which I think is really unsolved at this point. We don’t know.
Going back to this question, eventually there should be a regime where AIs are learning a ton from each individual instance of deployment that they have. Currently you could say there’s a fuzzy meta process by which models do improve from deployment. But I feel like it’s a very weak feedback loop. Do you see this on the horizon, where there’s this hive mind kind of learning that’s very rapid, and if so, how exactly does it happen?
John Schulman
I would say that whether we get a hive mind that learns from all of its deployment experience is in a big part about incentives, rather than being a technical question. Companies aren’t going to want to have the model provider learn from all of their deployment, because that might just reduce the advantage of their business.
Charlie O’Neill
I think the economics of this will pressure, not necessarily weight updates to one big common shared model, but modules that get subbed in. A very obvious example of this is a LoRA, but it might be something else.
There’s been a lot of work to try and fit an arbitrary context length into a fixed size. This is all the linear attention stuff. And cartridges, which are essentially KV caches trained to be very, very compressed to fit in a lot of information. That’s another example of something that companies may be willing to sign up for, if that gets subbed into the model and it’s not actually changing the base underlying model itself.
There are many different versions of learning from your data in real time. The latter ones are not really helping the big labs, because they are just these modules. But I think the economic pressure will force the labs to go down that path first before they can embark on this...
Beren Millidge
Which economic pressure, though? I feel like even if you have a bunch of cartridges or LoRAs or whatnot, you can still just take all these traces and dump them into the pre-training of your next generation of models.
Charlie O’Neill
Yes. It may be a more indirect form of learning that the big labs are getting. That’s obviously still really valuable to them. But I can’t imagine a world in which we start off with, “We’re going to just directly train this one big model on all the exact data that we’re getting.”
Beren Millidge
No, I think it will definitely go through stages, because this is assuming there’s one discontinuous event where suddenly we fix weight updates continuously. In practice, I think it’s much more likely to be that the cartridges and stuff allow you to specialize in deployments. Then you generate traces, you put that in your model, and three months later you come out with a model which is better at this stuff. You specialize it again, you consolidate it again, and then eventually we’ll just make this leap faster and faster.
Instead of releasing a model every three months, now it’s every week, and then every day, and then every hour, at which point we’ve basically solved it.
Charlie O’Neill
I think this is a good point as well, because you asked how far off the current paradigm we are from being able to do this. We’ve done a bit of research into this, and people have done a lot of research. At a really large scale, when you wash out enough noise and you have large enough batches, this outer-loop process of putting data into mid-training and creating our own environments does work in some sort of continual learning regime.
But the problem is, when you zoom in close enough at a micro level — I’ve got one model and I’m trying to update it again for a law firm or something, and I’m trying to do that very continuously with a relatively small amount of data — all the methods kind of break down a bit. If I SFT the model on just successful traces, off-policy or on-policy, eventually in the very iterative regime, when you’re doing hundreds of these micro-updates, you see catastrophic forgetting. You see forgetting of previous information learned on top of the base model that was much earlier on, and you see degradation of general capabilities.
On-policy distillation seems to push this horizon out a little bit, but it still eventually succumbs to the same thing. RL is good at getting capabilities in, but it’s not as good at getting knowledge in, this very explicit knowledge of, “Ah, okay, this person does this at this law firm, and this is a very specific process we find.” You have to pour in a lot of compute to create the right environments to get the knowledge in with RL.
Do you think the fundamental issue here — why you get worse at these other skills or there’s forgetting — is fundamentally an issue of capacity or an issue of techniques?
Charlie O’Neill
A little bit of both. I think SFT and even on-policy distillation can be way too destructive. The reason RL is so nice is because it changes a very, very small amount about the model. There’s a lot of evidence for why this is the case. It just tweaks it in this very, very small loss valley to get it into the right point. But that also then limits what you can do with RL, how much you can actually change the model.
So you’re saying the reason this isn’t a winner-take-all, potentially, is that it is just very hard to distill that much information into the base model?
Charlie O’Neill
Without ruining something, in an iterative fashion.
Beren Millidge
It’s easy to distill it into a different base model. This is where I think it’s mostly technique. It’s definitely not that there isn’t capacity. If you had some model with all this data, and you take literally the same-size model and pre-train it from scratch with all of the stuff in mid-training, it will be better. I think that’s a lot of what’s happening today.
There’s very much a bottleneck that stops us from just keeping training the same model forever, versus just getting all the data from the old model and training a new model from scratch. This is exactly as Charlie was saying: some combination of plasticity and catastrophic forgetting. If you just naively train on non-stationary data, because you’re adding new data as you go, this is messing with the data distribution, so the old stuff is just forgotten. We don’t really have good methods to stop that from happening.
So maybe in the limit you’re just bottlenecked by retraining the model from scratch with all this new information.
Beren Millidge
Yes, which of course is very expensive. Training a model from scratch is expensive.
But you’re going to do that anyways.
Beren Millidge
Not necessarily. Maybe eventually, if you have continual learning, you never train a new model. You just have a model and it keeps learning and expanding.
But there might be some deep technical reason why that’s very difficult.
Charlie O’Neill
That’s the question. I think we have pushed back how much from scratch we need to do. It is definitely possible now to take the pre-trained base and do very good mid-training on top of that, kind of continuously, plus some RL from different checkpoints that are later on in the training. That’s looking more like continual learning, but it’s certainly not the case of taking the most recent model, applying a couple of very small updates, and iteratively never losing anything.
Sorry, but I’m a bit confused, because isn’t this literally what happens during training? During post-training or something, you have a model that’s already gone through so much training, and then you distill some fork that’s been further RL’d. Isn’t that literally what happens?
Charlie O’Neill
But it’s still at a large enough scale, I think, that you’re washing out a lot of the noise, and you’re not just focused on one distribution, which, as Beren said, is the issue. If you’re just focusing on one task—
But in the eventual regime you’d be doing… There are billions of deployed instances. You’re learning from all of them at once, so hopefully there’s some washing out of noise from that.
Charlie O’Neill
Maybe at that scale, yeah.
Beren Millidge
As Charlie was saying, you can definitely do continual mid-training for a long time, and you can roll back to a checkpoint and give it new mid-training data. But at the same time, you can’t do this indefinitely. If you just keep continually training the same base forever, it asymptotes at some point. You can’t just learn new stuff in that base. This is why people end up training new bases. Otherwise you would just keep mid-training the same base forever.
01:00:33 – How much progress is explained by data?
Let’s talk a bit about data now. I’m generally interested in this question of how much of AI progress is just explained by data progress. That doesn’t mean it will necessarily be hard to automate, but that’s a separate question. Is there some data distribution which, if you trained current architectures on it, would result in a superintelligence that totally dominates human experts across every single field?
Charlie O’Neill
Are we talking about pre-training plus post-training data, environments as well? I think the existence of this is obvious. It’s just whether we can create the right environment to get there.
Beren Millidge
In the trivial case, we could just train it to output the Python file which trains the actual superintelligence. Just have that memorized in the weights.
Charlie O’Neill
Yes, there’s probably a ladder of RL environments that is possible to construct such that you would get an AI researcher which is at least as good as a human researcher. But the effort to climb each successive rung grows kind of exponentially. Those are the two things you have to trade off against as to how fast we’re going to hit that final rung where it’s better. I think that’s fairly clear.
We’re still relatively early in RL environment creation. There are a lot of asymmetries that we exploit in order to create good environments. One of the asymmetries which we’ve talked about before is that there are environments where it’s easier to go backwards than forwards. What I mean by that is, it’s very easy to define this complex data-generating process, and this is the latent variable you keep hidden from the model. You can generate arbitrarily complex environments, and the model has to do a lot of irreducible token spend and irreducible work to figure out what that data-generating process was.
There are asymmetries in terms of injecting information from the real world. Anthropic finds a bug through tens of thousands of humans and LLMs combined, and turns that into a very, very neat environment which a single LLM could theoretically find within a few million tokens. There are all these asymmetries which we’re cherry-picking, and we’re counting on this kind of task-horizon generalization.
But I think it’s just going to hit diminishing returns at some point, diminishing returns in how hard it is to create those environments in the first place, coming up with them, because you can’t necessarily just have these processes where it’s easier to go backwards than forwards. You actually have to sit down and construct something that looks like a long enough time horizon with humans, and it’s going to be a really complex task to create. Then there are also going to be the compute and time bottlenecks for the agent to actually do those tasks. I think you’re just going to start seeing this curve flatten out.
John Schulman
I saw something about how someone fine-tuned the Talkie model, which is only trained on data up to 1930, on modern coding agent data. It did better than Claude 3 Opus on SWE-bench. So this model that has no knowledge of code whatsoever can be fine-tuned on a moderate amount of data and behave better as a coding agent than this much larger pre-trained model, which is pretty crazy. It kind of shows you that once you have an example of the right expert behavior, it’s actually surprisingly easy to copy that into a relatively weak model.
Charlie O’Neill
But a counterexample to that is a paper recently where they trained a model up to fifth-grade maths, and also primary-school English and stuff, so it was a decent language model. They tried to RL it to do late high school and college maths. The gap was just too large. They couldn’t get it to climb at all. But if you did successive rungs of year 7 maths and then year 8 maths… and so on, you could obviously climb to year 12. Again, it’s just what is the distance between the rungs on those ladders, and how hard is it to create?
Beren Millidge
This just comes back to the RL signal problem. RL is not very good at exploring right now. If the model can’t get it in 128 rollouts, it’s very unlikely to get signal to progress. This is why in RL we need curricula, whereas in pre-training we don’t, because that’s not a problem for pre-training at all.
Charlie O’Neill
Again, pre-training data is different to post-training data. I imagine as we continue on, humans will be involved less and less, but that doesn’t change the fact that you’re bottlenecked on how much signal you can extract from the real world.
There’s a lot of signal in the world, and that’s true. There’s people doing spreadsheet tasks, there’s people doing legal tasks and all this sort of stuff. But at the capability frontier of where the models are at now, how many bits in the world are actually really relevant to improving the model’s capabilities? How many new maths problems are being solved that are just beyond the reach or grasp of the current models? How many new coding problems are being created or solved that are beyond the reach of the current models?
I think that’s why the diminishing returns kick in, because even the world as a whole is not giving you the bits that are useful for tipping you into the next basin of capability.
Beren Millidge
I totally agree with this. It’s really a question of where the signal is coming from. In pre-training, the signal is already in Common Crawl. For the tasks that you care about in pre-training, the problem is not getting signal at all. It’s filtering out all the noise that exists. That’s quite an automatable process.
But as the models get better, as we enter mid-training and post-training, the signal just doesn’t exist anywhere in the original data we have. No amount of filtering will get this. There’s no hidden proof of a Millennium Prize problem sitting in Common Crawl that we can just filter until we see it.
At that point, you have to get bits some other way, either from humans directly, asking them to write out their reasoning, or by creating environments where humans decide what environment should be created and what the objectives of these environments are, or some kind of training on the human data that exists in deployment. You have to get the bits from somewhere.
There’s a question of how much of the progress in pre-training is being driven by data. I did this investigation with Jerry Han, who’s a student at Princeton, where we trained all the recipes from 2019 till now pairwise with all the data sets from 2019 to now. You’re training GPT-2 on the newest data set, like Ultra-FineWeb. You train Delphi, which is the newest open source training recipe, on the Pile or some old data set. You do the whole grid. You see, getting to some level of capabilities, how much less compute does it take, across this grid?
You see that the data seems to explain something like a 12.0x compute efficiency gain, but the architecture improvements explain something like a 3.7x compute efficiency gain, at a very small scale. To the extent that that is true at large scale — that most of the pre-training compute efficiency gains are coming from better data — how much can that continue? Can you keep filtering data more and more and building more and more synthetic data? Do you have a sense of how much this kind of pre-training progress can continue?
Charlie O’Neill
My prior is that, again, the low-hanging fruit is somewhat exhausted. We got the internet as this big block, and it’s not like the internet is necessarily growing at the same rate. All the useful stuff on the internet isn’t growing at the same rate. We’ve probably got a bunch of 0.1% loss drops to go, but definitely not as many as have currently occurred.
But it’s also really interesting that you find this cumulative 33x improvement across both. I think it was Epoch or someone who estimated 3x a year since 2019, which would imply something like 37, over 2,000X improvement. So where’s that missing 100x or whatever coming from? That probably gives you a good signal of how much of this is post-training.
I think the explanation has to be that a lot of the compute efficiency gains are scale dependent, and we’re starting at extremely small scale. That raises a question of whether the data compute efficiency gains or the algorithmic compute efficiency gains have more scale dependence. I don’t know if you have a prior on that. We just didn’t have enough compute to investigate that question.
Charlie O’Neill
Just naively, theoretically, the scale dependence of the architecture is fairly well known, and you can fit a straight line to it. Whereas I would have no idea how to do that for combining pre-training plus post-training data and mid-training data.
Beren Millidge
Funnily enough, I feel like data is actually more important with scale. I feel like architectures are kind of a one-time thing.
Saying just an X% efficiency gain is kind of misleading, because what an architecture does is let you reach a qualitatively new regime which you couldn’t reach with the old architecture. Within that regime, obviously the data is the primary thing determining it.
But if we didn’t have even GQA, if we were doing full attention all day, it would be ridiculously expensive to do a million context. Because of that, we could never use the data which is actually at a million context, so we couldn’t get these capabilities. Even though if you just do a naive “how much does this do at 2K context”, where the architecture isn’t unlocking anything, then the data will look much more important than in some sense it is. It’s unclear to me that these things are really just multiplicative gains in this way.
I see. So what’s your take on the scale dependence of data?
Beren Millidge
On scale dependence, I think a lot of the mid-training and post-training data we have now actually gets better with scale, because a lot of it — the very long context horizon environment stuff — really requires big models to be able to make use of it. If you try and train your 100 million parameter model on SWE-bench traces, it’s not going to get anywhere. It’s not going to show you the same kind of improvement that you would get if you train an actual sensible size model on it.
Charlie O’Neill
It’s hard as well now because so many of the architecture changes — you look at Kimi, for instance, or DeepSeek — they’re doing these architectural modifications not just with dropping the pre-training loss in mind, but with how the models are going to be used in the real world. The inference efficiency, having some form of compressed attention in the DeepSeek models, is not necessarily geared around a fundamental trade-off improvement. It’s just, “Okay, we’re considering how the models are going to be used.”
One question I’m curious about, to understand the future, is how parameter scaling will go as we’re getting into more of an RL-heavy regime. You can look at open source architectures and see how fast parameters have been scaling. Maybe it’s roughly 2x every year for frontier open source models. To the extent that even frontier closed source models have 100B or 200B active parameters, do you think that keeps 2x-ing year over year?
Or, now that we’re in an RL regime where you also want to conserve compute on rollouts… Also, maybe there is a threshold effect where you have enough capacity and at that point increasing parameters arbitrarily doesn’t matter as much. Do you guys have a sense of, in 2030, how many active parameters a frontier model will have?
Charlie O’Neill
I think for the next few years, because we are so focused on doing longer and longer horizon rollouts for RL, where inference efficiency matters a lot, it feels like the models aren’t necessarily saturated on their ability to do that. The bottleneck is still the environments. So we might see a little bit of a plateau.
I have a feeling that Mythos and the GPT models are much smaller than the 10 trillion parameter range that people are talking about. Even just naively comparing them to open source models, you can probably back out that conclusion.
Probably for the next few years, I wouldn’t imagine a huge growth in the number of parameters. But again, there’s so many different things to trade off here. You decide the size of your model based on how much pre-training data you have, and then the difficulty of the RL environments that you’ve got to train on. You ideally want to get to the optimal point where you can get a decent pass@1 or something on the hardest environments you have. It wouldn’t make sense to make a bigger model pass there, because then you’re just paying much more inference than you need to.
So a lot of it depends on how quickly Mercor and the in-house teams can scale up the complexity of the RL environments they’re training on.
John Schulman
I would expect the models to keep getting bigger just because people are scaling up compute and the GPUs are getting bigger. But exactly how much they get bigger depends a bit on the scaling laws in non-obvious ways.
One thing is that I think data efficiency is going to be a bigger driver than compute efficiency of the exact architectures people use, now that we’re getting to the regime where we’re running low on high quality pre-training data. That might affect how sparse you want to make the model.
I also think we don’t understand sparsity that well. Parameters are a different resource than active parameters. Sparsity has definitely increased a bit, but it’s not clear that it’s going to keep increasing without bound. There might be some kind of sweet spot.
There’s an argument that sparsity should make data efficiency worse, because you might have to learn the same thing on multiple experts, though that’s debatable. I don’t think we have a good enough theory of scaling laws that we really understand why sparsity is helping, how much it’ll help, and if that’ll plateau at some point at a certain level of sparsity.
Sorry, can you spell out exactly what the implication of data efficiency would be on parameters? It sounds like you’d say there should be less sparsity, but what are the other implications on parameter scaling?
John Schulman
Just that with the scaling law, you’re not trying to optimize compute efficiency. You have all your choices you can make on the architecture. Each of these gives you a different scaling law. Traditionally, you would look at some kind of envelope based on compute. You would look at performance versus compute and take the envelope of the best models.
But if we’re making that decision based on data — we’re assuming we can spend a lot of compute, so data is on our x-axis instead of compute — then we just get a different set of optima, or a different set of models that are on that frontier.
Charlie O’Neill
I also don’t think that we’ve necessarily doubled the size of the models every year for the last few years. People have been training 1 trillion parameter models for at least a few years. There was even an open source one called Falcon. Liam from Periodic Labs, I think, posted yesterday on Twitter about how an early experiment was training a 1 trillion parameter model that was very, very sparse.
John Schulman
That was what they did before OpenAI, at Google, the Switch Transformer.
Charlie O’Neill
It was very, very good at knowledge but terrible at reasoning because it was so sparse. It feels like we’ve been playing in this 100 billion up to 2 trillion parameter range for at least a little bit. It certainly hasn’t been this nice linear increase.
Beren Millidge
I feel like there’s two things. As Charlie was saying, inference efficiency is super important for RL rollouts. This will really push down active parameters quite a lot.
I think the total parameters really depends a lot on the hardware as well. You really need very high memory bandwidth and VRAM size to actually be able to serve multi-trillion parameter models. Right now, people are still using a lot of H100s and stuff. As everyone moves to GBs and then Vera Rubins, we’ll get more of the ability to scale and actually serve and do large RL inference at larger scales.
The data question I think is interesting, because naively, larger models are much more sample efficient in the actual data points. Even if you’re not saturating the model, it’s still better to go bigger, because larger models generalize better and get to a better loss for the same amount of data. Right now I think we have a lot of data, and that’s not the constraint. Compute is. So we’re having smaller models which are very inference efficient. But if compute is no longer the bottleneck, it might come back to larger models which are undersaturated, but have this generalization ability because they’re much larger.
If you just look at the basic Chinchilla scaling law and you just maximize out parameters, it actually decreases the amount of data you need to get to the same loss very little. If you go to infinity on parameters, the amount of data you need I think goes down less than 10X, just because of the nature of the power law.
Beren Millidge
But we’re now on the way-too-much-data side of the Chinchilla laws. Right now we over-train models according to Chinchilla. So we could easily get back to a point where, as we’re running out of data, we move back to the Chinchilla optimal point, or even a bit on the under-training model side.
Charlie O’Neill
But surely, even with these new chips that come online, we’re just going to be so compute bottlenecked for the next few years that that won’t necessarily be the case.
Beren Millidge
This depends on the ratio you have of training and inference compute, really. If you’re super bottlenecked on data, not on compute, you should go bigger. If you’re super bottlenecked on compute, you should always go smaller. You can also use computer-generated synthetic data, so it’s one of these very hard things to predict.
John Schulman
I think part of the reason it took people so long to figure out the scaling laws in the first place was that if you don’t get all these things right, then you don’t get such a clean relationship. The beautiful straight lines on graphs hide a lot of complexity in how you have to make sure to scale every hyperparameter the right way, or parameterize your optimizer in a way that scales and where you don’t have to change your hyperparameters as you change the model size.
Charlie O’Neill
Bugs have their own clean scaling laws as well. Like with Kaplan forgetting the cosine annealing thing, or even just not considering embedding parameters, I think. That messed up the estimate at smaller models because embedding parameters are a decent size of the model.
01:18:03 – Why is RL working so well?
A bit on RL. A year ago, a lot of people were making this argument that RL will not be super successful at scaling for models. John, you wrote a research paper where you were pointing out that models learn one bit per episode when you RL. They learn, “Did I get the answer right or did I get it wrong?” Then I wrote some blog posts earlier this year where I was like, “It’s even worse than that,” because when the pass rate is low and the model is very unlikely to get the answer right, it learns almost nothing at all from an RL episode.
But I look at the models today, and they seem pretty smart. It seems to be the result of scaling up RL. Beren, you had a post a few weeks ago where you were trying to explain what’s going on. Why has RL been more successful than one would have naively thought?
Beren Millidge
I think the success of RL comes down to a bunch of different things. First, what is slightly underestimated is the mid-training. An awful lot of what we see as successes of RL actually comes from very, very good mid-training data, which is where we’re essentially doing pre-training but on synthetic reasoning data and the kind of environments that get the model warm-started for RL. This takes the model almost 80% of the way to the final RL checkpoint often.
Then what RL does on top of that is essentially tweaking the policy. This is one of the reasons why it doesn’t need as many bits as you would naively think. It doesn’t have to learn all of these behaviors from scratch. It needs just a few bits from these episodes, which you do get. The other thing that I point out in my blog is that these bits are extremely high signal compared to regular pre-training, which is why you need RL at all versus just SFT-ing on successful reasoning traces.
Because it’s exactly the bits about how to get the answer right.
Beren Millidge
There’s two things. Yes, one, it’s exactly the bits about how to get the answer right. But this is not exactly how you think of it, because in SFT, you have a trace. You have, say, a bunch of math reasoning and then the answer at the end. The bit is still there. You still SFT on the answer token.
What’s important is that the objective ignores all the other bits. In SFT, you have to try and match the exact reasoning tokens that the model produces. You’re essentially getting too many bits about the exact way this other model you’re training on reasons. For RL, you only get the one bit. That means that signal is not drowned out in the noise of all the other bits the model has.
It’s really a super dramatic increase in the signal-to-noise ratio during training, which is why RL is so dramatically efficient in terms of steps.
Charlie O’Neill
There’s been so much debate about what RL does to the model versus mid-training or SFT or whatever. Everyone talks about how pass@1 will go up, but pass@256 will go down. Very rare correct reasoning traces will be down-weighted and outweighed by a gradient signal from easier reasoning traces.
I think the simple way to view RL now is that if you have a large enough amount of compute to sample a large enough group size — such that your probability of getting a bunch of correct answers is past some not insignificant probability — then it will be up-weighted. To Beren’s point, mid-training and more pre-training — the pass@1, the starting point for RL — scales in a log number of pre-training tokens.
Can I ask some very basic questions? That answer makes sense, and maybe there’s empirical research which shows that this is what’s happening. But then I just look at the models themselves… I don’t know what’s happened. Maybe you can give me a sense of what is the basis of the AI progress over the last year.
Maybe it’s just up-weighting the policies which were going to do the correct thinking anyways. But it just seems like qualitatively, the models have gotten so much more capable. Maybe there’s no inherent contradiction there. But how do we square the relatively small impact this take would imply that RL would have with the actual qualitative capabilities the models seem to be gaining?
Beren Millidge
One thing I want to point out here is that it doesn’t necessarily imply that RL has a small effect. Even if you have a few bits and you only change the parameters a small amount, the actual impact on function space — the input-to-output mapping the model learns — can still be super dramatic. Even one bit can change your function space a lot. It can rule out half the hypothesis space, which is huge.
I don’t think it’s necessarily the case that small amounts of bits, small amounts of RL, once you’re starting from a really good point, means that you don’t have dramatic impacts in behavior. At least… not necessarily.
Charlie O’Neill
I think it comes down to two things. The first thing is that everyone was hoping that RL would generalize this reasoning across all these different domains. I don’t think we necessarily got this horizontal generalization. Just training on math doesn’t necessarily make you the greatest coder. You do have to do RL on code environments.
I think what we did get, though, is horizon generalization. The models just learned how to use more tokens for longer and still make progress on some sort of task. You can train on environments where they get longer and longer and then put them into a completely new environment. Yes, they may not have generalized the reasoning patterns which allow them to do well in that environment, but they’ve at least generalized the ability to continue on that task for longer, which is correlated with success. There was a paper called EdgeBench which showed that the rate at which models can work for longer is doubling every three months. That’s clear evidence of generalization.
The final way to think about it is, in pre-training, there’s this idea of quanta. You have this very smooth pre-training loss curve. When you look at what’s happening in the model, the model is learning all these very discrete tasks, and there’s all these emergent points where there’s a phase transition. It didn’t have induction heads, now it has induction heads. There’s tens of thousands, millions, probably hundreds of millions of these things. You average them all together and you get this very smooth loss curve.
To an extent, a similar thing is happening for RL. There is this very slow outer loop, as Beren mentioned. We will train a model and then RL it, and then in the next model iteration of training, we will dump a bunch of these synthetic reasoning traces into the mid-training data. We’re kind of hitting all these quanta for all these different tasks, and on an individual task level, it may look like a phase transition. You’re suddenly going from a 0.5% pass rate to a 90% pass rate on a particular finance task or Excel task or whatever.
But you average all these things together, plus the horizon generalization, and you kind of go, “Wow, we’ve got qualitatively better models.”
Beren Millidge
I think a lot of this as well is just… RL does generalize a bit. You get some transfer between math and code, or puzzles and math and this kind of stuff. Also, the sheer amount of environments people are targeting is just vastly greater. Before, when you tried to do some task which you do in your daily life, two years ago, the labs wouldn’t really care about this. They wouldn’t train the model for it. Now it’s just so much broader. They have a lot of environments targeting this specific thing.
01:24:54 – Move 37 and entropy collapse
Earlier in the conversation we were talking about RL in the context of causing this entropy collapse, or just concentrating probability on solutions the base model had already done, and causing relatively sparse updates in the policy.
But I think there’s also another story about RL, which is going back to the Atari games and then AlphaGo coming up with move 37, the super creative move. Because it was never initialized on human data, it can think in ways that humans are not even thinking and come up with extremely creative solutions.
Do you have a sense of when we should expect, or if we should expect, RL on LLMs to result in things like move 37, extreme creativity even beyond human creativity, because there’s just de novo initialization of intelligence?
Beren Millidge
A couple of things here. First off, I think that AlphaGo is using MCTS, which obviously does more exploration and stuff than regular policy gradients. But I also think that RL doesn’t necessarily reduce the creativity.
This is obviously qualitative, but if we look at the OpenAI-Hugging Face incident, these models were coming up with multiple zero-days at a time to break out of the sandbox. This is clearly some level of move 37 creativity already, which we just get from the general generalization properties of the LLMs. It’s definitely not the case that RL is totally destroying entropy, especially on long horizons.
John Schulman
One thing that people call creativity is just solving hard search problems. Move 37 is obviously an example of that, or writing some kind of poem that satisfies a ton of different constraints. That’s something AI is obviously going to be extremely good at, if trained for it.
Then there’s another way in which the diversity of the models’ outputs is a lot lower after RL, and they develop these tics. Even though the models seem like they’re good at writing, when you do some kind of distributional analysis, you find that they’re reusing certain themes all the time and they’re using the same character names all the time. You’re not getting the same kind of diversity that you get from human authors. You’re getting one really good style. So I think that kind of diversity has definitely been cut down by RL a lot.
In fact, since we were talking about distillation earlier, one thing that’s happening is that so many people are distilling, mostly from Claude, that all the open-weight models write the same way as Claude and have the same tics. This seems kind of concerning to me, that we’re having this monoculture emerge.
Beren Millidge
Again, I don’t think this is fundamental to RL as a method, though. The same with distillation. Even with distillation, you’re just training on the data. Just because your data is not super broad, that doesn’t mean the training method itself is somehow wrong. It’s a problem with the data.
I think a lot of the RL entropy collapse, for instance, is basically due to exploitation of fairly simple verifiers when you don’t have a huge diversity of environments. The writing, for instance, is presumably graded by some judge. The judge has some specific tics, and the model is learning to reward hack the judge, and that’s why it collapses. But this is really a problem with the judge. It’s not a problem with RL in general.
01:28:32 – Rapid-fire timelines
Okay, super rapid-fire predictions about the future. I want timelines on the following couple of questions. By when do we have models which… Here’s what it feels like to a user. You basically hire it as a drop-in remote worker for all kinds of white-collar work? Not just coding, but video editing, law, paralegal, et cetera. It’s literally an actual remote worker, with full computer use, with literally a month of seamless learning and operation, executing on complex projects that require interacting with other people, et cetera. Everything a human worker could do over a month.
Charlie O’Neill
If you mandate it to use a browser or whatever — rather than the firm setting up the information to be programmatically accessible — maybe a couple of years. But if it’s not browser-based — it can send Slack messages, it can do all this stuff — I’d still probably say around a year.
Beren Millidge
I would say maybe three years for the full generality. But to Charlie’s point, we will end up with a lot of people making their organizations easier for the AIs to use, and so you get 80-90% of the way there before that.
Sorry, but the diff between one year and three years there is just literally…
Beren Millidge
I think there’s going to be a long tail of miscellaneous stuff which some human can do, which will take the models quite a while to do.
Are you thinking of computer stuff or basic cognitive capabilities?
Beren Millidge
I think this really comes down to a question of how quickly we can solve this kind of online learning, and whether we can get 80-90% of the way there with compaction and writing files to yourself and stuff. That’s my big uncertainty. I really don’t know.
Charlie O’Neill
An example of something that it wouldn’t be good at is if I have to yell at someone to get something at work, or really push someone to get something done. The model just isn’t going to do that. It’s going to be too nice.
John Schulman
I’d say there’s a wide variation in quality of human remote workers. If you try to hire someone off of Upwork to do a software engineering project, there’s going to be a huge variation. It’s often quite hard to get them to do a good job or pay attention to all the feedback you’re giving. I would guess that in some cases, the pre-AI version of this was worse than what you can get now from existing AI.
I think it might end up being a little complicated, because to some extent we already have this for some not-so-high-quality work. But then obviously we’re not matching human level in certain higher-quality forms of work. But I basically agree with Charlie and Beren that maybe we’ll have some version of this in a year or so that’s okay. We’ll have that form factor, and it’ll be able to do some things really well, some things not so well, and things will be improving from there.
Charlie O’Neill
We shift the goalposts based on the very long tail all the time. I feel like you’ve used this example before of doing your taxes or something. This year, I literally just told Codex to go get everything I needed and send it to the accountant. There was this massive list of stuff it had to use computers to click through and download. It did it. It was perfect. A lot of this stuff it can already do.
Okay: give you 10x total productivity uplift. Basically, if it takes you a year to make a breakthrough now, you make a breakthrough every month.
John Schulman
I think I would just refuse to give you a scalar on this. We might already be past that in some types of work. Let’s say you’re trying to do certain types of math, and—
Oh, sorry. But for you as AI researchers trying to advance the state of AI research. How much are AI researchers sped up or uplifted?
Charlie O’Neill
Somewhere between 5-10 years?
Oh, really? Okay, that’s far away.
Beren Millidge
Really, you think it’s longer than for a general remote worker? Interesting.
I’m realizing you probably have a very different definition of a fully general remote worker. I could have specified that earlier.
Beren Millidge
This is true, because obviously an AI researcher can be a remote worker.
Charlie O’Neill
I’m picturing normal white-collar work over the period of a month. I think it starts to diverge a little bit past two months.
A very competent white-collar worker, but not necessarily a super creative researcher.
John Schulman
I would say two years.
Two years? 10x? Okay. How about you, Beren?
Beren Millidge
I can kind of see that, actually, because right now it’s already definitely more than 10X for coding stuff. So if it can do even one or two loops of experimental feedback, that would actually be massive already.
So 10x uplift of AI researchers within two years. If you plug that into a very naive model of AI progress and how much is coming from AI researchers, and there’s a 10x increase in their productivity, you have a radically accelerated pace of AI progress starting two years from now.
Beren Millidge
I think this will mean that AI progress doesn’t get bottlenecked on AI researchers’ ability to run small experiments. It gets bottlenecked on other things.
Of course. But it just happens 10x faster, which is a huge deal. That also helps the next thing, which gives you a 100x speedup, happen sooner, et cetera.
Charlie O’Neill
I’m happy to just take a bit longer on that one.
What’s the crux?
Charlie O’Neill
My capacity to absorb information and make the Bayesian optimal decision on the next experiment.
Beren Millidge
I’m assuming that you can delegate some of this to the AI. The AI is becoming decent at deciding. It’s run this experiment, it’s got this result, it runs the next experiment. If it can run two or three experiments in a row without crashing, then that is actually a big uplift.
Okay, final question. An AI which dominates top human experts across every single field of work that can be done over a computer. So not only AI research, but all cognitive work. Not just short-horizon work, but literally, if it takes three years or something, the AI will still do better than humans.
Beren Millidge
This is basically just ASI?
John Schulman
I would say 3-4 years.
The fuck? I mean that doesn’t seem wrong, but—
John Schulman
AI is obviously getting more attention. It’s one of the harder things, but a lot of energy is being put into it. It’s also not one of the hardest things for AI, because it involves a lot of code and math, which models are really good at.
For things that involve 3D and spatial stuff and physical stuff, I think that will take a little longer. If it’s mechanical engineering or something, and it’s not getting the most attention right now, that might take a little longer.
But it also does include fields where there is relatively little data because of the nature of the field, and it has to learn that data on the fly. For example, it has to become superhuman at being an engineer at TSMC or something.
John Schulman
So you would have to assume that you can give the AI the same onboarding material. Then something has to be solved about longer-horizon learning.
Charlie O’Neill
I’d say 5 to 10.
So basically, you think automating AI research is ASI-complete or something?
Charlie O’Neill
Yeah, I think so. I think there are so many things in the world where, even if you have some sort of memory system external to the model, and even if context length grows a little bit, there are just fundamentally things where, even if you could research the information or write notes yourself, you’d need more than a million-token context window today.
Beren Millidge
I kind of agree on the 5-year range, at least for the stuff that labs are focusing on. But I think there’s going to be a long tail of stuff which the AI could theoretically go out and learn about, but no one has bothered to do it and the compute hasn’t been allocated to that. So that might take longer for literally every single human expert.
Sorry, but by this I also included the ability to learn a new domain as fast as a human.
Beren Millidge
I think that’s not necessarily necessary, because the AI will have vastly greater experience than any human.
Thanks so much for doing this, guys. I feel like this was a great format for getting different experts to disagree and debate and discuss things together. It was very productive.
John Schulman
Thanks for having us.