学习范式与路线

强化学习消耗了 10,000 多张 GB300 长达四周的时间。因此，实际上花在强化学习上的 FLOPs 比预训练还要多。强化学习任务要复杂得多，因为你既需要利用各类智能体沙箱进行大规模推理，又需要同时进行训练。

reinforcement learning was a little over 10,000 GB300s for four weeks. So, actually there are more flops spent on reinforcement learning. Reinforcement learning jobs are, you know, more complex in the sense that you do both a lot of inference at scale with all sorts of sandboxes for agents and training.

语境：Misha Laskin 披露 Beam（500B 总参数、23B 激活 MoE）的算力分配细节：后训练强化学习在集群规模与运行时长上已全面超越预训练阶段。

支持 → source-41e2110c2ba788c4

支持 → dwarkesh-podcast-ajeya-cotra-20260901

事实证明，为了让强化学习在大规模下取得良好效果，你确实需要自己对模型进行预训练。你不得不这么做，这两者之间的耦合太紧密了，你必须两者兼顾。

it turned out you actually do need to pre-train your model in order to make reinforcement learning work very well at scale. So you actually kind of do need to do that the things are just so tightly coupled that uh you you do need to do both.

语境：Misha Laskin 解释为何 ReflectionAI 不能只做微调或后训练：基础预训练与强化学习表征存在深度协同，必须端到端闭环开发。

支持 → dwarkesh-podcast-ajeya-cotra-20260901

依然处于 GPT-2 时代，原因在于：要达到 GPT-3 的水平，我们首先需要少样本学习能在大量不同任务上真正奏效；其次机器人学至今依然极度受制于环境迁移——如果一个模型只能在你那台具有特定配置的机器人上运行，它能算得上是通用大脑吗？

still GPT-2 and here's why. I think for GPT-3 we need firstly few shot learning to work really well for a lot of different tasks and secondly also with robotics weirdly is still very subject to cross environment, right? If it just works on your robot with your specific setup, is it really a generic brain?

语境：Keerthana 评估具身智能现状，指出机器人模型尚未迈过泛化门槛：不同物理本体和硬件传感器的异构性，使得通用具身大脑远比纯数字 LLM 难以泛化。

支持 → a16z-fei-fei-li-20260904

支持 → the-ezra-jensen-huang-20260923

通过极少量的数据——仅用 200 个示例——你就能让它在许多截然不同的机器人身体上学会大量任务，因为你已经具备了对物理世界进行高度通用理解的基础底座，并懂得如何对其进行控制。

You can with very, very small number of examples, with 200 examples, you can learn a lot of different tasks on a lot of different bodies, because you already have the foundation to do a very generic understanding of the physical world, and also like how to control it

语境：介绍 Gemini Robotics 端侧小模型的少样本跨本体迁移能力：预训练模型赋予物理常识后，仅需几百个演示即可适配新本体。

支持 → sequoia-sutton-20260818

我们所认为的语言模型——也就是我们向其发起查询的那个东西——未来很可能实际上是一个蜂群、一个脚手架，或者某种扩展性极佳的奇特外周框架设计，只不过用户根本感知不到底层细节。

what we think of as a language model like the thing that we query might actually be like a swarm or like a scaffold or like some weird harness design that scales very well, but the user just doesn't see it.

语境：Alex Zhang 预判大模型形态演变：单一 Transformer 前向传播终将被底层多智能体递归与抽象脚手架取代，对外则维持统一的模型接口封装。

支持 → latent-space-thariq-shihipar-20260929

支持 → the-every-sam-altman-20260930

我这篇博文想要表达的核心是：我们应当重新思考外周框架（Harness）的角色，因为框架实际上能极大地提升模型的泛化能力以及有效利用的数据量。而且这种特性绝不仅限于递归语言模型。

what I'm trying to get at with this whole blog post is just that like we should rethink what the role of the harness is because harnesses can actually greatly increase the generalization capability of your model and the amount of data that it's given. And this is not exclusive to RLMs.

语境：反思社区对通用框架的轻视：优秀的外周脚手架引入强归纳偏置，能用小规模环境数据实现向更长、更难任务的递归归纳泛化。

质疑 → sequoia-sutton-20260818

人们所恐惧的超级智能版本是一个独立模型，它在所有任务上的表现都超越专家，且不依赖任何人。我看到的模式是：各种新类型的数据都会被拉到上游注入到最大的前沿模型中，然后高效的专用专家模型再从中被蒸馏出来。

And the version of super intelligence people fear is a single model that does every job better than the specialists and depends on no one. My answer was the pattern as I see it so far. New kinds of data get pulled upstream into the biggest model and efficient specialists get distilled back out of it.

语境：Nathan Labenz 论证全功能通用大模型与领域专用小模型的关系：单一全能教师模型在前沿统摄所有模态，专用模型仅作为下游蒸馏产物。

支持 → youtube-daniel-kokotajlo-20260908

有人曾问我，随着底层模型能力不断变强，提示词优化的重要性是否会下降？我的观点恰恰相反：模型越强，其指令遵循能力就越敏锐；当你给一个极其聪明的模型提供关于具体任务愈发精确、高质量的指令时，它在解决该任务上的表现就会呈指数级跃升。

people have asked me that oh as models get better the importance of prompt optimization will go down. I argue the opposite which is as models get better they will get better at instruction following and the more precise instruction about your task that you have to give to a very smart model the better that model will be at a uh solving your task

语境：Lakshya Agrawal 驳斥「模型越强提示工程越无用」的流行论调：更聪明的模型具备更精细的上下文理解力，任务指令与引导信息的精确度反而成为拉开性能差距的乘数。

支持 → all-in-satya-nadella-20260915

我们的核心思路是在文本空间中执行反思式优化：不再仅仅依赖 0 或 1 这种粗粒度的标量奖励信号，而是让大语言模型或智能体审视整个执行轨迹，并深入反思其中哪些步骤有效、哪些无效。

Our idea is to perform reflective optimization in text space where instead of only using the zero or one reward signal, we can have a language model or an agent look at the trace of the entire rollout and reflect on what worked in them, what did not work in them.

语境：Lakshya Agrawal 阐释 GEPA 与传统 RL 的本质区别：传统梯度回传抛弃了中间工具调用与报错诊断，而文本反思能直接提取高阶语义诊断信息实现跳跃式改进。

支持 → ai-engineer-lakshya-agrawal-20260926

我们真正关心的最终结果是系统整体能力的实际提升，而非特定具体层级。当然各层都有改进空间，而且平心而论，当最终走向拐点拐弯的 Level 3 终极递归自我改进时，所有不同层级之间必然需要建立起完整的自指涉闭环。

actually improve the capability of the system not uh a specific layer. uh of course there could be argument that there's more to be do on different layers and to be fair I think eventual level three recursive self-improvement like when we actually reach the inflection point I think there need to be this uh self-referential loop on all different layers

语境：Zhengyao Jiang 阐述 RSI 的系统论视角：不必执念于必须更新神经网络权重，只要系统工程层面跨层实现了端到端自主能力跃迁，即构成了实质意义上的自改进。

支持 → all-in-satya-nadella-20260915

我们常说的一件事是：随着我们不断扩展视频模型，仅仅通过 Scale 视频模型，我们意外地为机器人领域创造出了一个顶尖模型。我们在去年年中意识到，机器人实验室纷纷主动找上门来，希望将视频模型用于生成合成数据，并让我们针对机器人场景对视频模型进行后训练。

One thing we like to say is we as we scaled video models, we accidentally uh created a state-of-the-art model for robotics uh by just scaling video models. So we realized at some point um mid last year that robotics labs started coming up to us and kind of asking to use video models for synthetic data kind of asking us to post train our video models to work really well for robotics

语境：Anastasis Germanidis 披露 Runway 开发 GWM1 时的工程惊喜：视频规模化预训练直接沉淀出了对物理交互的通用感知表征，无意中成为机器人策略的核心基础设施。

支持 → cognitive-revolution-keerthana-20261003

如果你拥有一个交互式模型，你希望它能够生成反事实结果——例如如果我执行动作 A 与执行动作 B，它能生成同样逼真的不同结果。我认为这就是视频模型与世界模型之间的重大鸿沟：反事实推演能力。如果你想要一个出色的机器人世界模型，你必须能够极其逼真地模拟动作的失败。

if you have an interactive model, you wanted to be able to generate counterfactuals like if I take this action versus this action, you wanted to generate equally realistic outcomes. Uh so that's I think the big gap between video models and and and world models is that idea of the counterfactual generation. And if you want a great model for robotics, you want to simulate failure very well

语境：Anastasis Germanidis 辨析世界模型的技术本质：视频模型倾向于生成顺畅的成功画面，而具身策略评估与强化学习闭环则高度依赖对「失败」的反事实精确模拟。

支持 → cognitive-revolution-keerthana-20261003

这一时期的绝大多数视觉 AI 仅仅是靠像素效果来被评判的。如果你生成了漂亮的像素，你就做出了看似精美的软件，但这些像素仅仅是在尝试模拟观测现象。它们并不实际具备或表征现实世界背后的任何语义。因此，它们根本无法带来真正的真实性，无法作为规划的依据，也无法作为对真实物理世界的稳健仿真基础设施。

most of visual AI for this period has been just judged by the pixels. If you've got beautiful pixels, you've got a beautiful piece of software, but these pixels are trying to simulate observations. They don't actually have or represent any of the semantics behind the world. And therefore, they aren't good at realism. They aren't good for giving a basis to plan. They aren't good for actually having a robust simulation infrastructure of the real world.

语境：Christopher Manning 尖锐批评视觉生成与世界模型混为一谈的现状：像素级的逼真无法替代底层因果规律，缺乏因果表征的模型无法用于智能体规划。

支持 → latent-space-anastasis-germanidis-20260925

支持 → ai-engineer-deepak-pathak-20260924

这种神经符号方法你可以称之为保守路线：在动作的生成与控制中明确借助物理引擎和物理学知识。当你面对一张水的图像并希望理解其运动规律时，你是通过物理模型来预测其行为表现的。

this neurosyolic approach is you know you call a conservative approach if you will you know that this is absolutely using physics engines and knowledge of physics um to in its generation and control of movement, right? That if you're what you know when you've just started off with an image of water or and you want to understand how that's behaves, you're using a physics model to predict how it's going to behave.

语境：Christopher Manning 阐述神经符号系统在物理建模中的必要性：将已有物理知识显式嵌入生成与控制循环，而非寄希望于黑箱网络完全隐式重学物理学规律。

支持 → latent-space-anastasis-germanidis-20260925

我们的核心论点是构建所谓的全本体通用智能：任意机器人、任意任务、同一个通用大脑。无论是双足人形、四足机器狗、传送带上的机械臂还是灵巧手都无所谓。这种假设甚至比人类的控制能力更为通用，因为人类只能控制自己的身体。

our thesis is is to build what we call an omniodied intelligence any robot any task one brain okay any robot it can be a humanoid it can be a quadriped it can be a robotic arm on a conveyor belt or a dextrous hand doesn't really matter this even this hypothesis is way more general than uh one would argue humans are

语境：Deepak Pathak 提出 Skild AI 的跨本体通用基础模型路线：打破传统按单一形态定制机器人的范式，用单一通用策略统一多样化硬件。

支持 → latent-space-anastasis-germanidis-20260925

支持 → cognitive-revolution-keerthana-20261003

端到端是 AI 许多领域被大量使用的词汇。但当我谈及端到端时，我是指真正的端到端：模型直接读取摄像头视觉输入，并直接将动力施加到各个电机上。除了最后用来将频率从 100Hz 提升到 500Hz 的经典比例控制器（P）之外，中间没有任何模块。

Now end to end is heavily used term in in many areas of AI. But when I mean end to end, I really mean end to end. It reads directly from the cameras and it applies power directly to motors. So we use nothing uh in between except just a P at the very end to go from 100 to 500 Hz.

语境：Deepak Pathak 解释其控制策略架构：彻底抛弃传统的感知-建图-规划分层管线，实现从像素到电机电流的纯神经网络闭环直驱。

支持 → cognitive-revolution-keerthana-20261003

只要继续训练这些模型它们就会一直变聪明，事实并非如此。[…] 这正是为什么必须出现第二缩放定律的原因。[…] 第二缩放定律就是测试时扩展，也就是推理。

It is not true that if you just keep training these models they'll get better. […] Notice it is the reason why the second scaling law had to come along. […] The second scaling law is test time scaling, inference.

语境：黄仁勋解释前沿大模型能力进展的拐点：单纯堆叠预训练算力与数据的边际收益已在递减，推理阶段的思考时间与算力扩展成为维持模型进化的关键。

回应 → sequoia-sutton-20260818

TabPFN 是 AutoML 演进的自然终局：我们学习的是整个能在单次前向传播中执行的完整学习算法；根据你输入的具体数据集，它会在上下文内部拟合出完全不同的分类器。它不再是单纯的一个静态分类模型，而是一个能根据输入即时生成不同分类器的超网络。

TPFn is really this natural progression of AutoML where we learn this entire algorithm that is executed in a forward pass and depending on the data set that you feed in it learns a different classifier. So um it's not just one network that is one classifier but it's a a network that can learn different classifiers depending on the input

语境：Frank Hutter 阐述 TabPFN 的本质跃迁：从传统反复试错的模型训练演变为上下文学习算法，将整个机器学习拟合过程转化为 Transformer 的单次前向推理。

支持 → all-in-satya-nadella-20260915

我们依然可以在预训练阶段针对神经网络架构和学习率等超参数执行自动化搜索；但在测试与应用阶段，所有冗长繁琐的调参过程全部消失了，数据科学家终于得到了梦寐以求的工具——一个毫秒级极速且真正开箱即用的建模方法。

So we can still do a search over hyperparameters of the architecture of our neural network. We can search over our learning rates etc for the pre-training but all of that at test time falls away and data scientists get what they always wanted namely a just really fast method that works out of the box.

语境：Frank Hutter 总结表格基础模型为数据科学带来的效率颠覆：繁重的调参成本被前置固化到基础模型预训练中，终端使用者无需任何特征工程即可获得最强性能。

支持 → machine-learning-frank-hutter-20260923

预测模型就像是：你输入一些数据，得到一些输出……那仅仅是一个统计模型。而解释性模型才是科学真正追求的目标——它必须能够进行外推，因为它内部捕获了物理规律或对底层现实的某种真实描述。

a predictive model is like let's say you just have a you have some inputs and you have some outputs […] That's just a statistical model, right? A descriptive model is actually what science is trying to get to, which is okay, it should be able to extrapolate because it has sort of the physics or the actual some description of reality that's captured within it.

语境：Google Fellow John Platt 剖析统计机器学习与真实科学研究的根本分界：拟合历史数据的预测模型只能内插，而科学的核心是发现包含物理因果机制、具备跨领域极端条件外推能力的解释性模型。

支持 → youtube-terence-tao-20260828

限定 → youtube-demis-hassabis-20260911

这些 AI 工具就像直升机，能直接把你空运到指定的瀑布前，让你看一眼然后飞回去，但你对如何抵达那里一无所知。除了你具体要求的那个事物之外，你可能看不到任何其他有趣的现象。因此，尽管技术上你更高效地达成了目标，但也许有些东西已经失落了。

these AI tools, they can be like helicopters that will just fly you directly to this waterfall and you can see it and then you fly back, but you learn nothing about how to get there. You may not see any other interesting phenomena than the specific thing that you asked for. And so even though technically you achieve your goal much more efficiently, there may be something that is lost.

语境：陶哲轩提出著名的‘直升机隐喻’，反思 AI 辅助科学发现的潜在代价：自动化工具虽能高效求解特定目标，但跳过了人类徒步穿越数学丛林中附带获得的整体洞察与知识地貌认知。

回应 → youtube-demis-hassabis-20260911

支持 → the-economist-yuval-harari-20260826

这些在前向传播中由微分方程控制的系统具有高度非线性特征，且完全处处可微。因此你可以执行反向传播，并凭借其包含的多重反馈回路结构，极其优雅高效地从数据中完成表征学习。

these are highly nonlinear systems governed by differential equations in a forward pass and they are differentiable. So you can actually run them backward pass and you can basically learn them like really nicely because of the multiple feedback structures that they had.

语境：Ramin Hasani 阐释液体神经网络（LNN）的数学精髓：将动力系统微分方程与现代端到端梯度优化相结合，突破离散循环与自注意力机制的固有限制。

支持 → all-in-satya-nadella-20260915

细看 LFM-2 系列模型，它们拥有非常优雅且精简的架构设计：大体由 80% 的双门控一维卷积算子加上约 20% 的分组查询注意力（GQA）混合构成。这是一种高度精简、极速且高质的基础模型架构。

The LFM2 instances of the models, they are, if you look at them, they have a very elegant and simple architecture. So the architecture is basically 80% doublegated convolutions you know 1D convolutions plus some uh group query attention basically 20% of query. So the architecture is just a very simplified

语境：Ramin Hasani 公布 LFM-2 的算子混合配比：通过神经架构搜索找到能原生在 CPU 上高速跑通的自回归混合结构，极大削减对全自注意力机制的算力依赖。

支持 → latent-space-ramin-hasani-20260918

自回归模型在推理端本质上依然是完全串行的：计算只能从左至右逐个 Token 推进；在生成前 9 个词之前，你绝不可能提前生成第 10 个词。这种工作负载与 GPU 的并行硬件架构极其不匹配，受到极其严重的内存带宽限制。

auto reggressive models are still sequential uh the computation is one left to right one token at a time. You cannot generate the 10th token until you've generated everything that comes before it. That kind of workload is um does not map well to GPUs. that kind of workload is extremely memory bound.

语境：Stefano Ermon 指出自回归 Transformer 在推理时代的核心死穴：串行生成导致大量时间浪费在内存层级搬运权重上，算力利用率低下。

支持 → all-in-satya-nadella-20260915

你把绝大部分时间都浪费在了在内存层级结构中搬运模型权重上，实际执行的算术运算少之又少，这是自回归模型的根本痼疾。正如当年 Transformer 在训练端用并行计算取代了 RNN 一样，在推理端取代自回归串行架构的终极形态正是基于扩散的生成模型。

spending most of your time moving around weights across the memory hierarchy and you're doing very little arithmetic and that's a fundamental problem of auto reggressive models. And so what's the equivalent if you think RNN's transformers auto reggressive models the equivalent at inference time is a diffusion based is a diffusion model

语境：Stefano Ermon 梳理生成架构历史演进：类比 2017 年 Transformer 并行化训练淘汰 RNN，推理端必须走向多 Token 协同去噪的扩散范式以解锁单位算力智力上限。

支持 → no-priors-stefano-ermon-20260918

解决千禧年大奖难题（纳维-斯托克斯方程）的成果，绝不能归功于多智能体系统，我甚至连 10% 的功劳都不会分给多智能体。现实是 OpenAI 训练了一个非常强大的模型。我们能让该模型在超长跨度内运行并并行思考，但其根本原因在于我们拥有一个极其强大的通用基础模型。

The effort to solve a Millennium Prize Problem, this was not due to multi-agent. I wouldn’t even attribute 10% of the credit to multi-agent. The reality is that OpenAI has trained a very powerful model. We can get that model to operate over very long horizons. We can get it to think in parallel. But at its core, the reason why we’re able to do this is because we just have a general-purpose, very strong model.

语境：针对业界过度炒作‘万级智能体协同解题’的热潮，Noam Brown 澄清技术实质：多智能体外壳只是显眼包装，真正实现数学与科学发现质变的是底层通用基础模型的长程推理能力。

回应 → tbpn-sam-altman-20260929

支持 → youtube-demis-hassabis-20260911

从演化论视角来看，新视点预测正是大自然通过让动物运动所必须解决的问题。大自然赋予了动物眼睛，却没有给树木眼睛。为什么？因为只要你运动，你就会看到新的视点。无论你称之为 AI 完全性还是智能完全性，我们都坚信‘下一个视点预测’与‘下一个词预测’完全等价。

new viewpoint prediction is exactly what evolution had to solve by making animals move. Nature gave animals eyes, but nature didn't give trees eyes. Why? Because when you move, you see a new viewpoint. And that is whether you call it AI-complete or intelligence-complete. So we do believe very strongly that next viewpoint prediction is the equivalent of next token prediction.

语境：李飞飞从生物演化与智能起源阐释 World Labs 空间智能的核心哲学：运动与交互迫使生物感知三维空间的新视点，因此预测新视角是通向物理世界通用智能的本原路线。

支持 → youtube-demis-hassabis-20260911

承接 → sequoia-sutton-20260818

在通往空间智能的道路上，生成像素无疑只是初级的第一步，无数模型都已经做到了这一点。但生成真正具备空间上下文感知并扎根于几何与物理现实的像素，则是质的飞跃。而这正是 Atlas 迈出的最艰难一步。

generating pixels is definitely an early step, which we have seen with gazillions of models. But generating pixels that are truly spatially contextualized and grounded is absolutely another major step. And that is the very hard step that Atlas has taken.

语境：李飞飞区分 2D 视频生成与 3D 世界模型的本质差异：表层像素生成容易产生幻觉与物理不一致，真正的世界模型必须具备空间几何接地（grounded）。

支持 → youtube-demis-hassabis-20260911