🧠 阿头学 · 💬 讨论题

Scaling红利见顶，RL数据策展成千亿刚需还是营销话术？

作者以OpenAI内部视角断言Scaling Law已失效并预言千亿级RL数据市场，但其核心论证存在明显的循环自证与商业包装，实际反映的是AI产业从粗放算力堆砌向精准数据策展转型的必然阵痛。
打开原文 ↗

2026-07-30 原文链接 ↗

阅读简报

双语对照

完整翻译

原文

讨论归档

核心观点

LLM能力呈“尖峰状”分布，泛化瓶颈已现：模型在封闭基准表现优异，但在真实工程流中仍需人工兜底，证明单纯堆叠算力无法自动兑换通用生产力。
高质量RL数据的核心壁垒在于“反事实评分”：现有供应商仅提供成功路径答案，缺乏对探索性错误路径的可靠评分机制，导致模型训练极易陷入过拟合。
千亿级数据采购预言缺乏财务支撑：作者将实验室盈利压力直接推导为超1000亿美元的数据资本开支，无视算力基建仍占预算主导的现实，属于典型的销售恐吓。
科学推理与多模态工作流存在真实缺口：前沿模型在长程科学实验分析与跨模态图像理解上通过率不足30%，该短板独立于作者个人体验，是Agent走向生产环境的硬性约束。

跟我们的关联

对Neta意味着技术路线需从“拼参数”转向“拼数据质量”：下一步应建立内部“可评分反事实”数据构建管线，优先在长程推理场景引入带真值的探索性数据，替代粗放合成数据。
对ATou意味着投资决策需警惕“千亿市场”叙事泡沫：下一步在评估AI基础设施标的时，应剥离创始人光环，重点审计其数据产品的上游整合能力与独立基准验证结果。
对Uota意味着工作流设计必须保留“人类Deslop层”：下一步在部署AI Agent时，需绘制“尖峰-谷底”能力热力图，在模型能力断崖处强制设置人工审核节点，避免端到端自动化带来的隐性返工成本。

讨论引子

当“反事实路径评分”成为RL数据的核心壁垒时，初创公司如何在不依赖巨头算力与内部黑箱数据的前提下，低成本构建具备经济价值的可评分环境？
面对LLM“尖峰状”能力分布，企业应优先投入资源打磨单一垂直场景的90%+可靠性，还是继续追求多模态通用Agent的泛化能力？

今天是我在 @OpenAI 的最后一天。我很高兴，自己人生过去的八个月是在这里度过的！

我创办了一家新公司，专注于生产高质量的强化学习（reinforcement learning）数据集：

大语言模型（LLMs）的泛化（generalization）能力显然非常薄弱，即使在投入了大量资金和关注的领域，其能力也呈现出“尖峰状（spiky）”。例如，尽管经过多年、数十亿（甚至数百亿）美元的投资，连编程能力都未展现出“通用性（generality）”——即使每个模型解答 Codeforces 题目都比我强，或者将 C++ 移植为 Rust 都比我做得好，我仍然不得不手动为拉取请求（pull requests）“清理（deslop）”。
绝大多数具有经济生产力的能力在现有数据产品中都未能得到充分体现。首先，强化学习数据集（RL dataset）的设计本身是一门艺术，而大多数供应商并未亲自将数据整合进大规模训练运行的上游环节，因此对此并不真正理解。其次，更重要的是，大多数工作都具有高度的上下文依赖性（contextual），且不易被编码进可评分环境（gradable environment）；即使我们能观察到一条人类走过的、我们认为不错的“黄金路径（golden path）”，也很难判断那些替代性的、反事实的（counterfactual）路径究竟会产生好结果还是坏结果。

这里的基本前提是，由于我曾明确站在需求方一侧，并且从采购（procurement）到训练全程都深度参与，因此我清楚实验室需要什么、想要什么，而且我有能力提供。我还相信，未来几年数据需求将大幅增长，尤其是当前沿实验室（frontier labs）面临越来越大的盈利压力时；它们无法仅通过扩展（scaling）就“免费”获得相应能力；相反，它们需要在精准、有针对性的数据获取上投入超过 1000 亿美元。

我们的首批产品将聚焦于生物学和统计推理：

首先，是解决长程科学推理（long-horizon scientific reasoning）的数据集，借鉴了我和 @jeremyli__ 在 GeneBench-Pro 上的工作。前沿模型（Frontier models）仍然无法可靠地执行那些需要判断力、探索精神和自适应修正的“混乱”数据分析（在 GPT-5.6 Sol 上，GB-Pro 的通过率（passrate）勉强超过 30%）；为了解决这个问题，我们有能力生成数千个带有已知真值（ground truths）、可被可靠评分的高质量问题。（相比之下，现有大多数用于生物信息学（bioinformatics）的强化学习数据要么严重过度指定，要么严重指定不足，用来训练模型很可能会把你的模型练坏。）要将“可靠性差距（reliability gap）”从 30% 提升到 90% 以上，显然是科学加速所必需的；而且——尽管我对强化学习的泛化（generalization）持怀疑态度——这在所有可想象的数据集中，也是最能为模型整体推理能力带来泛化收益的数据集之一。
其次，是解决与日常工作流相关能力的数据集。想象一位科学家拍下某个实验过程或结果的照片——比如一块细胞培养板（cell culture plate）或一张 Western blot——然后向 Claude 提问。前沿模型在处理这类问题上仍然相当糟糕，尤其是那些包含多模态（multimodal）成分的问题。但它们显然是加速科学发现所必需的；在我们梦想实现科学自动化之前，必须先巩固这些基础的通用能力。

除上述两项之外，我们希望扩展到相邻领域（化学、材料科学等），进而延伸到具有更直接经济价值的领域，如医疗（healthcare）和白领办公（white-collar office work）。

我强烈建议有数据需求的实验室与我们联系。我们提供行业标准的价格和条款，而且正如我所说——我了解这个流程的运作方式，知道什么样的数据才是优质数据，也知道如何令人信服地向您证明，您能够将我们的数据无缝整合到训练流程的上游。我的私信（DMs）始终开放！

Today is my last day at @OpenAI. I'm glad to have spent the last eight months of my life working here!

今天是我在 @OpenAI 的最后一天。我很高兴，自己人生过去的八个月是在这里度过的！

I'm starting a new company focused on the production of high-quality reinforcement learning datasets:

我创办了一家新公司，专注于生产高质量的强化学习（reinforcement learning）数据集：

The generalization ability of LLMs is clearly very poor, with "spiky" capabilities even in areas that have received tremendous amounts of investment and attention. For example, despite multiple years with tens (if not hundreds) of billions invested, even coding capabilities don't demonstrate "generality" -- even if every model can solve Codeforces questions or port C++ to Rust better than I can, I still have to manually "deslop" pull requests.

大语言模型（LLMs）的泛化（generalization）能力显然非常薄弱，即使在投入了大量资金和关注的领域，其能力也呈现出“尖峰状（spiky）”。例如，尽管经过多年、数十亿（甚至数百亿）美元的投资，连编程能力都未展现出“通用性（generality）”——即使每个模型解答 Codeforces 题目都比我强，或者将 C++ 移植为 Rust 都比我做得好，我仍然不得不手动为拉取请求（pull requests）“清理（deslop）”。

The vast majority of economically productive capabilities are not well represented in existing data offerings. First, there's a certain art to the design of an RL dataset which most vendors, not having upstreamed data into large training runs themselves, don't really understand. Second, and more importantly, most work is highly contextual and not easily encoded into a gradable environment; even if we can observe a "golden path" taken by a human which we believe to be good, it's challenging to understand whether alternate, counterfactual paths produce good or bad outcomes.

绝大多数具有经济生产力的能力在现有数据产品中都未能得到充分体现。首先，强化学习数据集（RL dataset）的设计本身是一门艺术，而大多数供应商并未亲自将数据整合进大规模训练运行的上游环节，因此对此并不真正理解。其次，更重要的是，大多数工作都具有高度的上下文依赖性（contextual），且不易被编码进可评分环境（gradable environment）；即使我们能观察到一条人类走过的、我们认为不错的“黄金路径（golden path）”，也很难判断那些替代性的、反事实的（counterfactual）路径究竟会产生好结果还是坏结果。

The basic premise here is that I have a clear understanding of what labs need/want, having explicitly been on the other side and having been involved at every level from procurement all the way through training, and I'm able to provide it. I also believe that data needs will grow tremendously in the coming years, especially as frontier labs face increasing pressure toward profitability, and that they won't get the relevant capabilities "for free" through scaling alone; instead, they'll need to spend >$100B on precise, well-targeted data acquisition.

Our first products will be focused on biology and statistical reasoning:

我们的首批产品将聚焦于生物学和统计推理：

First, datasets that address long-horizon scientific reasoning, drawing on my work on GeneBench-Pro with @jeremyli__. Frontier models are still unable to reliably execute "messy" data analyses that require judgment, exploration, and adaptive revision (GB-Pro passrate on GPT-5.6 Sol scarcely exceeds 30%); to address this, we have the ability to generate thousands of high-quality problems with known ground truths which can be reliably graded. (In contrast, most existing RL data for bioinformatics is either massively over- or under-specified, and will probably break your model when you train on it.) Moving the "reliability gap" from 30% to >90% is obviously required for scientific acceleration, and -- despite my skepticism about generalization of RL -- is one of the most promising datasets conceivable when it comes to yielding generalization benefits for models' overall reasoning capabilities.

首先，是解决长程科学推理（long-horizon scientific reasoning）的数据集，借鉴了我和 @jeremyli__ 在 GeneBench-Pro 上的工作。前沿模型（Frontier models）仍然无法可靠地执行那些需要判断力、探索精神和自适应修正的“混乱”数据分析（在 GPT-5.6 Sol 上，GB-Pro 的通过率（passrate）勉强超过 30%）；为了解决这个问题，我们有能力生成数千个带有已知真值（ground truths）、可被可靠评分的高质量问题。（相比之下，现有大多数用于生物信息学（bioinformatics）的强化学习数据要么严重过度指定，要么严重指定不足，用来训练模型很可能会把你的模型练坏。）要将“可靠性差距（reliability gap）”从 30% 提升到 90% 以上，显然是科学加速所必需的；而且——尽管我对强化学习的泛化（generalization）持怀疑态度——这在所有可想象的数据集中，也是最能为模型整体推理能力带来泛化收益的数据集之一。

Second, datasets that address capabilities relevant to day-to-day workflows. Imagine a scientist snapping a picture of some experimental process or result -- say, a cell culture plate or a Western blot -- and asking Claude a question. Frontier models remain quite bad at these questions, especially those with multimodal components. But they're obviously required for acceleration of scientific discovery; before we can dream about automating science, we have to begin with shoring up these basic, generalist capabilities.

其次，是解决与日常工作流相关能力的数据集。想象一位科学家拍下某个实验过程或结果的照片——比如一块细胞培养板（cell culture plate）或一张 Western blot——然后向 Claude 提问。前沿模型在处理这类问题上仍然相当糟糕，尤其是那些包含多模态（multimodal）成分的问题。但它们显然是加速科学发现所必需的；在我们梦想实现科学自动化之前，必须先巩固这些基础的通用能力。

Beyond these two, we hope to expand to adjacent fields (chemistry, materials science, etc.), and then even further into fields with more direct economic applicability like healthcare and white-collar office work.

I strongly encourage labs with data needs to reach out. We offer industry-standard pricing and terms, and like I said -- I know how this process works, what good data looks like, and how to demonstrate to you, convincingly, that you'll be able to upstream our data into your training processes without issue. My DMs are open!

Today is my last day at @OpenAI. I'm glad to have spent the last eight months of my life working here!

I'm starting a new company focused on the production of high-quality reinforcement learning datasets:

The generalization ability of LLMs is clearly very poor, with "spiky" capabilities even in areas that have received tremendous amounts of investment and attention. For example, despite multiple years with tens (if not hundreds) of billions invested, even coding capabilities don't demonstrate "generality" -- even if every model can solve Codeforces questions or port C++ to Rust better than I can, I still have to manually "deslop" pull requests.
The vast majority of economically productive capabilities are not well represented in existing data offerings. First, there's a certain art to the design of an RL dataset which most vendors, not having upstreamed data into large training runs themselves, don't really understand. Second, and more importantly, most work is highly contextual and not easily encoded into a gradable environment; even if we can observe a "golden path" taken by a human which we believe to be good, it's challenging to understand whether alternate, counterfactual paths produce good or bad outcomes.

Our first products will be focused on biology and statistical reasoning:

First, datasets that address long-horizon scientific reasoning, drawing on my work on GeneBench-Pro with @jeremyli__. Frontier models are still unable to reliably execute "messy" data analyses that require judgment, exploration, and adaptive revision (GB-Pro passrate on GPT-5.6 Sol scarcely exceeds 30%); to address this, we have the ability to generate thousands of high-quality problems with known ground truths which can be reliably graded. (In contrast, most existing RL data for bioinformatics is either massively over- or under-specified, and will probably break your model when you train on it.) Moving the "reliability gap" from 30% to >90% is obviously required for scientific acceleration, and -- despite my skepticism about generalization of RL -- is one of the most promising datasets conceivable when it comes to yielding generalization benefits for models' overall reasoning capabilities.
Second, datasets that address capabilities relevant to day-to-day workflows. Imagine a scientist snapping a picture of some experimental process or result -- say, a cell culture plate or a Western blot -- and asking Claude a question. Frontier models remain quite bad at these questions, especially those with multimodal components. But they're obviously required for acceleration of scientific discovery; before we can dream about automating science, we have to begin with shoring up these basic, generalist capabilities.

📋 讨论归档

讨论进行中…