精选行业82

前 OpenAI 工程师称通过权重指纹发现中国实验室用 Blackwell 训练模型

原标题:前 OpenAI 工程师发帖声称通过模型权重初始化指纹分析,发现 DeepSeek、Moonshot、MiniMax 和阿里等中国 AI 实验室使用被美国出口禁令限制的 NVIDIA Blackwell 芯片训练最新模型。

indigo@indigox发布时间 (北京时间)

前 OpenAI 工程师 Will Depue 发帖称,通过分析模型权重初始化的 GPU 指纹,发现 DeepSeek、Moonshot、MiniMax 和蚂蚁集团的最新模型在 NVIDIA Blackwell 芯片上初始化,其中包括 DeepSeek-V4-Pro、Kimi K3 和 MiniMax-M3。

推荐理由这条转述了一项基于模型权重指纹的指控,涉及出口管制与算力来源,读者可据此了解争议的方法与证据链。

中文译文尚未完成,先为你展示原文。

前 OpenAI 工程师发帖声称通过模型权重初始化指纹分析,发现 DeepSeek、Moonshot、MiniMax 和阿里等中国 AI 实验室使用被美国出口禁令限制的 NVIDIA Blackwell 芯片训练最新模型。

其方法依赖 GPU 类型和随机种子产生的独特权重布局模式,这些模式在训练后部分保留,经数万亿组合搜索和真实 Blackwell硬件验证,与多个中国模型高度匹配。

分析显示这些实验室可能使用了数千块Blackwell GPU,涉及数亿美元规模,暗示出口管制存在漏洞,作者称结果获 OpenAI和 Anthropic 资深研究员独立验证。鬼故事越来越多 …

我有一个重大结果要分享。 我找到了确凿的取证证据,证明中国排名前四的 AI 实验室——DeepSeek、Moonshot、MiniMax 和蚂蚁集团——用 NVIDIA 最先进的 Blackwell 芯片构建了它们最新的模型,而这些芯片是被禁止出口到中国的。 这听起来很疯狂,但我发现几乎每个 AI 模型都带有训练它的 GPU 的隐藏特征。当一个模型首次建立时,其权重的初始值会以一种取决于具体 GPU 的模式排列。每一种 GPU 与种子数的组合都会产生一个独特的模式——一种用于创建该模型的硬件的指纹。这种模式的痕迹会在训练后保留到已发布的模型中。 再次强调,这一证据并非基于任何内部信息或泄露。它基于对各公司自己发布的模型权重的取证分析。 我搜索了超过六万亿种芯片布局与种子数的组合,并将它们与开放权重模型进行匹配。在此过程中,我找到了确凿证据,证明六个中国 AI 模型——包括 DeepSeek-V4-Pro、Kimi K3 和 MiniMax-M3——是在 NVIDIA Blackwell 系统上初始化的。其中四个模型匹配 Grace Blackwell,即 NVIDIA 旗舰级液冷机架式系统中的 GB200 和 GB300 部件。 随后我在真实硬件上验证了这些结果。使用未修改的 PyTorch 和我恢复出的种子,一台 Blackwell B300 逐比特复现了 DeepSeek-V4-Flash 未经改动的初始权重,而一台租用的 GB200 复现 Kimi K3 的相关性达到 0.999995。一台 Hopper H100 未能复现 DeepSeek-V4-Flash 的任何权重,而将种子改变一个就会破坏匹配。(顺便说一句,DeepSeek 的种子是 953049583。)该方法还被验证能将 24 个已知的开放权重模型匹配到开发者声称所使用的硬件。 这些结果是巧合的可能性极低。在六万亿种组合中,最佳的偶然匹配应该只能达到约 8 sigma(即比随机偶然产生的水平高出八个标准差)。真实匹配的范围从 49 到 1,385 sigma,即使是最弱的单项测试也达到 20。 这些是数据中心规模的训练运行。通过对发布时间表和模型细节的一些分析,我估计最大的 Kimi K3 和 DeepSeek-V4-Pro 各自使用了 2,000 到 3,000 块 Blackwell GPU。这种规模的集群每套购买成本约为 1.5 亿到 2 亿美元,仅这些实验室加起来就会构成超过五亿美元的受出口管制 GPU,还不算实验性算力或其他方面。我们很可能看到的是数十亿美元花在受出口管制的 GPU 上。 我对这一结果高度自信,并且它已得到 OpenAI 和 Anthropic 资深研究人员的独立验证。你可以在 http://unemploymentlabs.com/checkpoint-forensics 阅读完整报告。 通过这类 checkpoint 取证,我还发现了关于这些模型的更多信息,很快会有更多分享。
查看原文

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at http://unemploymentlabs.com/checkpoint-forensics. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

查看原帖