九原客 (@9hills)

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 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.

Today, @llama_index is an applied research lab for document OCR and document work. Our core focus is investing in building the best-in-class models+ agents at the pareto frontier of accuracy/cost/latency for document processing. To do this, we have to invest in all the different components around the research infra stack: 1. Building comprehensive and fair benchmarks. Today we have ParseBench, ExtractBench, and ExtractBench Pro. They're already the most popular benchmarks in parsing and extraction today, and we're just getting started. 2. Investing in data curation and synthetic data generation. 3. Investing in compute and our training pipeline. This has produced models like OCRia-1. 4. Investing in agent optimization - tune everything around the models to further hillclimb cost and accuracy. 5. Closing the loop based on feedback - whether from active usage of our product, implicit/explicit customer feedback, or automated verification. We are accelerating our capabilities at a crazy pace. Today, our work is documented in our research page: There is a lot more to come!

We've investigated a few messages about codex usage limits being different. That's not something we change without engaging the community and being transparent. What we did see is that when talking to affected users many were using sub2api. Converting a subscription into api traffic to then re-serve or share across many users is not something we support and this type of usage gets flagged by our fraud-prevention systems. You are completely fine if you use your subscription through Sign in With ChatGPT, either through the official clients or through one of the many OSS clients (Pi, OpenCode, ...) that support signing in with your account and using your included usage.

Main result. At xhigh every quant landed between 88.0 and 90.0% pass@1 on the full suite: AWQ INT4: 90.0% NVFP4: 89.3% GGUF Q4_K_M: 89.3% FP8: 88.7% NInfer: 88.0% Yes, the 4 bit quants scored above the FP8 baseline. McNemar says its a statistical tie, first and last place differ by three tasks out of 150. I started this run to show you how quantization eats quality. There is nothing to show. The gap between quants is smaller than the gap between reasoning presets.

越来越看不懂 claude code 了，为啥 bypass permission 优先用 shell 脚本来改文件？ 看过源码就知道，他那个 Read/Write/Edit 工具有并发修改检测，还有内容追踪和回滚的功能。用 shell 命令来改文件，完全无法追踪和恢复。 咋想的？