Dario's entire argument hinges on this paragraph, which falsely assumes that US labs still have a meaningful lead over Chinese labs. The gap is gone. Chinese models such as DeepSeek, GLM, and Kimi are now on a par with the top models from OpenAI and Anthropic.
Image from Dr. Eli David (@DrEliDavid)
16
20
2
129
15,098
US labs are 2-3 generations ahead of the frontier models that are public. You can deduce that from openAI's statement that the last mathematical proof was performed by a model significantly more advanced than Astra. The proof itself is about 100 pages and took months to verify.>>

Sep 13, 2026 · 1:44 AM UTC

1
37
RelevantRecentLikes
In addition to the aforementioned, the Chinese models are on par with the American ones only in benchmark results and not in novel real world tasks because they train on stolen American model CoT data and are benchmaxxed. An American pause would also set back the Chinese.
15