Keep pulling the thread on Nathan Lambert.
Chinese President Xi Jinping has committed to openness and open source as a national strategy.
Qwen announced that its next large model will be open-weight.
The model weights for Kimi K3 are expected to be released on July 27th.
Chinese AI labs may be more capital-efficient than their Western counterparts in turning capital into compute, data, and talent for model development.
The performance gap between closed-source and open-source models appears to be narrowing, contrary to expectations that the capital intensity of training would cause it to widen.
The acquisition of restricted AI chips by Chinese firms, through methods like smuggling or use in third-party countries, has substantially increased in the last 6 to 9 months.
LongCat released a model claimed to be trained entirely on Chinese-made chips, speculated to be Huawei's Ascend series.
The most advanced "frontier" AI models are becoming increasingly inaccessible, while "near-frontier" capabilities are becoming commoditized.
In a recent incident, Hugging Face had to use the less-restricted GLM model to analyze a cyberattack because the safety guardrails on models like GPT and Claude prevented them from performing the analysis.
Banning US companies from using top Chinese open-weight models would create a structural disadvantage for US cybersecurity defenders.
Chinese President Xi Jinping gave a speech committing to openness and open source as a national strategy.
Qwen announced its next major model will be released with open weights.