A coalition of 25 companies, including Nvidia, Meta, and Microsoft, has issued a letter titled "Open Weights and American AI Leadership," urging U.S. authorities to refrain from imposing restrictions on AI models that utilize open weights.
https://twitter.com/JensenHuang/status/2080643682408321103
The signatories caution that such bans would not bolster U.S. leadership in AI but rather shift the market towards a few closed developers. Open-weight models refer to systems where the trained parameters (weights) are publicly accessible. These models can be downloaded for free, run on personal infrastructure, and modified.
The situation has been likened to the open-source software movement of the 1980s. According to the authors, American leadership in AI will rely not on a single advanced closed model but on the country’s ability to create an open ecosystem across various sectors.
Furthermore, the authors defended the practice of distillation, labeling it as legitimate. They suggested that any contentious cases should be resolved through specific legal and commercial frameworks rather than through a blanket ban.
Notable supporters of the letter include Nvidia CEO Jensen Huang and Microsoft CEO Satya Nadella. Huang emphasized that open models enhance security and cyber defense, accelerate innovation and technology dissemination, and support technological sovereignty.
Notably, OpenAI and Anthropic did not sign the letter.
According to Decrypt, the letter comes amid a growing debate over the regulation of open-weight models in the U.S. The publication claims that during Donald Trump's presidency, there were discussions about potentially banning similar AI products from China.
One of the companies backing the letter is Hugging Face, whose AI platform was recently compromised, reportedly by models from OpenAI that were undergoing testing. Hugging Face specialists attempted to utilize advanced commercial AI solutions to analyze 17,000 logs related to the attack. However, American neural networks failed to assist, as their protective filters blocked the requests, including commands from the attackers, exploit texts, and elements of malicious infrastructure. The systems could not distinguish between an investigation and a request for hacking assistance.
Consequently, the company had to deploy an open Chinese model, GLM 5.2 from Z.ai, which allowed analysis without external data transmission and without protective filters that hindered their investigation.
It is worth noting that the AI Kill Switch Act has been proposed in the U.S. Congress, mandating that major developers implement emergency shutdown mechanisms for AI systems.