Jensen Huang calls AI distillation 'competition' as Bessent brands it 'theft'
Jensen Huang says AI distillation is 'competition.'
Nvidia CEO Jensen Huang pushed back on the White House's characterization of AI model distillation, telling CNBC it's simply 'competition.' Treasury Secretary Scott Bessent called the practice 'theft' in July and threatened sanctions against overseas firms that use it to extract capabilities from U.S. models. The dispute has become a flashpoint in U.S.-China AI tensions, with CISA accusing Chinese companies of 'industrial-scale knowledge distillation campaigns' and Anthropic alleging Alibaba and DeepSeek engaged in 'illicit distillation.'
I'd really prefer that nobody learns from our products, and we have the benefit of just cruising along. But, you know, frankly, competition makes everything better.
- captainbland
I think it would be totally incoherent to say that copyrighted information is essentially fair game to include in your model but the outputs of your model are privileged against being included in other models.
I'm not sure "competition" is necessarily the right word but I don't think that model creators and their political backers have a leg to stand on when complaining about distillation.
- bennettpompi1
I think that Jensen's position as the guy selling the proverbial shovels incentivizes him to take a lot of irresponsible positions (namely around safety) but imo he's fundamentally correct here.
- Eliah_Lakhin
Now when the big LLM companies create their products by accumulating a notable portion of copyrighted creative works (including computer programs) from Internet, it does not count as copyright infringement or competition (or "theft"). It is considered as "fair use". So, why training LLMs on other LLMs is not fair use too?
> If you don’t like that, if you don’t like people to use your products, all you [have to do is] know your customers, and disable the service
Considering the current situation, I understand Mr. Huang point, but that's not how the copyright framework assumed to be working from the beginning. It should protect both small actors (authors) and the big companies from unrestricted use of creative works. Now this mechanism seems to be practically dysfunctional.
And a big portion of this lies on shoulders of proponents of permissive OSS, who defend an idea of (almost) unrestricted use of their source code texts for many years, and long before mass LLM scrapping became a thing.
- chrsw
Would a 100% compatible and open source CUDA stack be “competition” too?
- HarHarVeryFunny
I wonder how the "Chinese AI is all just distillation" folk are going to cope now that the Chinese are all aboard the post-train via agents in custom training environments wagon?
Here's Xiaomi's discussion of this, plus their open-sourcing of 7000+ RL training environments.
https://www.alphaxiv.org/abs/2609.mimo-scaling-reinforcement...
https://mimo.mi.com/docs/en-US/news/latest/v2-6
Who needs a few of someone else's "vacation postcards" of their post-training experience when your agents can go on vacation themselves!