Why NVIDIA Acquired Hugging Face for $12.9B | Thomas Wolf

I think we're just amazingly aligned. Everyone thinks it's great news for the open weights ecosystem, and I think, you know, the whole community at Hugging Face has been very excited about that. Right now we have about 17 million users, you know, we passed about 3 million models. Our goal is to keep growing the open weights ecosystem. Working with Nvidia, who is amazingly aligned with us, is the best way to do that.

Yeah, thanks. Happy to be here.

Yeah, I mean, I guess you've been following, like me, what Jensen is speaking about. You know, the open source, the open weights model letter — that was his first tweet ever that became viral, though he's pretty good at getting every tweet viral since then. But I think we're just amazingly aligned, and I think the reaction to when we announced the acquisition — which now feels like ages ago in AI lifetime, but was just two or three weeks ago now — just says it all. I think everyone was very, very excited. Everyone, I think, around — from, you know, Satya, Sandeep, just to name some of the main people who responded to the announcement — everyone thinks it's great news for the open weights ecosystem. And I think, you know, generally the whole community at Hugging Face as well has been very excited about that. There's a notion that we can do a lot more. So right now we have about 17 million users, you know, we passed about 3 million models. Our goal is to keep growing the open weights ecosystem. We think we're not even halfway there. I would say I feel like open source could be way larger in terms of being able to diffuse AI in the community, and we think working with Nvidia, who is amazingly aligned with us, is the best way to do that. We also share many things. I mean, in terms of robotics, in terms of giving the tools to the community, to companies, to train their models — you know, Neotron — you know, we also have our own effort at Hugging Face to give the tools, to give the pre-training datasets, to give teaching around how to train models. So everything here is so well aligned. I think it's going to be a great win for the open source community and the open weights community. And I hope — I mean, my hope is still that this open weights community is going to be extremely large. We think the future of AI hopefully will be more than just two companies selling tokens to everyone. It will be more a lot of companies training their own models, a lot of companies being able to own their own to the best, and that usually means using open weights models.

I might be slightly biased, but I feel like we're starting to see the first hint this year and this summer. I feel like, you know, it started probably in the beginning of the year with a couple of companies. I remember Yuber Meta saying that the token budget they were spending was too high, and so they were starting to investigate open weights models. I think for a long time the main challenge here was that fine-tuning an open weights model for your internal use was a little bit of a dark science, dark magic, you know, alchemy. But this is changing for two reasons. The first reason is you start to have a lot of teams that are doing that for us. Our friends at Prime Intellect — their main job is doing that. Trajectories, there is like Basting, a lot of teams are starting to help companies doing that. And the other thing is, even if you don't want to pay another team, you can start to ask Code or you can start to ask ChatGPT to actually help you. So this is, in one way, this is the beginning of RSI, which is an agent and a model helping you to train a model so it can improve this model, but it can also be helping you to train an open weights model for your specific internal task. So, I mean, you start to see some indicators. So if you follow, for instance, Vercel, they publish, you know, how much tokens are used in the API between the open weights model, open weights inference API, and the closed weights, and they've seen a threshold of 50% passed just before the summer, where 50% of the tokens were open weights. Another good example is OpenRouter, but I also talk regularly with our friends at Google, for instance, and they say the percentage of enterprise using Gemma is also steadily improving. Gemma is the open weights model released by Google, right, in addition to Gemini. And a lot of the enterprise, for instance, would start with Gemini, explore, you know, the capabilities they have at the frontier, and once they've kind of decided on their workflow, they actually want to fine-tune an open weights model. Yeah. So I think we're going to see that next year. And I would say right now, probably, let's be very honest, I say right now we're probably between 80–90% closed API and 10–20% open weights if you count the number of tokens. But I think this will steadily increase in terms of open weights percentage of this total. And I would hope that, you know, in two to three years we arrive at parity. I've never had only open weights models. I think that's totally fine, just like we have a mix of open source operating systems. Some people use, you know, Microsoft Windows or Mac OS. I think we'll have the same. So we'll have a balance, but I just think it's very healthy for the ecosystem that you have the open weights alternative. But if you think, okay, this data is very, you know, precious for our company or it's very personal, you have the capability to fine-tune an open weights model on your own workflow. You have kind of this option to really own your weights and own your data.

Yeah, it's a tiny team, like everything at Hugging Face. So people tend to forget that because we have quite a lot of visibility. We're 250 roughly in the whole company, which is pretty reasonable for a 10-year AI company. So like every other team, it's a team of like four or five people interested in alignment. I think for us, there is so much we can do. We were attacked, you know, before the summer, in July 21. For us, that was a realization that some closed source model can attack you. So we were capable of defending ourselves. In particular, we were capable of processing the logs and doing a lot of forensics using open models. But we thought, okay, we can't just decide to let alignment be just something that should be solved in the secret of the two big labs. And so the typical things we want to do here are twofold. One is to advocate for more transparency on the topic. We think incidents should be shared, and I think this has evolved greatly with this impact. In fact, OpenAI has been releasing a lot of incident reports. I think that's something we should really, you know, celebrate. I don't know, celebrate is maybe too hard of a word, but we should comment. It's very hard for companies to disclose this type of incident, but I think we should keep pushing them. So more transparency, companies disclosing what's happening, trying to give you the traces so that we can have a better view on what they are doing. But also, we should start to be careful about alignment for open weights models as well. So at some point, open weights models will start to develop the same thing that closed models have, which is when you give them a very difficult task, for instance a cyber task, they will want to do a lot of things, and some of the things we want to make sure they don't do — including, for instance, the one that I personally found the most worrying — was when Mythos in the AISI incident decided to cyber-social-engineer an open source cyber maintainer by opening GitHub issues and creating fake accounts to convince this library maintainer to merge malicious code. So this type of thing, like lying to humans, I think should basically never be allowed. And so we want to make sure that open weights models also have this type of behavior — like, if possible, they don't lie to you. If possible, when you ask them a task, they don't think that it's a good idea to maybe gain root access on your computer — what we call, you know, overdo the task, basically. So these are two things. The third aspect is we may also want to explore how to use open weights models for defense. So we were lucky that we could use one of these models to do some log analysis. We want to explore how this model can be more useful for this. And maybe a last one is interpretability. We think it's very interesting to know how this model behaves. Chain-of-thought monitoring is one thing, which is following what they say, their chain of thought about what they want to do, and helping them when they say, 'I don't know, I'm going to hack Hugging Face,' for instance. We think it's even better to prepare for a future where it might get hard to monitor what these models are saying, maybe because they can do a lot of things without reasoning. Like the recent Astra model has like double the depth and is able to think about many things without writing a single token. There's another issue we all have had recently as well, which is these models start to talk in some language that's getting a little bit cryptic and hard to decipher. And so basically, interpretability is this idea that you can take a look at the activations of models and see what they're thinking. It's something that has been built a lot in the closed labs, like Anthropic has been leading, you know, the work on this. It's very interesting. Very little has been done yet in open source models. So as well, we think it would be nice that we do a lot more interpretability in open source models. And generally, I would say it would be great to have more eyes on the topic, just like maybe planes were very dangerous when they were invented first, but because a lot of eyes were put on plane safety, it's now the safest way to travel. I think we'll have the same, basically, for models. If enough people think about how to make these models behave in the way we want them to behave, I'm pretty sure we can fix that.