Thomas Wolf says Hugging Face hosts close to 4 million open models, with a new one published about every five seconds.
“There is close to 4 million of these models at the moment. There is one new model being published every five seconds.”
Person
Thomas Wolf
1 episodes · 21 quoted claims
Wolf says DeepSeek's January release drew major adoption and visibly shook the stock markets.
“one that I think got the most adoption and most visibility recently was DeepSeek, who was released in January and kinda crashed the stock markets at the same time that it was released.”
Wolf says Hugging Face has spent eight years building its platform with a community that shares and downloads models.
“And so over the past eight years, Hugging Face has been building this platform, growing it together with the community of people and teams who are both sharing and downloading models.”
Hugging Face's community has roughly 10 million users and also hosts datasets, not only models.
“This community is now roughly 10 million users and AI builders, how we call them, and we've expanded it as well beyond just model hosting to also host datasets”
Hugging Face offers Spaces, a low-code way to test models.
“and more recently also what we call spaces, which are simple low-code way to test all of these models.”
Wolf says Unix-style systems underpin enterprise software and cloud, though Linux was once seen as niche against Microsoft.
“Unix is really the basics of all enterprise software and all enterprise cloud.”
Wolf notes that macOS, a major competitor to Windows, is itself built on a Unix core.
“even Mac OS, which I think is probably the most widely diffused on, like consumer laptops nowadays, is one of the largest competitor to Windows based, is itself based on the Unix core.”
Hugging Face published a new 3-billion-parameter model small enough to run on a laptop or smartphone.
“we published a new model that's called a smaller M three, which is an extremely smart model, but this is the best one. But uh, at 3 billion parameters”
Hugging Face released the data and training recipe alongside its new small model.
“And we've decided to share at the same time all the data, all the recipe, all the knowledge on how to build these models.”
Hugging Face initially curated models manually but found it unsustainable given the pace of new uploads.
“In the beginning, we tried to have some curation manual for us, but with one new model every five seconds, that's just not really sustainable.”
Wolf says the most downloaded model category on Hugging Face is often speech rather than text.
“the most downloaded one will be like a speech model, and I get the LLM people or text people are very surprised”
Wolf describes an AI model as matrix multiplications, a few nonlinearities and attention blocks.
“In a way, it's just really this series of matrix multiplication, a couple of nonlinearities, and you have the attention blocks.”
Wolf says Cerebras builds very large wafer-scale chips that can host a full model, and Hugging Face works with them.
“cerebras is this example of let's be able to host the full model on just this very large wafer scale chips.”
Wolf attributes to Patrick Collison the view that the limiting cost of intelligence is electricity.
“And the last person I heard saying that was Patrick Collison, but I think a lot of people view this this way.”
Dario Amodei, CEO of Anthropic, wrote an optimistic essay predicting AI will enable major scientific breakthroughs.
“So in this case, started by this essay from Dario Amodei, the CEO of Anthropic, basically saying, Yeah, it's a beautifully written essay.”
Amodei's essay is titled Machine of Loving Grace.
“It's called Machine of Loving Grace, and it's basically saying that AI will enable us to do extremely important scientific breakthrough.”
Wolf says his former professor Alain Aspect won a Nobel Prize for experiments bearing on Einstein's disagreement with quantum mechanics.
“Alain Aspect, who got the Nobel Prize ten years ago, I think, if I'm not mistaken, for basically proving this disagreement that Einstein had with quantum mechanics”
Wolf says language models are trained to predict the most likely next word, so they miss unlikely ideas and drift toward the average.
“they are trained to predict the most likely next word in a sentence, which means they will miss word that are unlikely.”
Wolf organized a hackathon with his twelve-year-old son and friends using the vibe coding tool Lovable.
“my son is twelve, he's still interested in what I do.”
The hackathon participants were aged nine to twelve and built websites through prompts.
“they were between nine and twelve, and to imagine they will grow with basically just deciding that if they want to create a website, it's like just a couple of prompts away.”
Wolf said Hugging Face planned to repeat the kids' vibe coding hackathon worldwide in September and October.
“September and October, we want to redo this type of hackathon everywhere in the world.”