Without Open Source, Machine Learning Would Be Decades Behind

Sarah Guo and Elad Gil welcome Clem Delangue, co-founder and CEO of Hugging Face, on No Priors. The company’s public story by then was already a plot twist: it began as a talking-emoji companion and had become an organizing force for open-source AI research. Its models were in use at companies such as Apple, Salesforce, and Microsoft, and the stated ambition was to become the GitHub for machine learning.

The timing of the conversation matters. ChatGPT had been out only a few months, and the dominant Silicon Valley story was tilting closed: frontier training looked ruinously expensive, the moat was compute and capital, and a handful of labs would take the spoils. Clem’s platform was the counter-evidence sitting in plain sight — researchers and companies putting models, datasets, and demos in one place so others could download, fine-tune, and ship.

Clem grew up in France and ran an electronics shop on eBay that got so big he was one of the largest French sellers on the site — prolific enough that eBay gave him an internship. On the user-facing team he was sent to trade shows, which he remembers as some of the worst days of his working life. PayPal still belonged to eBay, they shared a booth, and PayPal users would come over and shout at him because accounts were blocked or money was being held.

At one of those shows he bumped into a man with big round glasses who looked, in Clem’s telling, very nerdy. The stranger noted that eBay had acquired RedLaser, a barcode-scanning company, to recognize objects and surface listings — and that it was the wrong bet. Soon, new algorithms would make barcodes unnecessary; you would just recognize the object itself. Clem’s first reaction was that the guy was crazy.

That night he looked the stranger up and found a graduate of a serious French engineering school with a small startup that had raised a bit of money. One thing led to another: Clem left eBay for that computer-vision machine-learning company. It was almost fifteen years before this podcast. A single encounter at a trade show rewrote his career, and Gil notes how often those early accidents send people in a completely different direction.

Clem started Hugging Face with co-founders Julien and Thomas, all of them passionate about AI and looking for something scientifically hard that was also fun. Heavy users of Alexa and Siri, they found voice assistants boring — weather, calendars, productivity. They set out to build an AI Tamagotchi, a sci-fi friend of the kind movies had promised, which Clem says is probably a lot of what people use ChatGPT for today. They did that for almost three years and saw billions of messages exchanged between users and the chatbots.

The shift came three years in, after the seed round. They had always been open-source people and had been publishing pieces of the work. When Transformer models started to land — BERT in particular — usage of their open-source library exploded, from a couple of people looking at it to hundreds of companies. It moved from mostly Thomas, to a few teammates, to everyone being more excited about the library than the original product. They raised a Series A on that early traction, which Clem reads as the signal to put most of the company on the new direction.

Guo compares the pivot to a Stewart Butterfield company: Flickr and Slack were both things that took off after something else was already built. Clem’s advice to founders is to get the ratio of exploitation to exploration right. Before product-market fit, too many startups change direction every week and learn little; after it, they freeze and sit in a local optimum. Hugging Face’s rule is to spend at least 30 or 40 percent of company effort exploring new things, and to give those experiments a real chance. Thomas often led them, but Spaces — ML demos that crossed 50,000 in the past year and a half — started with one teammate who thought they could build something cool.

Distribution started on Twitter. The team tweeted about the open-source work, the ML community there was already strong, and then it snowballed. Researchers uploaded models and got visibility; people in industry found Hugging Face through those models and asked more researchers to add theirs. Clem describes it as a classic marketplace network effect. Timing helped, he admits, but the operating rule was less glamorous: listen to the community and actually implement what it asks for.

One decision he thinks worked unusually well: they never hired a community manager or anyone in communications or PR. Interacting with the community, sharing work, and getting visibility was written into every role, including the most technical, specialized scientists. The result is an organization where talking to users is everyone’s job instead of something outsourced to a team. People in the community, he says, could talk to the builders directly, which made the interactions more meaningful.

Guo has heard that many people at the company run the public Twitter account. Clem goes further: everyone on the team has access and tweets from it. Most organizations, she says, are not capable of that kind of risk-taking. Hiring follows the same logic. Some companies treat community and open source as a means to something they are more excited about, and it shows. Hugging Face tries to hire people who would almost work for free on open source for the community and be happy about it.

Guo sees Hugging Face as a conduit from the pace of research into industry, and she also sees a huge distance between cutting-edge work on the Hub and the average enterprise’s machine-learning journey. Clem’s first correction is the time scale. His co-founder Thomas did a PhD in quantum physics, where a paper taking ten or twenty years to reach production was normal. In machine learning, work now reaches production in a year, a few months, a few weeks, sometimes a few days. That loop, he says, is what drives most of the speed of progress — which is why he keeps investing in open source and why proprietary models make him a bit worried. Strip out the last five years of open source, and “we would be like decades away from where we are now.”

Gil puts the closed-lab worry on the table: large Transformers get better as they get bigger, training is prohibitively expensive, so state of the art may lock up at DeepMind, OpenAI, or Anthropic. Clem refuses to pick a winner. Search has Elasticsearch and Algolia; databases have MongoDB and proprietary approaches; both will keep existing. ChatGPT was then ahead on text generation; open source was ahead on text classification, information extraction, and image generation with Stable Diffusion. The lead will flip in a few weeks. It is a race with dozens of racers, not a single finish line.

He does not believe in one model or one company to rule them all. The proof is on his own platform: 250,000 models, uploaded by almost 15,000 companies. They are not building models just to build models. If one model were better at everything, those artifacts would not exist. Specialized models are more efficient, cheaper to run, usually faster, and most of the time more accurate for the specific use case. Clem is more excited about a future where machine learning is available for everyone to build than about a concentrated, monopolistic world.

The Hub’s task mix, as he sketches it, is led by NLP — classification, extraction, generation — then text-to-image and computer vision, then audio (speech-to-text and text-to-speech). Time series is rising, including Uber’s ETA and financial fraud models, and biology and chemistry are starting to show more models, datasets, and demos. Sizes run from a few million parameters to 180 billion, and usage is distributed because the use case decides. Bloomberg wants the terminal real-time, so it uses smaller, faster models. A company that wants something general, from customer support to the meaning of life, and does not care about latency or cost, can go larger.

Asked what he is most excited about, Clem gives an unsexy answer: infrastructure. The field has not thought hard enough about what it costs to run these models or how fast they can go. He hopes this year brings more clarity so the community builds something healthy and sustainable. His nickname for the current mess is “cloud money laundering”: infrastructure cost almost disconnected from actual use cases. As the field matures, those two should align, and he thinks that alignment will enable the next phase.

Two wishes sit next to that. The first is more decentralized infrastructure and more online learning. Today you train a model once and maybe again six months or a year later, which he calls archaic — and which is why so many models do not know who the current U.S. president is. He wants to move from one big training run toward smaller, more regular ones. The second is data consent. Hugging Face worked on BigCode, an open code-generation effort that lets developers opt out of the training set. The Hub is also seeing more opt-in datasets, trained only on data whose creators consented — especially relevant for text-to-image, where artists’ work in the training data is already a live debate.

BLOOM came out of BigScience, which Clem calls the largest collaboration in machine learning to date: a thousand researchers from 200 organizations training a large language model completely in the open, with the runs and the decision-making public. Jean Zay, a French supercomputer, provided the compute. The project shaped Hugging Face’s thinking on ethics. He sees two main challenges in AI: concentration of power, and biases encoded in the models. Building in the open is more part of the solution than the problem — power is more distributed, and people affected by the biases, especially underrepresented groups, can be in the process. That is hard to do, he says, “behind closed door in the lab in Silicon Valley with mostly old white dudes.”

RLHF was already on the Hub about eight months before it became a popular talking point, and the company is leading an open-source library so others can fold the technique into their models and workflows. Diffusers forced a similar reshape when a new generation of models took off. The job of an ML startup, Clem says, is to keep the flexibility to evolve every few months and to empower the community to take advantage of each new wave.

Business models for ML startups, Clem thinks, will be a big theme as the field matures. Hugging Face’s own shape is a classic freemium: 15,000 companies on the platform, 3,000 paying. The line between free and paid is still being drawn. Large customers such as Bloomberg and Meta want security and compliance features. There is also demand for infrastructure — upgrading Spaces to GPUs, inference endpoints — and, further out, for help optimizing the cost of running machine learning. That last piece ties back to his infrastructure obsession.

Gil offers GitHub as an imperfect analogy: so many adjacent businesses, from on-prem enterprise to open-source supply-chain monitoring, that some got done late or not at all. Clem first defends GitHub — 100 million software engineers, a billion dollars in annual revenue; if there was a mistake, it was selling to Microsoft too early. The miss he does take seriously is compute. Once a platform is the starting point of projects, companies come there to find the model, the dataset, and the demo, then make infrastructure decisions. That makes the hub a natural gate for compute. Hugging Face is testing that earlier than GitHub did, with infrastructure products and collaborations with the three big cloud providers.

Zooming out, he is most excited about biology and chemistry. He borrows Andrej Karpathy’s framing: software 1.0 was humans writing code; software 2.0 is machine-learning-powered technology. Look at the high-impact sectors that this new way of building could change, and biology and chemistry sit near the top. Microsoft had just released BioGPT; Meta was pushing on protein generation and prediction; models, datasets, and demos in those areas were rising on the Hub. He refuses to predict which applications will win — he has been wrong too many times.

What he will bet on is a type of company: full-stack machine-learning companies that do not merely use ML but build it as the core of the product. He names Runway, Grammarly, Wombo, PhotoRoom, and Stability. In software 1.0 there were companies assembling a site on Squarespace and companies actually building technology. He expects the same split in ML, and he thinks the native builders — Runway’s video work is the example he points at — will be able to challenge incumbents because they can construct the models, not just call them.