Oct 17, 2024 · 1h 7m · mad
The $4.5B Platform Driving the Open Source AI Revolution | Clem Delangue, CEO, Hugging Face
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews Hugging Face CEO and co-founder Clem Delangue about the critical importance of open-source artificial intelligence, the paradigm shift toward Software 2.0, and Hugging Face's journey to becoming a profitable $4.5 billion platform powering millions of specialized AI models globally.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 18.9% of the talking time here. How this is scored →
Matt as informed peer 3.5 Guest teaching 3.7 Guest disagreement 0.9 Matt pushing back 2.2
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:01:35 Rejecting Lean Startup dogmas
Clement explicitly instructs AI founders to 'trash the book' on Eric Ries's Lean Startup methods, forcefully arguing traditional software playbooks are obsolete in AI.
Hardest push from Matt ▶ 8:36 Questioning open-source monetary incentivesMatt directly challenges Clement's optimistic take on open-source AI motives, asking why anyone would open source in a high-stakes market unless forced by platform defense like Meta.
Biggest teaching moment ▶ 15:31 China's dominance in open-source AI modelsClement reveals that China has surpassed the US in open-source AI releases, explaining how American companies like Grok end up building on Chinese open foundations.
Matt holds his own ▶ 21:57 Framing OpenAI enterprise graduation riskMatt synthesizes Clement's model spectrum thesis into a sharp industry insight: commercial LLMs like OpenAI face 'graduation risk' as enterprise clients mature and transition to smaller custom models.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Distributed Team Dynamics and Geographic Strategy | 2 | 2 | 0 | 1 | Matt asks about Clement's move from New York to Miami and Hugging Face's distributed team structure across Paris, New York, and SF. Clement explains their remote/hybrid culture and international orientation. The dynamic is polite and conversational. | |
| The Current State of Open-Source AI vs. Proprietary Models | 3 | 3 | 1 | 2 | Matt asks about the current state of open source AI vs commercial APIs, noting Clement's prior predictions. Clement outlines how the field shifted from open research to closed API models, but asserts open source remains competitive for narrow/specialized use cases. | |
| Incentives, Business Models, and Policy in Open Source | 4 | 4 | 2 | 4 | Matt presses Clement on monetary incentives, challenging whether companies have any real motive to open source when billions are at stake beyond corporate strategy like Meta's Llama. Clement reframes AI as science-driven and cites open-source commercial models like Red Hat and Elastic. | |
| Legislative Approaches and Defining the Open-Source Gradient | 4 | 5 | 1 | 3 | Matt mentions watching Clement testify before Congress and asks about legislative developments like California's bills. Clement explains the open-source gradient from full model weights/data transparency to partial openness. | |
| Geopolitics of AI and China's Open-Source Leadership | 4 | 5 | 1 | 3 | Matt asks if open-source deceleration is primarily a US phenomenon. Clement reveals that China is currently leading open-source AI publishing, forcing American platforms like Grok to rely on Chinese foundations. | |
| Large Models vs. Small Specialized Models | 3 | 4 | 2 | 2 | Matt prompts a discussion on large generalist models versus small specialized models. Clement dismisses the 'one model to rule them all' concept as a fallacy, noting Hugging Face reached 1 million public models. | |
| The Software 2.0 Paradigm and Code Repositories Analogy | 5 | 4 | 2 | 4 | Clement uses a code repository analogy to explain Software 2.0. Matt offers a sharp synthesis, asking if commercial LLMs face enterprise 'graduation risk' as users move toward smaller customized models. Clement agrees while adding nuance about technology cycle hype. | |
| Hugging Face's Origin Story and Pivoting to Transformers | 3 | 3 | 0 | 1 | Matt asks for Hugging Face's founding story. Clement recounts starting as an entertaining Tamagotchi AI chatbot before Thomas Wolf ported Google's BERT model to PyTorch over a weekend, initiating the pivot. | |
| Building a Community-Driven Platform | 3 | 3 | 0 | 1 | Matt asks when Hugging Face transitioned into a general platform. Clement outlines their community metrics, including 5 million builders, 1 million public models, and 300k spaces. | |
| Competitive Moat, Network Effects, and Profitability | 4 | 3 | 1 | 3 | Matt asks why prior attempts at creating 'GitHub for ML' failed while Hugging Face succeeded, noting Hugging Face has virtually no direct competitors left. Clement points to timing, luck, and founder consumer background. | |
| Venture Capital Dynamics and Hugging Face's M&A Strategy | 4 | 4 | 2 | 3 | Clement surprises Matt by mentioning Hugging Face is profitable. Matt reacts with surprise ('Meaning you don't burn five million a year?'). Clement discusses market rationalization and incoming M&A interest. | |
| Acquisition Highlights: Argilla and XetHub | 3 | 4 | 0 | 2 | Matt asks about recent acquisitions. Clement details acquiring Argilla (dataset annotation) and XetHub (model/dataset versioning infrastructure) and explains their decentralized acquisition strategy. | |
| Emerging Models and the Shift to Multimodal AI | 3 | 4 | 0 | 1 | Matt asks about trending models on the platform. Clement highlights non-text modalities like OCR, video generation, and an IBM/NASA climate prediction model. | |
| Hugging Face Spaces, Private Mode, and Collaboration | 3 | 4 | 0 | 2 | Matt and Clement discuss Hugging Face Spaces, evaluation leaderboards, and private enterprise mode (noting almost 1 million private models exist on the platform). | |
| Hugging Face Business Model and Enterprise Hub | 4 | 3 | 0 | 2 | Matt breaks down the pricing structure across tiers and enterprise hub. Clement details how enterprise hub, support bundles, and compute endpoints monetize the freemium platform. | |
| In-House Model Development and the Story of IDEFICS | 4 | 3 | 1 | 2 | Matt asks about Hugging Face's internal model development, referencing Bloom. Clement explains they train open models like IDEFICS to fill market gaps and encourage open-source dynamics in other labs. | |
| Hugging Face Culture and Angel Investment Insights | 4 | 5 | 3 | 3 | Clement discusses Hugging Face's unconventional culture, telling founders to 'trash the book' on Eric Ries's Lean Startup principles. Matt notes Clement's lack of traditional KPIs and asks about his angel investing strategy. | |
| Expanding the AI Builder Community and Long-Term Vision | 3 | 3 | 0 | 1 | Matt asks about long-term vision. Clement outlines expanding the AI builder community, prioritizing non-text modalities, datasets, and better evaluation methods. |
Statements from this episode (31)
Delangue: Open-source AI leads closed-source in most text applications
“Open source is actually heads of, closed source for most of the text applications today, especially when you have, like, a very specific, narrow use case.”
Delangue: OpenAI uses open source and is a Hugging Face customer
“As a matter of fact, open AI is using open source as a customer of Fergingface”
Delangue: AI will create open-source commercial giants like Red Hat and MongoDB
“And I would expect there to be even more in AI. We'll have kind of like the MongoDB, the Elastic, the Red Hat of AI, which are going to be like open source AI, AI companies.”
Delangue: Policy makers will require AI to be shared as public infrastructure
“I believe the policy makers will require it to be shared across as some sort of kind of like a fundamental infrastructure for all.”
Delangue: AI regulation should address present risks, not doomsday scenarios
“In general, for me, I'm in favor of the regulations that are creating more competition, more opportunities for everyone to build with AI, and tackling some of the end user risks of AI today, like the risk of misinformation, the risk of biases. Instead of focus…”
Delangue: Full open-source AI requires sharing weights, datasets, and training scripts
“So it's a gradient, right? Openness is a gradient from kind of like full open source AI, which would be considered kind of like a practice where you share the weights, you share the data sets that you use to train the model. You usually share the training scri…”
Delangue: China is likely the current leader in open-source AI
“China has been publishing much more open source AI lately. I would argue that they're probably the leader today of open source AI which is surprising to some, but for example, in, in video, they've been leading in, in open source AI.”
Delangue: France promotes open-source AI much more than the US
“France also has promoted open source much more than the U.S.”
AI will split between massive generalist and tiny specialized models
“So I think we're gonna end up in a world where there are both Right? On both sides of the spectrum, some extremely large, powerful, generalist, costly models for some use cases, all the way down to a very specialized, simple, optimized models, and depending on…”
Delangue: Every company and use case will have its own optimized model
“So the same way today, every company, every product has its own code base. Tomorrow every company, every use case is going to have its own model optimized.”
Delangue: Companies start AI cycles with large models, move to smaller ones
“We've seen a lot of companies who started their AI life cycle more with like a large model and then moving on to smaller models.”
Hugging Face spent over three years building an AI chatbot before pivoting
“Started building this Tamagotchi AI, right? So at the time it's like a Siri Alexa, but we wanted to make it entertaining. We did that for, ah, a bit more than three years”
Delangue: Hugging Face has 5M builders and 1M public models
“Today, as I mentioned, where we have five million AI builders using, using our platform, who collaboratively shared one million public models.”
Delangue: Half of Hugging Face's 1M public models downloaded recently
“Half of them have been downloaded in the past 30 days.”
Delangue: Users contributed over 200k datasets to Hugging Face
“They also contributed, I think it's more than 200,000 data sets to the platform.”
Delangue: Users built over 300k app spaces on Hugging Face
“And collectively they built over 300,000 spaces which are the apps, ah, on, on the Hugging Face platform.”
Delangue: Consumer background drove Hugging Face's community-first model over B2B enterprise sales
“Before we had some experience around consumer, and so I think it helped us to have a very community driven approach instead of maybe a more, like, enterprise driven approach. Or more traditional kind of like B to B approach.”
Delangue: Hugging Face acquired two startups in four months
“We actually acquired, acquired two in the past four months.”
Delangue: Hugging Face is profitable with most of $500M raised in bank
“We're lucky to be profitable. We raised a bit less than five hundred million so far. Most of it is still in the bank.”
Delangue: Non-text AI modalities will transform technology like NLP did
“It's gonna be audio, it's gonna be video, it's gonna be biology, chemistry, AI. These are, in my opinion the domains that in the next few years are going to change the world the same way NLP has in the past.”
Delangue: Open-source video generation will match Stable Diffusion's impact
“We're gonna have the stable diffusion model for video generation soon. There are a lot of, like proprietary models in in video generation, but the first companies that are going to release an open source video Generation model. I think it is going to have a tr…”
Delangue: Hugging Face has almost 1 million private models
“A big part of the usage of the Hugging Face platform now is in private mode. Almost as much as what is public. So there's almost one million private models on the platform.”
Delangue: Over 5,000 leaderboards have been created on Hugging Face
“I think we crossed 5000 leaderboards that have been created on the Hugging Face platform as a way to rank your models based on their accuracy and their performance.”
Most Hugging Face users will stay open source and free
“We think most of our usage and users are always going to stay open source and free and we need to build some sort of a freemium model.”
Microsoft and NVIDIA are massive Hugging Face enterprise customers
“When you think about Microsoft, when you think about NVIDIA all, all these companies are massive customers of ours.”
Over 2,000 companies, including NVIDIA, use Hugging Face's Enterprise Hub
“We have over 2000 companies using the enterprise hub today from kind of like Mercedes, Bloomberg all the way to larger ones like Nvidia, for example, that I was mentioning.”
Delangue: Traditional software startup rules are outdated for AI
“A lot of the things that we learn with traditional software are like outdated with AI.”
Delangue has angel invested in nearly 100 startups in past two years
“I've invested in almost a hundred startups in the past, past two years.”