ACQ2: Building the Open Source AI Revolution (with Hugging Face CEO, Clem Delangue)

2024-10-14 - source - Read full transcript

Ben Gilbert (host)David Rosenthal (host)Clem Delangue

Key insights

Hugging Face pivoted from a teen chatbot app to an AI infrastructure platform because of unplanned community demand for a single open-source contribution.

In 2016 Hugging Face started as a conversational AI 'tamagotchi' aimed at teenagers and raised its first two rounds on that idea. Three years and about $6 million in, co-founder Thomas ported Google's newly released BERT model from TensorFlow into PyTorch over a weekend; the tweet announcing it drew outsized attention, other researchers asked to add their own models to the resulting library, and that organic demand pulled the company into building what became the Hugging Face platform.

hugging-face-origin-story

Hugging Face functions as 'GitHub for AI' - a collaborative hosting and versioning layer for models, datasets, and apps rather than a compute provider first.

Delangue frames the platform's core value as collaboration features (commenting, versioning, bug reports, reviews) that let growing AI teams work together, not just storage. Companies now run thousands of internal (often private) users through the platform - about half of all models, datasets, and apps hosted are private - while the public side has over five million AI builders and crosses roughly one million public models.

hugging-face-origin-story

Openness in AI research has measurably declined since 2017-2019 as commercial incentives and safety arguments took hold, even though open research is what produced the current wave of AI progress in the first place.

Delangue argues transformers, BERT, and GPT-2 all emerged from an era when Google and OpenAI shared research openly, and that collaborative openness is what compounded into today's progress. He views the more recent 'existential risk' argument against sharing AI research as a repeat of historical fear-based restriction arguments (e.g., limiting access to books), used to justify closing down what used to be open.

open-vs-closed-ai

A non-decentralized path to AGI is the outcome Delangue fears most, which is why he treats open source as a safety strategy, not just a business or ideological preference.

He argues that concentrating advanced AI capability in a single company or organization creates the highest risk, whereas distributing access broadly - to companies, policymakers, nonprofits, and civil society - creates more transparency into how models work and don't work, which he believes leads to a safer outcome than restriction.

open-vs-closed-ai

Hugging Face avoided the capital-intensive 'foundation model' trap by design and is profitable on roughly $500 million raised, less than half of which has been spent.

Rather than compete on training frontier models or race compute prices to the bottom against hyperscalers, Hugging Face monetizes via a freemium platform, markup on bundled/integrated compute, and an enterprise Hub subscription - letting it stay sustainable without the massive capex burden that defines companies like OpenAI.

ai-business-models

The AI startup playbook is not the software lean-startup playbook, and founders who apply the old one will misjudge timelines and team composition.

Delangue argues AI progress looks like six months of no visible product improvement followed by a step-function 10x gain, unlike software's compounding 5%-better-every-two-weeks iteration model. He advises founders to prioritize a scientist co-founder over pure engineering talent and to expect training/optimization cycles measured in months, not days.

ai-startup-playbook

Delangue predicts most AI value will require companies to build and own their own models over time, not permanently rent capability through APIs.

He compares API-dependent AI companies today to the pre-broad-coding era when non-technical builders used no-code tools like Squarespace or Dreamweaver before software engineering became a widely learned skill. He expects companies to eventually build, fine-tune, and optimize their own models for their specific domains, constraints, and customers as AI capability becomes more accessible.

ai-startup-playbook

Delangue expects proliferation, not consolidation, in foundation models - arguing there could eventually be almost as many models as there are code repositories today.

He pushes back on the 5-8-dominant-players thesis, noting Hugging Face sees a new model, dataset, or app published roughly every 10 seconds, driven by models optimized for specific domains, latency, hardware, and use cases rather than one generalist model serving everyone. He expects a gradient between a few very large generalist models (e.g., ChatGPT-style) and many small, cheap, purpose-built ones.

ai-moats-and-competition

AI moats will likely resemble software moats (network effects, cost economies of scale) rather than being fundamentally new, and no one has proven a durable moat yet.

Delangue notes neither OpenAI nor Hugging Face had a unique data or resource advantage at the start - OpenAI scraped the same public web everyone else could, and Hugging Face's edge came purely from being community-driven early enough to build network effects, the same dynamic that has made GitHub and social networks hard to unseat.

ai-moats-and-competition

Delangue expects AI to expand who can build technology far beyond today's roughly 50 million software engineers, potentially producing more AI builders than software builders within a few years.

He estimates around five million people currently qualify as 'AI builders' (roughly Hugging Face's active user count) versus roughly 50 million software engineers/builders worldwide, and argues AI has a lower barrier to contribution than writing code - contributing data or domain expertise counts - so the builder population could grow 10x and produce more inclusive, broadly-aligned products.

ai-startup-playbook

Companies

Hugging Face - Delangue's company; described as the number one platform for AI builders, valued at $4.5 billion at time of recording, with over five million AI builders and a new model/dataset/app published every 10 seconds
GitHub - Repeated analogy throughout - Hugging Face is positioned as 'GitHub for AI,' the collaborative hosting layer for models and datasets the way GitHub is for code
Nvidia - Named as one of Hugging Face's investors and as a company with thousands of internal users on the Hugging Face platform
Salesforce - Named as a Hugging Face investor and as a company with thousands of internal platform users; also where early angel investor Richard Socher was chief scientist
Google - Hugging Face investor; also the source of the original BERT transformer model (released in TensorFlow) that Hugging Face's PyTorch port kicked off the company's pivot
Amazon - Named as a Hugging Face investor
Intel - Named as a Hugging Face investor
AMD - Named as a Hugging Face investor
Qualcomm - Named as a Hugging Face investor
IBM - Named as a Hugging Face investor
Microsoft - Cited as a company with thousands of internal users on the Hugging Face platform
OpenAI - Discussed extensively as the closed-source counterpoint; noted for open-sourcing GPT-2, for its early transformer/GPT research, and as the reference point for foundation-model capital intensity and business-model uncertainty
Betaworks - Hugging Face's earliest investor (New York), backed the company when it was still an AI chatbot/emoji-companion startup
Mistral - Referenced twice - founder Guillaume Lample contributed an early model (XLNet) to Hugging Face's library, and later as an example of a capital-intensive foundation-model startup
MOOD Stocks - The Paris machine-learning-for-computer-vision startup where Delangue worked before founding Hugging Face, which shaped his early interest in AI
Anthropic - Named (transcribed as 'Entropik' in the source audio-to-text) alongside OpenAI as a closed-source company that still relies heavily on open research and open source underneath

Techniques and frameworks

community-driven product development - Hugging Face's core method: build only what users in the open-source community ask for, which Delangue credits as the main driver of the company's growth from a single PyTorch model port into a full platform
locked-in compute / bundled infrastructure - Hugging Face's business model - rather than competing on raw compute price, it bundles compute with platform features (inference endpoints, Spaces GPUs) so tightly that switching to a raw cloud provider becomes more work than it's worth, letting Hugging Face charge a sustainable markup
software 2.0 - Delangue's reframing of AI as a new paradigm for building technology (training models on data instead of writing code), deliberately downplaying AGI/sci-fi framing in favor of treating it as an evolution of software
science-first founding teams - Delangue's advice for AI startups: prioritize a scientist co-founder over the classic software-engineer/lean-startup team structure, because AI progress comes from months-long research pushes toward step-function gains rather than iterative 5%-better releases

Summary

Ben Gilbert and David Rosenthal interview Hugging Face co-founder and CEO Clem Delangue for ACQ2, tracing the company's unlikely path from a 2016 teen chatbot startup (named after the hugging-face emoji) to the self-described "GitHub for AI" - a $4.5 billion platform used by over five million AI builders, backed by Nvidia, Salesforce, Google, Amazon, Intel, AMD, Qualcomm, and IBM. Delangue explains that Hugging Face spent three years and about $6 million building a conversational-AI "tamagotchi" before a spontaneous side project - co-founder Thomas porting Google's newly released BERT model from TensorFlow into PyTorch over a weekend - generated enough community demand that the company pivoted entirely, becoming the default hosting and collaboration layer for open models, datasets, and apps.

A large portion of the conversation is a defense of open source as both a technical and safety strategy. Delangue argues that the field has become measurably less open since 2017-2019, even though the openness of that earlier era (Google and OpenAI sharing transformer and GPT research freely) is precisely what produced today's progress. He frames existential-risk arguments against open-sourcing AI as a repeat of historical fear-based restriction (comparing it to controlling access to books), and says his real fear is the opposite: a non-decentralized path to AGI concentrated in one company, which he considers the highest-risk outcome. Distributing capability broadly - to companies, policymakers, and civil society - is, in his view, what makes the technology's future safer, not more dangerous.

On business model, Delangue is unusually candid that Hugging Face avoided the capital-intensity trap that defines frontier labs: the company has raised roughly $500 million over seven years, spent less than half of it, and is profitable, monetizing through a freemium platform, a markup on tightly-bundled ("locked-in") compute, and an enterprise Hub subscription rather than by training its own frontier models or racing hyperscalers on raw compute pricing.

The back half of the episode turns to predictions. Delangue argues the AI startup playbook has to be thrown out relative to software's lean-startup model - AI progress looks like months of invisible work followed by step-function gains, not compounding 5%-better iteration - and that founding teams need a science co-founder, not just engineering talent. Pushing back on Ben's expectation of 5-8 dominant foundation-model winners, Delangue predicts proliferation instead: near-parity between models and code repositories, with most companies eventually training and owning purpose-built models rather than permanently renting capability through APIs, similar to how software shifted from no-code tools like Squarespace and Dreamweaver to companies employing their own engineers. He closes on moats, arguing AI moats will likely mirror software's (network effects, cost economies of scale from compute providers) rather than being fundamentally new, since neither OpenAI nor Hugging Face had a unique structural advantage at the outset.

Throughout, the hosts draw repeated parallels to prior technology cycles - Web 2.0 API mashups, GitHub's early "curse of invisibility" as an infrastructure company, and the Yahoo/AltaVista-era uncertainty around foundation-model business models - while Delangue frames the moment as an opportunity to expand who gets to build technology, estimating the current ~5 million AI builders could grow toward the ~50 million software engineers worldwide, or beyond, given AI's lower barrier to contribution.

Notable Quotes

"I'm incredibly scared of a non-decentralized AGI. If only one company, one organization gets to AGI, I think that's when the risk is the highest." - Clem Delangue

"The best advice I give to people is to trash their lean startup book when they're starting an AI company." - Clem Delangue

"We might still be in the sort of Yahoo, Alta Vista era of foundational model companies." - David Rosenthal

"AI is the opportunity of the century to shake things up, break the monopolies, break and flag the established positions, and do something a bit new." - Clem Delangue