YC Root Access

Lessons from Building Open Source Libraries

During last month’s NeurIPS 2025 conference, YC’s Diana Hu sat down with Thomas Wolf, co-founder and CSO of Hugging Face to discuss his unconventional journey from physics and law to building one of the most influential open-source AI platforms. They discussed why open research accelerates innovation, the real challenges of turning AI demos into products, and how great open models and the application layer unlock the biggest opportunities for founders. Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.ycombinator.com/jobs Chapters: 00:00 — From Physicist to Hugging Face Founder 01:50 — Switching Careers 02:45 — How Hugging Face Was Born (Almost by Accident) 04:50 — The Limits of Closed Models 05:45 — Why Demos Often Don’t Become Real Products 07:05 — Fine-Tuning vs. Scaffolding: Startup Tradeoffs 08:40 — Turning Research into Widely Used Products 09:50 — Designing Great Developer Experiences 11:55 — The Future: Open Models and the App Layer

DHDiana HuhostTWThomas Wolfguest

Jan 16, 2026·14m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Thomas Wolf’s early career: physics research, depth, and choosing collaborators

   Thomas recounts his time in physics at Berkeley and on projects like laser fusion experiments and superconducting materials. He explains that picking the people he wanted to work with mattered more than the exact topic, and that desire for deep exploration shaped his later work.

2. 2

   Switching from academia to law: valuing time and focus

   After years in a PhD and post-doc track, Thomas pivoted into law because he loved writing and wanted a new challenge. He contrasts academia’s open-ended exploration with law’s strict accounting of time, which taught him to allocate attention more deliberately.

3. 3

   How Hugging Face began “almost by accident”: from game company to viral library

   Hugging Face started as a game company, and Thomas joined largely because he needed a job in the US. His exploration of deep learning turned into an open source library that spread quickly, leading the team to pivot the company around it and discover their mission along the way.

4. 4

   The open source mission: community, open science, and distributing power

   Thomas frames Hugging Face’s purpose as enabling a win-win ecosystem through openness rather than zero-sum competition. He argues that distributing power via platforms and community can catalyze innovation, allowing others to build major businesses on top.

5. 5

   Why open source accelerates AI: collaboration and faster iteration

   Thomas calls open source one of computing’s greatest contributions and emphasizes how it speeds progress by giving researchers a starting point. In AI, where many advances are incremental modifications, access to code and models prevents repeated reinvention and enables rapid iteration.

6. 6

   Exploration unlocked: using strong base models for unexpected use cases

   He highlights how open models let builders explore novel applications the original creators may never have anticipated. With a powerful pretrained foundation (often reflecting massive compute and data), developers can add components—like interactivity—to probe new product directions quickly.

7. 7

   Limits of closed models: constrained usage and domain mismatch

   Thomas contrasts open access with the restrictions of closed APIs, where you can mostly use systems as intended by the provider. If a use case falls outside the training domain (e.g., niche DSLs), lack of access makes it hard to adapt behavior meaningfully.

8. 8

   Why impressive demos don’t become real products: production is scaffolding-heavy

   Diana and Thomas discuss how easy demos can hide the complexity of shipping a reliable product. Thomas notes that both open and closed model approaches require heavy scaffolding—pre-processing, edge-case handling, and domain expertise—because out-of-the-box reliability is still rare.

9. 9

   Fine-tuning vs. scaffolding: startup tradeoffs and where to invest effort

   Thomas explains that closed-model teams often rely primarily on scaffolding, while open-model teams can choose fine-tuning or training—at the cost of operational complexity. For small startups, deciding whether fine-tuning is core to the product is a key resource-allocation decision, though emerging tools may reduce the burden.

10. 10

    Turning research into widely used tools: treat open source like product

    Thomas reframes his role: building libraries like Transformers and Datasets is product work, with the open source community as demanding customers. Because switching costs are low and expectations are high, teams must obsess over usability and make adoption effortless.

11. 11

    Designing great developer experience: onboarding, minimal abstractions, and taste

    He outlines two core principles: a delightful first-run experience and carefully chosen abstraction levels. Great libraries minimize friction, avoid forcing users to read docs, and balance ease-of-use with control—an opinionated design challenge that requires repeated self-testing and fresh perspectives.

12. 12

    The future: open models catching up and value shifting to the app layer

    Thomas argues open models are already nearing closed-model performance, citing examples like DeepSeek and Kimi as watershed moments. As parity increases, he expects more value to concentrate in interaction design and the application layer—areas where startups can compete without training frontier models.