Interview
Building the "App Store" for Robots: Hugging Face's Thomas Wolf on Physical AI
- An immediate surge in activity is anticipated for basic robotic building blocks, with startups already automating manual tests and physical tasks using the $100 SO-100 arm, while mass-market hardware for families is scheduled to launch starting "this summer" to move beyond the hobbyist segment.
- A transition is predicted where 100 to 200 million software developers become "roboticists" utilizing tools like "vibe coding" for devices such as the Richie Mini, enabling easy behavior tweaking for children and mirroring the shift seen in AI awareness.
- The consumer robot market is viewed as a bet for the first "ChatGPT moment" or "iPhone moment" to emerge in entertainment, fun, education, and physical learning rather than immediate enterprise deployment, with plans to rebuild an "app store of iPhone" model for sharing behaviors and integrating new VLN or chat models.
- Open-source world models are expected to drive a breakthrough in simulation data generation by fixing image generation coherence for video, facilitating a "galaxy of different form factors" via diverse datasets and solving generalization issues through user-shared recordings from various locations.
- Form factor diversity will be prioritized over humanoids in the near term due to high costs preventing mass adoption below car prices, with a progressive evolution toward humanoids only after tools are refined for smaller robots like the Richie Mini.
- Over a 10-year timeframe, the goal is to make robots accessible to many rather than remaining an elite pursuit, achievable through cheaper hardware options and a global community capable of building with AI as a creative tool rather than just consuming it.
- The industry will likely shift toward using both large foundation models behind routers and smaller local models on laptops depending on reasoning or speed needs, while open-source is expected to remain a winning solution despite a current "turbulence phase."
- Foundational model teams in the US are expected to emerge to challenge current frontiers, while Chinese teams will continue dominating open-source competition due to internal market pressures and negative hiring impacts from not open-sourcing.
- A resurgence of open sourcing in the West is anticipated as companies seek to quickly rise to the top or fill market gaps, though safety and reliability concerns regarding model behavior will drive demand for guarantees of consistent performance.
- Long-term trends include AI acting as an accelerator for scientific discovery to multiply human production by 10 to 1,000 times, and research focusing on developing models that can form strong opinions and disagree rather than simply agreeing with users.
- Ultimately, recipes to train AI models are expected to become as accessible as reading a book on general relativity, allowing everyone to learn and train intelligent artifacts.