Generalist AI's Pete Florence on building foundation models for robotic dexterity — and why humanoids aren't the whole story

Generalist AI's bet on robotic dexterity

Pete Florence spent a decade in robotics before founding Generalist AI — a PhD at MIT starting in 2014, then Google DeepMind, where he published extensively alongside co-founder Andy Zeng. The third co-founder, Andy Barry, came out of Boston Dynamics. The company has raised $400M at a $2B valuation.

Form factor agnosticism

Florence deliberately avoids building toward any single robot body. He frames Generalist AI as building the underlying engine technology rather than a specific vehicle — foundation models for dexterity that can run on humanoids, industrial arms, or whatever form factors emerge. Humanoids will be part of the future, he says, but the future is much bigger than humanoids alone, and billions of robots will take many shapes.

The dexterity unlock

The near-term commercial case isn't replacing robots that already work. It's unlocking tasks that no robot has handled before. Wire harnessing in auto manufacturing is the clearest example Florence gives — cables and flexible objects are trivially easy for humans but completely out of scope for traditionally programmed arms. The pattern repeating across industries, in his telling, is manufacturers who never considered robotics for a given application suddenly realizing these models might make it viable for the first time.

“We announced our Gen Zero model back in November — the first time in robotics that anybody had shown general scaling laws, where we can predictably advance performance with more and more compute and data. Gen One, announced in April, is starting to cross into levels of performance that we think are commercially viable for a good number of applications. Lived experience of the physical world is S tier data.”

Scaling laws in the physical world

Generalist AI released its Gen Zero model in November 2024, which Florence says was the first demonstration of general scaling laws in robotics — predictable performance gains from more compute and data. Five months later, Gen One followed. He draws a direct parallel to the GPT-2 to GPT-3 transition, when language models crossed into commercial viability for narrow applications like ad copywriting. He believes physical-world models are crossing a similar threshold now, where reliability, speed, and improvisational intelligence are good enough for a meaningful subset of real deployments.

Data tier list

On what actually produces useful training data, Florence is direct. Lived physical experience — real robot interaction data with high-quality manipulation tasks — is the only S-tier source. Everything else ranks lower:

• Internet video (YouTube etc.): B tier. Useful but not close to the top.
• Internet text / web crawl: Good foundation, analogous to reading about skiing before going skiing. Valuable but not sufficient on its own.
• World models as a synthetic data source: C tier, with Florence visibly uncomfortable going lower in front of the cameras. He acknowledges it's a promising area but says there are few proof points of synthetic data actually driving robotic capability.
• Motion capture: C tier for his purposes — useful for whole-body motion but not where dexterity data comes from.
• Simulation (Unreal Engine, physics-based): C tier.

The through-line is that data quality and what the robot is actually doing during collection matter more than the capture methodology itself.

Industrial before consumer

Florence expects industrial deployments to scale faster than home or consumer applications. The economic logic is straightforward — factories are structured, the ROI calculation is cleaner, and the tasks are more bounded. Consumer and home robotics remain on the roadmap, but he treats them as a later wave.

Generalist AI's structural bet is that owning the model layer — deployable across form factors and industries — gives it a larger addressable market than any single humanoid hardware play, and a faster path to revenue through the industrial applications already hungry for dexterity solutions no prior generation of robots could touch.