Writing · Robotics
Robots learn from people
Language models learned from trillions of words. Robots have no such library yet. In robotics the scarce thing is now data on how people work.

In 1988 the roboticist Hans Moravec wrote that “it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility”. Nearly four decades later his paradox still holds. A language model can pass a bar exam. A robot that folds laundry is still news.
The missing library
Part of the reason is data. Meta trained its Llama 3 language models on more than 15 trillion tokens of text, collected from public sources, most of it written by people for other people. There is no such archive for what hands do. Nobody has spent decades writing down the forces in a hand as it opens a jar.
The largest open effort so far shows the gap. In 2023 Google DeepMind and partners at 21 institutions pooled data from 22 types of robot into the Open X-Embodiment dataset: more than a million recorded episodes covering over 500 skills. It was a landmark for the field. Next to the text that trains a language model, it is a pamphlet.
Everyone is collecting
Money has followed the problem. Physical Intelligence, a San Francisco startup founded in 2024 to build general models for robots, raised $400m in November 2024 and another $600m a year later at a valuation of $5.6bn. Since then it has been reported in talks to raise again at a far higher price. It has named getting enough data, from many tasks and many robots, as one of its main challenges.
The methods differ. Some companies steer robots remotely and record every move. Others put people in motion-capture suits. In 2025 Tesla told the staff training its Optimus robot that it would rely mainly on video instead, with workers wearing camera rigs while they fold shirts and pick up objects, because video scales faster.
Where we come in
All of these approaches share one idea: the best teacher of physical work is a person doing it well. A mechanic feels a bolt about to strip. A nurse knows how to turn a patient without hurting them. That knowledge is what robots lack, and most of it has never been recorded.
That is why two of our companies work on it. Archaí Physics records how people really work, so that robots can learn from it. physical.wiki collects what objects do in the hand: their weight, their grip, how they open and how they break.
Hardware gets cheaper every year. Good recordings of skilled work, made with the consent of the people who do it, are still rare, and that is where we put our money.
Sources
- Hans Moravec: Mind Children, Harvard University Press, 1988
- Wikipedia: Moravec’s paradox
- OpenAI: GPT-4, March 2023
- Meta AI: Introducing Meta Llama 3, April 2024
- Google DeepMind: Scaling up learning across many different robot types, October 2023
- Open X-Embodiment: Robotic learning datasets and RT-X models
- Bloomberg: Robotics startup Physical Intelligence valued at $5.6 billion, November 2025
- Dataconomy: Bezos-backed Physical Intelligence raises $600M, November 2025
- eWeek, reporting Business Insider: Tesla’s Optimus has a new training strategy, August 2025
- TechCrunch: Physical Intelligence in talks to raise $1 billion again, March 2026
More writing
CapitalEfficiency is the new moatDeepSeek wiped $589bn off Nvidia in a day by doing more with less. The lesson for founders is older than AI: whoever needs the least money is hardest to catch.
FoundersA business plan is a test of judgementEisenhower said plans are worthless but planning is everything. We read business plans for the planning. Here is what we look at first.
PartnershipChemistry is due diligenceNoam Wasserman found that 65% of high-potential startups fail because of conflict among co-founders. So we check the people before we read the plan.