🤖 At the AI Startup School by Y Combinator, the strongest talk turned out to be a lecture on how robots work.
YC gathered its 2500 students from top CS programs for a conversation about AI for the first time—usually, such events focus on startups and pitches. However, one of the talks was given by Chelsea Finn, a Stanford professor and co-founder of Physical Intelligence. She explained how the company trains the pi-zero model: initially on a mix of data from different robots, then fine-tunes it on carefully selected demonstrations. The talk included an example—a robot learns to fold crumpled clothes and repeats this in unfamiliar apartments with items it has never seen before.
The second point is more interesting than the first—robots are taught to change plans right during task execution based on voice commands, not just follow instructions from the start. "Make a vegan sandwich, no cucumbers"—and the robot adjusts its movements on the fly because the language annotations for each action are generated by the models themselves.
Jensen Huang mentioned at CES 2025 that "the ChatGPT moment for robots is already here"—but this is more marketing than fact. The pi-0.7 model in April increased the success rate of unfamiliar tasks from 5% to 95% with voice prompts, which is impressive but still far from a breakthrough on the level of ChatGPT. Most in the industry estimate the real moment to be 3-5 years ahead—and if the forecast is correct, the physical world will change as quickly as chatbots did in 2023, just a bit later than what is said on stage.
Even then, distribution will be slower. ChatGPT spread worldwide in just a few months because it's simply code on servers. A robot needs to be physically manufactured and physically delivered—this is a completely different speed of scaling.
#robotics@rvnikita_blog #ai@rvnikita_blog #ycombinator@rvnikita_blog #chelsea_finn@rvnikita_blog