To build generalist robots, the most effective approach is pre-training foundation models on internet-scale video datasets, not just simulation or tele-operated data. This vast, diverse data provides a deep, implicit understanding of physics and object interaction that is impossible to replicate in controlled environments, enabling true generalization.

The Physical Intelligence thesis is that a foundation model learning from diverse data can achieve a "physical understanding" of the world, making it easier to adapt to new tasks than building single-purpose robots from scratch. Generality leverages broader data, which is ultimately a more scalable approach.

Computer scientist Rich Sutton's "bitter lesson" is evolving. The new frontier for AI performance isn't just more pre-training data; it's vast amounts of "experiential data" from real-world user interactions. Models post-trained on this experience data are beginning to outperform those trained only on static, human-knowledge datasets.

Contrary to the belief that AI requires perfect, clean data, the biggest opportunity lies in building technology that can find signals in messy, diverse data sets across different modalities and organisms. The tech should solve the data problem, not wait for it to be solved.

Figure is observing that data from one robot performing a task (e.g., moving packages in a warehouse) improves the performance of other robots on completely different tasks (e.g., folding laundry at home). This powerful transfer learning, enabled by deep learning, is a key driver for scaling general-purpose capabilities.

For physical AI systems like robots, data quality hinges on diversity, not just quantity. A robot trained to make a bed in one specific lighting condition may fail completely if the lighting changes or the bed is moved. This brittleness highlights a key challenge: training data must capture a wide variety of contexts and edge cases to enable real-world generalization.

Numenos AI found that unifying biological data without traditional borders, such as incorporating mouse data or cancer data for dermatological diseases, surprisingly increases the predictive accuracy of their models. This challenges the siloed approach to traditional research.

Dario Amodei views the distinction between RL and pre-training scaling as a red herring. He argues that, just like early language models needed broad internet-scale data to generalize (GPT-2 vs. GPT-1), RL needs to move beyond narrow tasks to a wide variety of environments to achieve true generalization.

Physical Intelligence demonstrated an emergent capability where its robotics model, after reaching a certain performance threshold, significantly improved by training on egocentric human video. This solves a major bottleneck by leveraging vast, existing video datasets instead of expensive, limited teleoperated data.