Keep pulling the thread on François Chollet.
Tesla is the only self-driving car company that has a large dataset of videos which include the corresponding driver actions.
The company Wave has raised $1.5 billion to develop its Gaia model, a world model-based approach for self-driving cars.
The DreamZero paper demonstrated training a joint model of state transitions and actions by initializing it with an open-source video diffusion model and fine-tuning on 500 hours of teleoperated data.
A 1967 study by Richardson demonstrated that a group who only imagined making basketball layups while blindfolded improved their performance by 23% compared to a control group.
To select a single move in a game, the AlphaGo system performs 800 Monte Carlo Tree Search (MCTS) simulations, each involving approximately 30 model invocations, resulting in about 24,000 total model invocations per move.
The AlphaGo approach does not scale to problems with large action spaces.
The DreamerV4 model was the first to successfully mine diamonds in Minecraft, a notoriously difficult task, by training a policy primarily on synthetic data generated from a world model.
Sam Altman believes that a new architecture will eventually be more performant than the Transformer.
The speaker predicts that world models will be the key technology to unlock full self-driving capabilities.
Current state-of-the-art AI systems, considered to be at the frontier of intelligence, are largely unable to solve the puzzles in the ArcGI test sets.
Multiple Y Combinator portfolio companies are actively working on various components of the robotics stack, including data collection, world models, action models, and physical embodiments.
Despite having a large dataset of human driving examples, Tesla's Full Self-Driving (FSD) system is not yet at the level of Waymo's performance and is not generally available.