Future Paradigms.

"Predicting the next token is discovering the hidden laws of physics."

Focus on Superalignment. As models evolve toward AGI, the compression of data becomes a world-model executable, enabling System 2 reasoning loops.

"The LLM is the new CPU; the context window is the RAM."

Envisioning the LLM OS: A future where AI manages memory, runs local code snippets, and interacts with peripheral APIs as a unified kernel for all computing.

"True intelligence requires consequence prediction, not just generation."

Advocating for V-JEPA: Moving away from generative autoregression toward predictive world-models that understand physical constraints and logic.