What's coming
Running the model & the world
The idea
Section titled “The idea”The open fronts are extensions, not replacements. Multimodality treats audio and images as more token streams for the same machinery. Agents stretch the loop from answering in seconds to carrying tasks across hours - which leans directly on context, tools, and honest self-limits. And use-time learning - models that keep improving after deployment - is this site’s closing question, because today everything learned is frozen into weights before you ever type (Karpathy’s preview at ~3:09-3:15).
One adjacent story is no longer just an open front: world models predict the next state, not the next token. Start at what is a world model; the JEPA spine runs through V-JEPA 2-AC (the planner) and LeVJEPA (the efficient encoder).
The story, animated
Section titled “The story, animated”More streams, a longer loop, a still-frozen memory - and one adjacent story handed off at the close.
Where it sits
Section titled “Where it sits”Worth having first: RL & reasoning models · Hallucination & tool use
Source-checked against Karpathy, “Deep Dive into LLMs like ChatGPT” (2025), ch. 3:09-3:15.