Video model scaling through non-semantic 4D vision tasks. Talks about scalling to 22 billion parameters

I'm catching up on current mechanistic interpretability work. Neel Nanda recommends this paper as a check on how well you understand transformers.

AI Village is an ongoing experiment in multi-agent systems, where computer use agents are given broad goals to work toward and anyone can observe their dynamics. The article highlights some interesting behavior from the Gemini family of models

Dario Amodei updating his current AI timelines.

RLM Author interview. Useful for intuitions behind the paper that don't make it into the write-up. Prime Intellect has adopted RLMs as a major research focus, calling it "the paradigm of 2026." The framing of context as environment rather than input feels like it'll become standard scaffolding for code agents and long-horizon tasks.

I'm seeing frequent backlinks to this as a potential next step from reasoning models, especially when it comes to tool use