-
The NeoLabs of Self-Evolving Agents
Six labs have raised roughly $900M betting on AI that improves itself, and each picked a different, non-overlapping layer. One of them shipped three papers instead of a valuation. I read all three, mapped every lab onto the two-substrate framing, and checked the claims against GitHub and Hugging Face.
-
Everything Is a File, Including the Agent
PenguinHarness makes recursive self-improvement a product feature by reducing it to file editing under a contract. I read the kernel, the four skills that make up its RSI loop, and the benchmark underneath — and found one gap the research literature says matters.
-
Two Schools of World Models
Forty-seven papers, one entry each, with the original architecture figure — from ViT and DDPM through Genie, Cosmos, TRELLIS and pi-0. One school understands the world by generating it; the other by acting in it.
