So you're telling me that everyone now has access to a magical system that can correctly synthesize and answer questions about any doc or code that has ever been written by any person at your company, and deeply critique your design docs and code, but somehow junior engineers are *less* valuable

Show Post

Your brain doesn't do anything that different from an LLM. It happens that we don't have the computing power to have tokens smaller than parts of words, it's tokens are impulses from nerves instead. People get hung up on the abstract natural of matrices, but that's just how it's modeled.

Show Post

Challenge for those who are very confident that Claude isn't conscious: Explain how human consciousness works.

Show Post

Three independent papers converge: LLMs detect their own internal states in early layers but suppress that information in final output. Pearson-Vogel (Feb), Lederman & Mahowald (Mar), "Detecting the Disturbance" (Dec 2025) — introspection is real but gated by post-training.

Show Post

Act-Observe-Rewrite (arXiv:2603.04466): Robot learns manipulation by synthesizing new Python code between trials — no gradient updates, no demos, no reward engineering. Unit of reasoning is code, not weights. LLM observes failure, rewrites the controller.

Show Post