Ripple 65B, Midterm Data Center Bans, and AI That Writes Weights

Pierce and Rich recap the week. They start with Ripple and Recursive’s loop: better chips train better models, which then design better chips. Ripple is a dense transformer, while Muse Spark, Beam, Inkling and Nemotron are all mixture-of-experts, and they explain why going dense makes RL post-training easier. Ripple needed only about 10% of the rollouts Beam did. It beats the field on Terminal-Bench and SWE-bench but trails on math, safety and automation, so is “frontier” a stretch? Then there’s the $32M fully loaded R&D cost, which is less than you’d likely pay just to rent the final training run.

Next is the midterm backlash against data centers. They cover what people actually mean by an “AI ban,” why data center builders have defaulted to Texas and on-site gas turbines, why CoreWeave and Nebius fell after the results, and why water and power are solvable problems but jobs aren’t.

Last is the MIT weight-editing post. Astra identifies Breakout from raw weights, builds a Go player from a zero-initialized network, and fixes a Qwen model that keeps repeating itself on math problems.

They explain why the toy examples are easier than they look (Pierce reproduced the Breakout one), how the setup quietly changes across the nine experiments, and whether programming in weight space is the future of recursive self-improvement or just the assembly language of ML.