[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop

Abstractly, the demonstration is that you can take a sufficiently smart model and through running a series of experiments across a compute cluster, train a model autonomously without human expertise or intervention. However, this is usually done by an involved process of writing PyTorch/Jax code and running compute and watching loss curves and reading evals and optimizing kernels and data and…. Well, last month’s guest Alex Zhang comes back to coauthor a MONSTER discovery: what if you just ask a smart model to directly read and edit the weights of a dumber model? Of course the exact thing we are really modeling is some variant of mechanistic interpretability (which could be done by hand), and right now unlikely to scale to more advanced tasks than simple games and RL policies, but the world has now gone from 0 to 1 on this very new RSI capability…. which will immediately put to use and explored due to its remarkable efficiency. Release: DeepSeek Harness published a long design article (article) and the dsh-superagent-prototype repo (repo link). An official English version followed (@tianyi). Architecture: One long-lived router session talks to humans and hands out work. Isolated workers run tasks in parallel, triggered by…