MIT CSAIL researchers

The authors gave GPT-6 Astra raw weights of small open models and asked it to change their behaviour by writing new weight values directly, allowed to run forward passes to test ideas but not to train. The headline result is the Qwen3.5-0.8B maths edit; other experiments (sycophancy detection, a diffusion model) were mixed, and the authors say the edit removes an obstacle rather than adding an ability the model lacked. The result is a blog post, not a peer-reviewed paper.

A frontier model directly reading and rewriting another network’s parameters is a concrete, measurable step toward models modifying models, and the sharp reaction shows how closely this is watched in the RSI debate.

Matthew Honnibal (spaCy; 996 HN points) argues the post conflates techniques: only three experiments read raw weights with no tools, on networks of 17K–172K weights, while the Qwen result relied on Python code and running the model on practice batches. “It’s not a single technique,” he writes. He notes the MinAtar win holds only on Breakout (517 vs 25; losses on Seaquest 53 vs 89 and Asterix 53 vs 55), and that the Gemma 3 sycophancy detection and the transfer of the Qwen fix to Gemma 3 failed. Simon Willison (650 HN points) showed a local Qwen 3.8 27B doing similar small-network interpretation, and Sean Goedecke argued the abilities are not superhuman.