ArticleAIMachine Learningllmmlneural-networks
Sparks of Weight-Space Programming
MIT CSAIL post showing GPT-6 Astra reading, fixing and writing neural network weights directly, with no gradients or training. One edit to 0.016% of Qwen3.5-0.8B's weights took MATH-500 from 7.9% to 48.1%.
Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab hand a frontier model raw weights and ask it to reason about them. Given only the 17,253 weights of a small policy network, Astra names the game it was trained on. With no data, loss or rollouts, it writes MinAtar policies in one shot that beat PPO at 100 million steps in all three games. Against 500 million steps it wins only Breakout.
The headline edit is the Qwen one. Astra got a Python kernel over the weights and a budget of forward passes on practice problems, traced the model's thinking loops to one direction in the MLPs of layers 15 and 19, and subtracted it from 131,072 weights. GSM8K went from 13.8% to 67.1% and IFEval from 21.0% to 62.3%.
Read the fine print on that one. It fixes loops rather than adding math ability, letting the model finish answers it could already work out. Most of the gain needs the full 50 forward passes (with 10 or 25 the loops barely went away), and on Gemma 3 270M, which rarely loops, Astra couldn't improve math at all. The authors call these very early results. It's a blog post, not a peer-reviewed paper, but the complete logs of every experiment are public.
Related
- SPACEPerplexity's open-source sandbox platform, the one behind every task in Perplexity Computer. It runs agent and user code in isolated Firecracker microVMs behind an HTTP API.
- MCP Events (SEP-3415)Draft SEP that adds an Events extension to the Model Context Protocol. A client subscribes to upstream events like a Slack message, a repo push or a PagerDuty incident, and the agent reacts when one arrives instead of waiting for a cron job or a user message.
- OpenSEOOpen-source, pay-as-you-go alternative to Semrush and Ahrefs covering keyword research, rank tracking, backlinks, site audits, and AI visibility, with an MCP server so agents can work with your SEO data.
- WhistleCactus Compute's on-device speech-to-text model in a single 16.9 MB file that runs on the CPU with no dependencies and mostly beats Whisper base at a fraction of the size.