AI can patch a small model without training
Hey hackers, fixing a model that repeats itself might mean changing its weights, not writing another prompt. Today’s lead explains a specific repair experiment, plus we’ve got a documented tool for letting your agents work with desktop apps.
In today’s edition
- GPT-6 Astra patches a small model’s reasoning loops.
An AI patches another model without retraining
Researchers used GPT-6 Astra to directly edit Qwen3.5-0.8B’s weights, the numbers that shape its behavior, to stop repetitive reasoning loops. The experiment demonstrates a targeted way to repair a small model without gradient-based training.
- With a budget of 50 forward passes on practice problems, Astra diagnosed the looping behavior and changed just 0.016% of the model’s weights.
- The researchers report that its MATH-500 score rose from 7.9% to 48.1% with sampling settings held fixed.
What it means Read this if you’re trying to repair a small model that gets stuck repeating itself. Start with the Qwen loop-fixing experiment, not the broader self-improvement predictions.
Source: rikiyatakehi
Quick hits
- Anthropic’s amended filing warns AI-assisted research could outpace safeguards; no current model has crossed its internal acceleration threshold, and policy changes remain under consideration.
- OpenAI’s internal Pando backend cut Git push latency by about 80% without changing engineers’ commands; external availability is under consideration.
- Anthropic opened Residency applications for established technical leaders seeking an AI-research transition, with paid six-week to six-month placements, no ML-background requirement, and a November 22 deadline.
GitHub repo of the day
agent-desktop: let agents operate desktop apps
- What it does: The maintainers describe a tool that lets agents read desktop controls and act without matching pixels, then check results with a fresh snapshot.
- How to start: On macOS, install with npm install -g agent-desktop and grant Accessibility permission using agent-desktop permissions --guide. Before first use, have your agent retrieve and follow agent-desktop skills get desktop and agent-desktop skills get platform, then inspect Finder with agent-desktop snapshot --app Finder -i.