AI Daily Digest · 2026-11-03
🔬 New AI Craft
1. Editing weights to suppress a small model’s reasoning loops
Using training-practice feedback, GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights within 50 batched forward calls, without gradient updates. With thinking enabled and a 32,768-token cap, the authors report MATH-500 rising from 7.9% to 48.1%, mainly by suppressing loops; a Gemma follow-up did not establish a gain.
https://rikiyat.github.io/blog/weight-space-programming
2. Use held-out tests and event traces to guide AI-generated fixes
A KDE handwriting experiment rejected its font-based training approach after holding out an entire font family, then used pointer traces to diagnose input bugs. The author’s 79 tests cover pure logic, leaving compositor-and-pen behavior outside automated coverage. The GPL-3 WritePad backend stays disabled by default, and the branch is not upstream.
https://www.collabora.com/news-and-blog/blog/2026/11/02/sunscreen-a-stylus-and-16-hours-of-pair-programming-with-an-ai
🛠️ Tools & Tips
(no items this day)
⭐ Open Source Highlights
(no items this day)
📰 Industry News
1. OpenAI introduces LifeVal to track performance on everyday tasks
🚀 Major Releases
1. LongCat-2.5 API launches for native multimodal and long-horizon agent tasks
