AI Daily Digest · 2026-11-08

🔬 New AI Craft

1. A single matrix edit improves synthetic support-intent classification
Giga used three ridge-regression solves to edit a final-layer matrix in Qwen3.5-0.8B, fitting recorded activations and selecting settings on a validation set. On 1,200 held-out synthetic conversations, the authors report exact-name free-generation accuracy rising from 39.9% to 80.9%, while general-English perplexity rose from 17.4 to 19.6.
https://giga.ai/news/one-weight-matrix

2. BioBasicsBench tests agent analysis against raw lab evidence
OpenAI and Ligo built 100 routine lab-analysis tasks using raw evidence such as gels, microscopy and instrument exports, scored against expert rubrics with partial credit. The reported evaluation uses five attempts per task and model setting; the authors document mistakes in controls and counting and caution that scores do not establish reliable laboratory judgment.
https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science

---

🛠️ Tools & Tips

(no items this day)

---

⭐ Open Source Highlights

(no items this day)

---

📰 Industry News

(no items this day)

---

🚀 Major Releases

(no items this day)