AI Daily Digest · 2026-10-28

🔬 New AI Craft

1. Test novelty judges across prompts and evaluation material
Ai2's benchmark exposes prompt-sensitive judgments of research ideas. Its mixed human/generated setting uses weak novelty labels. Teams selecting an evaluator can test multiple prompts on the material they expect to assess and examine how the labels were established before relying on a score.
https://huggingface.co/blog/allenai/novelty-bench

🛠️ Tools & Tips

1. Copilot SDK adds organization-owned service authentication
A public preview lets services use an organization-installed GitHub App rather than a signed-in user's credentials. The SDK accepts short-lived installation tokens after the app receives Copilot Requests permission. Requests are attributed to the organization and remain subject to its model and spending policies.
https://github.blog/changelog/2026-10-27-github-copilot-supports-server-to-server-authentication

2. Assisted permissions bring selective approval to Copilot tools
Copilot's CLI, desktop app and VS Code now offer a model-assisted permission mode. It automatically permits requests classified as clearly safe while retaining approval for others. Enterprise administrators can disable the mode separately; an unavailable or prohibited default falls back to manual approval.
https://github.blog/changelog/2026-10-26-github-copilot-assisted-permissions-generally-available

⭐ Open Source Highlights

1. LFM3 pairs downloadable models with specialization recipes
The LFM3 family spans text, vision, audio and combined modalities around a shared small-model backbone, with single-GPU customization examples. LFM3-1.7B uses the LFM Open License v1.0 with commercial-revenue restrictions. Its model card discourages agentic coding and complex multi-step reasoning, making task-specific validation important.
https://huggingface.co/blog/LiquidAI/lfm3

📰 Industry News

1. Anthropic reports early drug-development work with six partners, mostly preclinical

🚀 Major Releases

1. DeepMind announces Genie 4 with a research preview for trusted testers