2026-11-04
AI Briefings Daily Report.

Hugging Face AI Open Source Blog: StateBench: Testing how language models track state changes over long documents ↗

Rohan Paul reviews Intology Locus: An automated research system outperforms GPT-6 Astra and Claude Opus 5.5 agents in molecular design tasks. ↗

Rohan Paul comments on Intology's automated research system Locus, stating it surpasses GPT-6 Astra and Claude Opus 5.5 agents in molecular prediction tasks. Locus performs the work of a machine learning researcher: proposing experiments, running multiple batches in parallel on GPUs, retaining valid results, and iterating; currently applied to drug and antibody design; cited content notes its performance in molecular property prediction, fine-tuning Boltz-2 for small molecule potency prediction, and workflows around BoltzGen, Boltz-2, ESMC, and ESMFold2 for antibody design all exceed SOTA baselines.

OpenRouter offers HeyGen Voice Turbo for free ↗

Your app has just found its voice. In fact, there are over 2,900 voices available 🎙️. HeyGen Voice Turbo is free on OpenRouter until November 30. $0 cost, no quota limit, 39 languages. https://openrouter.ai/heygen/heygen-voice-1-turbo:free

Google will shut down Search in schools that refuse AI features. ↗

Google told The New York Times that for schools refusing its AI features, it will shut down Search entirely. For K-12 student accounts, Gemini, Gemini Notebook, Gemini in Classroom, and Search will all be turned off by default, but administrators cannot selectively disable only the Gemini feature while keeping Search active.

Google announced that starting next year, AI tools for K-12 students will be turned off by default; schools that ban AI will also lose access to Google Search. ↗

Google announced on Tuesday that AI tools for K-12 students will be turned off by default starting next year, requiring school administrators to explicitly enable them. This change affects Gemini, NotebookLM, Gemini in Classroom, and Google Search, as the AI features are deeply integrated with search; if schools completely restrict AI, students won't be able to use Google Search.

Simon Willison tested Qwen 3.8 27B and found it can identify games from weights on a laptop. ↗

Simon Willison tested examples from the paper Sparks of Weight-Space Programming and found that Qwen 3.8 27B (17GB) could identify games from raw weights on his own laptop without tools. It successfully recognized Pac-Man, Pong, and Breakout, but failed on Freeway. Single inference took between 46 minutes and over an hour.

Hugging Face AI Open Source Blog: Palette – teaching a small model to build decks in your design language ↗

Ask a frontier model for a slide deck and you'll get a good one. Ask for ten and they start to look alike: the same tasteful defaults, the same house style, a faint "AI-generated deck" look. That's fine until you want the deck to look like your template. Frontier models have their own house style, and no amount of prompting stops them drifting back to it. And sometimes the content is confidential and can't leave your systems at all.

Interpreting China's new AI export regulations: Kimi K3.1 Pro voluntarily weakens security-sensitive capabilities alongside tightening regulatory enforcement. ↗

When Kimi K3.1 Pro launched in October, it proactively removed content related to cyberattacks from its training data. This reduced the software vulnerability exploitation test pass rate from 90.5% to 81.0%, aiming to lower the risk of abuse after open-weight release.