AI News Roundup: A Genome Scan in 21 Hours, Math Under Fire, and a $3.1B Leaderboard

The last week of October landed a strange mix: AI doing real science, AI doing shaky math, and a benchmarking site becoming a unicorn. Here's the part that actually changes which tool you should reach for.

Anthropic pointed 950 Claude agents at vast genomic datasets and surfaced a previously unknown enzyme system, nicknamed ART, in about 21.5 hours - with CRISPR-like properties that still need lab testing and peer review. The interesting move isn't the speed, it's that a lab treated agents as a grounded search over real data, not a chatbot. If your work is research, that's the direction the tooling is heading.

OpenAI published 719 AI-generated math manuscripts, and researchers flagged gaps between the natural-language explanations and the formally verified proofs. It's the context-window problem in academic dress: a confident, well-written proof that doesn't hold. Keep the formal check outside the model, always.

On the practical side, Underdog released Saluki 27B, an open-source model compressed from roughly 54GB down to 7.89GB so it runs on a 16GB laptop, and it scored 88 of 120 on a tool-calling benchmark. Local, private, cheap inference stopped being a hobbyist dream. And the money kept moving: the Arena leaderboard raised a $200M Series B at a $3.1B valuation.

The throughline for buyers: Google introduced an enterprise Gemini agent with persistent memory and multi-agent orchestration, Mistral shipped 'Le Chonk,' a trillion-parameter open-weight model, and OpenAI priced GPT-6.1 Terra about 50 percent lower than its earlier Terra and Sol tiers. The expensive frontier is getting a cheaper, faster sibling, and the open-weight camp is closing the gap - so route the boring 80 percent to the cheap model and keep the premium one for the judgment calls.