Thursday, 12 November 2026
DeepMind's AlphaGenome 2 Reads DNA One Cell Type at a Time
DeepMind’s AlphaGenome 2 Reads DNA One Cell Type at a Time
The first AlphaGenome looked at a stretch of DNA and guessed its effects across dozens of tissues at once. AlphaGenome 2, released by Google DeepMind on November 11, asks a narrower question: take a million letters of DNA plus one condition — a tissue, a cell type, a disease state — and predict what happens there. It was trained on more than 100 million single-cell profiles from human and mouse samples across 68 tissues, and it can fill in what a lab did not measure from one measurement it did. DeepMind’s own numbers: nearly twice as many gene-expression-changing variants found as the previous best model, and on the TenK10K immune-cell dataset the true causal variant picked 30% of the time against 16% for specialised tools, with the direction of the effect right 43% of the time against 23%. As a demo it rediscovered a known variant that switches on the ETS2 gene in gut macrophages, a driver of inflammatory bowel disease. It is open to non-commercial research through the AlphaGenome API, in private preview for companies, and not validated for clinical use. Thirty percent sounds modest — until you remember the specialists manage sixteen.
Lutnick Was Warned About Jensen Huang’s China Tie Before Chip Approvals, Axios Reports
On Monday this site reported that Nvidia’s board had opened an investigation into whether Jensen Huang was targeted by foreign intelligence. The new part is who knew, and when. Axios reports that Commerce Secretary Howard Lutnick received a highly classified memo about Huang’s relationship with Alice Huang (no relation), whom U.S. intelligence officials suspect of working for China and whose ties to Megaspeed, a Singapore data-centre buyer of Nvidia chips, were under a smuggling investigation. The memo circulated before President Trump’s May trip to China — and before Commerce reportedly cleared about ten Chinese companies to buy H200 chips on May 14, though some licences came earlier. Trump says he may have been shown the warning but doesn’t remember it: “Certainly, it didn’t make much of an impact on me.” Lawmakers in both parties want closed-door testimony from Huang next week, and the board has hired former acting U.S. attorney Joon H. Kim to run its inquiry. Nothing in the reporting shows a licence granted because of, or despite, the memo. An export control is only as strong as the attention of whoever reads the warning.
Google’s Gemini Decision Answers Only Closed Questions — and Charges Nothing for Output
Google added a model that does not write. Gemini Decision (gemini-decision-1), launched November 11 with a new Decisions API, only answers questions with a fixed set of answers: yes or no, pick a category, or score something against a rubric. It reads text, images, audio, video and PDFs, and produces zero output tokens — so on the paid tier the whole bill is $0.10 per million input tokens, with free use on the free tier. Shared input is billed once per request however many questions you ask about it. For anyone running an agent that spends most of its day sorting, filtering and checking, that is a different price list from a chat model’s. The cheapest answer in AI is now the one with no words in it.