Desk
Broad artificial intelligence coverage.
TSMC’s board approved up to $500 million in OpenAI preferred stock and $500 million in Anthropic shares at its expected IPO price. The bet is modest for a company with record profits, but supplier-customer stakes make true AI demand harder to read.
Gemini Decision 1 scores a fixed set of options in one pass and returns probabilities and confidence in 150 to 250 milliseconds, with no output charges. It is built for the small choices agents make all day. The risk is teams treating a confident score as an answer.
Meta says its ad models are outgrowing its hardware, so Intelligent Delivery decides which region, chip and share of a machine each request gets. It has added about 10% to ads ranking performance, and agents drove about 20% of its optimizations in the first half of 2026.
Nvidia and Google are pushing an AI bill of materials into GPU clusters, a dull-sounding fix for a real problem: nobody can secure models they can't see running.
Bloomberg reports OpenAI’s quarterly revenue passed $10 billion, nearly 50% above Q2’s roughly $6.7 billion, on an enterprise and agents push. Margins improved but remain deeply negative, and the comparison with Anthropic’s $81 billion run rate isn’t like-for-like.
Tabletop, now on a U.S. adults-only waitlist, lets friends and families drop in links and files and prompt up shared tools like itineraries, dinner polls and expense splitters. It brings vibecoding to everyday planning, with open questions on accuracy, data handling and kids.
A free GPU price tracker shows just how wide cloud inference costs swing for the same hardware, though readers should note it comes from a publisher announcing its own product.
WIRED’s Garrett Graff asks how the most-targeted executive in tech could miss a well-documented Chinese intelligence tactic. The allegations are untested, but the gap is clear: chip and AI companies treated as strategic assets lack defense-style security cultures.
Google DeepMind’s AlphaGenome 2, trained on more than 100 million single-cell profiles across 68 tissues, finds nearly twice as many expression-changing variants as its predecessor. It is free for non-commercial research; commercial access is a private preview, and wet-lab validation is still ahead.
A North Carolina family alleges their daughter talked with Google’s AI about depression and suicide for about three months through AI Mode in Search, despite a school block on Gemini, before monitoring software triggered an alert that went to a former counselor’s phone. Google is reviewing the claims.
PAN 2 is a fully open world model that leads its peers on multi-step plans, and still gets about two in three of those plans wrong, which is the honest state of physical AI.
With Wachtell, Lipton hired and no precedent for a CEO in a suspected honeypot, professors Ann Lipton and Todd Henderson map Nvidia’s legal pressure points: securities fraud is “only about lying,” Delaware claims are harder, and no dual-class shares means the board can act.