
How Itaú is building generative AI for its entire customer base
Brazil’s Itaú is rolling ia.i toward its whole customer base by year-end — with IARA governance holding the regulated bank together.
Desk
Generative models for text, image, audio, and video.

Brazil’s Itaú is rolling ia.i toward its whole customer base by year-end — with IARA governance holding the regulated bank together.
MIT Technology Review sits down with Samuel King — the Stanford researcher whose generative model proposed blueprints for microscopic viruses.

YouTube Music adds Fully, Partly or No Gen AI labels to tracks — disclosure, not restriction, as rivals like Deezer and TIDAL cut AI songs from royalties.

Siemens EDA lifted its chip-characterization agent from about 55% to 95% on a verified benchmark by grading final artifacts, not individual steps — a lesson for every engineering agent.

Anthropic’s experiment shows small models can already set up inference on cheap IoT boxes. Full self-replicating AI worms are not here yet, but the path is no longer theoretical.

AI Studio projects can land in a private Play “Personal Library”—no public listing, optional friend sharing later, and a path to real developer publishing.

Meta’s GEM ads foundation model moves to one cross-surface brain—claiming 18.9% conversion and 12% CTR lifts while training at nearer-LLM scale.

Bloomberg via The Verge: Epic’s unannounced Disney shooter gets an “AI-first” mandate—less hand-written code, full AI art workflows, and a delayed March window.

Meta is putting liquid-cooled 102.4T switches, a multiplane fabric, and open Ethernet scale-up designs on the OCP table — the plumbing that has to keep pace with Muse-scale models.

Microsoft says its Maia 300 chip will become the main engine behind its own AI services from 2027, a serious bet that owning silicon beats renting it from Nvidia.
Microsoft is betting the next useful AI layer does not talk at all. Decision-1 picks from fixed options fast and cheap, but every number behind it comes from Microsoft's own scoreboard.
Qwen cut its open image model from 40 steps to 8 without shrinking it. Fast, capable and cheap, with one big string attached: the license.