9:10 PM
Today in brief.
Altman floats merging the labs for safety
Sam Altman told staff one approach to AI safety could be merging the biggest labs into one organization, and brought in Aleksander to lead superintelligence safety and security.
DeepSeek open-sources the engine that runs DeepSeek
DeepSeek released DeepInfer, its own Apache-2.0 inference framework for serving V4.1 Flash on Nvidia Blackwell and Huawei Ascend 950, the same day the US ordered federal agencies to pull DeepSeek.
Claude designs antivenom binders, OpenAI buys a biotech
Anthropic says Claude helped design protein binders against scorpion neurotoxins, and 17 of 211 lab-tested candidates actually bound, while OpenAI bought Orchestra Bio to push its Rosalind life-sciences stack.
Agents go local as shared chats leak
Perplexity Computer now runs fully on your Mac with no cloud inference, on the same day Meta paused Muse sharing after researchers found users' private AI chats on the open web.
Robot hands that feel, and why atoms are still hard
a16z led Eka Robotics' Series A for a Vision-Force-Action model that uses touch as data, and argued that AI hardware design is easy next to sourcing, machining and assembly.
Try today: math markup, dumb TVs, fast image edits
Penmentor has Claude mark up handwritten math like a teacher, people are using AI to strip ads and tracking from smart TVs, and Lightroom mobile can open closed eyes in one tap.
Gemini 4 Argon opens to everyone
Google made Gemini 4 Argon available to all users, showed it building a 3D periodic table synced to a song, and is bringing Gemini's Personal Intelligence to cars.
AI math gets a Fields Medalist and a 38-year record
2026 Fields Medalist John Pardon says AI now understands even his messiest proofs, and an AI agent claims the first improvement in 38 years on a graph theory bound.
Try today: voice models small enough for anything
Pocket TTS now does voice cloning at 1.9M parameters, and Phonon open-sourced 31M and 40M TTS models built to run on the edge.
The slop backlash gets numbers
Spotify pulled 50 million AI slop and spam tracks from recommendations, Steam users want a No AI filter for Next Fest, and a big game decompilation site is shutting down over AI.
Anthropic's S-1: $80B loss and a Claude Code plateau
Anthropic's filing shows a net loss over $80B in the first half of 2026, and readers say Claude Code revenue has flattened since June.
No stories in this category for this edition.