9:56 PM
Today in brief.
The AI hardware layer went vertical in one day
AMD bought a startup that etches models directly into silicon, Anthropic said it's building its own chips, and Nvidia showed a compute tray that assembles in a minute - everyone is trying to own their own substrate.
Frontier reasoning is now free and unlimited in ChatGPT
OpenAI just made GPT-5.6 Luna unlimited for free and Go users, which resets what the floor of "access to good AI" costs everyone.
+48 sources 4:20 PMAgent loops are becoming a design discipline
a16z published the clearest framing yet on why coding loops worked first and what it takes to make an agent know when it's actually done.
DeepMind's cyclone model buys 24 hours of warning
WeatherNext hit state-of-the-art on storm track and intensity in Nature, and DeepMind is open-sourcing it - a rare case where the AI story is straightforwardly about saving lives.
OpenAI's first Jony Ive gadget is a $300 speaker
The long-teased OpenAI hardware turns out to be a hockey-puck smart speaker in the $300 to $400 range - a much more modest first move than the hype implied.
Seedance 2.5 landed in CapCut and FLUX 3 is everywhere
30-second scenes with 50 reference images are now a consumer editing feature, and FLUX 3's public opening two days ago is already producing fake archival footage at scale.
A $2M book deal died because nobody could prove a human wrote it
An author won a 14-way auction, then lost the deal because he couldn't prove the manuscript wasn't AI-assisted - provenance is now a commercial requirement, not a philosophical debate.
Humans rubber-stamped one in three malicious agent commands
Across 40,000 runs, people approving what their AI agent wanted to do missed a third of the actual threats - the human-in-the-loop is much thinner protection than anyone assumed.
AI designed viruses that don't exist in nature
Genome models just generated working bacteriophage designs from scratch, which is the moment generative AI stops being about text and pixels and starts being about biology.
Data centers are losing local votes
Nashville used eminent domain to kill a data center next to its zoo, and the backlash is now bipartisan - the compute buildout has a permitting problem, not just a power problem.
AI's employment number finally went negative
S&P Global's read on the last twelve months shows more firms reporting AI-driven job losses than gains - the first clean negative in the aggregate data.
OpenAI's math results are being called research misconduct
Experts told Scientific American that OpenAI's recent math breakthroughs cross into research misconduct - the credibility fight over AI science claims now has a formal accusation attached.
No stories in this category for this edition.