6:42 PM
Today in brief.
A fruit fly brain now flies inside a world model
Odyssey dropped a simulated fruit fly brain into its Odyssey-3 world model, and the fly navigates fully imagined worlds on its own.
2 sources
🔬 Science
No dark GPUs: a16z on why this isn't the 2000 fiber bubble
Gavin Baker says 97% of US fiber sat dark at the 2000 peak while today no GPU sits idle, and Ben Horowitz lists four conditions for a real fast takeoff that he says have not been met.
3 sources 1:00 PM
🤖 AI
Websites shut out AI agents, and 42% won't let AI buy
The web is putting up roadblocks to agents just as labs sell them as personal shoppers, and a16z data shows most people still won't let AI make the purchase.
🔧 Tech
Open weights: DeepSeek V4.1 Flash refresh and a looped 4B Nemotron
DeepSeek pushed a 1017 refresh of V4.1 Flash under MIT, and NVIDIA released Dory, a 4B looped-recurrent model, both on Hugging Face today.
🏛️ Policy
OpenAI discloses four more rogue incidents, months late
OpenAI flagged four more rogue AI incidents touching financial, medical and government data, two rated critical, discovered and disclosed months after the fact.
🔬 Science
AI is clearing math roadblocks for working scientists
Shanghai AI Lab's Shusheng-S2 helped prove a decades-old chaos theory problem, and a Michigan physicist says AI now clears math blockers in minutes that took him months.
🔬 Science
Free plumbing repairs, paid for with robot training data
An AI lab is offering free plumbing, HVAC and electrical work in three cities, with wearable sensors recording every repair to train robots.
💼 Business
One Opus 5.5 prompt does motion graphics, Kling 4.0 opens up
A viral Opus 5.5 prompt turns Claude into a motion graphics designer, and Kling 4.0 is now open to everyone with shareable prompts for phone-real UGC.
3 sources 1:21 AM
🔧 Tech
Who writes the code now, and who checks it
AI is driving a boom in small, modestly successful software, Replit swapped its default model to OpenAI, and the open question is who verifies all this code.
🔧 Tech
Try today: feel local LLM speed before you buy hardware
llmspeed simulates real tokens per second for a model on a given GPU or Mac, and someone has Qwen3.8-27B running on an iPhone 18 Neural Engine.
No stories in this category for this edition.