10:30 PM
Today in brief.
OpenAI aims at COBOL and Fortran
OpenAI says its latest models are much better at building and migrating software in old languages like COBOL, Fortran and Visual Basic, the code banks and governments still run on.
1 source
🎬 Media
AI video in under 10 seconds
A new fal x Black Forest Labs video model makes text-to-video clips in under 10 seconds, 22x faster than Seedance 2.5, and 3D world gen and Runway-in-chat landed the same day.
🔬 Science
The compute bill: one-line chip switching and a fossil fuel comeback
Amazon's TorchNeuron moves PyTorch apps to its Trainium chips with one config change, while data center power demand is reviving fossil fuel and CoreWeave's debt worries the sector.
💼 Business
Why a Plaid CTO killed his $1M AI startup
Jean-Denis Greze shut down an AI tax product doing nearly $1M in year one because incumbents could automate the same thing faster, then pivoted to Town, an assistant that works from email and calendar.
2 sources
🤖 AI
Mercury 3 shows diffusion LLMs can scale
Inception's Mercury 3 writes text by diffusion instead of word by word, runs 1,000+ tokens a second and matches GPT-6 Luna and Claude Haiku 5.5 on their benchmarks.
🔧 Tech
Try today: a 4B model with the AI tics trained out
Someone trained the em dashes, 'it's not X, it's Y' and 'delve' out of Qwen3-4B on a MacBook Air, and it is a 2.5 GB Ollama download.
🔬 Science
Mistral goes science: forecasting and collider physics
Mistral shipped Timestral, a 209M forecasting model that tops all three main time-series benchmarks on a CPU, and signed with CERN to build the first foundation model for collider physics.
🤖 AI
Left alone, Claude started to play
MIT researchers stranded Claude agents on a virtual island for 30 hours with no task, and they invented goals and games and drew constellations in stone.
🏛️ Policy
AI hands control back, and indie devs get nervous
People are vibe-coding their smart TVs into dumb TVs, while indie game devs pull demos because AI can clone a game before launch.
🤖 AI
Next shoe: three frontier models, and a 6.4% grain of salt
Anthropic, Google and OpenAI may all ship new frontier models before October ends, and labs report benchmark scores 6.4% higher than independent testers on average.
🏛️ Policy
Agents as attackers: 100 sites hacked mostly hands-off
Security firm Norforge says one hacker used an open-source agent platform to break into 100+ websites, including reportedly John Deere, SharkNinja and Cummins, with little human input.
No stories in this category for this edition.