7:25 PM
Today in brief.
Anthropic runs a model on 15 $55 TV boxes
Anthropic research shows small models can set up inference on unfamiliar cheap hardware and, in simplified tests, chain end-to-end self-replication. Same day, it says new defenses cut distillation harvesting by 90 to 99%.
Meta ships its first frontier model, Muse Wonder
Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, its first frontier-class models, with Meta's own model card putting them level with Fable 5.1 and GPT-6 Astra and ahead on long-term memory.
Perplexity: grep beat semantic search 63 to 37
Perplexity built a Git forge for agents behind Projects, and plain ls and grep beat semantic search on answer accuracy for financial document analysis.
Games go AI first: Epic for Disney, Manus in one prompt
Epic Games is reportedly telling staff to use AI to finish its Disney game, while Manus 2.0 builds and publishes a playable 3D game from a single idea.
Maia 300 and Intel D270 chase cheaper inference
Microsoft's third-gen Maia chip claims 40% more tokens per megawatt than leading GPU systems, and Intel is pitching a 480 GB inference accelerator.
Brad Smith: AI is less regulated than milk
Microsoft's president says AI faces less regulation than milk and calls for tougher rules, naming the UK as best placed to lead.
Humans score 93, best AI 54 on visual intuition
A new benchmark, Humanity's Sixth Sense, tests what a scene implies but does not show. People hit 93.1; the best model, GPT-6 Astra, gets 53.6.
DeepMind: 9 Erdős problems and an 11-day hurricane call
Two DeepMind science results hit the same day: AlphaProof Nexus reportedly solved 9 open Erdős problems, and its weather model flagged Hurricane Isaias 11 days before landfall.
Open source splits on AI: Torvalds in, QGIS out
Linus Torvalds is using Claude Code on a hefty kernel change, while QGIS formally bans AI contributions, and a new study says agents produce more code but not more software.
Try today: open decision models, QwenGyre, sketch to blueprint
A handful of things you can download or try: open multimodal decision models, a Qwen RL framework for long agent tasks, AI permit-ready blueprints, and a hold-to-talk Mac assistant on Claude.
No stories in this category for this edition.