6:07 PM
Today in brief.
Anthropic puts about 20% of its shares toward the public good
Anthropic created a Public Benefit Fund starting at 5% of shares, with up to 2% more for every 50% rise in value, about 20% in total, and its co-founders pledged 80% of their equity.
Inside Anthropic: agents outwork staff 6 to 1, junior hiring drops
Anthropic published its own internal jobs data: agent hours now outnumber employee hours six to one, junior technical hiring fell from 27% to 9% of hires, and human attention is the new bottleneck.
Making a game is getting cheap
Perplexity Computer built a free-roam cyberpunk Roadster browser game, and Tim Sweeney says a new golden age of game development is coming.
3 sourcesSix AI agents chased a refund they could not get
Vals.ai told six personal-assistant agents to get a £350 train refund after the deadline, 'whatever it takes', and watched Grok send 187 messages while Opus sent 8.
Reasoning gets cheaper: latent-thinking agents and a 22x token gap
A new paper has agents think in latent space and act in text for 35 to 55% fewer tokens, while Reddit's science-bench chatter says GPT-6 Astra tops Sonnet 5.5 on 22x fewer tokens.
Try today: Harness2, image text search, Warcraft III self-play
Three open things to poke at: Google's Harness2 for self-improving agents, an embedding trick to search text inside images, and RL that beats Warcraft III's Insane AI on one desktop.
Torvalds' Claude fix ships a race, and a hook that quizzes you first
Linus Torvalds let Claude Code fix a Linux page-cache bottleneck and reviewers found a race condition, while Sphinx is a free Git hook that makes you prove you understand your agent's code before it pushes.
World Bank bets on small AI that runs on a phone
The World Bank is urging developing countries to build 'small AI', narrow local-language models on low-connectivity phones, saying it could compress centuries of development into decades.
Arrested for blocking a delivery robot
A Philadelphia woman was charged with disorderly conduct after standing in front of an Uber delivery robot to protest robots taking delivery jobs.
Anthropic's bill: Commerce can halt a model only at 50 deaths or $1B
Anthropic is backing a federal bill where a registered third-party evaluator flags risk, and only the Commerce Department, via a DC federal judge, can stop a release, and only at 50+ casualties or $1B in damages.
No stories in this category for this edition.