Three threads today: who owns the upside, who is building bodies, and who is allowed through the door.
Anthropic puts 5% of itself in a fund
Announced 18 October, subject to stockholder approval, with the stake able to grow as the valuation does. No governance structure, spending mandate or dollar figure has been published. Tying a public-benefit commitment to equity rather than cash grants is the unusual part - it dilutes holders instead of drawing on budget, which is why it needs a vote.
OpenAI is building a humanoid robot
Sam Altman told Fortune he expects a demo in 2027, while saying robots will not be on the streets for a few years after that. His framing: the brain is harder than the body. OpenAI has 27 robotics roles open covering actuators, sensing, thermal and supply chain, and prior form in the Dactyl research. Competitors named are Tesla's Optimus and XPENG's IRON.
The web is closing to agents
Schwab CEO Rick Wurster said on 16 October that the firm blocks large AI agents in the tens and twenties weekly, naming Meta's Muse, and asked who is liable when a credentialled agent gets it wrong. Read-only MCP access is planned; money movement is not. Jamie Dimon warned of rogue agents on JPMorgan's 13 October call. Beyond banking, The Verge reported Amazon has blocked Muse since 20 September, Yelp allows only licensed agents, and Delta has no third-party agent deals. Four sectors, one month, same answer.
Power and the pipeline
The DOE has $4.42 billion in loans out to restart Palisades (800 MW, Holtec), the Crane Clean Energy Center (835 MW, Constellation - the former Three Mile Island Unit 1) and Duane Arnold (615 MW, NextEra), for 2.25 GW aimed largely at AI data centres. Read it against Bernstein telling the FT that only about 35% of the announced US data centre pipeline is credible, and RegLayer counting 47-plus large projects delayed, withdrawn or blocked in 2026.
Training data
The Korean Publishers Society's No Books, No AI appeal had 158 organisations and 321 individuals signed as of 16 October, demanding verifiable records of what was trained on and rejecting silence as consent. Benchmark in the background: the $1.5 billion Bartz v. Anthropic settlement, approved 20 July 2026, covering roughly 482,000 works at about $3,000 each.
Two smaller ones
Three model IDs - aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha - were reported in OpenAI's API on 18 October, with no documentation, no changelog entry and no company comment. The original sighting is unverified; treat it as a rumour with a plausible shape.
And a correction to a story that ran widely: British Airways did not ban Meta AI glasses. Its conditions of carriage never mention smart glasses. Section 7a10 covers filming or photographing crew without consent on any device, and the penalty runs to refused carriage on remaining sectors.
Also shipping
Claude Code 2.1.302 landed 16 October with 63 changes. Fork subagents are on by default in claude -p, the Agent SDK, desktop and VS Code, and the starting permission mode changed for sessions that do not specify one. Standalone Claude Design closes 14 December 2026.