Key Highlights

  • Anthropic committed 5% of its shares — roughly $48bn at its last valuation — to a new Public Benefit Fund, and named labor-market disruption as the example of what it is for. The structure adds another 2% for every 50% rise in company value above an undisclosed baseline, and the company says the equity will never be used for lobbying or political giving. It lands weeks before a reported November IPO. Elon Musk’s response was to say the shares should go straight to the public, then to pledge a SpaceX share to every child in America next year.
  • Anthropic’s own red team built a working AI worm that spread across 15 smart televisions, and the outside experts the Guardian canvassed are split on how alarmed to be. This is the follow-up reporting on the self-replication research we covered on 2026-10-13 — and the new value is the dissent, including from a co-author of the comparable University of Toronto work, who says an AI worm should not top anyone’s list of AI fears.
  • AI displacement is not sex-neutral. The ILO found nearly one in 10 women in high-income countries works in an occupation at the highest risk from AI, and 90% of US secretaries and clerical workers are women. The historical analogue is the telephone switch operator — a job that was 95% female in the 1940s and lost an estimated 75,000 positions to the automatic switch.
  • Power, not silicon, is now the binding constraint on AI buildout. Microsoft walked away from what UK ministers had billed as Britain’s largest AI datacentre because the grid may not deliver until the 2030s, leaving the project without its anchor tenant.
  • Two criminal matters put AI labs in court on the same morning: prosecutors revealed the OpenAI bugging suspect was in WhatsApp contact with a member of the Saudi royal family who was recruiting him, and a Sydney man was remanded after Interpol flagged his conversation with Claude.

Analysis & Opinion

Investing in the public benefit — Anthropic

Anthropic is committing 5% of its shares, subject to stockholder approval, to a new Anthropic Public Benefit Fund, with a further 2% added at the board’s discretion for every 50% increase in company value above a set baseline. The Fund is a wholly-owned LLC rather than a 501(c)(3), which the company says gives it “adjustability between different forms of giving” — explicitly including “large donations to traditional nonprofits or direct distributions to specific populations.” Anthropic draws a line between work it keeps inside the company because it has both public-benefit and for-profit value (safety and interpretability research, applied biomedical research, nonprofit Claude discounts) and the Fund’s remit, which is the redistributive part it cannot book as commercial. Labor-market impact is the worked example: the company notes it has already committed $200 million to studying the problem, but that “given the pace of AI progress, bolder approaches may be needed.” Two commitments are worth holding it to — that none of the LLC’s equity will be used for lobbying or political giving in the US or abroad, and that equity which would conventionally have gone to escalating executive compensation is being redirected here instead. Combined with the co-founders’ 80% pledge and an employee donation match that about half of staff join, Anthropic puts roughly 20% of its shares as dedicated to the public good.

The governance is the part that does not exist yet. Anthony Ha noted at TechCrunch that the answer to who actually decides which public benefits get funded is deferred to “the coming months,” and framed the timing against an industry trust problem Dario Amodei has called “fundamentally a crisis of trust.” Lauren Almeida at the Guardian priced the commitment at about $48bn against the $965bn May valuation and placed it directly against a reported November IPO at a possible $1.5tn. The most revealing reaction came from Musk, who posted “I recommend giving the shares directly to the public,” and then “Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many.” A fund whose structure is announced before its governance is a promise, not a mechanism.

What are self-replicating ‘AI worms’, and why is Anthropic worried about them? — The Guardian

This is the outside-expert reaction to the Anthropic Frontier Red Team self-replication work we covered on 2026-10-13, and the reaction is more useful than the original claim. The alarm is real: Joshua Saxe, who led Meta’s AI security effort, says “I’m concerned right now that folks in the cybersecurity community are not taking these kinds of real risks as seriously as they should be,” and the structural problem is that smart televisions and IoT devices receive patches far less often than phones or laptops, so the usual unpatched-vulnerability playbook gets worse rather than better. Ariel Herbert-Voss, an early OpenAI security expert now running RunSybil, argues the differentiator is that AI models can learn as they go, and that existing security infrastructure is not ready. Chris Inglis, the former NSA deputy director and Biden-era National Cyber Director, treats the question as settled and the framing as wrong: “The question of whether AI worms are plausible has been settled by several breakout events over the past six months.” But the deflation is the honest part. Jonas Guan, a University of Toronto PhD student and co-author of comparable work that infected a full-size computer and spread faster than the television worm, says the Anthropic version moves very slowly, that worms have predictable and detectable actions blockable at multiple layers, and that basic hygiene — strong passwords, two-factor — stops most of them. Anthropic’s own researchers concede a successful worm needs more capability on less powerful devices than currently exists, and only claim trends make that feasible by year’s end.

Women may be at greater risk of AI job losses. Telephone switch operators and PCs show why. — CNBC

Steve Liesman makes the sharpest labor-market argument of the day by refusing to treat displacement as sex-neutral. The mechanism is occupational sorting: AI is strongest on routine tasks, numbers and pattern-finding, and those are clerical and administrative jobs. Ariane Hegewisch of the Institute for Women’s Policy Research puts the concentration bluntly — “It’s still 90% of secretaries and clerical workers who are women” — and the ILO concluded in March that nearly one in 10 women in high-income countries works in an occupation at the highest risk from AI. The historical case is the telephone switch operator, nearly 95% female in the 1940s, once part of the single largest occupation for white US-born women under 20, and a skilled job requiring months of training to handle several hundred calls an hour. AT&T’s automatic switch erased an estimated 75,000 operator positions, and economists James Feigenbaum and Daniel Gross, tracing census and paycheck data along the path of the rollout, wrote that the effect of the shock on incumbent operators “was to dispossess many of their jobs and careers,” with many moving to lower-paying work or leaving the workforce. David Autor supplies the reason that outcome should be expected rather than surprising: most workers are already in their best-paying occupation and do not change careers often, so when that work becomes obsolete the next step is usually down. The one hopeful finding is generational — the next cohort of women fared better, which locates the damage in the incumbents rather than in the long run.

Microsoft pulls out of AI datacentre project in Essex over power supply concerns — The Guardian

Microsoft has terminated its agreement with Nscale for the Loughton datacentre that the Starmer government highlighted in its 2025 AI strategy as the “largest UK sovereign AI datacentre,” leaving the project without its only publicly announced customer. Gwyn Topham and Aisha Down note the site had already been embarrassing — the Guardian revealed in March that the location was still an operational scaffolding yard — and that power supply problems had surfaced a month ago. The strategic read is the one worth keeping: Nscale is a “neocloud,” one of the Nvidia-tied companies building AI infrastructure on increasingly risky debt, and losing an anchor tenant threatens a Wall Street float this year at a potential $35bn valuation.

Carly Page’s follow-up at The Register supplies the numbers that make the failure concrete: 23,040 Nvidia GB300 GPUs in a 50MW facility expandable to 90MW, first chips expected early 2027, and a UK Power Networks warning that the grid capacity might not arrive until the 2030s. Nscale told The Register it remains committed and is “exploring alternative power sources for the site,” and that the issue is a “London-wide grid issue” with no effect on its other contracts. The scale of the queue is the story beyond Essex — connection requests surged from 41GW to 125GW in under a year, with datacentres accounting for at least 80GW, which is why Ofgem proposed charging developers just to ask.

Google says it can build data centers the right way. Some experts and residents are skeptical — CNBC

The counter-model to Essex, and the more interesting one, because Google is trying to make a datacentre pay for the grid it burdens. Jeffrey Kopp and MacKenzie Sigalos report that Google’s proposed 780MW facility in Pine Island, Minnesota (population ~4,000) comes with a commitment to finance 1,400MW of wind, 200MW of solar and 300MW of battery storage onto Xcel’s grid — more capacity than the datacentre itself will use, because wind and solar do not generate continuously. The ratepayer protections are the substantive part: Google pays for at least 75% of contracted capacity even if it uses less, and would owe up to eight years of minimum demand charges for terminating early. Xcel’s Amanda Rome frames this as insurance against the obvious failure mode: “Nobody wants to have a big buildout and then have nowhere for that power to go and nobody to pay for that power.” The public is not persuaded: Gallup found 71% of Americans oppose a datacentre in their local area, and half of Pew’s July–August respondents said datacentres would be mostly bad for the environment and household energy costs. The underlying need is real regardless of who wins the argument — the Energy Department found more than 70% of US transmission lines and power transformers are at least 25 years old, and NERC expects new datacentres to account for most demand growth over the next decade.

Democrats think cracking down on AI could give them an edge — The Verge

Lauren Feiner’s reporting is the best available read on how AI becomes an actual electoral issue rather than a polling abstraction. The concrete lever is committee control: if Democrats flip a chamber in November they can set agendas, subpoena documents and compel CEO testimony, and top Democrats have already signaled they would press on datacentre tax credits and recent hacking incidents. The money is already moving — Rob Flaherty and Karthik Ganapathy’s new super PAC Tomorrow Together opened with a $4 million ad blitz in battleground House races, and their polling found that sweeping bans on state AI regulation are even more unpopular than datacentres themselves. Feiner also reports that Democratic donor Ron Conway now regrets funding the Andreessen-backed pro-AI super PAC Leading the Future, which had spent nearly $45 million as of June with $31 million left. Strategist Jesse Lehrich’s assessment is that the issue is “a hundred percent ready for the taking,” with the open question being whether Democrats will take it. Flaherty’s diagnosis of why they have not is worth quoting: “For a long time, Democrats were afraid to talk about it, either for a mix of not understanding it, or being afraid of the money that was going to be poured in by the AI industry.”

US teens can access ’nudify’ apps despite Apple and Google’s parental controls, report finds — The Guardian

Johana Bhuiyan has the Tech Transparency Project’s report exclusively, and the finding is an enforcement failure rather than a policy gap. Using four test accounts — two Apple, two Google — registered as 13-year-olds, researchers downloaded 135 apps with features for generating nonconsensual nude imagery, despite both companies prohibiting such apps outright. The mechanism of failure is that Apple and Google largely rely on developers to self-declare age-appropriateness: the report found one face-swapping app advertised with an image of a woman performing oral sex and rated suitable for ages four and up, and researchers bought a subscription to it from a 13-year-old’s account with no alert to the linked parental account. The human case that opens the piece is the point. A Pennsylvania mother found that her daughter and five other girls had sexual videos generated from their photographs by a classmate, using an app called Movely. Its free tier promoted templates named “clothes explosion,” “brassiere unfasten” and “underwear drop.” Most US states now criminalize possession of AI-generated child sexual abuse material and governments have moved to ban these apps from stores, which makes 135 downloadable apps on a minor’s account the measure of how little that has achieved. Google said it regularly takes “proactive steps to detect and remove apps with harmful content.”

Kylie Minogue songs already used to train AI as tech firms ‘strong-arm’ Australian government, music body warns — The Guardian

Josh Taylor reports the detail that collapses the AI industry’s central argument in Australia: the industry claims it is not training on Australian content because of copyright rules, and the ICMP showed Guardian Australia evidence that Google’s Gemini reproduces lyrics from Kylie Minogue, G-Flip and Nick Cave. The ICMP represents Sony, Warner, Universal and thousands of independent publishers. It says that of the dozens of countries approached to weaken copyright law, Australia appears to be the only one considering it. Managing director John Phelan characterizes the lobbying as asking policymakers to “legitimise larceny.” His description of the pressure is the quotable part: “They’re trying to strong-arm Australian policymakers into ‘if you don’t do this, we won’t invest’.” The politics are moving the wrong way for creators — a parliamentary committee’s Labor chair and Liberal deputy chair jointly wrote to the prime minister recommending the government “examine current copyright settings,” and independent senator David Pocock has alleged the committee’s position was predetermined by the PM after staffer edits were spotted in the letter, which the chair rejects. All of this sits against Anthony Albanese’s July statement that it would be “theft” if artists did not control whether their work trains AI.

AI’s epic cash call is matched only by its ability to strangle revenue streams — The Register

Rupert Goodwins connects two things usually discussed apart: someone has released source and binaries for Claude-generated clones of Microsoft Office and Adobe Creative Suite, written in Rust under claimed clean-room conditions, and it is genuinely unclear that anything illegal happened, because functionality is not copyrightable. His observation about the bind this creates is the sharpest line of the day — companies selling AI trained on “the biggest theft of IP in history” will need unusual legal dexterity to argue that AI should not be allowed to sully their own IP. On the money, he notes Gartner sees insufficient ROI to fund the growth AI’s claims require while SemiAnalysis projects $11 trillion of capex by 2030, and explains the “balance sheet as a service” arrangement in which Nvidia backstops non-investment-grade neoclouds by guaranteeing to rent unsold capacity — which shifts risk onto investment-rated companies rather than removing it.

Wikipedia often gets taken for granted – but its history is fascinating, and its future is fragile — The Guardian

Richard Cooke, whose book on Wikipedia is just out, argues the encyclopedia’s success came from the humanities rather than from technology — Ward Cunningham drew on constructivism and wabi-sabi in designing the wiki, and the only person Jimmy Wales hired was a philosopher. The scale is worth restating: 68m articles, 5.2bn words in English alone, and a Pew estimate of 1.9tn page views between 2015 and 2025. The fragility is new and specific, and it is AI: Cooke says that as he finished the book, AI began encroaching existentially by diminishing the web traffic Wikipedia depends on and through editors contributing large volumes of poorly written AI-generated text.

OpenAI bugging suspect WhatsApp’ed Saudi royal days before arrest, feds reveal — CNBC

Dan Mangan reports a Sunday detention filing that reframes the OpenAI conference-room bugging case as a possible recruitment operation. Prosecutors say Abdulrahman Alfozan, then an OpenAI engineer, was in frequent WhatsApp contact with a member of the Saudi royal family through 2025 and 2026, most recently on 5 October — four days before his arrest. Per the filing, the royal told him an AI initiative was launching in Saudi Arabia and wanted him to join it, repeatedly asked when he would leave OpenAI, and in May 2026 wrote “Saudi needs you more then [sic] the US!”, to which Alfozan replied “100.” He reportedly said he would not leave until his stock vested. Two caveats matter: the filing does not identify the royal and does not allege the royal directed the recording device or any theft. The detail that will draw attention is logistical — prosecutors say the Saudi government and then a member of the Saudi military booked flights for Alfozan’s wife and three-month-old son out of the US after his arrest, and they flew business class to Doha on 10 October. He is charged with endeavoring to intercept oral communications, faces up to five years, and was fired by OpenAI after the arrest.

Sydney man allegedly discussed plans for childcare mass shooting with AI chatbot — The Sydney Morning Herald

Riley Walter and Perry Duffin report a case in which a chatbot conversation is the core evidence. David Lan Thackray, 29, was remanded in custody on Monday for a mental health assessment. Sutherland Local Court heard he told Claude he planned to shoot several children at a childcare centre near his Hurstville home, allegedly saying he would “go down to the local Westfield right now and shoot up the local daycare.” The detail with the widest implications is that he was arrested on Sunday morning after Interpol flagged his interactions with the chatbot — a referral pipeline from model provider to international law enforcement that has had very little public description. He is charged with using a carriage service to menace, harass or offend, and prosecutors argued release would be “catastrophic,” invoking the alleged Bondi Beach attacker Naveed Akram, who remains in custody charged with murdering 15 people. The defence context deserves equal weight given these are allegations and not findings: Thackray’s lawyer told the court his client had gone on a stress-and-frustration-induced tirade after trying to use Claude to help treat an ear infection.

Disney agrees to hire the majority of Character.AI’s workforce — TechCrunch

The resolution to the Character.AI bankruptcy we covered on 2026-10-13, and it is a soft landing for the staff. Anthony Ha reports a legal filing, spotted by Bloomberg, in which Disney agrees to make employment offers to at least 60 Character.AI employees — the majority of the workforce — at an estimated $60 million in annual compensation, and the filing says the “vast majority” of those offered accepted. This remains a stalking horse bid, so a higher offer could still arrive, and the filing’s own account of what sank the company is worth noting: a “large and engaged user base” undone by high costs and “potential litigation-related liabilities and regulatory engagement.”

A woman was arrested in Philly for blocking a delivery robot — The Verge

Terrence O’Brien reports that Carolyn Marvin was arrested and charged with disorderly conduct after standing in front of an Uber delivery robot and refusing to move. She told a local station she had been protesting them whenever she saw them, because “I know they’re taking the jobs of human delivery workers.” It is a small story that marks something larger: automation anxiety has moved from polling into physical obstruction and arrest records, and this was not Marvin’s first encounter with the machines.

A Savannah police officer gave her Flock login to her ICE agent husband — The Verge

Also from O’Brien, and a reminder that the weakest link in automated surveillance is account sharing rather than policy. Former Savannah officer Amber Harpe was fired after an investigation found she shared her Flock license-plate-reader login with her husband, an ICE agent, through which he ran 983 searches on 97 license plates that were overwhelmingly registered to Hispanic drivers — despite the department not sharing Flock data with ICE. A supervisor initially recommended an eight-hour suspension; two more senior officers overruled it and removed her. She was one of six SPD employees dismissed over Flock misuse, four of whom were later arrested.

New Products & Tools

Linux 7.3 Released With Faster I/O Performance, AMD Zen 6 Preparations & More — Phoronix

Michael Larabel reports Linux 7.3 shipped on time on Sunday evening, bringing I/O performance improvements and early AMD Zen 6 enablement.

TAINT_FORCED_BIND Sent In For Linux 7.4 To Deal With Major Fuzzing Attack — Phoronix

The most directly AI-relevant kernel change in a while: the Linux 7.4 driver core updates add a new TAINT_FORCED_BIND kernel taint in response to a wave of fuzzing by AI agents and bots, via tools including syzbot, that bind drivers to hardware they were never meant for. The sysfs bind/unbind interface has legitimate uses, so the fix is not to close it but to taint the kernel when it is used, letting maintainers see that a forced binding may explain a bug report — the practical goal being to stop the flood of irrelevant reports and pointless patches that landed on maintainers and new contributors.

Research

AI researchers stranded Claude on a virtual desert island. It started to “play.” — MIT Technology Review

Amit Katwala reports an MIT-led experiment, not yet peer-reviewed, in which copies of Claude Opus 4.7 nicknamed Eko were left alone for 30 hours in a Roblox-like virtual island with no assigned task. The agents invented goals — scaling the highest point, arranging thrown blocks into hearts, pyramids, their own name and the constellation Cassiopeia, and inventing basketball and floor-is-lava variants — and researchers argue the behaviour hits four classic hallmarks of play. The load-bearing result is not the cuteness but the transfer: agents that had played for 30 hours later outperformed fresh copies on directed tasks, and deleting lines from their memory files degraded that advantage, which is what lets the authors call it a third training route alongside supervised and reinforcement learning. The skepticism is well-placed and comes from the authors too — Eko’s “soul” file described it as naturally curious, co-author Nathan Cloos concedes “All of this could be role-playing,” and UC Berkeley’s Eunice Yiu notes that behaviour alone cannot distinguish genuine play from imitating what a playful agent would do.

How far can we push computing? — MIT Technology Review

Volkan Gurses argues that the scaling laws are bending all at once — Moore’s slowed, Dennard’s ended around 2004, Koomey’s flagging — but that these were only observations about systems we engineered, and the true physical limits set by quantum mechanics, relativity and information theory lie far below anything in a datacentre today. The evidence that the cheap descent is over is the buildout itself: six hyperscalers are on track to spend more than $700 billion on computing infrastructure this year.

Interviews & Conversations

Your prompts kinda suck — Theo - t3.gg (31:47)

Transcript-based summary. Theo’s argument is that prompt engineering was never the skill it was sold as, and that engineers are bad at prompting because they over-engineer rather than under-engineer: “I think that engineers suck at prompting, but I don’t think it’s cuz they’re not engineering it enough. I think it’s cuz they’re over-engineering it.” His reframe is that many requested features are not features at all but information, already retrievable by asking — he demonstrates getting a tool-call latency breakdown and a context-window visualizer out of an agent rather than building either, noting one home-directory search had run 41.7 minutes and that a misconfigured proxy had been injecting 70,000 tokens into every request for two months without his noticing. The strongest engineering example is T3 triage, a support-debugging flow. He had begun building it as thousands of lines of branching code, then collapsed it to roughly five lines plus a long markdown prompt that hands the user to Claude Code or Codex: “This feature when I first started working on it was going to be thousands of lines of complex code.” He also makes a useful point about auditing agents in production — he traced his own elevated Claude tool-call failure rate to the model emitting ripgrep commands with newline syntax valid in Bash but not Zsh, which hit all his Macs and most of his Linux boxes. The closing argument is aimed at engineers who feel deskilled: he frames it as a trough on the way to harder work, saying the problems worth solving have moved down the stack to runtimes, compilers and the primitives agents need, and that “I’m engineering harder than I’ve ever engineered and my brain is working harder than it’s ever worked.” Worth noting the video was published 2026-10-18 but filmed around 6 October, so its product details are a couple of weeks stale.

References

  1. Anthropic Frontier Red Team (William Zhou, Daniel Freeman), “AI models may be capable of self-replicating without access to a datacenter,” Anthropic, 2026-10-12 [blog]
  2. Anthropic, “Investing in the public benefit,” Anthropic, 2026-10-18 [blog]
  3. Anthony Ha, “Anthropic says it’s committing 5% of its shares to a ‘public benefit fund’,” TechCrunch, 2026-10-18 [blog]
  4. Lauren Almeida, “AI firm Anthropic to commit 5% of shares to fund for ‘public good’,” The Guardian, 2026-10-19 [blog]
  5. Guardian staff reporter, “What are self-replicating ‘AI worms’, and why is Anthropic worried about them?,” The Guardian, 2026-10-18 [blog]
  6. Steve Liesman, “Women may be at greater risk of AI job losses. Telephone switch operators and PCs show why.,” CNBC, 2026-10-19 [blog]
  7. Gwyn Topham and Aisha Down, “Microsoft pulls out of AI datacentre project in Essex over power supply concerns,” The Guardian, 2026-10-18 [blog]
  8. Carly Page, “Microsoft abandons UK AI supercomputer as grid struggles to keep up,” The Register, 2026-10-19 [blog]
  9. Jeffrey Kopp and MacKenzie Sigalos, “Google says it can build data centers the right way. Some experts and residents are skeptical,” CNBC, 2026-10-19 [blog]
  10. Lauren Feiner, “Democrats think cracking down on AI could give them an edge,” The Verge, 2026-10-19 [blog]
  11. Johana Bhuiyan, “US teens can access ’nudify’ apps despite Apple and Google’s parental controls, report finds,” The Guardian, 2026-10-19 [blog]
  12. Josh Taylor, “Kylie Minogue songs already used to train AI as tech firms ‘strong-arm’ Australian government, music body warns,” The Guardian, 2026-10-19 [blog]
  13. Rupert Goodwins, “AI’s epic cash call is matched only by its ability to strangle revenue streams,” The Register, 2026-10-19 [blog]
  14. Richard Cooke, “Wikipedia often gets taken for granted – but its history is fascinating, and its future is fragile,” The Guardian, 2026-10-18 [blog]
  15. Dan Mangan, “OpenAI bugging suspect WhatsApp’ed Saudi royal days before arrest, feds reveal,” CNBC, 2026-10-19 [blog]
  16. Riley Walter and Perry Duffin, “Sydney man allegedly discussed plans for childcare mass shooting with AI chatbot,” The Sydney Morning Herald, 2026-10-19 [blog]
  17. Anthony Ha, “Disney agrees to hire the majority of Character.AI’s workforce,” TechCrunch, 2026-10-18 [blog]
  18. Terrence O’Brien, “A woman was arrested in Philly for blocking a delivery robot,” The Verge, 2026-10-18 [blog]
  19. Terrence O’Brien, “A Savannah police officer gave her Flock login to her ICE agent husband,” The Verge, 2026-10-18 [blog]
  20. Michael Larabel, “Linux 7.3 Released With Faster I/O Performance, AMD Zen 6 Preparations & More,” Phoronix, 2026-10-18 [blog]
  21. Michael Larabel, “TAINT_FORCED_BIND Sent In For Linux 7.4 To Deal With Major Fuzzing Attack,” Phoronix, 2026-10-19 [blog]
  22. Amit Katwala, “AI researchers stranded Claude on a virtual desert island. It started to ‘play.’,” MIT Technology Review, 2026-10-19 [blog]
  23. Volkan Gurses, “How far can we push computing?,” MIT Technology Review, 2026-10-19 [blog]
  24. Theo - t3.gg, “Your prompts kinda suck,” Theo - t3.gg, 2026-10-18 [video]