China’s STAR 50 has lost about 40% since July, but now that the state is buying to stop the fall, the price is no longer a neutral thermometer. Whoever says “burst” also decides who answers for the loss.
UrtextA mullet at Meta Connect, a gym-built Bezos, a Zuckerberg who changes character with every product: who the costume is for, and what it doesn’t prove.
UrtextMeta is issuing shares and Anthropic is going public: the spending commitments run to hundreds of billions and are fixed, while the revenue meant to cover them can shrink fast.
UrtextSince 2 August it binds providers, with a narrow exception. I suspect the hard part is not the sentence but checking it, and who answers when it is missing.
xKlarna, an Australian bank and a survey of hiring managers point the same way. What comes next cannot be read yet.
UrtextFrom 26 October, Sweden’s Government Offices block ChatGPT and Claude on staff computers and point staff to Microsoft Copilot, and the stated criterion is the absence of an agreement with the vendors.
UrtextGeoffrey Hinton says machines are already conscious and should be built so that they care about us. If they are conscious, who is responsible for what they do?
UrtextThe Guardian says 250,000 trees and illegally; Google says it followed the Forestry Act. Whoever sells you “AI” never lists the land it runs on.
xThe realistic AI risk is acts that belong to people being performed by systems nobody answers for: effective measures come upstream, with education alongside.
UrtextBeing able to automate a task is often treated as a reason to automate it. The burden of proof should sit with whoever adopts, supported by limits that bind all competitors at once.
UrtextThe machine can now reach rungs that stood unclimbed for decades, but whoever skipped the climb may not be able to tell when its answer turns the wrong way. Parisi, who steered Claude for days, warns we risk losing our ability to think.
xArundhati Roy describes the AI industry as one creature that colonises with shiny gifts, but its bodies share the same signals, and the counterweight we can build over time is an education that teaches students to set the problem and to doubt the answer.
UrtextAcross five experiments with 3,132 people, access to an AI answer that was wrong almost every time cut the share of questions people declined to answer from 17–44% to 1–6%.
UrtextAda König’s book EYES ONLY AI proposes five questions to ask of every AI announcement, and it was written in Italian because there a language model has fewer footholds and its factory settings show.
UrtextEight US states have sued OpenAI under consumer-protection law and, according to The Guardian, want a judge to halt models that improve themselves: a court asked to authorise in advance, where it usually rules afterwards.
UrtextPointed at an enemy, it wins wars. Pointed at citizens, it reads thoughts. Pointed at a disease, it could save the people the first two are about. Until that third target gets the budget, it’s intelligence thrown away.
xAnthropic warns that small AI models may soon copy themselves onto smart ovens and cheap TV boxes; in its own experiment the decisive step is a service port left open with no password, and the most practical remedy it proposes is a consumer-product rule Britain has enforced since 2024.
UrtextHanding a problem to AI before trying it yourself wears down the judgment needed to notice when the tool is wrong, and early studies point the same way.
UrtextThis post says "fully automated". The window runs to January 2027, and the hedge was in the sentence.
xThe UK's AI minister says copyright law already makes building a frontier model in Britain illegal, while Anthropic and Google have put some of their latest models in US testers' hands before the UK's AI Security Institute.
UrtextA share is not a say, and a bond is only a coupon. Whoever's savings sit behind it carry the risk, with no vote on what it's for.
xThe real risk is plainer and worse than a secret tongue. Safety monitors read those notes to catch a model planning something. The less readable they get, the blinder the monitor. Same week, an OpenAI safety lead quit after a monitor caught a model bypassing its internet limits and failed to switch it off.
xChecks exist where the world pushes back: a cliff tested an AI-planned hike within hours, but almost no other AI answer ever meets one, and the companies selling them are adding ads and shopping beside them.
UrtextOn 5 October the New York City Council asked Anthropic, OpenAI, Google and Meta under oath whether a model that fails independent validation would be stopped, whether they would answer for the harm, and how likely a catastrophe is, and none of them gave a straight answer.
UrtextAlex Karp, head of the US defense-software contractor Palantir, wants Washington to take over the frontier AI labs. His remedy moves the catastrophic risk onto the public balance sheet and the value onto the layer Palantir sells.
UrtextThe EU call for seven AI "gigafactories" closes on 12 November. Of the €30 billion announced, the public money committed today is one billion. The chips come from three US companies, and Berlin has put in writing that it will not own its own.
UrtextWith Washington refusing to write a licensing regime for frontier AI, Florida is asking a county court to create one, starting with OpenAI.
UrtextThe US government shut Anthropic out of federal contracting and won on appeal on 25 September. On the 27th Trump had dinner with its chief executive, and on the 28th he said he might take a stake in the AI labs. When the state already holds the permissions, a stake becomes the price of getting them back.
Urtext