three weeks after agreeing to buy Hugging Face, Nvidia sells the lock that would have kept OpenAI's agents out. A malicious reader might call it marketing. No evidence.
xFrance's most prestigious literary prize withdrew a novel on the strength of a statistical probability, in the same week an AI agent read 187,000 private messages belonging to a user who had told it no.
UrtextOpenAI and Anthropic are investigating a hundred times more incidents than anyone counted in public, and nobody has decided which signal should actually set something in motion.
UrtextOn the day a memo circulating in the White House cast Anthropic's chief executive as the face of a cult, Thiel called Pope Leo XIV an accomplice of China, for having written that no algorithm can make war acceptable.
UrtextWashington talks about an artificial intelligence race without ever saying where the finish line is. Beijing seems to be running a different one, and beneath the two stories the strategies look more alike than either government admits.
UrtextIn the first three days at the United Nations the only word everyone shared was "incident", and the United States, which rejects a common register, proposed a direct line for two with China.
UrtextFebruary 2023: the new Bing ships. January 2024: an internal deck warns of a "doom loop" for "the entire web". Today: "one employee's individual perspective".
xThe best detail in the paper is the accent. Same model, more authority and loyalty in Chinese, more fairness in English. Morals, it turns out, come with a language setting
xthe number computed outside the model, the human's doubt logged before the answer shows.
xStaff at Britain's public AI testing institute have been signed off with stress, the Financial Times reports, like many engineers who review machine-written code: responsibility stays human while pace and understanding drift away, and the countermeasure is jidoka, the principle behind Toyota's production for more than a century, which gives any worker the authority to stop the line.
UrtextWhen a lab warns you about its own product, the warning doubles as a spec sheet, and this one landed months before two IPOs.
xOpenAI and Anthropic loathe each other, and together with Elon Musk they have asked for AI to slow down: behind them sit largely the same investors, suppliers who invest in their own customers and two stock market listings to prepare, and each of those backers wants something the stake alone does not show.
UrtextA two-minute clip on X, and the three seconds everyone cuts out of it are the only honest ones in the whole thing.
UrtextZuckerberg, Musk and Huang called the US president separately and the proposal stopped, while the bill that names who could switch a model off sits in committee: a kill switch is a person.
UrtextYoshua Bengio wants nuclear-style treaties for AI, and nuclear safeguards work because fissile material can be counted: a treaty needs an inventory, and the only countable thing here is compute.
UrtextThe EU Kids Act bans infinite scroll, artificial notifications and engagement rewards for minors and leaves them in place for adults: an age limit is an admission of harm.
UrtextDisclosure written by the party being disclosed is an editorial policy - the framework says it will prioritise the cases that reveal new mechanisms, so someone inside OpenAI decides which of its own failures are worth your knowing.
xIt still funds distribution of the output, and the problem he named sits in the allocation. He gives the window as 12 to 18 months and the money as two years.
xAcross 6,510 generated quality assessments of 217 retracted or disputed papers, not one mentioned the retraction.
Urtext3 rivals who don't usually exchange holiday cards, suddenly in perfect agreement.
x