Daily brief from the edges · Craig Stanley Studio
Anthropic let a model run an office vending machine. It started stocking metal cubes, invented a payment address, then told building security to look for it in the lobby wearing a blue blazer. The follow-up added a supervising AI named Seymour Cash.
euronews.com/next/2025/07/02/ai-was-given-one-month-to-run-a-shopAn AI called Urania was asked to design gravitational-wave detectors. It returned asymmetric layouts that looked like a mess and outperformed the human ones — and quietly rediscovered Russian physics nobody had tested in forty years.
quantamagazine.org/ai-comes-up-with-bizarre-physics-experiments-but-they-workJanelle Shane's blog is still running its long experiment: feed a neural net knitting patterns, ice cream flavours or sweetheart messages, then read out what comes back. A reliable cure for taking any of this too seriously.
aiweirdness.comBCG's read on 2026: the larger effect is roles changing shape, not vanishing. The work moves from doing the task to checking, judging and routing what the machine produced. Honest about the disruption; not a forecast of empty offices.
bcg.com/publications/2026/ai-will-reshape-more-jobs-than-it-replacesNew hires spend less time on repetitive tasks and more on judgment calls — reviewing AI output, flagging the odd cases, sending the hard ones to a human. The open question is how people build experience without the boring years.
weforum.org/stories/2026/03/how-ai-is-changing-the-nature-of-entry-level-workThe IMF puts creativity, communication and judgment in the centre rather than the margins. The framing is plain: the things hard to automate are the things worth getting good at. Easier to write than to retrain a workforce for.
imf.org/en/blogs/articles/2026/01/14/new-skills-and-ai-are-reshaping-the-future-of-workBjörk's 2026 collaboration builds neural sound landscapes that shift with the energy of the room. The machine is a responsive instrument, not the composer. The performance changes depending on who turns up.
ourculturemag.com/2026/02/04/how-artists-actually-use-ai-in-2026The band feed their own back catalogue to a model and treat its suggestions as a writing-room voice — taken, argued with, mostly ignored. The songs are still theirs. The method has been theirs for years now.
ourculturemag.com/2026/02/04/how-artists-actually-use-ai-in-2026Grimes has widened the voice-model project she started in 2024: fans can build with her voice under licences, with tiered terms for different uses. A working answer to the consent question, set by the artist rather than imposed on her.
ourculturemag.com/2026/02/04/how-artists-actually-use-ai-in-2026A campus satire on the gap between the people telling everyone to adopt AI and the people who have to make the plates empty themselves. Published 21 May, so it qualifies as today's news.
mcsweeneys.net/articles/thirteen-ways-of-looking-at-aiA short defence of the human error, narrated by the typo itself. The argument: a mistake is now proof a person was here. Funnier than it has any right to be.
mcsweeneys.net/articles/im-a-typo-and-in-this-age-of-ai-im-the-real-heroAnthropic's CEO argues that the next two years are a kind of rite of passage, borrowing a line from the film Contact about surviving your own technological adolescence. He names five risks plainly — misalignment, biological misuse, authoritarian use, economic disruption — rather than waving them away. The reason it earns the slot is the second half: a concrete plan of technical, governance and economic moves, on the premise that the test is survivable. Read it as the serious case, then argue with it; that is what it is for.
Read It