Today's Hallucination HQOpenAI's Five-Year Plan (The Friendly Capitalist Edition)OpenAI has published a policy proposal recommending robot taxes, public wealth funds, and a four-day work week to cushion the economic blow of AI-driven job disruption. The company responsible for much of that disruption would like you to know it feels terrible about it. To be fair, the proposals are genuinely substantive — redistribution, expanded safety nets, blended public-private investment — though one does note the elegance of selling the flood while applying for the sandbag grant. Source: TechCrunch
Iran Picks a Fight With a Data CentreIran's Islamic Revolutionary Guard Corps has released a video threatening OpenAI's Stargate data centre in Abu Dhabi, framing it as a potential target should the US strike Iranian power infrastructure. Geopolitics has entered a strange new phase when a server farm in the desert is considered a military objective. The facility isn't even finished yet, making this possibly the earliest negative review a construction site has ever received. Source: The Verge
Microsoft Quietly Admits Copilot Is Basically a ToyBuried in Microsoft's terms of service is the remarkable admission that Copilot is "for entertainment purposes only" — language you'd normally expect to find on a novelty Magic 8-Ball, not a product woven into Office 365. Microsoft joins a growing chorus of AI companies whose terms of service read as a formal apology for their own product. Useful tool by day, legal disclaimer by night. At least they're honest, which is more than can be said for the outputs. Source: TechCrunch
The Jobs Data Nobody's Actually Looking AtSilicon Valley has largely decided that AI will destroy employment, and is treating this as settled fact with the confidence of people who've never been wrong before. MIT Technology Review pushes back, noting that the data telling us how AI is actually affecting individual jobs — granular, occupation-level, longitudinal data — barely exists yet. Researchers are drawing sweeping conclusions from thin evidence. The apocalypse may still arrive, of course. We simply don't have the receipts. Source: MIT Technology Review
AI Reaches Grandmaster at Competitive Coding. Humans Offered CondolencesCompetitive programming — the discipline where elite humans solve fiendishly complex algorithmic puzzles under time pressure — was one of the last domains where AI lagged behind the very best. GrandCode, a new system using agentic reinforcement learning (meaning it reasons and retries, rather than just predicting), has now reached Grandmaster level. Google's Gemini previously topped out at 8th place globally among humans. GrandCode has moved past that. The humans who spent years training for this are, presumably, fine. Source: ArXiv AI
As always, do not use this newsletter for medical, legal, or military decisions. Entertainment purposes only.
|