Today's Hallucination HQAnthropic's Mythical Model Returns From Regulatory Purgatory
Two weeks of tense negotiations with the Trump administration, and Anthropic's Claude Opus 4 — known internally as Mythos 5 — has finally been cleared for limited release to a select group of organisations. "Limited" doing significant heavy lifting in that sentence. The government confirmed the arrangement via letter, which is a charmingly analogue way to regulate frontier AI. Progress, of sorts, though one suspects Anthropic's lawyers have earned considerably nicer holidays this year.
Source: The Verge
OpenAI Agrees to Government Restrictions, Then Immediately Complains About Them
OpenAI has quietly limited the rollout of GPT-5.6 following a government request — and then, with admirable speed, published a statement saying this sort of thing really shouldn't become normal. Their argument: keeping powerful tools from developers, businesses, and "cyber defenders" helps nobody. Which is a perfectly reasonable position to hold while simultaneously complying. OpenAI: principled in their objections, pragmatic in their behaviour. A very modern posture.
Source: TechCrunch
It's 40 Degrees in London and Your Brain Has Opinions About That
MIT Technology Review reports that scientists are actively investigating why extreme heat impairs cognition — memory, attention, decision-making, the works. Londoners, currently experiencing a genuinely dangerous heatwave, are presumably too warm to find this surprising. The research matters: as heatwaves intensify globally, understanding the neurological toll becomes rather urgent. Filing this one under "problems AI cannot solve by generating an image of a cool breeze."
Source: MIT Technology Review
Everyone Suddenly Wants to Make Their Own Chips (Nvidia Has Noticed)
OpenAI's custom inference chip, charmingly named Jalapeño, joins a growing roster of in-house silicon projects from Google, Apple, and SpaceX — all designed to reduce dependence on Nvidia, whose GPUs have essentially been the tollbooth on the road to AI. Building your own chip is expensive, slow, and genuinely difficult, which makes it notable that so many companies now consider it preferable to writing Nvidia another very large cheque. Jensen Huang remains, one imagines, unbothered.
Source: TechCrunch
AI Models Overthink Wrong Answers; Humans Just Give Up
A new paper from ArXiv presents a delightful finding: when large reasoning models get a problem wrong, they use more tokens than when they get it right. Humans do the opposite — we bail out early on things we can't solve. The models, bless them, keep grinding. The researchers call this "deliberation allocation." The rest of us might call it not knowing when to stop. Somewhere in that gap between human surrender and machine stubbornness probably lies the actual problem with AI reliability.
Source: ArXiv AI
We'll be back tomorrow, assuming the heat hasn't melted the servers.
|