Today's Hallucination HQClaude Wanders Into a Lab and Discovers Something Embarrassingly UsefulAnthropic's wet lab — yes, an actual physical laboratory, with pipettes and everything — has produced its first result: Claude autonomously identified a new enzyme system comparable to the machinery behind CRISPR, the gene-editing tool that rewrote modern biology. Whether this holds up to scientific scrutiny remains to be seen, but as opening moves go, "we accidentally made a biology discovery" is considerably more compelling than a new chatbot interface. Source: The Verge
Familiarity Breeds Concern, ApparentlyA new report has punctured the optimistic theory that people simply need more time with AI to stop worrying about it. Daily AI users remain just as uneasy about the technology as everyone else — and equally supportive of regulating it. This is roughly the equivalent of discovering that living next to a motorway doesn't make you more relaxed about speeding. Support for oversight isn't ignorance. It may just be paying attention. Source: TechCrunch
Google's Next Big Model Is Nearly Here, Which Is Precisely What They Said Last QuarterKoray Kavukcuoglu, the new head of Google DeepMind, used his first public appearance in the role to confirm that Gemini 4 is currently in testing and nearly ready for release — a statement that will comfort the many observers who'd begun to wonder if Google had simply lost the thing. The company has watched rivals ship flagship models at pace while Gemini quietly made tea. Gemini 4 may finally answer whether the wait was strategic patience or enthusiastic procrastination. Source: The Verge
Spotify Will Now Let You Argue With Its Algorithm DirectlyTaste Profile, Spotify's new feature rolling out to US Premium subscribers, lets you see exactly how the platform has categorised your musical identity — then correct it in plain English. This is the first time most users will be able to explain, in their own words, that they only listened to that one song forty times during a difficult week and should not now receive it as a personality. A small but meaningful step in the ongoing negotiation between humans and their recommendation engines. Source: TechCrunch
80 Billion Parameters Walk Into a Room, Only 13 Billion Bother to WorkTencent has released Hunyuan-A13B, an open-source language model with 80 billion total parameters that activates only 13 billion at any given moment — a Mixture-of-Experts architecture, meaning the model deploys only the specialists it needs rather than everyone at once. It's an efficient approach, and honestly a management philosophy many large organisations would benefit from adopting. The model is open-source, which at least means you're free to inspect which 13 billion showed up today. Source: ArXiv AI
As always, the models are getting smarter. Whether the same can be said for everyone deploying them is left as an exercise for the reader.
|