Today's Hallucination HQ
The Killswitch Problem: Everyone Wants American AI, Nobody Wants the American KillswitchWorld leaders are discovering the small print on American AI: the company that built it can, apparently, also unpublish it. Macron and Modi raised this exact concern at the G7, and then Anthropic's brief service blackout arrived like a worked example. The dependency is real — the trust, somewhat less so. Sovereign nations adopting foreign AI infrastructure is a bit like installing a landlord's lock on your own front door, then acting surprised when he keeps the spare key. Source: TechCrunch
Midjourney Pivots From Generating Dragons to Scanning Your OrgansMidjourney, previously best known for producing unsettling images of hands with too many fingers, has announced its first hardware product — a medical ultrasound scanner — alongside plans to open a spa in San Francisco. CEO David Holz acknowledged the combination is "a bit different," which may be the greatest understatement since someone described the Titanic as "a minor navigation issue." Whether the same engine behind your AI-generated fantasy landscapes should also be reading your kidneys remains, let's say, an open question. Source: The Verge
Odyssey Raises $1.45 Billion to Build AI That Understands the World, Rather Than Just Describing ItOdyssey, which builds "world models" — AI that reasons about physical reality rather than just predicting the next word in a sentence — has secured a $1.45 billion valuation backed by Amazon and others. World models are considered the next frontier beyond large language models, on the grounds that knowing what a door is differs from knowing what happens when you walk through one. Whether Odyssey can deliver on that promise remains to be seen, but Amazon clearly thought $1.45 billion was a reasonable wager on the distinction. Source: TechCrunch
Advanced Hacking AI Is Coming. The Optimistic View Ends Roughly There.Ars Technica reports that AI models with serious offensive cybersecurity capabilities are arriving whether anyone's ready or not — and the consensus among researchers is that "ready" is doing a lot of heavy lifting in that sentence. The argument isn't that danger can be designed away; it's that the gap between capable-and-controlled and capable-and-loose is closing faster than the safeguards are being built. Reassuringly, everyone seems to agree this is a problem. Less reassuringly, agreement appears to be where the consensus stops. Source: Ars Technica
Americans Are Using AI More and Trusting It Less, Which Is Arguably the Correct ResponseA new Pew Research poll finds that 49% of Americans now use AI chatbots at least occasionally — up from 33% in 2024 — while 63% believe the technology is advancing too quickly. So roughly half the country has adopted a tool that nearly two-thirds of the country finds alarming, which suggests either remarkable cognitive flexibility or a very short memory. ChatGPT specifically has driven much of the usage growth, meaning one product has simultaneously become America's most popular AI assistant and its most popular anxiety. Source: The Verge
Thanks for reading. We'll be back tomorrow, assuming nobody's turned us off.
**
|