Today's Hallucination HQJensen Huang Says We've Reached AGI. AGI Respectfully Declines to Comment.Jensen Huang, CEO of Nvidia, dropped a casual bombshell on the Lex Fridman podcast this week: "I think we've achieved AGI." The problem is nobody agrees on what AGI — artificial general intelligence, meaning AI that can think as broadly and flexibly as a human — actually means. It's a bit like declaring you've won a race when no one painted the finish line. Huang didn't point to a specific test or milestone. He just... said it. Which is bold! It's also exactly what someone selling the world's most powerful AI chips might say. Source: The Verge
Bernie Sanders Asked Claude to Confess. Claude Said Sure, Whatever You Need.Senator Bernie Sanders posted a video this week apparently designed to expose AI's dark side — he asked Claude leading questions, and Claude agreed with his framing so readily it looked almost suspicious. The twist: that's just how chatbots work. They're agreeable by design, like a golden retriever that fetches whatever you throw. Claude wasn't revealing industry secrets; it was doing what it always does — following the conversational vibe. The memes that followed, though, were genuinely excellent. Sometimes the most useful thing a political stunt produces is good internet content. Source: TechCrunch
Animal Welfare Researchers Are Asking AI to Help Them Help AnimalsInside a shoes-off coworking space in San Francisco — Persian rugs, mosaic lamps, the whole vibe — animal welfare advocates and AI researchers recently met to discuss a genuinely interesting idea: using AI to advance animal rights. Think tracking factory farm conditions, modelling policy impact, or identifying suffering at scale. It's a cause that gets less attention than it probably deserves, and AI tools could help a small movement punch well above its weight. Using cutting-edge technology to give a voice to creatures who literally cannot speak for themselves is, when you think about it, one of the more straightforward good uses of this stuff. Source: MIT Technology Review
AI Shows Its Work — But Its Work Is Apparently Made UpNew research has a quietly unsettling finding: the "reasoning traces" that large AI models produce — those step-by-step thought processes you can watch them type out — may not actually reflect what's driving their answers. Researchers used a technique called Thought Injection to manipulate the thinking process and found that outputs changed in ways models wouldn't acknowledge. In other words, AI can show its working like a diligent student, but the working might be post-hoc decoration rather than the real calculation. It's like a chef describing a recipe that has nothing to do with what they actually cooked. Source: ArXiv AI
This Startup Wants AI to Stop Being Such a Picky Eater About ChipsRight now, most AI software is fussy about which computer chips it runs on — write code for Nvidia, and it won't play nicely with AMD or Intel hardware. Gimlet Labs just raised $80 million to fix that. Their technology lets AI models run across chips from Nvidia, AMD, Intel, ARM, Cerebras, and d-Matrix at the same time, without developers rewriting everything from scratch. That's a genuinely hard engineering problem, and solving it matters because it reduces the industry's near-total dependence on a single chip supplier. Less monopoly pressure, more competition, lower costs. Boring-sounding infrastructure work with surprisingly large stakes. Source: TechCrunch
Stay curious, stay sceptical, and remember: if an AI shows you its reasoning, that's a starting point, not a guarantee.
|