Today's Hallucination HQOpenAI's Jalapeño: Finally, a Chip Worth the Heat
OpenAI has quietly entered the hardware business with Jalapeño, its first custom AI chip — and the benchmarks are rather flattering. Tested on SemiAnalysis' InferenceX benchmark, it beat current state-of-the-art silicon on two fronts: more tokens delivered per user, and more throughput per kilowatt of power consumed. In other words, it's faster and cheaper to run. A company that spends money like a sailor on shore leave has apparently decided to build its own till.
Source: TechCrunch
AI Makes You Lazier at Thinking. Study Confirms What You Suspected.
MIT Media Lab researchers asked participants to evaluate news headlines and images over four weeks — some with AI assistance, some without. Early on, AI users were 21% more accurate. Sounds wonderful, until you learn their independent critical thinking quietly deteriorated over the same period. Using AI to stay informed, it turns out, is a bit like hiring someone to do your press-ups: the results look fine on paper, but you're getting weaker by the day.
Source: MIT Technology Review
Robotics Startup Adds $1 Billion to Its Valuation. Just for a Laugh, Really.
Generalist, the physical AI robotics startup, has raised a $200 million extension that pushes its valuation to $3 billion — up from $2 billion just a few months ago. Adding a billion dollars to your valuation in the time it takes most companies to approve a budget request is either a sign of extraordinary momentum or extraordinary enthusiasm. Possibly both. The robots aren't shipping in volume yet, but the numbers certainly are.
Source: TechCrunch
College Students Pick Gemini for Essays. Google Frames This as a Win.
A blind test of ChatGPT, Claude, and Gemini on college-essay writing found students preferred Gemini's output — a result that will delight Google and mildly irritate everyone else. The study, run by StudyArena, gave no indication whether the essays were actually good, only that students liked them most. Which is reassuring, in the same way that finding out your dog prefers one brand of biscuit to another tells you very little about nutrition.
Source: Hacker News
Anthropic Teaches Claude to Remember Things. Users Quietly Relieved.
Anthropic has added shared memory across Claude's chat and Cowork products, meaning users no longer need to re-explain their projects, preferences, and context every single session like some kind of Groundhog Day for professionals. The update sounds modest. It isn't. Briefing an AI from scratch each morning is the modern equivalent of introducing yourself to your own PA every Monday — technically functional, deeply dispiriting.
Source: TechCrunch
As always, the AI was confident, occasionally correct, and blissfully unaware of the difference.
|