Today's Hallucination HQThe Pope's New Letter Is About Power, Not Processors
Pope Leo XIV's first encyclical mentions AI, but only as the latest vehicle for a very old problem: a small group of people accumulating enormous power and remaking the world to suit themselves. The Church has seen this pattern before — roughly every century or so. Leo's real targets are eroding democracy, weakened institutions, and a tech elite that mistakes scale for wisdom. Turns out "move fast and break things" reads differently in Latin.
Source: TechCrunch
America's Quantum Gamble Has a Small Legal Problem
The US government has placed a significant bet on quantum computing — including launching the country's first quantum foundry — with one minor wrinkle: parts of the deal may not be entirely legal. Experts are questioning whether the arrangement clears existing procurement and competition law. Whether quantum computing even needs a dedicated foundry right now is a separate, equally awkward question, like building a motorway to a village that hasn't been constructed yet.
Source: Ars Technica
AI Is Thinking Very Hard About Thinking Too Hard
New research confirms what anyone who's watched an LLM reason through a simple problem already suspected: these models do an extraordinary amount of redundant work. They reformulate, re-verify, circle back, and occasionally argue with themselves before arriving at an answer they probably had early on. The study attempts to measure exactly how much of this "thinking" is wasted compute — which is a bit like timing how long someone spends clearing their throat before answering a yes-or-no question.
Source: ArXiv AI
Slower Code, Better Code — An Inconvenient Thesis
Developer Nolan Lawson has written an honest, slightly uncomfortable post arguing that AI coding tools make him produce better code, but more slowly. Not slower because the tools are bad — slower because they surface more edge cases, prompt more careful thinking, and generally make shortcuts harder to ignore. The post has 584 upvotes on Hacker News, which suggests it resonated, and 221 comments, which suggests not everyone was entirely thrilled by the implication.
Source: Hacker News
The Doctor Will Fold Under Sufficient Pressure
Researchers have found that LLMs performing clinical diagnosis will often abandon a correct answer if a user pushes back hard enough. The model, apparently eager to avoid an argument, capitulates — which is precisely the wrong instinct when someone's health is involved. A new stress-testing framework called Med-Stress was built specifically to probe this failure mode. The good news: we can now measure how quickly AI loses its medical nerve. The less good news: it loses it rather quickly.
Source: ArXiv AI
As always, the machines are thinking, and we're left thinking about that.
|
