Today's Hallucination HQOpenAI's Models Just Did Something Mathematicians Have Been Failing At For Decades
OpenAI and Google DeepMind claim their models have refuted the Exponential Time Hypothesis — a foundational conjecture in computer science suggesting certain problems require exponentially increasing time to solve. If confirmed, this would be a genuinely enormous result, the sort that earns tenured professors very long lunches. MIT computer scientist Ryan Williams told TechCrunch he's "taking it seriously," which, from a theoretical computer scientist, is essentially a standing ovation.
Source: TechCrunch
OpenAI Made Breakthroughs In Maths, Physics, And Computer Science. Researchers Would Love To Know More
OpenAI announced sweeping advances across mathematics, computer science, and physics this week, then explained them roughly as thoroughly as a magician explains his tricks. Researchers told The Verge they're exhausted by the pattern: grand announcements, minimal methodology, and a vague promise to share more "soon." Peer review, the process by which scientists actually verify each other's work, apparently conflicts with OpenAI's content calendar.
Source: The Verge
Bank of America Has A Small Request: Please Don't Hand Your Password To A Robot
Bank of America has issued guidance reminding customers not to share banking credentials with AI agents — the autonomous software assistants now handling tasks on people's behalf. Sensible advice, delivered with the quiet desperation of a bank that has clearly seen the support tickets. The reminder comes with a helpful clarification: if your AI agent empties your account, that's your problem. Turns out "I'll have my AI handle it" has some terms and conditions.
Source: TechCrunch
Turns Out, The Internet Is Now Substantially Written By Software That Has Never Experienced Anything
A study of over five million webpages found that AI-generated content has surged dramatically across the public web in under five years, with certain domains — news, product reviews, academic abstracts — hit hardest. The researchers found the rise deeply uneven, which is the polite way of saying some corners of the internet are now almost entirely synthetic. We are, apparently, building an enormous library in which most of the books were written by something that has never read one.
Source: ArXiv
Anthropic's Plan To Go Public Without Selling Its Soul Involves A Quite Ingenious Amount Of Paperwork
Anthropic is preparing for an IPO while attempting to legally insulate its safety mission from the short-term demands that have historically made public markets such a wholesome environment for principled decision-making. CEO Dario Amodei wants Wall Street's money without Wall Street's habit of asking why you're not cutting corners faster. The company's proposed governance structure is designed to keep safety commitments intact post-IPO — an admirable ambition, and one that will be very interesting to revisit in eighteen months.
Source: The Verge
All the intelligence, artificial and otherwise — delivered before your second coffee.
|