The Hallucination HQ
AI news that's actually fun to read — 2026-05-21
|
Today's Hallucination HQOpenAI Solves an 80-Year-Old Maths Problem, and the Referee Actually Agrees
OpenAI's latest reasoning model has disproved a geometry conjecture that stumped mathematicians since 1946 — and crucially, the same researchers who caught OpenAI fabricating a maths breakthrough last time have reviewed it and called it legitimate. Two independent validations in one story: the maths checks out, and apparently OpenAI has learned what peer review is for.
Source: TechCrunch
The Clueless Intern Is Getting Slightly Less Clueless
AI agents — software that takes instructions and actually does things, rather than just chatting politely about them — have spent years promising to be your capable personal assistant while performing closer to someone's first day at work. That's shifting. Google, OpenAI, and others are now chasing OpenClaw, a viral open-source agent platform, which is the sort of sentence that would have sounded entirely made up eighteen months ago.
Source: The Verge
Anthropic Rents Its Brainpower From Elon Musk, at a Modest £1 Billion a Month
In a deal that raises several questions and answers precisely none of them, Anthropic — the safety-focused AI lab that has loudly positioned itself as the responsible adult in the room — has agreed to pay xAI $1.25 billion per month for computing power. That's Elon Musk's infrastructure running Claude's neurons, which is either a pragmatic business arrangement or the most ironic partnership since cats started endorsing dog food.
Source: TechCrunch
YouTube Now Lets You Insert Yourself Into Strangers' Videos, Which Is Fine
Google has added an AI remix feature to YouTube Shorts, powered by Gemini Omni, that lets users restyle any clip — or drop themselves directly into it. Tap the remix icon, write a prompt, and Gemini will cheerfully reimagine someone else's content with your face in it. The creative possibilities are genuinely interesting. The moderation challenges are, one suspects, even more interesting.
Source: The Verge
AI Can Now Write the Whole Research Paper. Whether It Should Is Another Matter
A new benchmark called ResearchArena has arrived to ask an uncomfortable question: AI systems can produce complete academic papers end-to-end, but can they produce good ones? Feasibility and quality, it turns out, are not the same thing — a distinction that has also eluded a significant number of human researchers over the years. The study provides a systematic framework for evaluating agent-generated science, which feels urgently necessary given how many papers are apparently already in the pipeline.
Source: ArXiv AI
|
|
As always, none of this newsletter was peer-reviewed — but then again, neither was roughly half of it.
|
|
Was this forwarded to you?
Subscribe here
Written with AI assistance, edited by humans. •
Privacy •
Unsubscribe
|
|
