Almost half (44%) of our tested claims are now expired in today’s models. Another 34% have weakened significantly, 16% persist, and 6% have strengthened relative to the original claim. Persistence varies by topic. Five of nine reasoning claims and six of eleven sycophancy-and-deception claims are classified as expired. Three of ten claims about models as judges, reviewers, and raters persist or strengthen.

Stanford researchers retest 50 widely cited LLM findings - most fade within 15 months via x.com

Cusack succeeds in delighting his readers with Momo, too, but only if traditional definition of “delight” is expanded to include hard-earned enlightenment. The story piles on enough facts, half-facts, and non-facts to create cognitive dissonance, which only then leads to something delightful: the moment when the reader asks where the truth is. Or more precisely, how much they’re being lied to.

Evan Brown on John Cusack's disorienting graphic novel Momo via x.com

Everything posted this week — 1 link.

What’s the Future for Pure Math Research in the Age of AI?

One might ask if at some point pure mathematics will somehow be “finished”. And to this we can definitively answer that it will not. Computational irreducibility ensures that there will always be an infinite sequence of mathematical facts—and surprises—to be discovered. And there will also be an infinite number of pockets of computational reducibility, representing new mathematical concepts and structures.

Many beautiful thoughts here in Stephen Wolfram's latest essay on AI's role in pure mathematics.

LLMs need an insane amount of text. Dessa learned language from hundreds of times fewer words (in fairness, LLMs also have to learn C++ and whatever the hell Angular is).

And it had to learn far more than language. Accents, pauses, breathing, sarcasm, humor, inflection, emotion, when to cut in: every one changes what someone means. Say “I didn’t say she stole my CUDA code” out loud, stressing a different word each time. Same words, different meaning every time.

Dessa is Block's speech-to-speech model, trained on human conversation alone - no text, no pretrained models via x.com