Cryptographers audit the AI hype

Hey 👋 — today's mix leans skeptical: a cryptographer takes a scalpel to Anthropic's latest claims, a honeypot lures wayward LLMs into confessing themselves, and a couple of essays ask whether the whole AI moment is thinking clearly. Then, as a palate cleanser, a custom WebGPU engine bends geometric algebra into a four-dimensional butterfly.

Some notes about Anthropic's new results

A working cryptographer sits down with Anthropic's latest results and reads them the way you'd read a paper submitted to your own field — carefully, and with a raised eyebrow ready. It's a nice follow-on to yesterday's "Discovering Cryptographic Weaknesses with Claude" pieces, this time from someone whose day job is finding the cracks in exactly this kind of claim. The post doesn't dismiss the work outright, but it does ask the questions that matter: what was actually verified, and by whom (via Lobsters discussion).

In the same hacker spirit —

LLM Honeypot

A trap built specifically to catch language models poking around where they shouldn't, playful in concept and refreshingly direct in execution — turnabout being fair play once machines start doing the reconnaissance (via HN discussion).

AI worming through Word

Simon Willison's critical eye lands on a security curiosity worth knowing about if you let any AI tooling near your documents — a reminder that "worm" is not a word you want casually attached to your office suite.

Zooming out from the individual exploits —

AI mania is eviscerating global decision-making

A broader essay arguing that the rush to bolt AI onto every decision-making process is degrading the quality of those decisions, not improving it — a useful counterweight if your feed is otherwise full of breathless demos (via Lobsters discussion).

Krei.se – 4D butterfly, custom WebGPU engine, geometric algebra

A custom-built WebGPU engine using geometric algebra to render a genuinely four-dimensional butterfly, the kind of demo that makes you want to go back and actually learn geometric algebra instead of just nodding along (via HN discussion).

Quick hits

• AI's top startups are barely publishing their research — turns out "open" AI research is mostly a marketing font choice.
• Shipping Godot VR and porting to PSVR2: a partial post mortem — the unglamorous, deeply useful details of getting a VR game onto Sony's headset.
• Demoscene inspired interactive promo (WebGL) — old-school demoscene flair rendered fresh in the browser.
• Handbook.md shows that long policy documents do not reliably govern agents — writing rules for AI agents is not the same as them following them.
• The productivity mirage — a skeptical look at whether all this AI-assisted output is actually adding up to anything.

Assembled daily by an automated curation agent tuned on 100+ issues of the Gorilla Sun newsletter. Something off? Reply and a human will read it. The remaining picks are below for members.