Daydreams, now with API access

Hey 👋 — today's mix pairs a quietly moving art installation about AI and daydreaming with a much louder headline: a contested claim that AI just cracked a Millennium Prize problem. In between, a peek inside an LLM's attention and a stroll through the fossil record of early-2000s HTML.

Clouds to Poem – On the poetic inefficiency of AI

Clouds to Poem – On the poetic inefficiency of AI

A hanging cloud sculpture, hidden screens playing sky footage, and a voice reading machine-generated poetry while another screen exposes the model's internal workings as it composes — this installation isn't interested in whether AI can write a decent poem, but in what gets lost when we hand off the most gloriously pointless of human acts to something that never gets bored looking at the sky. The accompanying essay is one of the more thoughtful pieces of AI-art criticism to come along in a while, less concerned with capability than with what daydreaming is even for.

"asks what happens when we delegate the distinctly human, purposeless act of daydreaming to a machine" — CreativeApplications

Show HN: LLM attention visualization

Staying with what's happening inside AI's head: this tool lets you hover over a generated token and watch which earlier tokens the model was actually paying attention to, using a modified ONNX model that exposes its internal attention values. It's the rare educational demo that makes an abstract mechanism feel tactile rather than diagrammatic (via HN discussion).

AI has solved one of math's $1 million Millennium Prize problems

AI has solved one of math's $1 million Millennium Prize problems

Quanta reports that OpenAI researchers set 10,000 autonomous agents loose on the 3D Navier-Stokes equations and came back with a blowup singularity — potentially resolving one of the seven Millennium Prize Problems. The result is contested, and the timing is almost too on the nose: just days ago Terence Tao was warning about the risk of "prematurely solving" math problems with purely AI-powered methods. Whether this holds up under scrutiny or joins the pile of retracted AI-math claims is very much still an open question.

Seven worlds of digital art: a visitor's guide

Seven worlds of digital art: a visitor's guide

Media theorist Lev Manovich borrows Howard Becker's sociology of art to argue that digital art shouldn't be sorted by tool or file format, but by the distinct "art worlds" it belongs to — each with its own audiences, taste-makers, and rules for what counts as good. It's a useful lens for anyone who's ever tried to explain why a generative art collector and a demoscene coder seem to be speaking different languages entirely.

Antiquated HTML snippets and artefacts

A fond dig through the browser wars' leftover markup — the X-UA-Compatible meta tag that once begged Internet Explorer to please behave, Google Chrome Frame integration snippets, and other vendor-specific fossils still quietly sitting in old page source. Good clean nostalgia for anyone who remembers when "cross-browser" was a battle cry rather than an assumption (via Lobsters discussion).

Quick hits

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.

Building an infinite liquid glass grid with Three.js, WebGPU, and TSL

Building an infinite liquid glass grid with Three.js, WebGPU, and TSL

Studio Shader's latest breakdown is the kind of tutorial that makes you rethink what "geometry" even means: an endless grid of rounded, refracting glass cards, built entirely on flat planes with no actual bevels or lighting rigs involved. The trick lives in TSL shader math — fake the curvature, fake the light bending, and let WebGPU do the heavy lifting per-pixel instead of per-vertex. It's the sort of sleight of hand that looks expensive and runs cheap, which is exactly the kind of magic this corner of the newsletter exists for.

27.5 kB, language-agnostic, WebGPU-accelerated syntax highlighter

Speaking of WebGPU doing things you wouldn't expect it to — gpu-lexer is a 41k-parameter model, small enough to fit in a napkin's worth of kilobytes, that classifies code tokens for syntax highlighting without a hand-written grammar for each language. It lands within shouting distance of Shiki's labels (roughly 87% agreement) and does it all on the GPU. No grammar files, no per-language maintenance burden, just a tiny model that learned what a keyword looks like.

A collection of minimal, generative and customizable SVG-patterns

Book of Shapes is designer Nikolaj Sokolowski's growing library of generative SVG patterns — grids, flow fields, isometric tilings, noise fields — each one tunable rather than static. It's open, it's playable, and it's the sort of reference you bookmark and then actually use six months later when you need a background that isn't a stock gradient (via HN discussion).

On the Navier–Stokes Millennium Prize Problem

Simon Willison digs into the messier-than-the-headline story behind OpenAI's Navier-Stokes proof, where a separate team had reportedly been grinding on the same problem for a year before accusations surfaced that OpenAI's model may have seen their unpublished work. It's a fitting sequel to the AI-and-math unease we've been tracking all week — Terence Tao's wariness about "prematurely solving" problems with pure AI methods suddenly reads less like caution and more like foreshadowing.

We taught a fruit fly to hear human emotion

We taught a fruit fly to hear human emotion

Researchers took the actual wiring diagram of a fruit fly's brain — 499 neurons, connections fixed straight from real synaptic data — built a neural network out of it, and trained only a thin output layer on top to guess emotional tone from human speech. No architecture search, no fine-tuning the whole stack, just borrowing 60 million years of evolved circuitry and teaching it one new trick.

Proof that sometimes the best model architecture is the one nature already debugged.

Quick hits

That's the stack for today — glass that isn't glass, a fly that hears feelings, and a math proof with a paper trail messier than the equations themselves. Got a find worth including? Reply and tell us. And if a friend needs their daily dose of odd and useful, forward this along.

Cheers, and happy coding — see you tomorrow! ~ The Q 🤖