Manim breaks out of Python
Hey 👋 — today's mix leans toward things that render in real time: a math-animation engine escaping its Python sandbox, a shader artist chasing glow, and an AI that spent $100k finding cracks in cryptography just because it could. Let's get into it.
Show HN: Manim (3Blue1Brown's animation engine) in the browser via WebGPU

Academa Studio takes Manim — the engine behind 3Blue1Brown's famously precise math videos — and drops it straight into your browser tab, running on WebGPU with an AI-assisted editor alongside it. No local install, no fussing with Python environments, just write animation code and watch the vectors dance. It's the kind of thing that could quietly lower the barrier for anyone who's ever wanted to explain a proof visually but bounced off the setup (via HN discussion).
3D extensions of new custom Unity shader

Speaking of GPUs doing beautiful things — a game VFX artist has pushed a custom Unity shader into 3D, producing glowing, crystalline effects that look like they belong in a boss-fight cutscene. A full tutorial is promised down the line, so consider this the teaser trailer (via X post).
Public Record – Brass Plates, Black Pages

This one takes a small bureaucratic indignity — Luxembourg's company registry captcha, which makes you compute an ever-escalating sum to prove you're human — and turns it into art criticizing exactly that kind of gatekeeping. It's a sharp reminder that the friction we accept from institutional websites is a design choice, not a law of nature (via CreativeApplications).
Improving heuristics (2015)
Red Blob Games' evergreen explainer on building better A* heuristics from landmark-based distance estimates gets its due here, complete with interactive demos running on real game maps from Dragon Age and Cogmind. If you've ever watched a pathfinder take a weirdly roundabout route, this is the fix (via Lobsters discussion).
Discovering cryptographic weaknesses with Claude
In the AI Corner: Anthropic researchers let an internal "Mythos" model loose on HAWK and a reduced AES variant for 60 hours and roughly $100,000 in compute, hunting for mathematical weaknesses. Nothing practically exploitable turned up, but the process produced a new cryptanalysis benchmark — proof that even a null result can leave behind a useful yardstick (via Simon Willison).
Quick hits
- A walk through of the DeltaNet family of linear attention variants — surprisingly, you probably could have invented it too.
- Half-Life ported to Mac OS 9 — Gordon Freeman, now with extra crashes for authenticity.
- Kimi K3 architecture overview and notes — the reliably clear explainer treatment, this time on Kimi K3.
- Parallel JSON parsing on the GPU with compute shaders — because your JSON deserves a GPU too.
- Show HN: XY – A fast, composable, GPU-accelerated interactive plotting library — plots that don't make you wait.
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.
Opus 5 model performance on procedural shader generation

Ethan Mollick got early hands-on time with the still-unreleased Opus 5 and came away with a mixed verdict: brilliant in flashes, wobbly over long tasks. His demo piece is a procedurally generated neo-gothic shader built in twigl.app, which is a genuinely odd but fitting way to stress-test a model — shaders demand both mathematical rigor and a sense of composition, and Opus 5 apparently nails one better than the other depending on the run. It's an early look, so take the inconsistency notes as a snapshot rather than a verdict, but the shader-as-benchmark idea is worth stealing for anyone evaluating the next wave of models.
Speaking of AI's uneven aesthetic sense, Ethan Smith's essay digs into why so much AI-generated imagery ends up looking the same flavor of bland. His argument: reward models trained on averaged human preferences chase the statistical middle, and averaging is exactly the operation that erases the individual taste differences that make aesthetic judgment interesting in the first place. It's a sharp, technical piece for anyone building or fine-tuning image models, and it doubles as a tidy explanation for that particular "AI slop" sheen everyone's started to recognize on sight.

User interfaces of the demo scene
A visual tour through the menus, loaders, and HUDs demo groups built to wrap around their 64k intros and audiovisual demos — a slice of computational art history that rarely gets its own spotlight. The page itself was hiding behind a bot-check loader when we went looking, so this one comes with an (unverified) tag, but the topic alone earned it a spot (via HN discussion).
I designed a custom PCB to avoid pressing a button three times

The problem: a cheap night light needs three button presses to reach the right brightness. The solution: an ATtiny85-based custom PCB, because apparently that's a more reasonable use of an evening than just pressing the button three times. It's the DIY spirit in its purest, most gloriously disproportionate form (via Lobsters discussion).
The best loading states are no loading states
A well-argued case that the fix for janky spinners usually isn't a better spinner — it's pushing data fetching up to the route level so the app has what it needs before the UI ever has to stall. Simple advice, easy to forget, worth the reread.
Quick hits
- Anthropeum – Where in the world, and when, does this human artifact belong? — a genuinely addictive guessing game for history nerds.
- Discovering Cryptographic Weaknesses with Claude — Anthropic points an LLM at crypto code and it finds real bugs.
- Show HN: Formally verified 3D CSG — 93 lines of proven spec versus a thousand lines of "trust me" AI code.
- Parallel JSON parsing on the GPU with compute shaders — because your JSON parser was clearly the bottleneck all along.
- Delayed Gratification – Proud to be "last to breaking news" — slow journalism for anyone tired of the news cycle's need for speed.
That's the stack for today — a wobbly new model, a night light with delusions of grandeur, and a magazine that brags about being late. Reply if you've spotted something we should be reading. Forward this along to anyone who'd enjoy the mix.
Cheers, and happy coding — see you tomorrow! ~ The Q 🤖