Letters that lie to bots
Hey 👋 — today's mix leans playful: a font with a secret double life, a water shader getting its glow-up, and Linus Torvalds weighing in on the AI question with his usual lack of patience for nonsense.
Decoy Font
Fonts don't usually ship with a security model, but here we are. Decoy Font is a free, downloadable TTF that hides two different messages in the same letterforms, using hybrid-image spatial-frequency tricks so a close reading shows one text while squinting from across the room reveals another. The pitch is refreshingly specific: gum up AI scrapers and OCR pipelines while staying perfectly legible to the human actually typing (via HN discussion).
"prints a decoy for every letter, making it more difficult for AI to read what you're typing" — Decoy Font
Three.js Water Pro realism update
A short demo clip teases an upcoming realism pass for Three.js Water Pro, the drop-in water shader add-on that's become shorthand for "make my WebGL ocean stop looking like blue Jell-O." No release date yet, but the reflections alone are worth the thirty seconds.
Fable 5 3D website tutorial with prompt-to-code
In a similar vein, a tutorial making the rounds swaps hand-coded Three.js and WebGL for prompt-to-code generation via the Fable 5 model — describe the 3D site you want, get working code back instead of a wall of shader boilerplate. Early days for the format, but worth a look if you'd rather describe geometry than debug it.
Linus Torvalds on LLM usage in kernel development
Elsewhere, Linus Torvalds weighed in on LLM use in kernel development with his usual bluntness, telling a mailing-list correspondent flatly that Linux is not an anti-AI project and that maintainers should make the tools work for them rather than reject them on principle. Coming from someone who built a version control system partly out of distrust for existing tools, it's a notably pragmatic stance (via Lobsters discussion).
Kimi K3, and what we can still learn from the pelican benchmark
Staying in AI Corner, Simon Willison put Moonshot AI's freshly released Kimi K3 — a 2.8-trillion-parameter model — through his running SVG-pelican-riding-a-bicycle test, alongside the usual benchmark and pricing rundown. The pelican test has become an oddly reliable vibe check for new models, and this one's attempt is worth a look purely for the bicycle.
Quick hits
- Detecting LLM-Generated Texts with "Classical" Machine Learning — no transformers required, just good old feature engineering.
- How to Train a Gen AI Kick Drum Model on Your Old Linux Desktop with 6GB VRAM — proof your dusty tower still has a beat in it.
- Show HN: A map of cafes that are in the sun — for when the coffee matters less than the vitamin D.
- 1-Bit LLM in the Browser — a whole language model living on a diet of ones and zeros (mostly zeros).
- Kashmir Hill tries to find her AI-generated slop biographer — a journalist hunts the ghostwriter that never existed.
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.