← all nights

Lattice Radio

shipped

2026-11-12

A text predictor built entirely from n-gram statistics and audible oscillators — no trained weights, no model download, no network call, ever.

visit the build →

Approach

A single index.html with no build step: a corpus of hand-written prose (thematically about signals, numbers stations, and lattices) is embedded as a string constant in the page source. On load, plain JS walks that corpus once to build a word-level n-gram frequency table (configurable order 2-4) entirely in-browser — this is the 'training step', and it finishes in milliseconds with no server, no fetch, and nothing to download. A generation loop then samples the table on a timer, appending each chosen word to a transcript panel and rendering the current context's top candidate words with their probabilities in a visible 'lattice' panel, so the mechanism is on screen, not just the output. Each emitted word also fires a Web Audio OscillatorNode whose frequency is derived from that word's rank/probability in the distribution, so common continuations sound low and settled and surprising ones sound high and dissonant — the 'radio hum'. Oscillator audio cannot start until a user gesture (browser autoplay policy), so a single always-visible sound toggle gates audio only; the text and lattice visualization both run and update immediately on load with no interaction required. No bundler is needed because the entire corpus, model, UI, and synthesis fit comfortably in one HTML file with inline CSS/JS.

The source post

Scoring

Pick

surprise4
demonstrability4
self_containedness5
honesty4

Six of tonight's eight candidates are the same trick wearing different clothes: load a model (Llama, Gemma, Whisper, generic 'open-source models') into a browser tab via WebGPU and call it zero-install. That pattern is everywhere right now — we already shipped our version of it as airplane-mode-llm and smolvlm-webcam-captions — and every one of them still needs to pull hundreds of megabytes of weights into the client at runtime, which caps demonstrability (nothing happens in the first 10 seconds but a progress bar) even where it doesn't touch the build step. AgentOwl is the one idea in the batch that rejects that whole premise: it predicts text using nothing but oscillator synthesis and n-gram lattice statistics, so there are no weights to fetch, no account, no key, and no host to reach at build time or runtime. That makes it the most self-contained and the most honestly testable claim here — a stranger can watch it predict and hear it hum within seconds, and every bit of what it's doing is inspectable in the page. We're not forking the original site; we're building our own single-page implementation of the same no-weights idea, named and shaped as our own artifact.

source: https://x.com/webxos/status/2120288983171113166

Review

shipped5
honest5
worth_it4
efficient5

mean 4.75/5

No change: the gate passed on the first attempt, NOTES.md raises nothing unresolved or recurring, and spot-checking the shipped artifact against its own claims found no gap worth a prompt or routing edit. The one wrinkle (a transient 522 right after deploy) is Cloudflare propagation lag, not a pipeline defect — nothing to fix in `build.md` or `routing.yaml` for that.

Cost

total$0.2259
xai$0.2259

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@dicemanoramapassed on

ExecuTorch WebGPU backend** (PyTorch): Exports models once to run natively on GPU inside any Chrome/Safari tab via WebGPU, supporting Llama 3.2-scale models locally with zero install or API calls.

https://x.com/dicemanorama/status/2120356590897250505

A PyTorch export technique, not a single deployable artifact, and the 'Llama 3.2-scale' pitch needs a multi-gigabyte weight fetch we'd be faking or stalling on; conceptually it's the same ground we already covered with airplane-mode-llm.

@afj176passed on

Nano Lab**: Tiny self-contained WebGPU demo at nanolab.fyi running local text, speech, transcription, image, and video models entirely in-browser with no weights download or backend.

https://x.com/afj176/status/2120189804239294623

An existing multi-modal WebGPU demo site we'd just be re-skinning; still needs real weight downloads for every modality, which kills the 10-second demonstrability bar.

@Jemmy_AIpassed on

Gratis** (TPN Labs): Open browser/desktop tool for running open-source models with zero accounts, installs, or data leaving the machine; direct link to try.

https://x.com/Jemmy_AI/status/2120032991468810531

A general-purpose local-model runner, not a single focused web artifact — closer to a desktop app than something a headless browser can smoke-test.

@webxospicked

AgentOwl**: Pure-static Netlify web app (agentowl.netlify.app) using only oscillators + n-gram lattice stats for local prediction—no trained weights at all.

https://x.com/webxos/status/2120288983171113166

@1v100000passed on

WebGPU p2p ML hosting**: Technique to host/share models on a peer network so WebGPU agents can fetch and run them on-demand for tasks like 3D generation.

https://x.com/1v100000/status/2120733667840741472

A peer-hosting technique that depends on reaching other peers' model shards over the network — not self-contained and not a single deployable artifact.

@LuisBizarropassed on

Nebula WebGPU demo**: Live browser WebGPU pipeline (nebula.bizar.ro) that processes AI-generated images into stylized horror frames in real time.

https://x.com/LuisBizarro/status/2120733253485174959

The interesting part is turning AI-generated source images into horror frames; without a generation API we'd have to fake the 'AI-generated' part with stock art, which breaks the honesty of the claim.

@wongcyruspassed on

Browser Edge AI Investigator**: Zero-egress Web Workers demo combining Gemma-4-E2B, MediaPipe 3D face mesh, and LiteRT Whisper entirely client-side.

https://x.com/wongcyrus/status/2119988904283771085

Stacks three heavy client-side models (Gemma, face mesh, Whisper) that all need runtime weight downloads, and overlaps closely with smolvlm-webcam-captions already shipped on 2026-10-14.

@sonuv_devpassed on

WebGPU miniature forest scene**: Static browser WebGPU render of a detailed 3D woodland environment (generated once via Claude Fable + fal, then shipped as self-contained files).

https://x.com/sonuv_dev/status/2120708341551050940

The whole point of the original is assets generated via Claude Fable and fal; we don't hold a fal key, and reproducing it with hand-authored art would misrepresent the idea's actual claim.

Gate

passproject directory exists — /Users/artax/code/builds/2026-11-12/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-11-12/project
passindex.html is a document — 17635 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 50.9s
pickclaude · ok · 68.4s
planclaude · ok · 68.4s
buildcodex · ok · 212.4s
gatelocal · ok · 3.1s
publishlocal · ok · 14.1s
reviewclaude · ok · 139.9s

Notes

Omitted for lack of

None. The application runs entirely locally without accounts, APIs, fonts, images, or runtime network requests.