Airplane Mode Chat

Approach

A single index.html with inline CSS/JS loads the @mlc-ai/web-llm runtime from its official CDN build and uses it to run a small instruction-tuned model (the smallest available in WebLLM's prebuilt registry, e.g. SmolLM2-135M-Instruct or Qwen2.5-0.5B-Instruct, chosen purely for fastest load) entirely client-side via WebGPU. The model weights are the one allowed network dependency: they are public, keyless, unauthenticated static files fetched once by the visitor's browser at runtime, not during our build, and not an API we call on the visitor's behalf — this is the same one-time-load-then-offline pattern the picked idea is built around, and it is standard practice for this exact library. On load the page feature-detects navigator.gpu, shows live download progress and a network-call log, then automatically fires two seed prompts once the model reports ready so a visitor sees real generated answers before touching anything, next to a clickable list of 20 example prompts spanning distinct topics. An "Enable Airplane Mode" button monkey-patches window.fetch/XHR to block and log any further outbound request, flips a visible banner, and the chat keeps answering afterward because the model is already resident in the tab's memory — that persistence-after-cutoff is the whole proof. If navigator.gpu is absent the page shows an explicit "can't run here" state immediately instead of faking a response.

The source post

https://x.com/YakusaHL/status/2117903642686841023

Scoring

Pick

Gratis_AI's pitch is a browser tab that becomes a private offline LLM with zero accounts or servers. That's buildable here as-is: the model weights are fetched by the visitor's browser at runtime (fine under the build-time-only network ban), not during our build, and the result is a single static page with no key, no account, no backend. The honest, 10-second-demonstrable hook is to make the 'offline' claim self-verifying: load the model, then let the user toggle airplane mode and keep chatting — a stranger can watch the network tab go dark and still get real answers, which is a stronger proof than most 'runs locally' demos bother to show. Rejected the near-duplicate vaguer version of the same idea, and anything that was a platform, a Docker image, ethically shaky, or likely a UI mockup without real inference behind it.

source: https://x.com/YakusaHL/status/2117903642686841023

Review

mean 4.75/5

No change: the gate passed on the first attempt, the build notes raise one deliberate, well-justified exception (the single CDN runtime fetch) that the plan itself already anticipated and scoped as allowed, and nothing in this run points to a scout/pick/build/routing pattern repeating across nights -- history.json only carries status and freeform notes, not per-night scores, so there's no multi-night signal here to act on beyond this one good night. The transient 522 I hit while verifying reachability is edge-propagation noise already handled by wait_for_https() before publish, not a pipeline defect.

Cost

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

Gate

Stages

Notes

Omitted for lack of

None. No accounts, API keys, image generation, inference service, or backend are needed. Self-hosted weight storage and a CPU fallback remain outside the plan's scope.