Three small macOS demos that try out typed decision models: models that return probabilities for yes/no, multiple-choice and rating questions instead of generating text. They compare four backends:
- FluidUse: on-device models on the Apple Neural Engine. CUA-S1-FORMS (706K parameters) for form filling, laya (322M) for general typed questions.
- Jev: TypeSafe's hosted System One model.
- Kev: open Jev-like models served locally with the same API.
- Claude (
claude-opus-5), in one demo, as a general-model baseline.
Each demo is a SwiftPM app in demos/ with its own README. Every demo has a headless
self-test that prints its numbers.
| Demo | What it shows |
|---|---|
demos/form-ghost |
Loads a sample form (checkout, passport renewal, W-4, clinic intake) and overlays what CUA-S1-FORMS would type into each field, with confidence halos, timings, and a personal database of your details. |
demos/tetris-heatmap |
laya, Kev and Jev play Tetris side by side on one seed. Every landing is scored with "Is this a clean placement?", drawn as a heatmap, and every model call can be inspected. |
demos/settings-patrol |
A local recreation of GitHub's user and org settings. Type a vague request ("stop emails when my CI breaks") and laya, Jev or Claude narrows it down to the exact setting, showing each step. |
- An Apple silicon Mac on macOS 14 or later, with Swift 6.
- The first run downloads the FluidUse Core ML models from Hugging Face.
- Keys go in a
.envat the repo root (git-ignored) or the environment:JEV_API_KEYfor Jev, in the Tetris and Settings Patrol demos.ANTHROPIC_API_KEYfor the Claude engine in Settings Patrol.
- For the Kev column in Tetris: uv and about 23 GB of memory for
Kev-4B. Start it with
demos/tetris-heatmap/scripts/kev-server.sh.
cd demos/tetris-heatmap && swift run -c release TetrisHeatmapAll numbers are from one M5 Max Mac. They are small samples (30 Tetris pieces, 15 settings queries), so treat them as anecdotes rather than benchmarks.
Form filling. CUA-S1-FORMS takes about 2 ms per field. From each form's sample profile it got most names, addresses and contact fields, skipped many fields whose labels didn't match, and was confidently wrong now and then (a ZIP code into "Promo code" at 99.9%). A personal database raised the fields filled per form from 9/12/4/7 to 13/18/6/10, mostly by matching field labels in code before asking the model.
Tetris (30 pieces, same seed):
| laya (on-device) | Kev-4B (local) | Jev (API) | |
|---|---|---|---|
| Time per piece | ~90 ms (≈22 calls of 4 ms) | ~855 ms | ~400 ms |
| Lines cleared | 8 | 3 | 1–5 across runs |
| Memory | ~0.4 GB (whole app) | ~23 GB | — |
| Cost | $0 | $0 | ≈ $0.0037 per 30 pieces |
No model sees the board. Code computes each landing's features and describes it in one sentence; the model judges the sentence. This is a latency demo, not a good Tetris player. Kev-0.8B was also tried: 5.6–8.7 GB of memory but slower (2.5 s per piece) and much worse (0 lines).
Settings search (15 vague queries against ~450 GitHub settings):
| Engine | Right setting ranked #1 | In top 3 | Time per query |
|---|---|---|---|
| laya (on-device) | 3/15 | 6/15 | ~0.7 s (~90 calls) |
| Jev | 14/15 | 15/15 | ~0.7 s (1 request) |
Claude (claude-opus-5, low effort) |
14/15 | 15/15 | ~4 s (1 request) |
laya fits a question and all its options into 256 tokens, so it can only compare a few settings at a time. Jev takes up to 255 options per question and sees every setting at once.
- The GitHub settings pages are a local mock written from memory of GitHub's UI, not scraped from GitHub.
demos/tetris-heatmap/Sources/TetrisHeatmap/TetrisSimulation.swiftis vendored from FluidUse (Apache-2.0).- Personal data in the demos is fictional (
example.invalidaddresses,000-…SSNs).