70–500ms
End-to-end
193.6×
Faster
444.6×
Cheaper
0
Hallucinations
Jev’s published figures. This page shows the part you can watch.
Real time, not sped up. Jev’s and the chat model’s published completion times on the same System One task.
A frontier-intelligence function call: unstructured state in, typed decisions out. Named for Kahneman’s fast System 1 — and for Jevons, whose paradox says making something radically cheaper only multiplies its use.
| Chat model | System One | |
|---|---|---|
| Output | A string you must parse | A typed decision |
| Sampling | Token by token | Every field in parallel |
| Training | RLHF — human preference | RLCD — calibrated decisions |
| Failure | Can hallucinate anything | Constrained to the type |
- 01Detect
An obstacle crosses the decision zone.
- 02Serialise
The world becomes eight numbers — no history, no pixels.
- 03Ask
Once per obstacle, never per frame. The server owns the prompt.
- 04Answer
JUMP, DUCK or NONE comes back. Nothing else is accepted.
- 05Act
If the obstacle is still live it fires; if it has passed, the reply is dropped as stale.
You can throw an obstacle at the runner mid-run and watch Jev answer it.
- Decision
- JUMP · DUCK · NONE
- Payload
- Eight numbers
- Cadence
- One call per obstacle
- Round trip
- 130–290 ms, measured
- Inference
- Served via Groq
- Stack
- Next.js 16 · React 19 · Canvas 2D
The runner, cacti and birds are Chrome’s dinosaur at 2× scale, redrawn as rectangles in code — none of Google’s image files are used, so the whole scene repaints when you switch themes.
U3lzdGVtIE9uZSDCtyB1bnN0cnVjdHVyZWQgc3RhdGUgaW4sIHR5cGVkIGRlY2lzaW9ucyBvdXQ=