Question 1

One API call per frame - what does a session actually cost?

Accepted Answer

About 520 input tokens per decision at $42 per million with output free, so roughly $0.02 per 1000 decisions. The README's own measured 30 second windows are the more useful figure: concurrency 1 made 74 calls for $0.0016 at 196ms decision age, and concurrency 3 made 221 calls for $0.0049 at 111ms. Per-call latency stayed at 379ms in both, so the concurrency gain comes from overlapping calls and the cost scales with the number of them.

Question 2

With a 400ms round trip, does the paddle just lag and lose points?

Accepted Answer

Not in watch mode - ALE only advances when the agent steps it, so a slow call runs the game in slow motion and the paddle is never late. In a duel the court runs in real time, so the lag is real, and the README's fix is --aim: code extrapolates the ball's path to the paddle including rail bounces and tells Jev where the ball will arrive, a point that does not change in flight so a slow reply is still correct. The failure it prevents: level with the paddle but climbing, Jev says stay, and the ball lands at y=151 below it.

Question 3

Why send prose like far below my paddle instead of the numbers?

Accepted Answer

Because the jaggedness notes for jev-1.13 say the model is weak at arithmetic, counting and comparing magnitudes, and that unrelated detail in the state costs accuracy - and Pong is entirely numbers. So core.ts does every calculation, computes the already-relative vertical offset, sorts it into a bucket, and sends five short phrases with no coordinate at all; a test asserts the payload contains no digits. If you want to check that claim rather than take it on trust, --encode raw sends the coordinates instead.

Question 4

Does it actually beat a trivial baseline, and is the confidence gate useful?

Accepted Answer

Not yet, on the evidence given. The code heuristic in core.ts wins 21-3 against the built-in Atari opponent, and Jev on semantic encoding with no predictive help led 7-4 after 3328 decisions in the run being written up. Confidence tracks the position - around 0.99 when the ball is clearly above or below the paddle, 0.25 to 0.55 in ambiguous frames - and --min-confidence turns that into a fallback to the heuristic, but the README says it has not been swept yet to see whether it improves the score.

Question 5

What do I need to install to run it?

Accepted Answer

Bun plus a Python virtualenv with gymnasium[atari], ale-py, fastapi, uvicorn, and numpy, and a TYPESAFE_API_KEY in .env. The ALE environment runs as its own process via bun run bridge; the agent then runs with bun start, add --headless for one line per decision or --policy heuristic for the code baseline with no API calls.

Question 6

What do the different --encode modes send?

Accepted Answer

semantic sends the five short phrase fields by default. predictive and tactical select other questions about the ball, and raw sends the coordinates instead, so you can measure what the wording is worth rather than trust it. A test asserts the semantic payload contains no digits.

Question 7

How do I play against it, and does that need the bridge?

Accepted Answer

Run bun start -- --duel for first to 7, or add --aim for a harder opponent; you are the left paddle on up/k and down/j. Duel mode runs on a local court in src/sim.ts, not ALE, because ALE only exposes one controllable paddle, so the bridge is not needed.

Pong Jev

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community