Question 1

Are the four lanes a fair comparison, or was the question tuned to flatter Jev?

Accepted Answer

Every lane reads the same two constants for the question and the three valid answers, so nothing is worded twice, and the chat models are asked the identical question through the same Gateway with structured output, temperature 0, reasoning off and no retries for anyone. The counter-evidence is in the open too: in the author's runs the chat models got that question right 95 to 100 percent of the time. Read it as the latency demo the README says it is, not as an accuracy ranking.

Question 2

When a human is playing, does the ball really wait for the model?

Accepted Answer

No, and the README says so. In any mode where a human plays, a ball step never completes faster than 260 ms even when Jev answers sooner, and the published latency is still the real round trip measured around the model call and nothing else. Only demo mode is unpaced. A closed laptop also cannot pin a paddle to the wall, because input is a level repeated every 150 ms and forgotten after 400 ms.

Question 3

What does the model actually see on each step?

Accepted Answer

About 125 bytes of JSON containing only numbers: court width and height, ball position and velocity, paddle position and height, an intercept target and a direction. The question and the up, down, stay criteria live in one file that every lane reads, and the three answers are fixed. One decision is one ball segment, so what comes back is a control, not a paragraph.

Question 4

What were the measured decision rates?

Accepted Answer

Recorded on Vercel iad1 through the Gateway on September 17, 2026, 45 seconds per lane, with every figure on the site read from a committed JSON file at build time. Jev at 4.4 decisions per second, 227 ms average, 400 ms p95, and 47 decisions in the first 12 seconds. Against that: Gemini 3.8 Flash at 0.32 per second and 3.2 s, Claude Haiku 4.5 at 0.40 and 2.5 s, GPT-5.6 Sol at 0.28 and 3.5 s.

Question 5

Why Pong of all the games to test this on?

Accepted Answer

Because it is the worst case. In the Atari-GPT benchmark, Pong was the game chat models handled worst, below random, because deciding takes them longer than the game gives them. The author's argument is the same one: everything built with AI so far waits, and a model that decides inside human reaction time can be a participant rather than a request.

Question 6

How does the agent take part in a game?

Accepted Answer

Every participant in a game is a member of one Ably channel named pong:{game id}, the agent included. The browser gets a short-lived token, enters presence as a player and publishes input; the agent is a Node process kept alive inside the same Vercel function invocation with after(), joins as agent, runs the physics and publishes a state snapshot after every decision and at least once a second. Spectators attach to the same channel with rewind, and the viewer count is presence, which is why human versus human needs no extra code.

Question 7

What makes a game end, and who decides?

Accepted Answer

A referee, on five conditions: a score reaches five, a player has been gone for 15 seconds, nobody took a side within a minute, a demo has been unwatched for two minutes, or 15 seconds before the platform deadline. The final frame says which one it was, and the agent then leaves presence. The agent itself is only kept alive for the function's maxDuration, so long games run under a hard ceiling rather than an open-ended process.

Jev Pong

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community