The pong demo is giving "Point to Jev" before it's even reached my paddle on mobile. I also don't really understand the frontpage model-speed comparison: it looks like you're slowing the ball down to compensate for the models' decision-making speeds? If so I think a more effective metric would be N-x, e.g. we slowed this game down 8-x (compared to Jev's game) for e.g. Sol to stay competitive. Or, just watching them fail would be cool too, though I know it's just a simulation and no real inference is being ran (except for Jev, maybe).
lol, playing against jev the ball routinely disappears dozens of pixels away from my paddle and it just awards itself the point anyways. Big Obama awarding Obama energy.
It's pong where the ball moves one step per model decision.
Four lanes, one model each, same serve, same rules.
Every step the model gets the same tiny JSON state (ball, paddle, predicted intercept) and the same question: up, down or stay. One answer moves the ball one segment; eight segments cross the court. No speed up / adjustments for real play, so a lane's speed is that model's decision latency.
Numbers, recorded on Vercel (iad1) next to the gateway, 45 seconds per lane: Jev averaged 227 ms (p95 400 ms). Gemini 3.2 s, Haiku 2.5 s, GPT-5.6 Sol 3.5 s.
In the first 12 seconds: 47 decisions vs 3, 2 and 2.
The pong demo is giving "Point to Jev" before it's even reached my paddle on mobile. I also don't really understand the frontpage model-speed comparison: it looks like you're slowing the ball down to compensate for the models' decision-making speeds? If so I think a more effective metric would be N-x, e.g. we slowed this game down 8-x (compared to Jev's game) for e.g. Sol to stay competitive. Or, just watching them fail would be cool too, though I know it's just a simulation and no real inference is being ran (except for Jev, maybe).
lol, playing against jev the ball routinely disappears dozens of pixels away from my paddle and it just awards itself the point anyways. Big Obama awarding Obama energy.
The game is broken for me. The ball just sits in the middle.
Hi HN, Matt here, co-founder of Ably. I built this after Jev landed on Vercel's AI Gateway: https://github.com/ably-labs/jev-pong
It's pong where the ball moves one step per model decision.
Four lanes, one model each, same serve, same rules.
Every step the model gets the same tiny JSON state (ball, paddle, predicted intercept) and the same question: up, down or stay. One answer moves the ball one segment; eight segments cross the court. No speed up / adjustments for real play, so a lane's speed is that model's decision latency.
Numbers, recorded on Vercel (iad1) next to the gateway, 45 seconds per lane: Jev averaged 227 ms (p95 400 ms). Gemini 3.2 s, Haiku 2.5 s, GPT-5.6 Sol 3.5 s.
In the first 12 seconds: 47 decisions vs 3, 2 and 2.
Bit of fun! Enjoy