← sources

My AI learns to clear Super Mario Bros 1-1 in 15 mins and it is not PPO based

source: reddit (reddit:deeplearning)published 2026-10-03 07:49ingested 2026-10-04 02:36

retained text

I tested Adapt-1, a non-LLM learning and reasoning system by Rei Labs, by having it learn and play Super Mario Bros, and it performed quite well.

I tried it on World 1-1, starting untrained. It learned a reactive policy from its own play in about 36 minutes of gameplay, then cleared the level with learning off.

With Machina, Adapt-1's sequence engine. Starting untrained, it found a button sequence that reaches the flag after 403 attempts, in 11 wall-clock minutes.

Full thread: [https://x.com/hsrvc\_/status/2106025501752234112?s=20](https://x.com/hsrvc_/status/2106025501752234112?s=20)

Code, the exact data, traces, clips and a step-by-step guide with costs are all public: [https://github.com/hsrvc/adapt1-mario](https://github.com/hsrvc/adapt1-mario)

document fetched 2026-10-04 00:23 · canonical https://reddit.com/r/deeplearning/comments/1wwhfuq/my_ai_learns_to_clear_super_mario_bros_11_in_15