AI hits top 1% at Super Smash Bros. Melee trained on 80k hours of human replays in 16 hours
FinanceYF5 · x · 2026-10-23
Eric Gu's HAL is a 247M-parameter causal Transformer trained purely on 80k hours of human Melee replays, reaching top 1% on ranked ladder after just 16 hours on one B200.
Key takeaways
- Reads 256 frames of game state + controller inputs (no pixels) and autoregressively decodes multi-frame action chunks, replanning below 10 Hz to hide inference and netplay delay for real-time online play.
- Return conditioning works surprisingly well: requesting a constant 99th-percentile return every frame yields stable play.
- Derives Melee pre-training scaling laws: model FLOPs should scale faster than data (M∝C^0.58, D∝C^0.42), contradicting the "just add data" intuition.
- Observation delay hurts more than model size helps; a numerically inexact KV cache still works.
- HAL can mimic pro players, adapt difficulty to opponents, and is playable against humans via Slippi with a GameCube controller. Compute supported by ajhinh and Modal.
Related event: AI Trained on 80,000 Hours of Human Gameplay Reaches Top 1% in Smash Melee(2 posts)→
More from Fun
- The AI business loop in one tweet: tech creates problems, then sells tokens to fix them — srimisra · 2026-10-23
- Watching AI Play Factorio and Build a Factory System for Free — rachittshah · 2026-10-23
- Claude takes user's joke rule literally and delivers meme-style 'GO TO SLEEP' lockscreen — venturetwins · 2026-10-23
- "All the 200 IQ Autists I Know Want to Be Rappers" — So He Trained His AI to Do It — ivan_bezdomny · 2026-10-23
- Palantir popup merch already hit by scalpers in Seoul, resale prices soar — eliano · 2026-10-23
- Palantir Seoul pop-up draws big crowds on day two — eliano · 2026-10-23