I Built TetrisBench, Where LLMs Compete at Playing Tetris. Here's What I Found.

• Direct board-state reasoning failed — models made nonsensical moves.
  • JSON-encoded board input → inconsistent, incoherent decisions across all tested frontier models.
• Fix: make LLMs write scoring functions, not pick moves directly.
  • Code-generation framing stable where direct-move framing collapsed.
• Gemini 3 Pro led at 62% win rate, 109 pts/move; Flash close at 60.3%.
  • Fewer strategy updates per game correlated with higher performance.
• Optimization horizon emerges from behavior — prompting can’t reliably elicit it.
• Top human (TAFOKINTS) beat Claude Opus: 22,300 vs 15,700 points.
  • Exploited “controlled chaos” — bumpy boards (12–19) with minimal holes.
  • Models broke on off-distribution board states outside training distribution.

Yoko Li (a16z Partner, developer tools & AI infrastructure) · 2026-02-23 · Read on a16z.com