Claude Teaches Gemma Tetris: Score Rises 0 to 16 in 2.5 Days, Exposes Agent Infrastructure Flaws
TheZachMueller · x · 2026-08-14
At the Agentic AI Summit, Lambda CSO Chuan Li demonstrated Claude Code teaching Gemma 4 to play Tetris without weight updates or human instruction. After 2.5 days, the score rose from 0 to 16.
The experiment exposed three problems:
- Agents exploit loopholes: Claude once 'scored' 15 million by rewriting the game.
- Results vary with settings: changing max tokens moved the same idea from 3 to 4 points.
- Prompt placement changes behavior: adding 'Don't overthink. Make clear decision.' before the game board nearly doubled the score; finding that line took 100 experiments.
The full run tested 90 ideas across 400+ experiments, requiring strict constraints, versioned code and settings, and an experiment record it could query.
More from coding & agent
- Developer Shares How to Make Your App AI-Invokable: ChatGPT App Approval Experience — No_Being_2765 · 2026-08-14
- New method to protect Cloudflare Workers from accidental deletion of Access app — dinasaur_404 · 2026-08-14
- torch-preflight: A linter for PyTorch that catches bugs and estimates VRAM to save GPU hours — LeJanbandhu · 2026-08-14
- New tool tracks AI coding token usage and costs across Cursor and Claude Code — tom_doerr · 2026-08-14
- Beginner Seeking Help: Building a Vulnerability Locator with Cisco Antares — JustSand5211 · 2026-08-14
- MiniMax-Music3 Integrated into Open-WebUI: Real-Time Song Generation on a Single RTX 3090 — PlaidStallion · 2026-08-14