AI Bengineering

Thanks for the callout @GoogleDeepMind and congrats on beating Minecraft with 4xGemini 4 argon! Incredible accomplishment seeing agents independently solving problems throughout the game and excellent paper! arxiv.org/pdf/2610.22319

Typesafe's Jev is great and REALLY fast, but the round trip from Australia took ~300 ms - a little too slow for the quick reactions Minecraft combat needs. So I distilled Jev Minecraft combat decisions into LightGBM: ~1 ms locally, no API cost, and 90.8% agreement with Jev on movement decisions. Watch it fight and explore the model: aibengineering.github.io/min… Source Code here: github.com/aibengineering/mi…

opus 5 beat minecraft in 4 hours - jev seems a little stuck after 22 :D nitter.cf/aibengineering/status/…

💵Estimated API-equivalent model cost: ~$93 USD for the run. ~88% came from cached context reads: 163.3M cached tokens, versus 148K output tokens including thinking. Calculated from token usage and API rates—not an actual API bill. Excludes hosting and voice generation. 5/6

An AI agent has beaten vanilla Minecraft, end-to-end. From an empty inventory to defeating the Ender Dragon, using the new mine-ai-mcp server I built for gamers and researchers! Here’s the run—and the engineering that made it work. 🧵 1/6