Mrs. Nesbitt

Someone I know asked Astra to play Factorio. It wins, but it takes forever and guzzles tokens. If you ask Astra instead to “design a system to win Factorio,” it builds software that runs 1000x faster. Mechanistic software gives you cheap leverage, and LLMs can write mechanistic software on demand.

there's better ways to do this. See this? Fully independent WoW agent that runs off Haiku, happily running around levelling up, costs very little, handles realtime actions. The LLM can just be the brain and design it's own harness as it plays. In this case, start it off with something good at writing code (like Opus), then go to Sonnet when it's seeming confident, and then Haiku or a local model when it's doing well. I'm also experimenting with using decision models like Jev for fast decisions where it's smoother than trying to have the classic game AI run everything with a million if statements, and in Minecraft training small policy models with RL as the nervous system You don't sit there and conciously think through every muscle movement and direction your tendon needs to move, you set up a goal and intent, and your nervous system has learned to interpret that. It's interesting to let LLMs alone drive things raw as an eval, but it's not very practical or economical