Post #1540615

a friend of mine told me they use custom scheme syntax instead of MCP with gemma, since it costs fewer tokens, and as they say "it knows scheme well". the other token optimization they use is to rollback context on errors and inject apriori knowledge into the think section ("actually, this won't work because ") both good ideas. they also told me they were initially quite enthusiastic about letting llm's code, but have since given up on it b/c it causes amnesia. #devlog