artlu's Bear Blog

Too Early Optimization

when it comes to LLMs, I am the über-efficient tokens user. I have mastered one-shot agentic tasks at the exact optimal cost-benefit tradeoff

this includes minimizing effort, trading upfront thoughtfulness / thoroughness against ongoing monitoring, and reducing casino pull aspects of non-determinism

it has everything to do with the state of LLM tooling, and almost nothing to do with real-world goals and constraints. And ignores the pace of LLM capabilities improvement

yes I'm winning at LLMs. but probably losing at future LLM usage

when tokens were more subsidized, I probably should have taken the practice swings to become a transcendent all-time GOAT like Shohei, rather than honing my swing to be a hall-of-fame singles-and-triples OPS maxxxer like Ichiro