Hello, world

I don’t like LLMs

Let me explain.

I think LLMs are fantastic tools, I use them every day. They help me prototype faster, research broader, test more extensively and occasionally…send me barking up the wrong tree (usually with a high degree of confidence and gaslighting).

As a systems engineer who has spent 5 years worrying about cache miss rates, memory layouts and inefficient branches, I’ve developed a sense of when something may be doing more work than required. This instinct is constantly firing when I see people using LLMs and agentic flows for work. I’ve noticed inefficiencies both in my own workflows, and in other people’s (often technical people).

Some of this can be user error, a lot of it is just a nascent and rapidly evolving technology.

That said, this is not really user error. A lot of people using something like Claude Code or Codex for work now have not really been technical before, and often even the technical people end up using these tools inefficiently.

What irks me specifically

Long winded dumps to LLM bots, where much less input would suffice.
Inefficient use of context windows.
Constant repeated state building (this is a big one).
Using the most powerful models for the very simplest tasks (because we have quota).
Non determinism.

One option to side step the impact of inefficient model use is to make serving these LLMs extremely efficient, where naive usage does not cause much wastage, and while we have made great strides in the pursuit of this (KV cache, batching, model routers, etc) - we’re not quite there yet. LLMs are expensive to serve and to run.

Another option in the meantime is that we build simple tools that help both the human and the model out within the pursuit of doing work.

What do I love?

Reused work (yum)
Caches
Small local LLMs made to do more than they could in a vacuum with robust, SIMPLE, tools
Determinism
Did I mention determinism?

So what is this place?

This is where I’ll write about those simple tools as I build them, and about the thinking that leads to them. Some posts will be about the tools themselves, some will be about the habits they’re meant to fix (mine included).

You can find the things I’ve built so far on the building page. If any of this resonates, or you think I’ve got it completely wrong, I’d love to hear from you.