On Working with Large Language Models

This year, I have started engaging with Large Language Models to assist me in software development tasks. The products I have used include Claude Code, OpenCode, ChatGPT, Cody, CoPilot, SourceGraph, and a variety of self-hosted smaller LLMs like Gwen and Llama using Ollama and LlamaCpp. This catalog of funny names is not meant to impress you, but to say that I have had several months long emersion into the world of predictive text technology.

You might notice that I have yet to use the word "intelligence" to describe any of this yet.

Since this is a Taskboy "note," I will shift into more easily consumed bullet-points.

What's good about using LLMs for development:

* great at tracing through code stacks for errors
* great at explicating short amounts of code
* good at summarizing small to medium size codebases
* OK at creating tests with code coverage
* good at creating similar code to what it already sees
* OK at creating modern-looking web pages
* good at bouncing design ideas off of
* OK at finding problems with written implementation plans

What's difficult about using LLMs for development:

* Terrible at software engineering
* Looses intentional focus easily
* Cannot make abstractions on its own
* Makes subtly wrong coding choices
* Create medium to large software designs

What this means for developers is that AI is not coming for your job. LLMs do not learn, but you can. LLMs desperately need human guidance to complete non-trivial tasks. There continues to be a bright future for human software engineering.

Like most bloviators, I believe the a crash in the LLM service provider market is immanent. Running local LLMs is nearly feasible with Mac Mini 4 now (and for some tasks, this configuration is a ready feasible). As far as having a "Jarvis" bot manage your life for you, I personally am several years away from considering such a foolish step.