Where I'm at With AI

It's been over three years since I first mentioned LLMs on this website.

At the time we were still laughing at ChatGPT for not knowing how many "R"s there are in "strawberry". Now we're talking about OpenAI accidentally hacking RubyGems and HuggingFace. So things have progressed.

My take from that first LLM piece has basically held during that time; AI remains an extension of the will of a user, and lacks any internal goals of its own which would form the basis of autonomy. (For convenience I'm still going to keep saying "AI" rather than "LLMs" or anything else that might be technically more correct.)

What hasn't held during that time is my own stance towards using AI in my own work. Initially I was fully opposed to any use, on the grounds that I'm generally against new things and I actually like writing code, thank you very much. Within a year or two that stance had shifted towards "I'm interested, but only if I can run it locally". By now I've resigned myself to a light amount of Codex usage in my personal coding projects.

I continue to never use AI for my writing.

My reluctance towards AI is not based entirely on hipsterism (or at least, I'm able to construct a coherent post-hoc justification.) As a world we have been taking progressively larger steps towards this utopia where all jobs are automated away, while still requiring everyone to find a job or starve. How this paradox will resolve is unclear, but it is likely to be painful. AI is yet another step towards this future, and I'm loath to participate.

The other factor repelling me from AI is its widening of the gap between those who can and can't. The standard of what constitutes programming, along with many other disciplines, is changing to only encompass AI-produced, or at least AI-enhanced, work. What I fear will happen is that obtaining a job (scarce as they might become) will require being skilled in using the latest and greatest models and tools. These jobs will go to those few who have access to these models and tools in the first place, and leave everyone else locked out.

This second concern is what drove me to lean towards locally run models. But what has become clear since then is that getting anything done with a local model requires dropping $1k+ on a graphics card and RAM, and doing so out of a concern for accessibilty is kind of moot. I may still do it one day, for fun, but at the moment it's not in the cards (or budget.)

At the same time... does it matter? Does my using or not using AI have any impact on anything? My problems with AI aren't really with AI itself, but with the larger attitude which brought it about, and even if we all successfully boycotted AI to death that attitude would remain. Without a real answer to, or willingness to enage with, the overall direction our world is headed in, it's fairly pointless to be taking hard stances on intermediate steps.

And while this internal debate has gone on I've been working at an AI-maxi company for over a year, and so have been fully exposed to that whole world. We use the latest and greatest models and tools, and I've developed the skills to use them. What I've learned is that with a sober attitude, and some practice, I do get better results than I would on my own. All summed up it doesn't make sense to me to not bring this improvement to my personal work as well. If I retain the hope that my projects are in some small way useful to the world, shouldn't I take all advantages available to see them through?

So for the time being I've accepted a certain amount of AI into my life, even outside of work. I'm mostly using Codex Luna high, which strikes a nice balance of being efficient and, given proper handholding, not completely fucking stupid. I would not describe my personal workflow as "vibecoding", it's more like working with an apprentice; I know what changes need to be made and how, I describe it, the agent does it, I review it, repeat. Where I'm unsure, I still work through the code myself.

------

Published 2026-09-20

This site is a mirror of my gemini capsule. The equivalent gemini page can be found here, and you can learn more about gemini at my 🚀 What is Gemini? page.