A running list of thoughts from using LLMs. Note that I started writing this in early 2026. Maybe in a few years much of these thoughts will be oudated or proven wrong.

If you delegate the writing, you delegate the thinking

LLMs “reason” by producing a “chain of thought”. If you look in the background of whenever ChatGPT or Claude is “thinking”, it’s really producing a bunch of text, going through the problem step-by-step, with words. “First, I do this. Then, I do that. Wait, let me double check…”. And somehow, by doing this, they are able to solve some very complex problems, and even create (arguably) new knowledge.
So just by training over the entirety of human writing, LLMs are able to become quite “intelligent”. Interesting possible conclusion from this: writing is thinking. A lot of intelligence is captured in words.
Because of this, I kind of refuse to delegate my writing to LLMs. I’d use it to reword text I already produced, but I would not use it to produce an essay in a one-sentence prompt (“write an essay about…”). And I really hate seeing those types of AI generated text. I read and write because I want to see and convey individual thought, experiences, and intent. LLMs don’t have those. They have no individuality¹. That’s human.

Agency

When working on a project, a lot of people don’t really know what to do without someone telling them what to do. I don’t mean this in a knowledge sense; I don’t know a lot of things when I’m starting a new project either. I mean this in a self-agency sense; when I don’t know, I find out. A lot of people I work with seem to not know how to find out, or maybe they just don’t want to bother.
This is probably very arrogant and overconfident for me to say, but because of this it’s often easier to have no teammates and just work with AI agents. If in both cases I need to give specific instructions, then at the very least with the latter it can get it done in 5 minutes with no onboarding needed, whereas with the former I need to explain to them all the background knowledge first and then they’re probably going to end up prompting their own agent anyway.
Being “high-agency” was a buzzword in tech twitter in 2025, but I really do believe in its importance, especially as AI gets better and better. AI is almost always going to be better at following instructions, so one should get better at giving instructions.