Chatbots are social media of work. In Meta’s platforms you can pretend to be social and chase meaningless dopamines hits that appear similar we get from genuine interaction, whereas with AI agents you get dopamine hits from similar to doing good work and achieving your goals, while actually you just fried your brain doing making money for huge corporations. Similarly both you can use for genuinely good things, but one has to be extremely disciplined to do it and conscious of the trade off.

I think the answer is simply to not use LLM’s to generate much anything at all. When writing code I only use Claude chat (in separate virtual desktop on a browser) only when I can’t grok the documentation or the bug really kicks my ass. I rarely want it to even write the code, just to explain what I am doing wrong.

When I write the initial idea might be just me having a discussion with Claude (“What exactly was Marcia Williams’ hold on British PM Harold Wilson”) about a topic I am interested in and want a quick overview of the literature, but if I end up writing about it none of it is generated.

Claude just helps me to refine my thinking like a rubber duck that has in its palmate tips most all of information saved online. It is simply an extension of my intellect. The thinking and the work remains my own.