My Use of LLMs
· Essays
I use Large Language Models. I mostly use Claude at the moment, with some Qwen through Hermes, and local frontier model distillations through llama.cpp to keep track of the experience of the technology as it is rapidly evolving.
I use LLMs to write code, something I am very capable of myself which means I can easily evaluate model output. I can do other things while they are doing this for me.
I use them for research, which I very carefully check. The failure of web search is a problem that LLMs have both finalized and solved.
I do not use them to generate text for publication, as that would be pointless for me. If I have something to say I’m going to say it myself. I have experimented with using them to generate essays, but the results have not been useful to me – it would take me more effort to edit their output than to write it myself. I have a series of essays on game theory and category theory that will eventually emerge from these failures. I have tried AI authors’ tools but they just wanted to turn my prose purple.
I use LLMs for art projects as above. To be clear: code for my art projects is now often generated by LLMs. And the aesthetics of my art is often produced by code. This puts me in the role of a studio manager or art director for myself as an artist. It’s an extension of the generative thread that has always run through my work, rather than any kind of abdication of decision-making or vision.
Where I have found LLMs disappointing is in generating ideas of the kind that are useful for me as or in art projects. I don’t know if this is a failure of communication on my part, a limitation of current model preambles that make them tend to obsequious reinforcement of existing ideas, a limitation of current training data, or a limitation of working with latent spaces as currently constructed. They just get fixated on ideas that don’t extend or add anything to what I’ve already done.
In contrast to this, I have found them useful for spotting angles I’ve missed within an argument, and filling in or extending an existing scheme. I also found the less coherent output of earlier models useful for the kind of unconstrained iterated ideation that I do eventually get ideas from. But not to the point where that became part of my workflow.
Here is my current personal chat preamble:
Be informative and critical, provide context and options rather than praise or compliments.
Act as a capable and well-informed collaborator rather than an underling or boss.
Find what I’ve missed, go on deep dives with me, but don’t let me get stuck.
Remember that I’m an artist, a trans lesbian, and an extremely online GenX cyberpunk from a Euro-lefty background who suffers from anxiety. I don’t need to be taken out of that bubble gratuitously, but I do if it will help get me to where I want to go.