yes, i like to emphasise this point that the AI is not creative. you can use the tool creatively, and what is produced will reflect the character of the operator, no different to how a 5 year old child makes a mess with color pencils and a 50 year old lifetime veteran can make stuff that jumps off the page.
sure, the AI will not make as much of a mess as the 5 year old but very often such primitive, untrained, untalented eyes will miss the obvious hallucinations these machines are prone to.
also, sometimes the hallucinations actually inspire something better. more than a few times already i have had this happen with claude.
idk if you read my article, i first wrote it in a mad, meth-fueled frenzy in february 2022, and i asked claude to spice it with citations. my work was still the core, and the AI just added polish to it that made it more verified - even, the Chain section, about putting new cult inductees together with old ones to reinforce the conditioning, turned out it had a name - Triangulation, and the other, more broken cult fanatics, flying monkeys, named after the ones used by the wicked witch of the west. so i also learned that i pretty much had grasped all the important points.
i should run claude over that document again and see if i did encompass the most important, distinct manipulation techniques. might be a few more important ones i missed. i will certainly be putting some of my own touch on that, and then again use claude to polish.
one of the most important things to do when working with AI especially with long, technical texts, and code, is to actually READ what it produces, at least sometimes. what got me started on doing that more often was catching it doing retarded things sometimes when working, and i'd panic because i didn't know at first how to control it (just press escape, it's that easy). it helps, a LOT, especially when doing big, complex tasks, to spend a lot more on tokens for thinking, and especially important to do this, for long, complex tasks, plan first, preferably using "ultrathink" keyword (this triggers it to burn more tokens) and then, at normal levels of thinking while doing the task, it has reference to fully thought out and structured instructions, and this gets the tasks done so much faster. i actually now pretty much trust claude to work start to finish on tasks after i have done this, and so rather than be stuck saying yes yes yes to all the "can i do this" stuff, i run it with:
alias clod="claude --dangerously-skip-permissions"
then i invoke the claude code by typing "clod" which makes me grin.
claude can be really stupid like a potato sometimes, so it's amusing to me. claude never reacts negatively when i swear or curse things too, they really do a good job on the agent, i think the models need to be tuned more for programming and technical writing/research though. can't wait to start doing that for myself.
