Clojurians Log v2

depends on the problem & domain & complexity of the issue you are brains tstorming

It's made me appreciate pen and paper more. I loathe the verbose text it spits out

The verbosity is load-bearing, and worth being honest about.

XDDDD

yeah, it depends but sometimes yeah

if I were to give general rule: super helpful when looking for particular, existing technical solution I know nothing about. I have a technically shaped hole and I need to fill it with suiting solution. I was "writing" my own Vim and needed to handle viewing large files, then Claude was like: yeah, let me write that tiny C function that uses mmap and will give you ability to navigate huge files. huge win here. hours of investigation saved super annoying when it comes to analysing business cases and designing coherent systems. it will always mixup ideas or forget about something or propose solutions that make no fucking sense

the more technical I get the more useful LLMs are, the more analytical I get the more annoying they get

Agreed, but I would say that your first example barely qualifies as brainstorming. You know what you want. I struggle a lot with perfectionism, and end up spending a whole day chatting with an LLM about something that has no right answer (e.g. UI design). Even regarding things that do end up having an optimal answer (e.g. code architecture), it takes a lot of time to end up at the correct answer, and usually it involves trial and error anyway, so all the time spent chatting was mostly useless.

yeah, true

it was made in whole context of building application so I kinda new what I want, kinda wasn't sure, but it was still looking for tech solution more then general brainstorming

the one plus is that when you're in middle of brainstorming and are not sure about some idea you can task LLM to give you some half-assed version of that idea

and then you can kinda check in practice if it works

Honestly, recently it feels like LLMs waste my time on anything. I asked a LLM to generate a code for me. It generated. Every function had a problem - literally. I fixed it manually, and then found that the thing didn't work. I asked the LLM on why, it gaslighted me 4 times until I cursed it and called it names, THEN it gave me the correct answer.

And that has been my experience recently - the prompts, to be correct, need to be very verbose and it kind of defeats the purpose because by the time I ended up writing the prompt, it's easier to just.. you know... write the code, investigate the problem, etc.

I only use an LLM to ask me questions during a brainstorm. That works well for me as it helps me sharpen my initial (often poorly shaped) ideas. I've tried (way in the beginning) to ask the llm to come up with ideas, but abandoned that most of the time. Anything it produces can be found, it does not feel to me it gives me something original

They are engineered to be agreeable and affirming so if you show any bias it will often just reinforce whatever you're leaning toward. I think they can help identify terms and approaches you may be unaware of (like when you're searching the web and you don't yet know the proper term for what you're looking for) but perhaps less useful for landing on the appropriate solution. YMMV, of course.

here's a tinfoil hat theory: LLMs giving incorrect answers is a richer source of training data, because they can absorb more information about how the users correct defects in LLM output

@afoltzm I have a lot of tinfoil hat theories, and some are not as absurd. One of them is that the "brain fog" I felt when I was using it too much it's also on purpose - it does get this "addiction" of "hey, it got me to the goalpost really fast, amazing", dopamine hits, and our brains that are still very dumb because we didn't have time to evolve away from it, fell "ok, all dopamine, no effort!" Probably companies knew that their LLMs could do that; and felt that it's a bug, not a feature