>>7116
people don't like llms for a few reasons

1: hallucinations
even the really fucking large Claude models hallucinate shit, let alone free-tier chatgpt
I do wonder if some of the bigger issue is just token selection at the end lmao; the model clearly does the work thinking, selects the wrong token out of chance (token selection literally involves an RNG to keep the model from being stuck in a loop), then keeps selecting high probability tokens after that aren't based in any fact, it just has to keep saying something after due to having already said something. Almost all dumb LLM errors I've seen follow a similar pattern.

2: environmental issues
some of the issues are overblown, but the real issues are that energy and water are being diverted away from other users for data center build out that isn't even in the ballpark of profitability (please remember that OpenAI and Anthropic are selling access to their high end models at significantly below cost)

3: creatives
LLM writing kinda really sucks, and LLMs basically distill the works of others into a heavily compressed pile of weights.
I don't care that much about the latter (I think IP as a concept is gay and AIDS), but goddamn, LLM output is cliched in ways that are uncanny. I personally have to fight this constantly. You can prompt around a lot of it (a surprisingly good idea I saw was to tell the LLM to follow a specific writing style manual, killing off the shitty LLM-speak tone), but most people don't.

4: coding issues
huge models like the high end Claude models are better about this but you have issues with code quality and design quality
models tend to get confused in projects written without someone experienced in the loop, and context size issues become an eternal problem for larger projects as the LLM acts like it has seen 100% of the code and writes in a function that already exists
also, you will have the non-zero (and honestly, surprisingly high) chance of any model you gave shell access to running a command that fucks shit up, directly related to the token selection issue mentioned above regarding hallucinations (eg, a dumb command that breaks things ends up being surprisingly high probability during token selection)
oh, and Claude gets weirdly lazy sometimes lmao

5: someone else's computer
most AI shit is out of your control and sends everything you do with it, fully analyzed by machine minds, directly to companies with massive data centers lol
they can build a profile on you that pre-AI Google and Facebook would find hard to match lmao
it isn't even paranoia, there is a 100% chance that your chat logs are being used for model training at the bare minimum, let alone generating extremely high quality profiles of its users
you might be able to dodge this with care if you don't spend all your time asking the AI questions, but good luck lol
also most people still use Google and thus the point is moot, we live in the era of post-AI Google and they have 100% built a profile on you already