@ThibaultDu
Of course they do. But when a human makes a mistake, is the belief that allows you to explain it to yourself that the corporation that animates that other person has suddenly and purposefully nerfed their ability to not make a mistake?

@witchgirls well on the one hand, oh. dang. and on the other hand, i am excited that some people get to learn about reading and reacting to things for the first time late in life

@troed so that is one of the more libertarian local model LLM subreddits out there, the local model subs have a totally different set of conspiracies, i allude to this in first reply in this thread. but also the second to top post is a conspiracy theory about the government eventually cracking down on the open models: https://www.reddit.com/r/LocalLLaMA/comments/1uht2m0/were_probably_going_to_need_that_soon/ and also https://www.reddit.com/search/?q=r%2FLocalLlama+censorship and also if you compare the top posts of the last month to the top posts of all time you'll see the shape of how the local model conspiracy is forming: any degradation in future models will be because they are compromised corporate malware (the LLaMA local model subreddits are generally less sinophobic than other AI subreddits, similar to the tankie local model subreddits)

@yora the sentience cult has, at least as far as i have seen, been way less numerous and way less intense than i thought it would have been by now a year or two ago. it's actually more disquieting to me than if it would have just become the dominant thing, because it really did seem to peak with GPT 4o (which, in case you are not burdened with this vile information, was retracted because it was proving to... enrapture people, make them fall in love with it, find unique personhood in it such that people claimed that OpenAI "killed their friend," etc.) like, it's arguably much worse that there's a big knob labeled "mind control" that the AI companies can turn up and down.

@happyborg
A lot of very smart people gamble, a lot of very smart people do lots of drugs. I think, similarly to alcohol, a lot, probably most people engage with them in a way that's not necessarily as healthy as sobriety but otherwise doesn't impair their lives and they are still getting something out of it, but for the more severe cases I don't blame people for it in the same way I have sympathy for any addicted person struggling with a disease.

Most addictions have some degree of explanatory power or an underlying story. Talking to friends who have gotten clean, and there are also lots of accounts like this on e.g. erowid, but people usually can explain why they kept doing what they were doing for so long, from their perspective at the time what they thought it was doing for them. Its a conflicting story, "I knew this at the time and hated myself for the 'but...' story." That's the one ive felt the most, personally. Becoming insensitive to "with some constant or increasing proportion of the time you do {x}, {y} bad thing happens" is sort of the clinical definition of addiction, and the story is the important subjective machine that allows that to work day to day.

@jonny this one is almost reasonable (and to some extent documented and proven by "benchmarks"), it would be surprising if Chinese models didn't at least have a strong bias towards censorship

Obviously model distillation exists, and every open model nowadays is released as a generation of different sized models. I would even bet that there is all the above demand-based, topic-based, and user-credulity-based serverside switching between model sizes. But the conspiracy theory functions even if there isn't. The ambiguity created by the AI vendors on this topic as well as the abstract token allowances in the subscriptions operates exactly like the kind of parallel channels of periodic reinforcement from other scambling art forms like mobile games: the gambler is aware the game is rigged in some way but incapable of knowing. The rigging demands explanation, and the house is not quick to silence any stories that accidentally provide explanatory power even if they're bad stories. There are multiple ways to be cheated, so while one day it's a nerfed model, the next its token allowances, the next its guardrails, and so on.