the secret is that if it's a bad idea but funny i will almost always do it
Como todos los hombres de Babilonia, he sido procónsul; como todos, esclavo; también he conocido la omnipotencia, el oprobio, las cárceles. very sane ai newsletter: verysane.ai
4,185 followers 6,884 following 40,890 posts
view profile on Bluesky
the secret is that if it's a bad idea but funny i will almost always do it
if you can point me to a source i will cherish the screenshot forever
it helps if you turn off new follow notifications lmao
also thanks i forgot you aria you're on it now too
i can actually remove you but i assume discovery is part of the point
the top two are in my threads and don't count, absolutely disqualifying. five full days???
if i talk to you and you're not on the list know that this was completely arbitrary and makes no sense
i am just going to constrain myself to people i think are net positives and clearly ai-themed which lets me be completely arbitrary
maija can ask to be removed but i am just assuming social media default is more interaction and not less
i should remove him from the pack it was just too funny not to do
yes, but also they are occasionally correct when no one else is. this is an infuriating dilemma. it would be less difficult if they had not become a cult
fuck it why not
lol. actual safety work, ie, something that 90% of rationalists would not consider safety work
they are also wrong about roughly everything so maybe i should expect the refutation to also seem completely wrong
*all the security doors, two posts back
under *what circumstances* would it be safe does this question in any way relate to whether it is omnipotent, and if so, how
is this safe.
i am going to hook a neural network up to a hospital. it will control every respirator and every IV drip in the place. it also has access to all of the security, and the fire control system. and all the security cameras. also it's in an active war zone so i have a mortar on the roof, which it uses
okay: and that is completely beside the point
lmao thank you i was trying to add two of those (i will not say which two) and blanking on their names
"what the thing is likely to do" and "how capable is the thing" are not related to each other, they're different questions, "the thing is omnipotent" is not a predicate for anything other than the fact that it makes the work important
I don't believe in AI consciousness, but from the discussions about Phil of Mind that go on around here whenever it gets brought up, it seems like Marxists now oppose Materialism?
the omnipotence is entirely beside the point
the overlap of "people who are somewhat AI people" and "people i remember are on here off the top of my head" is it turns out not huge
going to eat dinner, who belongs on this list. i haven't seen a good ai starter pack in forever go.bsky.app/2N4PAzz
i decided that getting into this trivial an argument with two people i sometimes talk to was annoying enough to be transitively blockable
having read this now i think every single thing in it is wrong
i think we were sort of scoffy right up until release day. i was medium impressed by gpt-3 but in retrospect not as much as i should have been
i admire your optimism
ironically it does not matter because essentially everyone leveling major criticism has always taken a "tell people to not do this at all" approach and been ignored completely
openai coopted this was by saying "we formed a non-profit that says safety, and you can trust us" and there was no pre-existing work about who you should trust to make momentous decisions handling the god machine, so nobody could say "you are clearly not trustworthy" by any real criteria
for ai specifically? figure out how to make the things better. which technically they did do sort of but they wanted it to be pure math which is the entirely wrong field. put some effort, any effort at all, into the concept of governance and what good governance looks like, because the way that
well this does help me to understand why Team Warning is still a-warning away even though it is very clearly not working and not going to work
going to guess that you have to have something wrong with you to look that specific horror in the eye, that forms a heckuva pattern
for like. more than a decade? you're gonna just keep saying it while you watch them slow-walk towards the stove while making eye contact with it
the fact that the only person who articulated this clearly was pro-crushing is one of the strange paradoxes. also everyone ignored him basically
i mean the thing that has been apparent for at least a decade and that probably should have been obvious from first principles is that yelling "this thing that nobody has done yet is very dangerous" is never going to work
i'm definitely medium-long run positive (as in, that it will work, not in that it will be good for society) and i have been watching people make whiffed predictions long enough to see a good number of them come and go. and then also nobody predicted gpt-4 really
we are terrible at predicting progress in the field. just like nothing but whiffs the entire way
2001 was written with the consultation of world-class experts about how good ai would be in 2001. lmao
yes the fact that this is a completely failed prophesy should prompt re-calibration but mostly has not
a meaningful weapon leveraging current tools would start at DJI and work upwards, they are starting at mil-whatever and working downwards. it seems fundamentally incorrect to me. but maybe they'll have a product anyone wants to use some day
honestly they mostly seem to be making some fucked up hybrid thing that is not meaningfully better than DJI and that splits the difference on price
yes. this is also not my construction but i cannot remember who else described him this way. it is just literally true. during the 90s he's some strange anti-Deleuze and post-dark-enlightenment he's just some nrx dude
basically nick land went from esoterically and apocalyptically right-wing to "reads the daily mail" right wing when he got older and stopped doing drugs, it's not that interesting
it is sort of reinventing nick land from first principles, though
he also coined the term "hyper-racism" and thinks it's a good thing! truly he contains multitudes basically imagine if eliezer yudkowsky read Deleuze and developed a fetish for the apocalypse, and also capitalism
they are also very resistant to being told "you seem to have arrived at a timeline that is exactly as long as openai's runway, are you sure nothing about openai nudged you in that direction". it is very weird
like eg my ai 2027 piece is basically that it's openai's business plan. and that's true. i think if openai did not exist the shape of the predictions in that document would not exist. openai is paying people to make these things seem true but that doesn't mean those people don't believe it
i am positive a bunch of them are very serious about that but incapable of meaningfully connecting it to any useful actions in the world
a bunch of core people are, a bunch of them are not. they have had a major influx of careerist weirdos for years
no, i don't actually see them as human either. but the weird errors and misapprehensions i see people making and am confused by perpetually are all explained by figuring that people cannot stop seeing them as more or less human
the notion that you'd 1) cause the founding of deepmind, 2) not make any real money on it, and 3) think that AI was very scary and nobody should make it ever is insane. this is a completely nonsensical set of things to do together.
yeah. just, also, in my view a lot, like a looooot, of this is otherwise oblivious people carrying water for basically purely financial interests while being oblivious about it
this is in spite of a certain 90s author who had an amphetamine-powered deleuzian body of work about how "technocapital" was inevitably going to devour us all
you may notice that the rationalists hyped how dangerous AI was for many years without apparently thinking very much of the fact that they were also asserting it was valuable
not like. by asking it. presumably by reverse-engineering it or something
yeah that's a fair qualifier. all the computational/logical work seems like it's basically gone though, if your theory doesn't sit nicely inside an llm the llm will be able to answer any empirical question you can raise better than you can
yep. oh welchlabs on youtube also good. anyway once you know How To Neural Network, or if you decide to skip that, you can probably figure out how to use embeddings (we turn input into a vector, and the vectors are meaningful) and the sorts of nonsense people do with them. glhf
yeah there's like. a whole cult or three doing that 24/7
this is one of the reasons many of us have gone slightly insane yes. i set up a resources page (the first one is 3blue1brown lol): www.verysane.ai/p/learning-ai once you get past basics you can read books or you can just go dig into specific subjects and come back to reference stuff tho.
the entire legal system agrees with this, so we're probably fine there. i think you'd have to assign copyright to a human and claim they had some effort in because human effort defines copyrightability
like. unironically. NLP as it previously existed is completely gone.
linguists are either 1) ai people now or 2) building a vast fortress of sophistry from which they can claim that this is all meaningless
yeah, at a workflow level (ie from the process end) they are totally different, but on the other end it's the same gizmo basically. that brush is probably doing a reallllly small diffusion model pass over a small area.
@theophite.bsky.social is probably the cleanest example of "someone who is doing something fantastically bizarre to stable diffusion to get specific results that he likes" (the fact that he is here is a bonus) and it still doesn't resemble a normal artist workflow at all
there's the opposite end where serious power users for image models are doing some incredibly convoluted and specific shit that is no longer at all just pulling a lever, there's still a wiiiide gap in the middle between "the model is a tool" vs "the model is the primary medium" tho
it's a high dimensional space where distance corresponds to meaning and nobody really knows what that means
so if you work with them, they are remarkably good at giving correct answers, way more than they should be from either internet pretraining corpus or just regurgitating which answers were marked "correct" in the finetuning corpus. but they also fail catastrophically in strange ways
yes, the finetuning process will have strictly separated things-openai-wants-the-model-to-say from things-openai-does-not-want-the-model-to-say on a smallish curated corpus, and this generalizes reasonably well but not at all perfectly, so it is in fact "trying to say true things" in some way
there's a separate piece of software for this actually. it is very entertaining
i realized suddenly that a b2b SaaS that had a meaningful moat and actually good coding automation would be the ultimate centralization of software engineering as a field, undoing all the progress since IBM lost the crown anyway i'm gonna go do something else now
the google search integration specifically is an abomination that should not exist, other search integrations are much less bad
lmao yes this seems accurate for what people are doing. i convinced i was right, it must be true
this is accurate
the llm is unique compared to previous techniques in that the concept of truth is in there somewhere and you can actually sort of use it as an optimization target. but this is impossible to explain to sane humans. and it is also not at all possible to optimize for it perfectly
mathematically they are very similar and require similar skillsets! as a description of the research area this is a very sane grouping. however,
“When fundraising it’s AI, when hiring it’s ML, and when implementing it’s logistic regression” was a current joke in good standing like 4 years ago
i prefer demon as the most accurate description of their role in the world.
this is why i don't feel too loopy by mentally classifying them as some sort of demon or empty vessel that speaks. there's not really a lot of good options immediately available
their relationship to truth is weirder to that but that's a lot better than assuming they are truth-machines
and also, a lot of people think you mean an LLM any time you say "AI", even though everyone is rebranding all of ML and data science as "AI" now.
its mistakes are not like humans though, they are way worse and often more surprising. so no matter how anodyne what you are doing is, you have to double-check if being correct really matters
if you're aware that it's just giving you a first-pass and you have to check everything you're a power user. congrats. you are exonerated
basically OpenAI made the term "AI" cool, supplanting "ML", which was previously how you distinguished your approach as being statistically driven and not useless (which is what AI was not), and now anything that has any sort of complex math in it is "AI" when it was "ML" or "data science" before
amusingly literally the same people invented both things
it's funny because depending on if you mean "ai-powered as in terminator" the answer is effectively never, and if you mean "ai-powered as in capable of doing automated target acquisition" we already have that because yolo is in people's pipelines
i ... yeah. idk. i get the impression that this is not far off from "human whom you think is smart and trust a lot", it's only a little askew, whereas what the thing is is WAY askew from that
god that's so long
yeah. which, amusingly: this is power user behavior. you understand and know that the thing can be made to change its tune. but then you fall right back into using it wrong by taking the bot agreeing with you when poked as a meaningful thing
roughly same
tragically we released them to the public so i think we are just going to get a big natural experiment on whether people settle into trusting them a reasonable amount eventually
to consumer? maybe. idk. i could see other cases but that's the only one that's really mature. for b2b code automation is 1000% the leading contender for big money use though
yes that is the conclusion i am leaning towards. or it needs to be a very strictly specialist tool. or people need to acclimate a LOT to the thing before it can be used safely
i still have lasa blocked lmao
shit. this is getting the longform treatment. i don't wanna do that but fine
i decided this deserved a main timeline post because it does unify a lot of previously bizarre responses to LLMs into one theory bsky.app/profile/segy...
agreed
i suspect, unfortunately, that she is correct and most harm will not be legible. it is likely to require someone to do and publish a good expert review and sort of shame either the industry into adding more warnings or some regulator into making them