|
Silver Alicorn posted:AI is worthless if you can't explain how it works not sure this is quite the correct standard. a thing like chatgpt inference is on one level extremely trivial in what it does; there's a set sequence of of roughly half a trillion arithmetic operations applied to the input, you are free to ask about any one of them. rather one would need tools to tell it to do things and and a reasonable assurance that it'll do something like the instructions, and not something else. e: one might of course want an explanation of where all the constants come from, but i've for practical reasons taken to acting like the models are things that fell from space and all we can do is analyze them. to some extent cynical, since economic and other practical constraints make large models work out that way, and likely will for the foreseeable future. Cybernetic Vermin fucked around with this message at 14:53 on Jan 9, 2023 |
![]() |
|
![]()
|
# ¿ Sep 1, 2024 00:51 |
|
just make Goodbye Lenin As A Service. squeeze that last buck from the boomers.
|
![]() |
|
rotor posted:like no its not replacing us today but if you can't see it on the horizon you need glasses total time spent writing code for a software developer is still quite low, so it is not necessarily much of a problem. chatted with one of the copilot engineers at a thing couple of months back, and they made some pretty good points about it. the most compelling that obviously for testing it is good if it can be derived from the specification entirely independently from the person writing the "real" code. i.e. let a future copilot just spew a billion fairly readable tests for your code based on comments and documentation. even if they are bad/wrong that will still be a something like a double-check for free. also automation should be good, and it hitting a well-off segment of society is very good as it makes it more likely that society will change to make something of this good. Cybernetic Vermin fucked around with this message at 21:25 on Jan 9, 2023 |
![]() |
|
SporkOfTruth posted:at least 3 things in this sentence are not only unverifiable, but also a wild hazard to anyone or anything that depends on that hypothetical code there is literally no way they can be worse than human-produced tests is the thing
|
![]() |
|
SporkOfTruth posted:mm, the very common "holders of capital will invest it in the thing for the common weal" from the history of never it is a nice place for it though, if the plan was to roll back automation as the vehicle for ending capitalism we should probably demechanize agriculture. that sounds loving horrible though, capitalism needs to be fixed in a way more fundamental way than trying to preserve jobs under it, so i'm pretty ok with accepting the accelerationism of white pmc people getting automated. that said i don't foresee ai as it stands doing much other than, as noted, something like guided fuzzing. Cybernetic Vermin fucked around with this message at 23:12 on Jan 9, 2023 |
![]() |
|
don't know much about law, but suspect that this is the same sort of deal as in programming where people automatically think about the 2% of the time when one is writing actually important core logic/algorithm code, which wont get automated because it is both the interesting and most important part. as opposed to the other 98% which is dreary repetitive stuff which takes all the longer and has more errors because it is dreary and repetitive. that is, the actual oral arguments being the actual human interaction core part of the proceedings i doubt there's any push for adding a bunch of ai nonsense to it. if i understand correctly though there's a ton of searching and collating documents, producing exhaustive filings, and summarizing up the relevant information in preparation involved, which all seems more amenable to ai (good automatic summarization and semantic search both seeming pretty plausible with current tech) and the kind of thing humans are pretty bad at.
|
![]() |
|
rotor posted:idk poo poo about the field but I suspect liability will be a serious issue here and the first couple times an AI produces an inaccurate summary or makes up case law out of whole cloth there will be some fireworks. no clue about the actual legal bits either, *but* with some limited knowledge about the language models i do suspect summarization and retrieval are places where the "making poo poo up" and wild inaccuracies can be tamed without some great leap forward. i.e. a matter of cutting text down rather than proper generation, making it possible to trace what in the original documents it refers to. for a start mostly giving an indication of where to look, no doubt missing some stuff, but again one where i suspect it'd brute-force do better than a human pretty quick.
|
![]() |
|
i increasingly suspect that a lot of what is happening with chatgpt is a good old eliza effect. with a bunch of gpt-generated text sure, but i generally suspect that there's a bunch of hard logic layered on top which crudely parses a couple of classes of questions (e.g. "but in the style of x", "explain y", "that's not correct, should be z") tied up with rigid ways of (re-)querying the model. which is significant mostly because that part is back in the land of the fragile and labor-intensive, not in itself representing very real progress. though it does improve on the past gpt demos by *not* just randomly generating garbage.
|
![]() |
|
jemand posted:I thought you were going to go for the Eliza effect of humans attributing more capability to systems that interact via chat than they're actually showing. That, or just anthropomorphizing the "reasoning" behind the outputs as human like when it's not. that's roughly what i do mean; specifically that the main new thing impressing people here is actually a few surface-level mechanical rules. the actual text and general information recall being a previous known, but chatgpt seems way slicker but maybe not for a very deep or sophisticated reason. jemand posted:Also, I've had some difficulty in either finding or believing the size on disk the gpt3+ model versions take. Latest i have seen is 800GB, which is actually larger than it takes to store the entirety of some versions of the massive common crawl dataset. I do wonder what fraction of the observed performance would have been possible by efficiently organizing this data for search and layering on top of it some summarization, text mixing/recombining, and style/format translation capabilities. Functionally, with a less than 3-1 compression ratio of training tokens to model weights and the known capability of these models to memorize training elements, this may very well be what is actually happening, just obfuscated within the mass of opaque model weights. 175 billion parameters in fp32 for 700 GB, plus some more surrounding stuff rounds it up. that kind of tells part of the story of what is happening though, as each fp32 parameter on average likely encodes fairly little information. experimentally as models get larger they tend to be more amenable to chopping off more bits on the parameters (into bfloat16 habitually, but int8 also showing up a lot). to some extent they are no doubt memorizing a lot, but a lot of subtle stuff clearly winds up being represented still. compare to rule-based text models it is already significant enough that llm's have a pretty clear abstract representation of a lot of grammatical rules; e.g. you can prompt one with made-up words and it'll use it in a way that makes a fair bit of sense both grammatically and with some world knowledge. the hype making that not seem so impressive, but it was so far out of reach 5 year ago.
|
![]() |
|
eschaton posted:I get the impression this has been a problem in the translation community for a while now, nobody wants to pay for a skilled translator when a lovely neural net Chinese Room regurgitation looks like it gets enough of the point across enough of the time and why spend a penny more than absolutely necessary, ever? at the same time it is hard to view automatic translation as a negative, as i think it is a very modest estimate that there's 10x more communication happening between people not sharing a common language today than there were before there was decent automation on that. i remain forever unsure on automation in general, it has all the potential of being a great good, and trying to *prevent* it will not do anything to fix the problems of capitalism. we are certainly not in a state currently where scrambling to save it from a change seems very worth it. luckily there's a lot of time to consider yet, as quite few of the things imagined imminently automated will be.
|
![]() |
|
for images they do do that already, there's a watermark in images from stable diffusion and similar which contains just the single bit "this was made by ai" to weed those out of future sets. pretty tricky for text though. i generally expect though that the internet getting flooded with (even deeper) garbage will be a bigger problem for people trying to use the internet than it will be for future training of models. even now you can just run all training text through gpt-3 and discard anything with a truly minimal perplexity and get a pretty strong filter. even when it rejects something not ai-made it will happen to be a piece of text that doesn't really add to the whole, so there's no downside.
|
![]() |
|
Geebs posted:Who's to say all of you aren't just language models created to post for my own amusement you're not getting a lot for your money if that's the case
|
![]() |
|
Pollyanna posted:legit q what exactly are chandler and el goog trying to make use of chatgpt/lambda for??? i don't find it very useful, but reports suggest that googles attempts at answering questions directly above the search results see a lot of "engagement", so, idk, if people like that i do guess chatgpt is that but more bigger.
|
![]() |
|
it gets funnier the more one thinks about it too, because google for sure can't retaliate and point out errors that bing makes. it is not even that some blatant errors are unexpected for most people i don't think, it is just that they put the error in the marketing materials. good clean funy computer
|
![]() |
|
in the interest of saving the cyberpunk thread from this, without just shutting down conversation, i will state the entirely unironic and honest belief: if natural language processing is a thing that can possibly be defined distinct from general artificial intelligence, that has probably now happened. which is probably not a very interesting statement to most people.
|
![]() |
|
that it is a perfectly valid position that nlp short of gai is impossible. it used to be a very common, possibly majority, opinion, but it has eroded a bit of late. i honestly can't argue about it, while i believe we've now found a meaningful tier short of gai, it is not like it is objectively established.
|
![]() |
|
distortion park posted:slack does autotranscripts of some videos now and it's very bad i expect this will be real bad too. summarization is increasingly not terrible, but that's running on like coherent edited texts, doubt it'll understand or prioritize things sanely in a poorly transcribed freeform conversation. has a lot of room to be hilarious though.
|
![]() |
|
stuff like translation and summarization is a lot easier to keep on topic and task though, basically only concerned with expressing things already in the text. free-form conversation has a bunch of things against it in that regard (e.g. humans will try the limits of the thing, it is actually *tuned* to drift away from stuff (or it'd just relentlessly repeat itself and what you say), and generally ingesting text it has generated leaves a lot of room for issues to reinforce themselves) but, again, i highly doubt it'll be very good at telling the point of the meeting from all other things that go into human communication. many meeting summaries starting with the point "why isn't loving audio working in teams again?!".
|
![]() |
|
bing so loving serious all the time ![]() quote:For example, let’s say your monthly income is $2,000. According to the 50/30/20 rule, you should spend no more than $1,000 on needs, $600 on wants and $400 on savings and debt repayment. If you are currently spending $3,600 on candles, that’s way beyond your budget for wants. You need to cut back on that expense and allocate more money to your needs and savings.
|
![]() |
|
Silver Alicorn posted:just coming in to say that stable diffusion is ugly and bad and you are all fools for thinking it is in any way impressive. go look at a real painting in drawing this parallel between stable diffusion and legitimate art, implying they are in the same space and fair game to compare, you're putting way more significance into ai models than most people do.
|
![]() |
|
on a related (though not ai) topic i feel sometimes like one huge visual development of the last decade has gone by a little bit unnoticed: how every visual medium is now packed with gorgeous sweeping drone shots of stuff like cities, landmarks, and stunning nature. one would think it'd get tired, but it remains delightful.
|
![]() |
|
another strange ai take, by hinton in the nyt, worrying that irresponsible availability of ai tools will flood the world in generated images and text (no doubt true) and that will make it impossible to tell what is true. where before the normal way of understanding what is reality was to read sufficiently much of the internet and whatever there was the most mass of was the truth.
|
![]() |
|
continuing on the theme of people warning about llms and similar while ridiculously hyping their abilities, right in nature: https://www.nature.com/articles/s42254-023-00581-4 specfically the quote "As we rush to deploy LLMs into scientific practices, it is important to remember that science is a human enterprise and LLMs are tools" which only makes me ask "who the gently caress is 'we' here?" like, the point of the entire article is ethics, careful consideration, and nuance, but it *starts* from the premise that it is a legitimate idea that you'll do science with llm's (as distinct from treating them as an object of science, which is fine and good).
|
![]() |
|
echinopsis posted:because it’s inevitable ?? lmao what is?
|
![]() |
|
echinopsis posted:doing science with ai doing science using llm's is what is being talked about, and it is not inevitable, in that i struggle to come up with a single scientific question not asked *about* llm's that would be well answered using an llm. but with all this inevitability going on i'm sure you can propose something
|
![]() |
|
Beeftweeter posted:i think a lot of the stories we've seen over the past couple of months prove that even if it's wildly inappropriate for the task people will try and use a LLM to do whatever they're trying to do otoh a lot of people seek fundamental truths about the world using seances and ouija boards as well, but we don't view *that* as an important debate about ethics and best practices in science, it simply isn't science. again part of what makes the nature article infuriating, that tone of "obviously we'll llm up science any moment now, but have you considered this ethical nuance?". it is not at all obvious.
|
![]() |
|
Beeftweeter posted:lol that's actually not too bad. i'm kinda wondering with the WGA strike going on now if we're gonna see a bunch of studios start using LLM output for scripts nah. good writers are dirt cheap. there's excellent people out there who'd do anything to break into that gig. great writers are expensive, but if you're shopping for great you've long left the realm of ai garbage
|
![]() |
|
for those not keeping track btw i am easily yospos biggest llm fan (partially because i am paid to be). just been a few pages of nonsense hype that's easy to slam.
|
![]() |
|
relentless hype about ai more specifically.
|
![]() |
|
ibm is just firing people while using overblown ai hype to play pretend with shareholders about it.
|
![]() |
|
Hed posted:From Wsj extremely unexpected and yet very dumb, i will necro beeftweeters comment from months back (entirely correct, i just think the "might cause" was overly optimistic) Beeftweeter posted:i mean, if anything, google just doesn't have to do anything at all to appear more competent than microsoft right now. the bing ai failures are extremely public, they've tried to defend it, nerf it, tweak it — but it's still quite bad Cybernetic Vermin fucked around with this message at 12:43 on May 8, 2023 |
![]() |
|
![]()
|
# ¿ Sep 1, 2024 00:51 |
|
MononcQc posted:everyone is aware that google's search results have gotten worse over time, but the thing they're trying to save here isn't search, it's the ad profit, and it very much feels like the latter was already harming the former for a while. i mean, the only thing that as a corporation makes sense to save is ads since that's their entire business. the new here is that google had a solid 20 years of basically understanding that they couldn't gently caress too much with what search does like, any average product manager would have fallen into the trap of going "how do we drive engagement with the page", making you stay in search and click on things and whatnot, so google has been nicely conservative not messing with it. but now here's bing chat to spook them into doing something extremely dumb (at least my suspicion is that this started with bing chat and spiraled into the "what do kids like these days" engagement thing).
|
![]() |