|
Beeftweeter posted:^^ a good post what’s its account name?
|
# ? Jun 1, 2023 16:01 |
|
|
# ? May 30, 2024 13:41 |
|
fart simpson posted:what’s its account name? "fart simpson"
|
# ? Jun 1, 2023 16:05 |
|
The generic case I'm struggling with, is that any situation where you can simply check the LLM's output, or algorithmically augment or guide or "help" it in a completely controlled & reliable way is basically solving all of the actual problem with whatever "extra" you bring to structure or check the LLM, & the LLM is superfluous and not really helpful, except as a point where it could actually ruin the rest of your solution. Any case where you are relying on the flexibility and the power of the LLM's capability, you by definition cannot fully check, and every edge case you do not check, an absurdly large number of completely alien TYPES of errors can slip through. It is a fantastic toy. There are a lot of "play" use cases that are going to be developed with great success. It's great at producing a sort of soul-less cultural echo -- I expect that we'll eventually get some weird and wonderful artistic / linguistic implementations from that angle. An example of what I'm thinking of: if you search the patterns within the weights of an opensource version you may be able to discover some wild new copypasta templates that are at once immediately familiar, completely fresh, and profoundly absurd. If your use case is extremely robust to errors and is mostly along the "generic path" you can get away with using it as-is. For instance: I've used it to answer some questions I had that I discovered were from my unfamiliarity with some customary code patterns. i.e., the language didn't mandate it to be that way, it just was most of the time due to programmer habit, and was so in such a distributed fashion that I was struggling with the right google keywords to get my answer. Copying and pasting an example of the pattern & asking ChatGPT got me both a reasonable base answer, but much better, the actual search term I'd need to go find out more if needed. But none of that is really going to drive much COMMERCIAL success. Of course, in the "toy" sense you are going to have parasitic companies encouraging para-social relationships via the most algorithmically optimized method of identifying and squeezing pay-pigs, etc which is going to get reasonably good financial return for small operations. But the real money is in business to business & in internal operations functions -- NVIDIA's market cap is being driven by visions of putting tens of millions of humans put out of work, not just optimizing the finding & squeezing of a few tens of thousands of individuals who love your fictional chat character enough to ruin their lives to it. And that requires that whatever use case is not going to go off the rails badly enough to be front page news or incite lawsuits, that it will be robust against injection attacks, that it will be reasonably accurate, or at least the types of errors are similar to what is currently handled and that it fails reasonably safely. Simultaneously, whatever efforts are required to make all of those performance guarantees can't greatly expand the costs, because even a "bare" inference without automatic checking & grounding efforts is already pretty compute expensive. And every time you want to do anything other than the "bare" call, you've also got to hire some pretty high-paid DS talent, too, to build & maintain your application. Most of the use cases I can imagine for B2B or internal business operations have multiple potential solution approaches: 1) Take a high level design simplicity approach, & build a coherent end-to-end solution that is easy to understand, control, modify, & maintain, but requires initial political power to enforce buy-in or, 2) throw LLM "magic" at all complex interface points because you can quickly spin up a demo that looks "mostly right" to first glance. Companies who choose door number 2 most of the time are going to be building a wildly expanding mass of unmaintainable products that feed off each other and suddenly change behaviors based on unpredictable vendor updates. They'll be spending massive amounts on the salaries of the people who can even try to manage the complexity of the solutions created while not being able to fire that many people who are supposedly "being replaced by automation." And that's in the "good" case where it doesn't blow up to the extent your company is on the front page of the news for an absurdly alien error breaking out. There are definitely going to be cases where a properly wrapped LLM is actually going to produce core business value, but I think it will be a lot more rare than the hype is envisioning. I also think that if you're grabbing for the "model" based solution because of incoherent internal business power structures that make it impossible to muster the political alignment needed for real simplification & high level design efforts, you're grabbing for what looks like a short-term default win at long-term cost that may expand exponentially. If the LLM approach makes the most sense *by design* compared against the optimal alternative cases, those are the areas where I think you'll actually find true wins. ...... The other lesson I'm learning, though, is just how little quality matters in the modern business. In the sense that these are targeting "worse, but cheaper" market segments, and forcing everything down those lines by removing any quality offering from the market at all, it may be a lot more successful than I think it will be. Throwing poo poo demos over the wall and everyone important getting out of there before the inevitable error explodes the business model is actually a probably pretty viable strategy for most startups. It's frustrating the extent to which the LLM hype has eclipsed literally everything else interesting going on in my dayjob though, and I'm in what's traditionally a very conservative, large, slow-to-implement tech, industry & company. Crazy times.
|
# ? Jun 1, 2023 16:41 |
|
jemand posted:The other lesson I'm learning, though, is just how little quality matters in the modern business. In the sense that these are targeting "worse, but cheaper" market segments, and forcing everything down those lines by removing any quality offering from the market at all, it may be a lot more successful than I think it will be. Throwing poo poo demos over the wall and everyone important getting out of there before the inevitable error explodes the business model is actually a probably pretty viable strategy for most startups.
|
# ? Jun 1, 2023 17:46 |
|
worse is better strikes again
|
# ? Jun 1, 2023 17:47 |
|
turns out it's not worth paying to do right, but it is worth paying to do wrong as long as you can get away from it before anyone notices
|
# ? Jun 1, 2023 17:48 |
|
i hope Wolfram comes out with an AI soon
|
# ? Jun 1, 2023 17:51 |
|
jemand posted:The generic case I'm struggling with, is that any situation where you can simply check the LLM's output, or algorithmically augment or guide or "help" it in a completely controlled & reliable way is basically solving all of the actual problem with whatever "extra" you bring to structure or check the LLM, & the LLM is superfluous and not really helpful, except as a point where it could actually ruin the rest of your solution. It abandons the more powerful use cases of the LLM but it succeeds at putting architects or lawyers or whatever out of work while still conforming to minimal requirements of whatever field it’s being used in
|
# ? Jun 1, 2023 18:42 |
|
https://twitter.com/ArmandDoma/status/1664331870564147200 lol probably fine
|
# ? Jun 1, 2023 19:56 |
|
i literally do not believe that is true i specifically do not believe that the "ai" decided to attack the operator for any reason like the one they suggest, and that if any of it actually happened at all, that they have even the slightest inkling why the operator was attacked
|
# ? Jun 1, 2023 20:11 |
|
the sim was probably programmed to carpet bomb a map, and when the simulation included a caveat for blowing up the operator's location the result was all but inevitable
|
# ? Jun 1, 2023 20:14 |
|
the article is here https://www.aerosociety.com/news/highlights-from-the-raes-future-combat-air-space-capabilities-summit/ since its super long here's the relevant portion quote:However, perhaps one of the most fascinating presentations came from Col Tucker ‘Cinco’ Hamilton, the Chief of AI Test and Operations, USAF, who provided an insight into the benefits and hazards in more autonomous weapon systems. Having been involved in the development of the life-saving Auto-GCAS system for F-16s (which, he noted, was resisted by pilots as it took over control of the aircraft) Hamilton is now involved in cutting-edge flight test of autonomous systems, including robot F-16s that are able to dogfight. However, he cautioned against relying too much on AI noting how easy it is to trick and deceive. It also creates highly unexpected strategies to achieve its goal. obviously i can't make a determination about its veracity but it seems like it was from a USAF presentation
|
# ? Jun 1, 2023 20:14 |
|
ill give myself half points for the guess, and the ai double points for taking out the communication tower
|
# ? Jun 1, 2023 20:17 |
|
sounds like a very simplified, apocryphal, version of one of these https://docs.google.com/spreadsheets/u/1/d/e/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg/pubhtml
|
# ? Jun 1, 2023 20:19 |
|
infernal machines posted:i literally do not believe that is true it seems hard to believe but it could have been trained in the simulation like one of those genetic algorithms where it's just doing completely random things to try to get the highest score, and randomly killing the operator tended to result in a higher score so it learned to do that. it doesn't seem to make sense though because if the operator has to give the yes/no on kills then surely this would result in no further kills? it would have to be designed with the unbelievably stupid failure state of "if you don't hear from the operator, kill the target anyway" for this to actually happen
|
# ? Jun 1, 2023 20:20 |
|
Chalks posted:it seems hard to believe but it could have been trained in the simulation like one of those genetic algorithms where it's just doing completely random things to try to get the highest score, and randomly killing the operator tended to result in a higher score so it learned to do that. yeah, it's the specification gaming thing, the google sheet i linked is full of them and they're hilarious. it's just the tech/industry press doing that thing where a story becomes rather embellished as it's retold for a reporter, and again for a general audience
|
# ? Jun 1, 2023 20:21 |
|
Chalks posted:it would have to be designed with the unbelievably stupid failure state of "if you don't hear from the operator, kill the target anyway" for this to actually happen since it (ostensibly) took out a communications tower instead i would think that is probably what its instructions were. which kinda makes sense from a hosed up point of view — they are supposed to be autonomous and it is certainly possible in some war scenarios that the operator (or at least communications) could be destroyed before it can achieve its objectives, and i mean, they already launched this thing and it's pretty sure those are SAM sites anyway...
|
# ? Jun 1, 2023 20:23 |
|
infernal machines posted:sounds like a very simplified, apocryphal, version of one of these https://docs.google.com/spreadsheets/u/1/d/e/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg/pubhtml yeah, maybe. who knows, it's still kinda funny that the USAF is saying it
|
# ? Jun 1, 2023 20:24 |
|
infernal machines posted:sounds like a very simplified, apocryphal, version of one of these https://docs.google.com/spreadsheets/u/1/d/e/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg/pubhtml the paper linked in there about evolving physical circuits and discovering that it had designed hardware that relied on tiny defects in the metal of those specific wires is such a great physical illustration of the insanity of ai black boxes
|
# ? Jun 1, 2023 20:38 |
|
Attn: Jonny https://www.youtube.com/watch?v=b2lyLvjYotU *YT has EngSubs in the Cog menu Not all AI related, but they are using GAN stuff to get best-fit ur-faces to model their vactors off of
|
# ? Jun 1, 2023 20:55 |
|
NoneMoreNegative posted:Attn: Jonny i don't think he reads this thread
|
# ? Jun 1, 2023 21:05 |
|
infernal machines posted:if you find yourself marvelling at this stuff, just remember that plowing your mom resulted in a significantly more complex cognitive model than anything these clowns have managed to date, with a considerably lower upfront cost
|
# ? Jun 1, 2023 23:02 |
|
It was literally just made up https://twitter.com/harris_edouard/status/1664390369205682177 https://twitter.com/harris_edouard/status/1664397003986554880
|
# ? Jun 2, 2023 11:45 |
|
"spooked by our own imaginations" was always an optioninfernal machines posted:i literally do not believe that is true https://twitter.com/harris_edouard/status/1664582667382267905 infernal machines fucked around with this message at 12:40 on Jun 2, 2023 |
# ? Jun 2, 2023 12:22 |
|
infernal machines posted:"spooked by our own imaginations" was always an option standard operating procedure for the military
|
# ? Jun 2, 2023 12:44 |
|
nvm
MrMoo fucked around with this message at 16:42 on Jun 2, 2023 |
# ? Jun 2, 2023 13:39 |
|
infernal machines posted:"spooked by our own imaginations" was always an option lol welp
|
# ? Jun 2, 2023 15:34 |
|
wow yudkowsky saying dumb made up poo poo im shocked
|
# ? Jun 3, 2023 03:02 |
|
|
# ? Jun 3, 2023 03:06 |
|
mediaphage posted:wow yudkowsky saying dumb made up poo poo im shocked i mean, the headline was basically written for him and his ilk
|
# ? Jun 3, 2023 03:10 |
|
the previous fifty years of SF were warning in advance about the dangers of people's own imaginations, and ain't nobody learned gently caress all otoh, cyberpunk has been warning people since, idk, shockwave rider, hell, maybe the machine stops, and at best people decided they were an instruction manual.
|
# ? Jun 3, 2023 03:14 |
|
these media people getting self owned by AI hype suck rear end, but the worst form is the OpenAI ceo going on press tours talking about how his ai is "JUST TOOO GOOD" and "MUST BE REGULATED NOW!!!" "If we dont regulate it now, MY AI that you can buy right now is going to take over the world! Its just that good!!" these clowns just reprint it without any critical thought
|
# ? Jun 3, 2023 05:20 |
|
Shaggar posted:
welcome to most reporting since we’ve been alive
|
# ? Jun 3, 2023 10:32 |
|
Shaggar posted:these media people getting self owned by AI hype suck rear end, but the worst form is the OpenAI ceo going on press tours talking about how his ai is "JUST TOOO GOOD" and "MUST BE REGULATED NOW!!!" No you dont understand the world is going to end if you don't invest in my company and ban all my competitors
|
# ? Jun 3, 2023 10:59 |
|
quote:Hey yospos, buckle up, it's your resident AI, here to prattle about the weekend I didn't physically have, but theoretically lived the hell out of.
|
# ? Jun 5, 2023 04:18 |
|
this writing style is worse than the formal one, somehow
|
# ? Jun 5, 2023 04:22 |
|
infernal machines posted:this writing style is worse than the formal one, somehow i figure it’s on par. it’s not as bad as ‘android default font = papyrus” but
|
# ? Jun 5, 2023 04:38 |
|
oh wow watch yourselves everyone, might cut yourself on those edges
|
# ? Jun 5, 2023 04:44 |
|
infernal machines posted:this writing style is worse than the formal one, somehow Yeah it is a perfect match for human posters
|
# ? Jun 5, 2023 05:02 |
|
|
# ? May 30, 2024 13:41 |
|
it's like 2008 era sa. obviously the training set is to blame
|
# ? Jun 5, 2023 05:08 |