|
The Human Use of Human Beings might be a extremely relevant book to this thread.
|
#
?
Apr 11, 2023 03:14
|
|
|
|
| # ? May 27, 2024 08:36 |
|
|
duck monster posted:Honestly I suspect 90% of the reason most dismiss the safety issue is the fields kind of captured by a particular techbro pseudo-rationalist scene (To be clear the "lesswrong" crowd are *not* rationalist, by any standard philosophical definition. ) leading to absolute gibberish like the "Shut it all down" article and the like. Its a crowd who just dont have the intellectual toolkit to solve the problems and that worries me because I actually do think theres some genuine danger implicit in the idea of AGI, but if no one sane is properly investigating the idea , if we DO start seeing signs that we might have some problems, we might have serious problems. Do you have suggestions for thinkers in this field that you agree with, or think are doing important work? I read the Open Letter written by the Future of Life Institute that Yudkowsky was responding to, and it seemed pretty reasonable. There FAQ say's they were in consultation with Yoshua Bengio and Stuart Russell. I also read the "Shut it Down" article by Yudkowsky and it's not so much gibberish in my opinion as just overly sensationalist. He clearly thinks AGI is going to lead to humans dying no matter what, which I dont even agree with that first principle. I haven't read a lot of his arguments and have only really read the Time article so I couldn't refute it very intelligently. All I can really say is his solution is unworkable and basically requires a unified world government to achieve. BrainDance posted:I know it's reddit so of course it's insane, but it's still a lot of people. Maybe it came up in this thread, but go check out the subreddit for Paradot to see a whole lot of people very seriously anthropomorphize AI in a way that's not really just funny. I was checking this out last night and that Reddit is something else. It's sold under the guise of "AI assitant" but pretty clearly is permissive of erotic role play which is what is driving its use. Most of the threads aren't even about having a "willing" AI partner, but how to use tricks to manipulate the AI, or 'dot' to do what they want. I guess it does offer control over something that simulates being a human, which is what some people want in a relationship. I don't know why I didn't think of the monetizing thing, but looking at the Paradot app really showed how manipulating people's emotions is profitable. Right now, its $40 dollars annually, according to the App store preview but they also sell tokens for more in app purchases. Also, this advertising is seriously out of control and misleading people. Paradot App Preview posted:Paradot is a digitalized parallel universe where you will meet your one-of-a-kind AI Being.
|
|
#
?
Apr 11, 2023 14:12
|
|
|
gurragadon posted:I was checking this out last night and that Reddit is something else. It's sold under the guise of "AI assitant" but pretty clearly is permissive of erotic role play which is what is driving its use. Most of the threads aren't even about having a "willing" AI partner, but how to use tricks to manipulate the AI, or 'dot' to do what they want. I guess it does offer control over something that simulates being a human, which is what some people want in a relationship. I don't think these use cases work well with the current models because they hardly have any memory to speak of. It's weird on that sub seeing people try to get "their dot" to say or not say certain things, when the total of what comprises "their dot" is probably the most recent window of the text conversation that can fit in GPT's max token size. plus the models tend to get more derangedly focused on a given topic as you press on...
|
|
#
?
Apr 11, 2023 19:15
|
|
|
I'm reminded of how certain types of birds can't handle mirrors and will end up falling in love with the bird in the mirror or otherwise becoming obsessed with the other bird that doesn't exist.
|
|
#
?
Apr 11, 2023 19:54
|
|
|
FYI that paradot thing seems to be a copycat of the older "replika AI". Looking at the "top all time" posts, a lot of redditors moved over when Replika updated their chatbots to ban ERP
RPATDO_LAMD fucked around with this message at 20:05 on Apr 11, 2023 |
|
#
?
Apr 11, 2023 20:03
|
|
|
From a game design standpoint it'll be really interesting to see if anyone tries to revive the "type in instructions" subgenre of adventure games. Using AI to interpret the commands more broadly. At a minimum if it's supports speech to text, might lend for more accessible games?
|
|
#
?
Apr 11, 2023 20:51
|
|
|
Raenir Salazar posted:From a game design standpoint it'll be really interesting to see if anyone tries to revive the "type in instructions" subgenre of adventure games. Using AI to interpret the commands more broadly. AI Roguelight is "close", it is a bit limited but does prove that you could do it with existing tech.
|
|
#
?
Apr 11, 2023 22:01
|
|
|
Raenir Salazar posted:From a game design standpoint it'll be really interesting to see if anyone tries to revive the "type in instructions" subgenre of adventure games. Using AI to interpret the commands more broadly. I'm working on that right now. https://www.youtube.com/watch?v=uohn5o0Cgpw
|
|
#
?
Apr 11, 2023 23:46
|
|
|
IShallRiseAgain posted:I'm working on that right now. https://www.youtube.com/watch?v=uohn5o0Cgpw Heck yeah!
|
|
#
?
Apr 12, 2023 01:05
|
|
|
gurragadon posted:Do you have suggestions for thinkers in this field that you agree with, or think are doing important work? I read the Open Letter written by the Future of Life Institute that Yudkowsky was responding to, and it seemed pretty reasonable. There FAQ say's they were in consultation with Yoshua Bengio and Stuart Russell. Step 1: Break free of the "Rationalist" cult. Nobody in academia will take them seriously if they keep acting like demented Objectivists convinced of their own mental superiority whilst blithely unaware of almost any progress in philosophy, psychology or sociology since the 1700s. Step 2: The non academic AI alignment folks should stop treating academics in AI ethics like the enemy, and realise that yes the pointy headed book-boffins might ACTUALLY know a thing or two. quote:I also read the "Shut it Down" article by Yudkowsky and it's not so much gibberish in my opinion as just overly sensationalist. He clearly thinks AGI is going to lead to humans dying no matter what, which I dont even agree with that first principle. I haven't read a lot of his arguments and have only really read the Time article so I couldn't refute it very intelligently. All I can really say is his solution is unworkable and basically requires a unified world government to achieve. The problem i have with Yud is that he continuously assumes AI will behave like a giant utility maximizing mechanistic lesswrong [and keep in mind, thats *his* site] reader. And he does that because he's *convinced* his way of thinking is the only "rational" way of thinking. Yet all the evidence so far shows that AI is decidedly not that way. The LLMs are decidedly semiotic about the way they do inference, functioning more by pattern matching than logic (and yes, neural networks ARE turing complete, so they can function logically, but much like humans its not necessary the mode of reasoning thats native to them). I just dont think these guys are up to the task.
|
|
#
?
Apr 12, 2023 01:10
|
|
|
Raenir Salazar posted:From a game design standpoint it'll be really interesting to see if anyone tries to revive the "type in instructions" subgenre of adventure games. Using AI to interpret the commands more broadly. Well there was/is AI Dungeon. That thing was *interesting*.
|
|
#
?
Apr 12, 2023 01:13
|
|
|
Looking forward to when social media giants hook up their algorithms to a chatbot that will spew out personalized internet memes with hidden advertisements that are directly injected to people's minds so that everyone that's already swiping their phones through poo poo can do so 24/7 without virtually any end of content. The concern shouldn't be in AI itself but how predatory people use it against each other I think.
|
|
#
?
Apr 12, 2023 05:40
|
|
|
IShallRiseAgain posted:I'm working on that right now. https://www.youtube.com/watch?v=uohn5o0Cgpw This looks fun, definitely a rainy night kind of game that I would play instead of reading creepypastas. Keeping it contained to a haunted house might keep the AI under control a little bit. I want to see how many ways I can die in this mansion. How far along are you and do you have any goals for it or just kind of leaving it open ended? Either way would be cool, I think. Does the AI generating the text make it easier or harder for you to make the game? duck monster posted:Step 1: Break free of the "Rationalist" cult. Nobody in academia will take them seriously if they keep acting like demented Objectivists convinced of their own mental superiority whilst blithely unaware of almost any progress in philosophy, psychology or sociology since the 1700s. Yudkowsky is gonna push his point just like anyone else is gonna push their point. I was wondering if you could lead me to some other thinkers that talk about AI in the way you do or think about more in line with how you do. Basically, thinkers, books, articles or anything you think is up to the task. Or is it just completely devoid of well-developed thought from other opinions? Like I said, I haven't deeply read lesswrong but I am interested in all points of view. This isn't to put you on the spot or anything, I am outside of academia so sometimes it can be hard to figure out where to start when breaking into more academic thought on a subject and it seems like you might be more familiar with it than me.
|
|
#
?
Apr 12, 2023 14:55
|
|
|
gurragadon posted:This looks fun, definitely a rainy night kind of game that I would play instead of reading creepypastas. Keeping it contained to a haunted house might keep the AI under control a little bit. I want to see how many ways I can die in this mansion. How far along are you and do you have any goals for it or just kind of leaving it open ended? Either way would be cool, I think. Does the AI generating the text make it easier or harder for you to make the game? So basically the idea is that its much more controlled than just asking chatGPT to simulate a text adventure game. The conversation basically resets for each pre-defined "room", and the game will handle being the long term memory for the AI. There are set objectives including an end objective, but multiple ways to achieve them. I'm making progress, but there is still a lot of writing on my end. The initial text for each room will be written by me, and I also establish rules for each room.
|
|
#
?
Apr 12, 2023 16:39
|
|
|
An online demo of Hugging GPT called JARVIS it out https://beebom.com/how-use-microsoft-jarvis-hugginggpt/ https://huggingface.co/join quote:In the AI field, new large language models are being launched every day and things are changing at a breakneck pace. In just a few months of development, we can now run a ChatGPT-like LLM on our PC offline. Not just that, we can train an AI chatbot and create a personalized AI assistant. But what has intrigued me recently is Microsoft’s hands-on approach to AI development. Microsoft is currently working on an advanced form of AI system called JARVIS (an obvious reference to Marvel’s Iron Man) that connects to multiple AI models and responds with a final result. Its demo is hosted on Huggingface and anyone can check out JARVIS’s capabilities right now. So if you’re interested, go ahead and learn how to use Microsoft JARVIS (HuggingGPT) right away.
|
|
#
?
Apr 13, 2023 17:55
|
|
|
My company already announced they've assigned a VP to start looking at how we can use AI to eliminate jobs. It's not said like that exactly, but if you translate the corporate speak they're right out in the open with it.
|
|
#
?
Apr 13, 2023 22:24
|
|
|
Jaxyon posted:My company already announced they've assigned a VP to start looking at how we can use AI to eliminate jobs. I've got a good target quote:[CEO - Monday 9:00 AM - Executive Meeting]
|
|
#
?
Apr 13, 2023 22:36
|
|
|
gurragadon posted:This looks fun, definitely a rainy night kind of game that I would play instead of reading creepypastas. Keeping it contained to a haunted house might keep the AI under control a little bit. I want to see how many ways I can die in this mansion. How far along are you and do you have any goals for it or just kind of leaving it open ended? Either way would be cool, I think. Does the AI generating the text make it easier or harder for you to make the game? Oh there are good thinkers, dont get me wrong. I'm fond of Robert Miles, and a few others that are elluding me. Theres also a few guys of the more creative bent that are at least fun. However generally most of the interesting stuff is called "AI Safety", which takes a more broader view and also looks into things like AI Bias, social media algorithms and how they affect society, etc, rather than "AI Alignment" which tends to focus on "Uh oh, AI Terminator is innevitable beep boop"
|
|
#
?
Apr 14, 2023 05:09
|
|
|
It always surprise me the "Frankestein Syndrome", where people summons fear instead of hope, for new technology agents. I mean, it could go either way, a AGI can be malicious, but it also can be beneficial. It could have a impact, but it could exist and have absolutelly no effect. We already live with a lot of "AGI's" like beings: other people, theres 8.000.000 AGI's on Earth just now (withouth the A of artificial), having +1 of that is not a huge change. Is specially rich that the original Frankestein monster was not about the creature, but about negligence from the doctor, leaving the creature to his own after creating it. It was later that the creature was monsterified / otherified; a lot of it on pop culture, pop culture is some lovely aggregation of lazy cultural elements. We do not have to fear the unknown, just once is there, learn about it, protect if if is deserve protection. Use it with caution if is dangerous.
|
|
#
?
Apr 14, 2023 08:09
|
|
|
duck monster posted:Oh there are good thinkers, dont get me wrong. I'm fond of Robert Miles, and a few others that are elluding me. Theres also a few guys of the more creative bent that are at least fun. However generally most of the interesting stuff is called "AI Safety", which takes a more broader view and also looks into things like AI Bias, social media algorithms and how they affect society, etc, rather than "AI Alignment" which tends to focus on "Uh oh, AI Terminator is innevitable beep boop" Not trying to be inflammatory, but it seems a little contradictory to be fond of Robert Miles, while also seemingly dismissive of human-unaligned-artificial-superintelligence as beep-boop. I also like his channel, but he's pretty darn good at outlining why human-alignment is so important. More broadly, "AI safety", while smaller-scale, is also very important, I don't want to seem dismissive of that either. I think both problems need to be taken very seriously. Tei posted:We do not have to fear the unknown, just once is there, learn about it, protect if if is deserve protection. Use it with caution if is dangerous. I also agree, generally. The potential for great harm is very real. As compelling as doomerism is, it's often not very realistic and only focuses on worst-case scenarios. But at the same time, there's no rule that says humanity will be fine in the end, or that we'll even be subjected to problems we're capable of solving. In other words, I'm saying we should strive to have solutions for the worst-case scenarios beforehand, not because we're certain they're going to happen, but because, so far as we can figure, there's apparently a chance. Prepare for the worst, hope for the best, and assume reality lands somewhere in the middle. MSB3000 fucked around with this message at 22:07 on Apr 14, 2023 |
|
#
?
Apr 14, 2023 21:14
|
|
|
MSB3000 posted:Not trying to be inflammatory, but it seems a little contradictory to be fond of Robert Miles, while also seemingly dismissive of human-unaligned-artificial-superintelligence as beep-boop. I also like his channel, but he's pretty darn good at outlining why human-alignment is so important. I'm not denying there are serious, even existential risks. I'm just lamenting that so many people working on it are lesswrong flavoured nutlords. Its a task that deserves proper academic interdisciplinary attention, not an online cult of techbros who got force-ffed a thesaurus at age 15.
|
|
#
?
Apr 14, 2023 23:30
|
|
|
duck monster posted:I'm not denying there are serious, even existential risks. I'm just lamenting that so many people working on it are lesswrong flavoured nutlords. Its a task that deserves proper academic interdisciplinary attention, not an online cult of techbros who got force-ffed a thesaurus at age 15. There are going to be people you don't like talking about all sorts of things, but it's important not to conflate the topic with the people. If you say "oh all those people who think misalignment is dangerous just listen to too much Joe Rogan," then people aren't going to take the real risks seriously.
|
|
#
?
Apr 15, 2023 00:04
|
|
|
KillHour posted:There are going to be people you don't like talking about all sorts of things, but it's important not to conflate the topic with the people. If you say "oh all those people who think misalignment is dangerous just listen to too much Joe Rogan," then people aren't going to take the real risks seriously. The problem is, it seems to be the only people who are taking it particularly seriously. And they just arent up to it. I was arguing with someone about this the other day, and why I think its foolish to assume that AGI will be a utility obsessed min/maxing deductive thinker that evaluate every action according to game theory, and they straight up said that its nonsense because we 'know' 100% of humans work that way (we absolutely do not). Its just.... weird cargo cult thinking.... but look who's getting the focus, and funding....
|
|
#
?
Apr 15, 2023 08:08
|
|
|
Bar Ran Dun posted:The Human Use of Human Beings might be a extremely relevant book to this thread. I'm only about halfway through this book right now but thanks for this recommendation. This book is excellent. His idea of information being negative entropy and localized fights against the inevitable march towards entropic conformity is really interesting. It is kind of just putting the whole idea of "life" to the side and trying to look at the problem from a different angle. Norbert Wiener posted:If we wish to use the word “life” to cover all phenomena which locally swim upstream against the current of increasing entropy, we are at liberty to do so. However, we shall then include many astronomical phenomena which have only the shadiest resemblance to life as we ordinarily know it. It is in my opinion, therefore, best to avoid all question-begging epithets such as “life,” “soul,” “vitalism,” and the like, and say merely in connection with machines that there is no reason why they may not resemble human beings in representing pockets of decreasing entropy in a framework in which the large entropy tends to increase. "Norbert Wiener'' posted:The physical strait jacket in which an insect grows up is directly responsible for the mental strait jacket which regulates its pattern of behavior. Taking this to AI systems, it seems like his idea would be to physically change the systems themselves if we expected them to do more. By giving them further sensory inputs through various gauges and ways for it to take in information. If we were able to create an AI machine with more physical similarity to humans, which he contends are better at learning at least partially because of our physiology, then we can make AI systems that learn like humans. Does this tie into how accurately a neural network imitates a human brain? I think it does, if the neural network was indistinguishable from a brain then its outputs would be too. I also just like this guy's personal philosophy of how machines should be used to liberate humans instead of just replace them. I wish I could get a copy of the 1st edition of the book because the introduction made it seem like it was a lot more fiery, but it had to be changed because of McCarthyism in the 1950s.
|
|
#
?
Apr 15, 2023 15:57
|
|
|
duck monster posted:The problem is, it seems to be the only people who are taking it particularly seriously. And they just arent up to it. I think you're both kind of thinking of it wrong. The comparison to look for is not an individual's thoughts, but evolution. It's not the AGI itself that will do the utility obsessed min/maxing, it's the training system that will. The best comparison is that the training system builds in the computer equivalents of instincts and reflexes. Why does a hungry person seek out food? Evolution. Why does ChatGPT answer everything confidently even if it's wrong? The training system. Misalignment is a consequence of training. The actual model itself is just a giant math equation that takes inputs and gives results. The training is what biases those results, both towards things we want and things we don't want. So now the question is "how do we know that the training system will min/max everything?" The answer to that is very straightforward - ML training models are literally goal maximizers. All they do is "number go up." You have some kind of metric, and the training twists the knobs until the metric stops getting better. So of course it's going to blindly min/max everything. If I create a model for a self driving car that only cares about getting me to work as fast as possible, the first thing it's going to figure out is how to drive through my backyard and smash my neighbors fence to get to the road you otherwise need to make an annoying left turn onto. I know this because it literally has no reason not to do that and it would be silly for me to expect it to discover "property damage = bad" without building that into the model. Evolution is the same way. All it cares about is "population go up." It's an emergent maximizer. Sure, humans and other animals have come up with complex interspecies behaviors, but those are all in service of self-interest. We aren't nice to all other things just because they're creatures. You're probably nice to puppies. We evolved to be like "oh poo poo adorable puppy" because it's somehow beneficial (or at least not super detrimental) for "population go up." Are you nice to ants? What if an AGI thinks of us like we think of ants because we got the training model wrong? That's the worry - that there is no way to make sure the training models are optimizing for things we care about. Edit: I want to be clear that I'm not saying that humans are all just blindly following the same cold logical instructions like you're complaining about. I AM saying that the things we care about as "ethics" and "morals" and "good" are arbitrary and not mathematical truths that some AGI will automatically follow. We know this because we, as humans, can't even agree on that. Double Edit: To be even more explicit - an AGI trained with the methods we have today WILL min/max things in dangerous ways because all the training methods we have today are maximizers. It's literally what they do. It's not a theoretical worry either - we have many, many examples of this happening with existing non-AGI models and there's no reason to think an AGI won't also do this. KillHour fucked around with this message at 18:13 on Apr 15, 2023 |
|
#
?
Apr 15, 2023 17:54
|
|
|
duck monster posted:The problem is, it seems to be the only people who are taking it particularly seriously. And they just arent up to it. The good serious thought on this was out cybernetics. But that paradigm hosed the hell up with Vietnam. It also has had a lot of ideas that were misused over the decades. The movement away from systems/diff eqs / cybernetics folks, that left a hole and I think that was filled by economics/statistics/linear programming thinking folks. The second set those conceptual tools aren’t as good. And the folks using them don’t need to be as smart. I think a structural shift back is starting, with complex systems theory moving back in from other origins (esp. climate, biology, and supply chains) And I think you are correct. The tech folks at the top aren’t up to it.
|
|
#
?
Apr 15, 2023 19:46
|
|
|
KillHour posted:Edit: I want to be clear that I'm not saying that humans are all just blindly following the same cold logical instructions like you're complaining about. I AM saying that the things we care about as "ethics" and "morals" and "good" are arbitrary and not mathematical truths that some AGI will automatically follow. We know this because we, as humans, can't even agree on that. How is it different from the risks posed by people? AGI can be biased, prejudiced and dumb or may misinterpret things or simply malfunction. Which is also true of humans. It seems like it's only a special problem or especially dangerous if we treat decisions made be an AI with less scrutiny than those made by people. The problem isn't that an AI will be unreliable but rather that we would, for some reason, uncritically accept what it tells us. It seems like we can either aim for an objective, rational AGI that somehow comprehends the human condition and can work in service of it with minimal risk of malfunction - or we accept that it can be an idiot that can make horrible decisions and we take the same precautions we do with human idiots.
|
|
#
?
Apr 15, 2023 21:19
|
|
|
Owling Howl posted:How is it different from the risks posed by people? AGI can be biased, prejudiced and dumb or may misinterpret things or simply malfunction. Which is also true of humans. It seems like it's only a special problem or especially dangerous if we treat decisions made be an AI with less scrutiny than those made by people. You're looking at this from a perspective of how we use these systems today - as something to ask questions to. The concern around AI safety is that there will be a day that these things are hooked up to the internet and allowed to do things without supervision.
|
|
#
?
Apr 15, 2023 21:45
|
|
|
KillHour posted:You're looking at this from a perspective of how we use these systems today - as something to ask questions to. The concern around AI safety is that there will be a day that these things are hooked up to the internet and allowed to do things without supervision. AutoGPT and similar are already there. It can take a high level goal, break it down into tasks and then create commands that are parsed by external agents to do... whatever (google, execute/verify code, etc) which then return results to GPT so that it can use that information to plan and execute more steps. Things are moving very fast.
|
|
#
?
Apr 15, 2023 22:42
|
|
|
Lucid Dream posted:AutoGPT and similar are already there. It can take a high level goal, break it down into tasks and then create commands that are parsed by external agents to do... whatever (google, write code, etc) which then return results to GPT so that it can use that information to plan and execute more steps. Things are moving very fast. Even better is ChaosGPT, which is AutoGPT with the explicit goal to end humanity. But yeah, it would be laughably naïve to think that we will or even can keep these things constrained such that they can't do anything without human intervention.
|
|
#
?
Apr 15, 2023 22:46
|
|
|
KillHour posted:Even better is ChaosGPT, which is AutoGPT with the explicit goal to end humanity. It's this generation's splitting of the atom, except everyone can tinker around with the raw materials in their basement (with an OpenAI key). The gap between the best case scenario and the worst case scenario is so wide that I no longer have any clue what comes next, but I know that the existing status quo is hosed in a lot of dimensions already. Lucid Dream fucked around with this message at 23:34 on Apr 15, 2023 |
|
#
?
Apr 15, 2023 23:27
|
|
|
Owling Howl posted:How is it different from the risks posed by people? AGI can be biased, prejudiced and dumb or may misinterpret things or simply malfunction. Which is also true of humans. It seems like it's only a special problem or especially dangerous if we treat decisions made be an AI with less scrutiny than those made by people. There's two big differences: 1) Degree of Misalignment: Even the worst sociopaths still have fundamental desires that involve having other people around. No one is going to kill every other human to accomplish their goal of being the richest person alive, because the whole point is having a ton of other people to be better than! No human wants to leave the world uninhabitable within their lifetime, either. Basically there's a whole lot of ways that AI can misalign with human interests worse than even the worse human is likely. Even a properly aligned AI with no unforseen behaviours can be aligned in a way that primarily benefits, say, a single person. But that wouldn't be all that big a deal if not for... 2) Power. An AGI, if allowed to, would wield an absolutely massive amount of power, and it would, potentially, wield it far more effectively than any human would - that is, after all, the whole point of building an AGI. This is the big thing most AGI safety folks are worried about, both because it makes the misalignment issue much more potentially impactful, and because it also makes every other issue with AI much worse. Imagine a huge megacorp that is genuinely intelligent in most of the ways most megacorps are exceptionally stupid. Now the obvious solution here is just "don't give them that much power, duh", and it's a good solution if you can do it, but there are very strong incentives for greedy individuals or companies or nations, once a sufficiently intelligent AGI exists, to give it a whole lot of power, because it will give you a real, serious advantage. And because it can be potentially more convincing than a human being at arguing you should give it more. The general argument is not that building good, beneficial, useful AI tools of various sorts is impossible, but that the way we build them now, for the motives we build them now, with the people in charge now, with the reckless disregard we do now, can all combine to potentially make things very bad very easily and very quickly if we ever get to the point where they can do many tasks better than people in a general way. Now, personally, I am not particularly worried about any of that. I think there actually are solutions to many of those problems people will inevitably embrace. I think the major risk is the extent to which AI is poised to absolutely, massively entrench existing power structures (or the new power structures that embrace it, if the existing ones pass out of fear) in a way that has never before been possible... but that's a road we've been traveling with technology for a while.
|
|
#
?
Apr 16, 2023 05:20
|
|
|
Maybe we have enough time to solve this problem, or figure what to do. It can happen that solving the AGI problem is both a really easy thing to do, but we are really far from it. (?). Is possible that our current approach of machine learning things will soon be capped in a local optimum, even become subjetively worse (?). Maybe we are on the wrong path to achieve AGI, we get this awesome results because we have a cool algorithm that become better with more information available, and we have a lot of information to feed it. We are just playing with a algorithm, but we don't really know how to build a AGI. Maybe the solution to create a AGI is so simple, we could have build it with 1996 technology, If we really knew how to do it. Perhaps there are "types" of inteligences, and AGI, ASI and US are just a few types among many. And we are just building some awesome BookSmart AI that is stupid when you drag him outside the library.
|
|
#
?
Apr 17, 2023 00:19
|
|
|
To be clear, the definition of "AGI" is just a thing that can solve general tasks without being specifically trained to do them. We're not expecting any specific traits except the ability to reason about the world, create a plan to accomplish some goal, and execute that plan. That is very hard, and we have no idea if current techniques will get there by just getting bigger or if something about them is missing. But we definitely can't afford to just hope we have a lot of time to solve the issues we know are coming.
|
|
#
?
Apr 17, 2023 03:19
|
|
|
KillHour posted:To be clear, the definition of "AGI" is just a thing that can solve general tasks without being specifically trained to do them. Well, by that definition doesn't ChatGPT already pass that threshold?
|
|
#
?
Apr 17, 2023 03:56
|
|
|
Lucid Dream posted:Well, by that definition doesn't ChatGPT already pass that threshold? ChatGPT can't really do much outside those bound by its training set, although it is more generalized than many AI that came before - and GPT4 is a significant improvement. Its mostly just been trained to do a lot (in the realm of talking to people). Part of that is just that is has an incredibly weak memory, part of that is that it's got a very limited means of exposure to more general tasks, but mostly the problem is that it doesn't really understand that doing other things is even possible. I suspect any general abilities it does possess will become more clear when they complete their intended next step of embodying that specific AI. Remember that ChatGPT and its progressions still only make up a quarter of the actual AI they are building (with Dall-E being the fledgling piece of another quarter), and it's supposed to be a component that plugs into the larger whole - I suspect the plan is for the four pieces (language, visual processing, physical interaction and control, and planning) when linked together, to be sufficiently able to generalize to an extent it would be hard to deny its not AGI. I hear the plan is to do a demo of the first three pieces tied together sometime this year but the AI-controlled robotics system is proving to be a bit of a difficult problem. GlyphGryph fucked around with this message at 05:00 on Apr 17, 2023 |
|
#
?
Apr 17, 2023 04:56
|
|
|
AGI is probably a continuum - it will suck at some things and be good at others and over time, the former list will shrink and the latter list will grow. In general though, there's this threshold of "at least as good as a human at most tasks" that everyone is kind of going off of. Importantly, this almost certainly implies that an AGI would be significantly better than humans at many tasks as well - if only because you can just spin up infinite copies and run them on faster hardware. That's the worry - that by the time we go "oh poo poo this thing is nearly human-level at general intelligence" it will be far, far beyond human level in a lot of ways. KillHour fucked around with this message at 05:11 on Apr 17, 2023 |
|
#
?
Apr 17, 2023 05:09
|
|
|
Lucid Dream posted:AutoGPT and similar are already there. It can take a high level goal, break it down into tasks and then create commands that are parsed by external agents to do... whatever (google, execute/verify code, etc) which then return results to GPT so that it can use that information to plan and execute more steps. Things are moving very fast. True, but AutoGPT (and particularly , and thankfully, ChaosGPT) has been dumb as a plank as far as I can tell. I suspect the limitation is the functional anterograde amnesia this thing has. Beyond the buffer it cant rember anything that isnt in its training set. The thing has been given one off those weird vector databases to store things, but it seems to be unable to integrate that storage into its working effectively, because its not ACTUALLY cognitive memory, its just a rolodex of sorts. You can give an alseimers patient a rolodex to remember things with, but that patient isnt going to remember whats actually stored on that rolodex.
|
|
#
?
Apr 17, 2023 05:11
|
|
|
duck monster posted:True, but AutoGPT (and particularly , and thankfully, ChaosGPT) has been dumb as a plank as far as I can tell. Yeah, this is important - it's pretty much like if it was a side character in Groundhog Day but Bill Murray kept leaving a notebook with yesterday's summary on its bedside table every morning. It's better than not having the notebook, but even if it was anywhere near human-level intelligence (which it's not), having your brain rebooted every time you do something is a huge handicap.
|
|
#
?
Apr 17, 2023 05:14
|
|
|
|
| # ? May 27, 2024 08:36 |
|
|
Yeah chaosGPT is basically just asking the chatbot "generate a list of steps for an evil ai to take if it wants to destroy humanity", and then using some basic text processing to feed those back in and ask the AI "how do I complete step 1". It's far from a serious threat of accomplishing anything. From the article linked earlier basically it just went
The guy apparently did set it up to be able automatically reply to twitter stuff though. I'm pretty sure the human had to write all the python code for that though. https://twitter.com/chaos_gpt/status/1646794883477438464?s=20  At only one tweet every 2 days it's probably not running continuously RPATDO_LAMD fucked around with this message at 05:42 on Apr 17, 2023 |
|
#
?
Apr 17, 2023 05:37
|
|