|
Jeb Bush 2012 posted:this is either meaninglessly true or clearly false, depending how you define the state space of the supposed "markov chain" Ok, how about "GPT uses a statistical model to mimic previously observed word relationships in order to imitate common human sentence structures in its training set"? It's literally a chinese room.
|
#
?
Apr 7, 2023 19:02
|
|
|
|
| # ? May 31, 2024 13:47 |
|
|
my dad posted:it helps skip the agony of the blank paper. Jack Trades posted:There's no AI there. ChatGPT is just a stochastic parrot. These are very good and descriptive phrases by the way, I'm gonna steal them for the future.
|
|
#
?
Apr 7, 2023 19:06
|
|
|
While I don't like the dumb hype around ChatGPT, calling something a Chinese room isn't the diss you think it is because that particular thought experiment is bullshit.
|
|
#
?
Apr 7, 2023 19:06
|
|
|
my dad posted:While I don't like the dumb hype around ChatGPT, calling something a Chinese room isn't the diss you think it is because that particular thought experiment is bullshit. yeah, I'm pretty skeptical of a lot of claims made about chatgpt but stuff like "it's just a model trained on text data" doesn't tell you anything I've actually seen the chinese room though experiment raised more often by chatgpt super-optimists, because if the question is "does something that perfectly mimics a human count as intelligent if it works in a weird way" you're already talking about something amazing. the problem is that the premise isn't true and the way chatgpt works is very relevant to knowing if it will ever be true (I'd be surprised)
|
|
#
?
Apr 7, 2023 19:34
|
|
|
Brendan Rodgers posted:4X AI is so bad that these Markov chains don't have a high bar to clear at all. https://youtu.be/vozVq2puMYw
|
|
#
?
Apr 7, 2023 19:43
|
|
|
The fact that so many people have been fooled into thinking that GPT is even slightly intelligent just because it can string together words grammatically and in an authoritative voice is pretty sad.
|
|
#
?
Apr 7, 2023 19:53
|
|
|
Clarste posted:The fact that so many people have been fooled into thinking that GPT is even slightly intelligent just because it can string together words grammatically and in an authoritative voice is pretty sad. Counterpoint: ChatGPT actually is very intelligent, because the way it strings words together into a coherent whole without having an referents or understanding of the content is actually an incredibly difficult thing to do, its just a very narrow sort of intelligence. It's only mimics real meaning, sure, but it does it incredibly well, and they way they do it is incredibly cool. GlyphGryph fucked around with this message at 20:36 on Apr 7, 2023 |
|
#
?
Apr 7, 2023 20:33
|
|
|
As an example:What is a solid strategy for mission 3 of Myth 2 Soulblighter posted:Mission 3 of Myth 2: Soulblighter is called "The Watcher". In this mission, you must defend a dwarven stronghold against waves of enemies that will attack from multiple directions. Here are some tips for developing a solid strategy: I'm assuming this is how people who actually know their field feel about GPT writings on their stuff. Still a neat toy though
|
|
#
?
Apr 7, 2023 21:23
|
|
|
GlyphGryph posted:Counterpoint: ChatGPT actually is very intelligent, because the way it strings words together into a coherent whole without having an referents or understanding of the content is actually an incredibly difficult thing to do, its just a very narrow sort of intelligence. It's only mimics real meaning, sure, but it does it incredibly well, and they way they do it is incredibly cool. It's certainly an impressive toy, but it's not intelligent. Calling ChatGPT "intelligent" is like thinking your TV is intelligent because it marvelously shows you pretty pictures. It also only mimics meaning "well", if you're an idiot. The stuff they vomit out is a good source of comedy, though. I'll give you that.
|
|
#
?
Apr 7, 2023 22:01
|
|
|
GlyphGryph posted:Counterpoint: ChatGPT actually is very intelligent, because the way it strings words together into a coherent whole without having an referents or understanding of the content is actually an incredibly difficult thing to do, its just a very narrow sort of intelligence. It's only mimics real meaning, sure, but it does it incredibly well, and they way they do it is incredibly cool. https://en.wikipedia.org/wiki/Chinese_room Whether or not this qualifies as intelligence is a very hotly debated topic in the realm of AI right now wiegieman posted:It could only repeat text descriptions of decisions that human beings had already posted to the internet. This is not true though. The mechanisms through which gpt works aren't fully understood by literally anyone on this earth, but it is definitely not simply repeating posts humans have made. It learned from the posts that people have made, but its scope is way beyond what those posts said. deep dish peat moss fucked around with this message at 22:12 on Apr 7, 2023 |
|
#
?
Apr 7, 2023 22:08
|
|
|
Jeb Bush 2012 posted:this is either meaninglessly true or clearly false, depending how you define the state space of the supposed "markov chain" Yeah it's smug nonsense.
|
|
#
?
Apr 8, 2023 00:20
|
|
|
deep dish peat moss posted:This is not true though. The mechanisms through which gpt works aren't fully understood by literally anyone on this earth, but it is definitely not simply repeating posts humans have made. It learned from the posts that people have made, but its scope is way beyond what those posts said. It's learning form the amalgam of all posts people have ever made. It learns what kinds of responses usually follow what questions (ie: the format of the posts, the language used, etc), and spits out on the fly the equivalent of a form letter with some mad libs style substitutions based on what you specifically asked for. The fact that it's generating a form letter out of the average of all English texts is sort of impressive, to be sure, but ultimately it's just spitting out common phrasings that it's learned from other places. Clarste fucked around with this message at 00:42 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 00:38
|
|
|
Veryslightlymad posted:
Off topic, but I find almost every public facing discussion around this infuriating as an archaeologist. A large portion of that is that depending on the reporting its not clear at all about the timeline being discussed.
|
|
#
?
Apr 8, 2023 01:11
|
|
|
ChatGPT is about as intelligent as a Civ AI (that is, it isn't), it's just better at doing what it was designed to do than any Civ AI.
|
|
#
?
Apr 8, 2023 01:36
|
|
|
Clarste posted:It's learning form the amalgam of all posts people have ever made. It learns what kinds of responses usually follow what questions (ie: the format of the posts, the language used, etc), and spits out on the fly the equivalent of a form letter with some mad libs style substitutions based on what you specifically asked for. The fact that it's generating a form letter out of the average of all English texts is sort of impressive, to be sure, but ultimately it's just spitting out common phrasings that it's learned from other places. It's more about interpretation of data. It can interpret data in ways that humans never could and spot patterns in it that Humans wouldn't be able to. So while the responses might be synthesized based on human-written text, it's the connection between data points and response that separates its "intelligence" from the data it was trained on. Doctors are going to be one of the first jobs to be fully replaced by AI because it is literally impossible for a doctor to keep up with the speed of medical research. A ML-driven Civ AI would absolutely crush 99.99% of players right now if that was its goal, because it could correlate the data from across millions of games to devise an unbeatable strategy and prepare itself for every counter that players throw at it. That kind of AI wouldn't be fun to play against for the vast majority of players. That's not because it's a "tactical genius" or whatever, but just because this theoretical AI has watched every Civ game ever played. Some players would be able to beat it, sure, but over time the tactics they used would stop working. deep dish peat moss fucked around with this message at 01:55 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 01:51
|
|
|
Super Jay Mann posted:ChatGPT is about as intelligent as a Civ AI (that is, it isn't), it's just better at doing what it was designed to do than any Civ AI. ChatGPT is a pair of learning networks built on a complex human feedback foundation trained against a vast and immense dataset. It's not an AGI in any shape or form, but arguing that it's no more intelligent than the Civ AI requires a definition of intelligence that I can't imagine is very useful. The Civ AI is never going to get better at the game unless you go in and change the code with explicit intent by someone who understands the game very well. The vast majority of the reason ChatGPT does what it does so well has nothing to do with any type of explicit design or human understanding of the problem space. Its because it can learn and improve without any human involvement at all. What metric of intelligence leaves an algorithm capable of learning and adapting and improving itself as "no more intelligent" then a fixed, unchanging heuristic algorithm?
|
|
#
?
Apr 8, 2023 02:27
|
|
|
If you define intelligence as the ability to predict the outcome of actions and use those projections as a basis for decision-making (which I think maps pretty well to a human understanding of "intelligence"), then a neural net is just as non-intelligent as a pre-programmed set of decisions. Both are less intelligent than monte-carlo tree search. Interestingly, it turns out that neural networks alone are pretty mediocre at playing games, even games with very simple rule sets. But using them to pick candidate moves to plug into MCTS, essentially "grafting on" the ability to predict how specific moves will turn out, is what makes them wildly successful.
|
|
#
?
Apr 8, 2023 02:48
|
|
|
deep dish peat moss posted:It's more about interpretation of data. It can interpret data in ways that humans never could and spot patterns in it that Humans wouldn't be able to. So while the responses might be synthesized based on human-written text, it's the connection between data points and response that separates its "intelligence" from the data it was trained on. Doctors are going to be one of the first jobs to be fully replaced by AI because it is literally impossible for a doctor to keep up with the speed of medical research. There can be medical AIs finding hidden correlations and making crazy-but-correct diagnoses that no human doctor would, but that's not ChatGPT. ChatGPT just spits out words, and has no data other than words. You put in words and words come out, and they all mean nothing because they don't refer to any external reality. It's just playing games with words, and would be terrible at both medicine and Civilization. Edit: Did you watch the video about ChatGPT playing chess? It has no idea what the rules of the game are, just what the format of a chess move looks like. It constantly does illegal and nonsensical moves until it arbitrarily decides that the game has gone on for whatever length of time it considers reasonable and declares checkmate with no basis whatsoever. An AI that only plays with words is less than useless and probably a distraction from making actually useful things. Clarste fucked around with this message at 03:00 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 02:52
|
|
|
ChatGPT actually is pretty useful already for several purposes, although obviously that utility isn't worth the massive cost. But the newest version of GPT already does more than just play with words - it does visual processing as well (although quite poorly, for the time being). Generalized game playing and robotics controllers modules are both on the roadmap to eventually be included and are being actively worked on as well (form what I understand the robotics controller is pretty far along, although I've not seen any word of progress on the gameplaying bit - Meta's been way ahead of them there and still has a long way to go). Will it get there? Who knows. But their goal isn't ultimately to build a chatbot. deep dish peat moss posted:Whether or not this qualifies as intelligence is a very hotly debated topic in the realm of AI right now It absolutely isn't, and it really couldn't be? First of all, it is considered irrelevant by literally everyone in the AI field I've ever talked to. They honestly, sincerely do not give a poo poo about, for what should be incredibly obvious reasons. Your link explicitly says this and also explains why, if you care to read it. Because, y'know, it's not about intelligence, but about consciousness, and AI folks aren't trying to build conscious beings - just ones that behaves in an intelligent manner that can be used to solve problems. GlyphGryph fucked around with this message at 03:45 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 03:24
|
|
|
Clarste posted:The fact that so many people have been fooled into thinking that GPT is even slightly intelligent just because it can string together words grammatically and in an authoritative voice is pretty sad. well yeah who can blame em, for decades they've been asked to accept the proposition that SV business dudes are extremely intelligent on exactly the same evidence A Wizard of Goatse fucked around with this message at 04:58 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 04:44
|
|
|
Neural networks have clear applications for game AI. This has been true for awhile. I have no idea why people are talking about ChatGPT though. The input isn’t text, and the output isn’t text either, so what do LLMs have to do with generalized game playing? I guess in theory you could write a 4X where diplomacy happens by conversing with a chatbot. That would be a cute gimmick, but I don’t think it would make sense in any game that isn’t similar to Diplomacy.
|
|
#
?
Apr 8, 2023 05:29
|
|
|
Let's go back to text parser interfaces.
|
|
#
?
Apr 8, 2023 05:34
|
|
|
wasn't there a while with some earlier less locked-down version of chatgpt where people were doing exactly that, feeding Zork parser commands in and letting it generate the game world on the fly? I would assume the logical games application for this stuff, inasmuch as there is one, isn't super sophisticated AI opponents that can work through a game's mechanics especially well but turning some 3D analogue of Midjourney loose on the assets so units all look unique and maps look way more intricate and incidentally every game for the next couple decades looks like you put every other game in a blender and dumped the homogenized slurry into Steam A Wizard of Goatse fucked around with this message at 06:20 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 05:41
|
|
|
nrook posted:Neural networks have clear applications for game AI. This has been true for awhile. I have no idea why people are talking about ChatGPT though. The input isn’t text, and the output isn’t text either, so what do LLMs have to do with generalized game playing?
|
|
#
?
Apr 8, 2023 06:24
|
|
|
GlyphGryph posted:It's not an AGI in any shape or form, but arguing that it's no more intelligent than the Civ AI requires a definition of intelligence that I can't imagine is very useful. You are absolutely correct, defining intelligence in terms of computer algorithms is, in fact, not very useful, because that's not what computers do. What computers do is compute, it's in the drat name. They're logic processing engines designed to perform math quickly and efficiently. That's incredibly useful, even transcendently so in many applications, but that's not intelligence. Jack Trades posted:Let's go back to text parser interfaces. This but unironically Super Jay Mann fucked around with this message at 07:06 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 07:03
|
|
|
oh god another thread I'm subbed to is now incredibly misinformed people arguing about AI here's a thread for it, guys: https://forums.somethingawful.com/showthread.php?threadid=4027671
|
|
#
?
Apr 8, 2023 08:08
|
|
|
I mean, the best game AIs i've seen for actually complex video games have been stuff like this, where, 30 years after the original was made, the person programming the AI has an understanding of optimal tactics because there's been so much play. The best examples of this are stuff like the AI for Dominion or the Brutal AI for openxcom https://openxcom.old.mod.io/brutal-ai. 4x games are much more complicated, and take longer for players to figure out the optimal lines so that they can program an opponent to play them.
|
|
#
?
Apr 8, 2023 10:51
|
|
|
OT for the rest of the discussion, but I've never understood how the Chinese Room thought experiment is supposed to be an argument against anything. "Can you point to the specific, single neuron in your brain that knows how to speak English? No? Gotcha, clearly you can't really speak English!" -- that's what the argument sounds like to me. Complex things can do things their parts alone can't do, news at 11. Tuna-Fish fucked around with this message at 12:37 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 12:10
|
|
|
GPT is mostly neat in terms of its computational optimization that allows it to digest much larger training sets than traditional methods. At the end of the day it's just trying to build a correlative model. It's handed a corpus of whatever gazillion documents and what it builds is stuff like "when a sentence starts with 'GPT is mostly', the next word might be 'neat'. This is why it's such a good bullshitter. By construction, GPT cannot produce a player that is better than its training set. It's an unsupervised learning algorithm designed to *replicate* its training set. If somehow you hand a GPT algorithm billions upon billions of Civ matches and sort out the "not following the rules of the game" problem, you'll end up with an exactly average player. The only way to get GPT to beat 99.9% of players is to feed it a dataset of only the very best players, and good luck finding a sufficient large dataset of that size, especially considering that it'll get invalidated every time you patch the game. There's various ways to leverage ML techniques to do clever AI stuff (for example, having the AI play against itself), and this might work. (Though to my knowledge it's only been done with games with much more restrictive sets of moves. Consider that Chat GPT operates using about 20kb of conversational memory. Now think how big your Civ savegame files are.) GPT isn't a magic spell that solves all problems. Fangz fucked around with this message at 12:47 on Apr 8, 2023 |
|
#
?
Apr 8, 2023 12:39
|
|
|
GPT is actually supervised and has an alignment layer specifically built to exceed average. But it's also not a game AI, so I don't know why we're still talking about it instead of CORSICA, the one I brought up earlier, that involves a language mode actually layered on top of a game playing model to create a very functional strategy AI thats capable of talking to you.
|
|
#
?
Apr 8, 2023 14:15
|
|
|
thhe best use of chatgpt for 4x games would be having it write teh dialog in the diplomacy screen when the other empire are mad at me
|
|
#
?
Apr 8, 2023 14:19
|
|
|
Megazver posted:oh god another thread I'm subbed to is now incredibly misinformed people arguing about AI AI in 4x games has been a problem since day one. So while yes, it’s adjacent I feel it’s still an appropriate discussion.
|
|
#
?
Apr 8, 2023 15:10
|
|
|
Mayveena posted:AI in 4x games has been a problem since day one. So while yes, it’s adjacent I feel it’s still an appropriate discussion. Game AI? Sure. Whether ChatGPT is markov chains, maybe not?
|
|
#
?
Apr 8, 2023 15:18
|
|
|
Megazver posted:Game AI? Sure. Whether ChatGPT is markov chains, maybe not? It'll all come together eventually right? 
|
|
#
?
Apr 8, 2023 15:26
|
|
|
You know what game had a sick AI. Offworld trading company.
|
|
#
?
Apr 8, 2023 18:46
|
|
|
Not really a 4X game, though. I respect the design, but I didn't enjoy it. But Old World also had solid AI. Soren Johnson designs and balances all of his games by making MP versions first, which seems to be the trick.
|
|
#
?
Apr 8, 2023 18:52
|
|
|
ChatGPT is a really fun technology that costs 2200 beakers and allows you to have some great fun watching a lot of people absolutely critically fail a Turing Test.
|
|
#
?
Apr 8, 2023 22:48
|
|
|
Megazver posted:Not really a 4X game, though. I respect the design, but I didn't enjoy it. But Old World also had solid AI. Soren Johnson designs and balances all of his games by making MP versions first, which seems to be the trick. I can't recall the specifics, but my understanding is that this actually caused some issues. There were some things vastly preferred by anyone playing SP that it took heroic effort to convince the 'mp mafia' to support because of the different perspective. Old World's AI has some good points, but it's ultimately in an awkward place design-wise. Adjusting the difficulty up gives the AI more starting cities, so in the early game the AI will significantly outnumber the player. The player effectively has control of whether you get into early wars with the AIs so this isn't a death sentence like it would be in some games, but since the player is still a lot better at growth and fighting it means that every game is just a curve situation where you are outmatched early, grow hard, and hit a point in early-midgame where you are superior to the AIs and go dunk them and the game is effectively over. The high # of starting cities also means that the AI easily bullies the Tribes (sort of enhanced barbarians) and makes a hobby of clearing their city sites, rendering the tribal mechanic somewhat pointless.
|
|
#
?
Apr 8, 2023 23:03
|
|
|
TheDeadlyShoe posted:since the player is still a lot better at growth and fighting it means that every game is just a curve situation where you are outmatched early, grow hard, and hit a point in early-midgame where you are superior to the AIs and go dunk them and the game is effectively over. 4X_AI.txt
|
|
#
?
Apr 9, 2023 00:44
|
|
|
|
| # ? May 31, 2024 13:47 |
|
|
orangelex44 posted:4X_AI.txt I mean, in theory it's possible to have an AI curve similarly to a player, through a combination of good programming and intelligently distributed bonuses (as opposed to just letting them start with more or get other flat bonuses). Some day I hope to see it, too.
|
|
#
?
Apr 9, 2023 02:44
|
|
