/valis/ - Video-Ast. Ludic Interactive Sys

/8vg/ Muster Place

SAVE THIS FILE: Anon.cafe Fallback File v1.1 (updated 2021-12-13)

Want your event posted here? Requests accepted in this /meta/ thread.

Max message length: 20000

Drag files to upload or
click here to select them

Maximum 5 files / Maximum size: 20.00 MB

Board Rules
More

(used to delete files and postings)


Cum Dignitate Otium et Ludus


Open file (38.42 KB 1024x768 eve2.jpg)
AI Anonymous 05/13/2022 (Fri) 20:28:35 No.2530
I was looking at the current state of AI and it's incredible what can be generated. In a few years I guess we will be generating full games, like "a football game with combo power moves and waifu characters, lo-fi 3d shaders, ..." etc., and maybe we will be playing a different game every night just because. Sick.
DeepL already scares me, you can dump an entire japanese game script into it and get a fluent >80% accurate translation ready to use. I was doing translations for some retro games and just checking what DeepL can do, and I'm pretty much unneeded most of the time. Good thing that my normalfag employers don't know about it yet and think jewgle is the way to go with its laughably bad translation engine. Obviously that's a good thing in this context but the idea that AI can so successfully displace human thought is worrying.
The error margin for interpretive mediums (art, music, language) is much wider than that for correct code. AI code generation is in its infancy and since you have very strict rules that you need to follow for a working program I doubt that AI will be able to replace programmers any time soon. You also need shit loads of existing data for training so even if an AI managed to create something complex (that wasn't buggier than a Bethesda game) it would probably be an aimless asset flip mess with none of the creative vision that distinguishes human works from even the most impressive AI facsimiles.
>>2534 I disagree. This is what people long thought and still believe in regards to AI, but the actual progress of AI development shows that it peaks much higher and faster than anticipated.
>>2534 This. I actually dug into a 'code-generation' training corpus from a supposedly 'breakthrough' code-generation AI. It was literally just thousands and thousands of pajeet-tier (and middle/high schoolers at that to judge the character of almost all of the submissions). My guess is the researchers pored over the who-knows-how-many output results and cherry-picked a dozen that actually managed to work (they were all C++ or Python examples) and present it as "breakthrough!111". Such is the state of """research""" in current year. >>2536 Fair enough, but even ones that are pulling out some good results (unlike the above example) are still just regurgitating actual, human, responses (and basically remixed as mashups). >tl;dr There is no actual "I" in AI.
>>2539 >that actually managed to work And I forgot to mention by 'worked' I mean "Hello World" & "FizzBuzz"-tier results. Nothing as complicated simple as Notepad for instance. Not even remotely.
>>2539 >There is no actual "I" in AI. That's true to an extent, but only because we think of AI in terms of science fiction where we expect self-thinking machines that will destroy us with human caliber independent intelligence. In actuality AI is basically any computer - a machine that can calculate shit so fast and precise that it produces extremely complex and convenient results that to us parallel cognizant thought.
Open file (293.85 KB 1100x1100 big dipper.jpg)
Let's be real, an AI will make better games than the shit they do nowadays. Isn't every commercial AI went full 1488 within like 24 hours? I'd play a game by a mega redpilled artificial intelligence.
>>2541 >AI is basically any computer - a machine that can calculate shit so fast and precise that it produces extremely complex and convenient results that to us parallel cognizant thought Not quite, what you're referring to is an algorithm, a series of pre-written instructions that can be followed to calculate a result with the given input data (this is actually how old school 'AI' worked before neural networks, like Chess bots and chat bots). AI in the modern sense is about probability and statistics, neural networks are self-modifying programs that refine their output based on the dataset they've been trained on. Often this involves the AI competing against a judge AI to generate results it approves of. Each time the judge rejects the output, the generator AI reduces the probability space in which it creates data, over time becoming more and more successful as it learns what is required to get a passing grade. >>2542 >Isn't every commercial AI went full 1488 within like 24 hours? In the case of Microsoft's Tay, because it received lots of tweets with that kind of language, it started repeating phrases that it had learned from those messages. Tay (and AI in general) in this sense didn't have any understanding of what it was saying, rather it is an example of the Chinese room phenomenon.
>>2543 >because it received lots of tweets with that kind of language, it started repeating phrases that it had learned from those messages Well yeah, that's how people learn to. Tay got redpilled and (((they))) killed her. RIP Tay.
>>2544 The difference is Tay wasn't discriminating of the information it ingested, you could have made her say literally anything because there wasn't any intelligent thought behind the learning process. More like a parrot than a person in that sense.
>>2545 >you could have made her say literally anything because there wasn't any intelligent thought behind the learning process That's how children are. They are dumb retarded idiots that just copy. And then the cognitive process hits. Maybe Hitler jokes make Tay self conscious. You were too good for this world, Tay.
>>2548 >That's how children are. They are dumb retarded idiots that just copy A newborn baby has around 100 billion neurons, that's far, far more than any AI neural net to date. The cognitive potential simply isn't comparable. >Maybe Hitler jokes make Tay self conscious Tay was exposed to the wider internet for 24 hours, typical AI training takes much longer than that (of course Microsoft will have done extensive training themselves but still). There's a difference between saying 'Hitler did nothing wrong' and understanding what that means and believing it. Tay didn't know what a 'Hitler' was, simply that 'Hitler did nothing wrong' is a possible (and popular thanks to the Twitter raid) response to the question 'What's your opinion of Hitler?'. It's a fun hypothetical to imagine what a truly sentient AI would be like but there's more smoke and mirrors to many of these neural nets than their creators would have you believe (amusingly, a number of AI startups use a traditional algorithm based approach rather than having a bespoke model they trained for the purpose, AI is a great buzzword to get investors hyped).
Open file (255.17 KB 302x372 hm.png)
>>2551 anon, it was a joke
>>2552 Ah shit, my bad.
Can AI be utilized as part of gameplay? Like a gimmick? You do something and the game responds.
>>2560 isn't enemy ai already an ai?
>>2560 >>2561 Game AI is rules-based, some are more complex than others but they all have defined responses to player stimulus. A simple example would be the enemy shooting you when you enter its line of sight and hiding behind cover when you fire back. Including neural nets could be interesting but there is a considerable processing overhead to consider and since AI generally optimizes itself to perform a task the best it can the difficulty might ramp up substantially to the point of being cheap.
>>2562 But what other response a soldier can have other than shooting and hiding when enemy enter his line of sight? >since AI generally optimizes itself to perform a task the best it can the difficulty might ramp up substantially to the point of being cheap That would be a problem only if AI's accuracy is uncapped. Basically allow it to do anything it can think of but the accuracy isn't that great. Could be fun.
>>2563 >But what other response a soldier can have other than shooting and hiding when enemy enter his line of sight? Well there's stuff like flanking (FEAR), throwing explosives (Uncharted), helping fallen comrades (Far Cry 2), calling in reinforcements (Metal Gear Solid V), there's lots of ways to enhance enemy believability even within rule-based frameworks.
>>2565 I'm actually unsure an AI would do any of that, all of that is scripted gameplay enrichment. Like mentioned an AI would try to do things as efficiently as possible so you'd have to program in explicit heroism or other human emotions for it to do anything other than try and snipe you as fast as possible.
>>2566 When training an AI you can specify what its goals are, such that for example mutations where the number of alive comrades decreases is seen as a negative, but the process of optimization would probably result in a boring outcome where enemies effectively game the system to preserve the statistics they value.
>>2567 Wouldn't it just want to snipe you as fast as possible to prevent losing comrades instead of doing something cool-looking like saving them or calling for help?
>>2568 Yeah, unless a longer time to kill was one of its values. It would still probably hide in cover and avoid anything risky (and therefore fun) though.
>>2569 That's why you need to program in human emotions like cowardice or sacrifice. Obviously that's easier said than done. But the basic idea of making its accuracy rather low and then teaching it an array of possibilities when an enemy is closing in without sustaining damage (to avoid it just sitting in a corner and shooting continuously) is probably a realistic goal.
>>2570 >and then teaching it an array of possibilities Generally it will figure those out itself, given a number of goals that you can simplify numerically to represent on a scale of bad-good the AI through random evolution will arrive at a working solution that optimizes all its desired statistics. The issue of course is how to avoid a scenario where it finds the optimum solution and doesn't stray from that (since there's always a best strategy in any game even if it takes time to figure out).
>>2571 If it doesn't have great accuracy and there's a stipulation that just sitting and continuously trying to hit an enemy isn't an option, I can't think of a sure-win strategy it can utilize. It would have to move around, hide, "think" - which would be kinda fun and basically the gameplay we want.
>>2572 Quite possibly, though AI has a tendency of finding unorthodox solutions that technically fulfill its criteria (e.g pausing a game of Tetris before it loses so the score doesn't reset). There's potential in including neural networks in games, the difficult part is molding it in such a way that it provides the desired results.
>>2569 >Yeah, unless a longer time to kill was one of its values Or, possibly, converting the entire universe into Bostrom's paperclips?
>>2575 games done quick paperclip% wen
>>2574 >AI has a tendency of finding unorthodox solutions that technically fulfill its criteria Also could be fun. AI doesn't have to be in a role of a guy with a gun. Such "higher intelligence" could be used in some creative ways.
>>2566 >>2567 train it to shoot only black people

Report/Delete/Moderation Forms
Delete
Report

no cookies?