General Gaming News and Discussion

  • Guest, it's time once again for the massively important and exciting FoH Asshat Tournament!



    Go here and give us your nominations!
    Who's been the biggest Asshat in the last year? Give us your worst ones!
C

Captain Suave

Caesar si viveret, ad remum dareris.
5,256
8,953
Mist said:
Linking this again.


TL;DR - 70 years of AI research has demonstrated that trying to make AI 'smarter' via clever tricks and cognitive hacks produced effectively 0 results. People are under the mistaken assumption that this new batch of AI is new tech, but Deep Learning is actually a return to a very old approach that just wasn't feasible at the time. "The Bitter Lesson" basically says that the only way to make anything resembling 'smart AI' is to just throw a ton of data and compute at the problem--aka there are no shortcuts to smart. The reason the latest round of models work as well as they do is because they collected basically all of the coherent data they could for the domains they cared about, then shoved it through a tremendous amount of compute to distill it.

To build models twice as smart, they'll likely need ~20x the data and ~20x the compute. The compute is 'easy' so long as we continue building chip fabs and avoiding a war over Taiwan. But there's very unlikely to exist 20x the high-quality data you'd need, unless humans keep producing new, high-quality data.

Do you see the trap here? As the internet gets more filled with mediocre goyslop, it becomes harder to build future smarter models using this method, with a high probability of leading to a doomloop of mediocrity as the models starve on their own shit. And 70 years of research has shown that none of the shortcut methods work.
Click to expand...

While the math is old, the engineering of it is new because we haven't had computing power or data scale to execute on this level until very recently. As far as whether it represents any kind of actual intelligence or pegs to some metric of "smartness", I don't think that's relevant to whether the output voice quality will be good enough for video games. I'm not talking about progress towards true general AI, I'm saying that these more limited tools will continue to improve in output quality.

The reason these models use vast quantities of data is that most of it SUCKS because it comes from the random cesspool of the internet. The signal to noise ratio is laughable. Curating a purpose-built, high-quality data set that is still large enough for training on the narrow task of something like voice replication will take a lot of human time and that's why no one did it. But someone will do it in the future, and the resulting model will be better.
 
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
popsicledeath said:
I didn't say a society programmed by algorithms and AI is good, you dusty potato. I said it sucks and is bad for everything, but that won't have any bearing on its success or reception.

For thinking you're so smart, you have really shitty reading comprehension.
Click to expand...
Na, I got it, we're on the same page basically. You've just got no fight left in you. I see how it is.

Do not go gentle into that goyslop night, friend.
 
  • 1Worf
Reactions: 1 user
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
Captain Suave said:
While the math is old, the engineering of it is new because we haven't had computing power or data scale to execute on this level until very recently. As far as whether it represents any kind of actual intelligence or pegs to some metric of "smartness", I don't think that's relevant to whether the output voice quality will be good enough for video games.
Click to expand...
It's already good enough for video games, it's just not good enough for good video games and likely never will be. But if we let companies shove shitty video games at us, that's on us.
 
C

Captain Suave

Caesar si viveret, ad remum dareris.
5,256
8,953
Mist said:
and likely never will be
Click to expand...

And that's where I disagree. That future is coming, and fast. I don't think intelligence is actually required to generate dramatic voiceovers with whatever you think of as "human connection", just some more targeted training on more curated data and secondary modules/features around the main model. Generalization doesn't matter. Averaging doesn't matter. We're not looking for NEW modes of expression in human voice in the same way you'd want new thought from a sentient LLM. There are no new modes. Human acting has been working the same space for thousands of years. We just need something to help pick from performance points within the existing known landscape.

In less than a decade we'll have off-the shelf products where I can type in a short prompt describing some context and direction, provide a script, and I'll get a result that indistinguishable from what I'd get from human actors. It'll take some labor to direct, just like with real people, but will save a lot of the overhead of needing studios, scheduling, etc, etc. Properly parameterized it will run in real time for virtual agents and/or NPCs in games.

And we'll all love it because it will be good.
 
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
Captain Suave said:
In less than a decade we'll have off-the shelf products where I can type in a short prompt describing some context and direction, provide a script, and I'll get a result that indistinguishable from what I'd get in human actors. It'll take some labor to direct, just like with real people, but will save a lot of the overhead of needing studios, scheduling, etc, etc.
Click to expand...
I think DickTrickle DickTrickle was correct in that doing this for any game with a lot of dialogue, the prompt engineering would just be harder than hiring real people.

But sure, with 10 years and enough GPUs to melt a battleship, yeah, someone could probably Derek Smart themselves their own one-man game. That doesn't mean they should. Working with teams of actual human beings creates compromises and generates unique ideas that ultimately produce better results.

And yes, I believe it's technically possible some superstudio could arise that would use large teams of humans AND powerful AIs to build something truly inspired, but if people are willing to buy slop instead, that will never happen.
 
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
Captain Suave said:
I doubt it's that much, but either way it's only got to be done once and then scales at basically zero marginal consumption for output in perpetuity.
Click to expand...
No that's not how that works either. Large, powerful models need a lot of compute during pre-training and fine-tuning, but they also require less large, but still large, compute instances to produce output from prompts. GenAI is expensive on both ends, training and runtime, vs other lesser forms of AI, which is why you see enterprise adoption still very, very low.
 
C

Captain Suave

Caesar si viveret, ad remum dareris.
5,256
8,953
Mist said:
No that's not how that works either. Large, powerful models need a lot of compute during pre-training and fine-tuning, but they also require less large but still large compute instances to produce output from prompts. GenAI is expensive on both ends, training and runtime, vs other lesser forms of AI, which is why you see enterprise adoption still very, very low.
Click to expand...

ChatGPT maybe, but not all models need to be that large and you can run image and audio generation on local hardware no problem with execution times faster than you can get a cup of coffee. I've done it.
 
Last edited:
popsicledeath

popsicledeath

Potato del Grande
7,547
11,831
Mist said:
Na, I got it, we're on the same page basically. You've just got no fight left in you. I see how it is.

Do not go gentle into that goyslop night, friend.
Click to expand...

Fight? What fight is to be had? I can hope for the betterment of society in one hand and Fecebook in the other. The problem is we know which fills up first, and increasingly which one we'll be forced to eat.

Demoralized people check out and rely on appeasement of the dissonance they feel. Demoralized people also pretend they're fighting the good fight, when really there is no fight. Staring with understanding and introspection into the abyss and coming to some form of acceptance of the inevitable and how to survive it isn't giving up the fight. There is no fight, that's perhaps just a thing people believe for copium.

Or, sure, start trying to explain to the people in your life that everything is a scam. Can't get most people to not have retarded TV setups or not wear their masks alone on a hiking trail, but don't give up fighting to inform them how AI sucks and is detrimental.

There is no fight to fight.

And interesting perspective can be gained from people interested in these subjects by asking them how Fahrenheit 451 ends. Obviously not a perfect test, but interesting. A lot of people, even who love the book and the genre and discussing such topics, don't remember how it ends or remember strange false endings. Have you read it?
 
Ambiturner

Ambiturner

Ssraeszha Raider
16,043
19,530
Pyros said:
The "problem" of removing all the shitty voice acting jobs that are not very well paid where they just do 1, maybe 2 takes and done with unknown voice actors is that it removes the jobs the newbie voice actors need to pay their bills until they train well enough to be good. Sure if they remove them tomorrow you have all these good voice actors already so it's fine, but there won't be anyone to replace them if every shit low tier job is being replaced by an AI.

The quality isn't going to get better either, the reason a lot of voice acting is shit is just cause the producer doesn't give a shit or the suits decided there was only going to be x$ to do the voice acting when realistically you'd need 3x or 4x to actually do a good job, and AI isn't gonna help on that since it'll likely be the same thing where you can use a good AI or a shitty cheap AI, and then not do any other work behind to make sure it's good or consistent.

The companies pushing AI for this type of stuff aren't the small indies trying to just make their first game either, it's the big companies just trying to cut more costs so they can give more money to their shareholders and CEOs. You can bet Ubisoft and EA will add AI text as soon as it's semi viable just so they can keep churning out the same game every year but for even less cost.
Click to expand...

The idea that they should have to pay people to do a shitty job so that they can maybe somebody not be shitty doing that job for someone else is a terrible argument for stifling innovation.

More likely you'll get AI for background characters and random quest shit with legit voice actors for the bigger/more intricate roles.
 
  • 1Like
Reactions: 1 user
popsicledeath

popsicledeath

Potato del Grande
7,547
11,831
Mist said:
It's already good enough for video games, it's just not good enough for good video games and likely never will be. But if we let companies shove shitty video games at us, that's on us.
Click to expand...

Yeah, guys, we just gotta keep fighting for good games!

Sorry, you're like 15 years behind the times.

Not saying you have to like the shit being shovelled in our direction. Not saying you're not right, it is on us, but we're often responsible as a society for things that aren't our fault as individuals. Not saying don't support the good and shit talk the bad.

What I'm saying is you seem to be struggling to come to terms with reality and things that are perhaps ironically inevitable because people are fighting some outdated fight because it's so hard to come to terms with reality.
 
TBT-TheBigToe

TBT-TheBigToe

<Rickshaw Potatoes>
<Gold Donor>
10,195
30,080
Ambiturner said:
The idea that they should have to pay people to do a shitty job so that they can maybe somebody not be shitty doing that job for someone else is a terrible argument for stifling innovation.

More likely you'll get AI for background characters and random quest shit with legit voice actors for the bigger/more intricate roles.
Click to expand...
AI generated and maintained radiant quests like in Skyrim, the Brotherhood repeating quests in Fallout 4, shit like that, would be kind of cool if they can generate enough branches that it doesn't always amount to "go to place and do thing" like it currently does. I mean, it can't really end up worse than the lazy repeating quest designs we have now...

can it?
 
Kirun

Kirun

Buzzfeed Editor
19,236
15,625
Mist Mist is that boomer in 1999 at CompUSA that sold you your first PC - "You'll never use more than 100 MB of hard drive space!"

The future is now, old man.
 
  • 1Worf
Reactions: 1 user
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
popsicledeath said:
Sorry, you're like 15 years behind the times.

Not saying you have to like the shit being shovelled in our direction. Not saying you're not right, it is on us, but we're often responsible as a society for things that aren't our fault as individuals. Not saying don't support the good and shit talk the bad.

What I'm saying is you seem to be struggling to come to terms with reality and things that are perhaps ironically inevitable because people are fighting some outdated fight because it's so hard to come to terms with reality.
Click to expand...
I buy very few video games, because I've been burned so many times. 95% of the stuff I play comes out of my grandfathered-forever Humble Bundle sub.

The only full-price release day games I've bought since become a full-time corpo drone are both Horizon games (great, okay), Destiny 2 (bad), Diablo 4 (fucking terrible), BF2042 (bad but eventually good) and that might be it idk.

EDIT: I lied, X-com 2 (fucking excellent), Bloodstained (very good).
 
Regime

Regime

LOADING, PLEASE WAIT...
<Aristocrat╭ರ_•́>
16,866
40,093
Mist said:
I buy very few video games, because I've been burned so many times. 95% of the stuff I play comes out of my grandfathered-forever Humble Bundle sub.

The only full-price release day games I've bought since become a full-time corpo drone are both Horizon games (great, okay), Destiny 2 (bad), Diablo 4 (fucking terrible), BF2042 (bad but eventually good) and that might be it idk.

EDIT: I lied, X-com 2 (fucking excellent), Bloodstained (very good).
Click to expand...
Didnt you tattoo I ❤️ Mass Effect Andromeda on your arm?
 
  • 1Worf
Reactions: 1 user
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
Regime said:
Didnt you tattoo I ❤️ Mass Effect Andromeda on your arm?
Click to expand...
No, and I didn't buy that on release either. I got that for 10 bucks or less like 2 years later, and played it fully patched.

And I still never finished it because Horizon Zero Dawn released on PC and made me immediately regret enjoying ME:A.
 
Daidraco

Daidraco

Avatar of War Slayer
10,054
10,381
Mist said:
No, and I didn't buy that on release either. I got that for 10 bucks or less like 2 years later, and played it fully patched.

And I still never finished it because Horizon Zero Dawn released on PC and made me immediately regret enjoying ME:A.
Click to expand...
Ive tried, just out of my love for the franchise, to finish Andromeda like four times now.. each time, I just.. totally forget about playing the game. Never make it further than the water lizard people planet, which is what .. like an hour or two in?

I get what you're saying about AI though - it'll be years before it replaces generic NPC quest dialogue and voice lines. Its just so.. rough and unrefined right now. So yea, I agree - we're going to get a whole shit load of "that" level of quality before we ever get anything worth a fuck. Procedurally generated content, even in triple A titles, to this day, is still a fucking shit show. The more I learn about language models though, I just dont understand why we call it "AI".. as it just doesnt seem to fit the definition of what "I" think AI is.
 
  • 2Like
Reactions: 1 users
Mist

Mist

REEEEeyore
<Gold Donor>
31,199
23,382
Daidraco said:
The more I learn about language models though, I just dont understand why we call it "AI".. as it just doesnt seem to fit the definition of what "I" think AI is.
Click to expand...
Right. This was my evolution. It was extremely neat for the first 4-5 months after GPT 3.5 smashed onto the scene, but as I started working with our chatbot developers at work and setting up Bedrock instances, and then dug in doing a lot of my own actual research... its really just very clever (and often unreliable) mimicry software built by pushing huge stolen datasets through tons of compute. The people who originally called it a "stochastic parrot" were correct, but the most accurate description is "bullshit bot." And the people who are the most susceptible to bullshit seem to be the people eating it up, which is unfortunately ~80% of the population.
 
mkopec

mkopec

<Gold Donor>
26,228
39,934
I would take a fucking robo voice with no emotion or acting at all over no voice so I dont have to sit there and read reams and reams of dialogue and other writing in some RPG games.

On that note, some streamers have some AI voice shit in their streams, one that comes to mind is Gladd, wich has several AI based text to speech plug ins for his chat. One of them is a pirate and it sounds fucking good. You basically pay a few bucks and it takes your text to pirate speech on his stream.
 
C

Captain Suave

Caesar si viveret, ad remum dareris.
5,256
8,953
Mist said:
its really just very clever (and often unreliable) mimicry software built by pushing huge stolen datasets through tons of compute. The people who originally called it a "stochastic parrot" were correct, but the most accurate description is "bullshit bot."
Click to expand...

I don't actually disagree with any of that, except it's not clear this is going to end up legally classified as infringement or how good/useful the output will eventually get.
 
You must log in or register to reply here.