/sci/ - Science & Math

>>9847924
More impressive than Deepmind TBh

>>9847924
>video games
That's pretty cool, but when is AI gonna do cool stuff like take over the world?

>>9847935
Already happened

>>9847924
A lot of what you posted is pure nonsense.

>>9848144
I'm literally posting things from the blog post lmao

Human-level AGI by 2020. Thngs are advancing faster and faster. Singularity deniers BTFO.

Impressive. I'm curious to see if this technology is integrated into video games in the near future. You could have genuinely challenging enemy AI in single player games.

>>9848239
AI is already better than people at games. A such game would not sell because players won't be able to complete the game.

>>9848246
Nothing really stops you from tuning the difficulty of an AI. Just use a version frozen at the specific skill level you want.

a bigger near future challenge is the fact that there just isn't mass access to tons of TPU resources

>>9848225
That's barely over a year from now. We're not even close to being close.

>>9848246
>AI is already better than people at games.
No it isn't. When you up the difficulty in a strategy game the ai just "cheats", eg. has more resources than the human player

>>9848246
>>9848265
t. brainlets

Conventional AI systems that are within games, built by the developers, have complete knowledge of the game and are built on very simple rules. For example a simple AI in a game could patrol random points, and chase you when it sees you. All of this can be done EZ PZ by any amateur and is NOT machine learning or neural networks. The game playing AI systems you've been seeing online learn to play games like humans, by observation and experimentation.

>>9848204
Yes and it's garbage

>>9848239
Not really, AI in games is meant to play a certain immersive role. Too have it be competitive you end up with CSS sort bots hopping around getting instant headshots.

>>9848347
Ok genius

>>9848239
>>9848256

Yeah I mean hasn't it been learning starcraft for like a year now?

Although it would be cool if they put the bots into random quake arena matches and checked to see whether people could distinguish them.

Jesus this graph scares me.
Considering it has no advantage over humans in terms of inputs/outputs to learn from (watching a screen/mouse n keyboard) and shows linear improvement which far exceeds humans. This AI could replace humans in everything they do. And I'm a little worried in the reward mechanism. The only programed reward is who has the most flags after 5 minutes? What happens when this thing breaks the sandbox and cuts power to enemies computers in order to win?

>>9849009
>What happens when this thing breaks the sandbox and cuts power to enemies computers in order to win?

That's not how any of this works at all. Don't be hyperbolic. And what do you mean, the linear improvement is awful ~100k games to match an average human.

deep reinforcement learning is a greater bubble than dotcom

>>9849378
No.

>>9848360
It really is though. Shit is laughable.

>>9849033
Agreed, you can see in most of the scenarios the AI's didn't shoot unless the enemy was either trying to kill them, or holding the flag and stopping them with scoring. And sure taking a year of non-stop games to beat the average human isn't impressive, but it only took two years to beat the top 20% of players. It's also pretty well known that in terms of elo the vast majority of gamers bottom out and stay there, while the agent just gets better and better, and after 4 years it can beat pros as often as pros beat everyone else. So take any job that's about as complex as playing CTF, after two years the AI will be as good as your average workers, and after four years would have perfected the job beyond human ability.

>>9849617
The scary thing - the speed mentioned is not actually human years. A GHz processor makes a lot of subjective timing changes possible. Also, parrallel.

And another thing - this knowledge isnt going away. It will only improve. Any bot - say a fpsrpg playing one - can use the massively refined knowledge of a beyond-world-class Ctf playing both and use it to evolve. The same holds true for other learned an refined behavior, at virtually no cost.

>>9849009
>And I'm a little worried in the reward mechanism.
the reward is just maximize flags, an AI is just an extremely non linear function along with an algorithm to find a global optimal. It cant do anything beyond what the algorithm can optimize and it wont 'want' to do anything that doesn't maximize the function you give it.

>>9849625
Well I mean each game is 5 minutes long, and the graph is number of games vs elo. 100,000 games played consecutively is almost a year.

>>9849769
Who says they were consecutive?

>>9849625

> Any bot - say a fpsrpg playing one - can use the massively refined knowledge of a beyond-world-class Ctf playing both and use it to evolve

No. Not yet.

>>9849845
Who said you could speak out of turn bitch

>>9849625
No you're wrong the "knowledge" is completely useless. Machine learning has practically 0 transfer knowledge ability right now. It's only good at what you train it for. You gotta wipe it out and retrain it for anything else.

>>9849857
>>9849991

I did mean that in the future sense. I however have personally done experiments where the top few layers are retrained in CNN's, and even have converted an fMRI-retinotopy experiment classifier using CNN's to work in machine vision.
Mind you to generalize this is harder, but it is not impossible. Once you find neural representations of goals - or their place in the network - it's very much feasible to try to swap those goals or replace them and make use of the prior-achieved network architecture.

But you are correct, not yet.

the day an AI beats Sunlust on UV is the day I sudoku myself

>>9848225
LOL, we would be lucky to get AGI by 2040 you fucking idiot.

>>9848270
>All of this can be done EZ PZ by any amateur and is NOT machine learning or neural networks.
implying machine learning or neural networks can't be done by any pajeet with a computer nowadays

>>9851773
True DESU senpai, but the bleeding edge of research is not a brainlet's game.

I did my undergraduate thesis on some basic method for image-based learning on games and it's bullshit, reinforcement learning will only ever give us lap dogs, butler-robots and so on, you get the deal. Humans are not born tabula rasa, we do not learn everything through experience, read some Kant.

>>9852719
>Kant
>Knowing even a singular iota of science

Would it be possible to make this kind of AI "human-like", i.e. introducing the ability to make mistakes? Ideally without just telling it to aim wrong every now and then

>>9852719
I think soon we'll see AI's move over to go beyond just reinforcement. Cause you're right, there needs to be a hierarchy of knowledge that they can inherit and build upon.

>>9852737
Nah, you just purposefully go in and decrease accuracy and speed. Mistakes can happen just from the AI not being trained properly or the algorithms just being limited tho.

>>9848246
It’s not just about difficulty. Imagine if dialogue with other characters was unlimited and they never repeated phrases over and over or forgot about your previous actions. Imagine if those characters could interact with each other and learn information. So if you’re in an RPG and you’ve committed a crime, the authorities can personally track you based on sightings of you in certain locations. Idk about combat, but just this change in the characters’ personality would make games a lot more immersive and enjoyable.

>>9849009
>1300 elo is "expert"
the ai is good but still a long ways away from beating the average pub gamer.

>>9852824
That sounds like an awesome idea for a gta/sims open world game. Instead of the peds/npcs basically being window dressing, they could actually be smart enough to properly simulate a real world experience. On the other hand, a game like that would be very addicting. I mean, why would I want to experience the real world when a game world of that kind would be way more thrilling??? Still, I wonder how far out we are from technology of that nature.

>>9853027
Also real fucking important to consider if you make agi to the point it can understand the world like we do, theres a solid chance you made consciousness. Better fucking be careful about how feelings work and all, could end up with a westworld scenario. Unlikely to be as consequential because its virtual but it would basically be a silent holocaust

>>9853027
>On the other hand, a game like that would be very addicting. I mean, why would I want to experience the real world when a game world of that kind would be way more thrilling???
I think you’re exaggerating a bit. Sure, it would make current game technology seem boring in comparison, but imagine if people in the 90’s were given a video of 2018 games... they would probably say the same as you. But if people do get addicted, then it’s their decision, if they’re hurting anyone then it’s themselves. Everyone should have some discipline, but if I’m going to play 30 minutes a day I want that 30 minutes to be the best it can be.
>>9853074
The characters definitely won’t be conscious, although they will be programmed to seem that way. It’s so immersive that you connect with characters but if you were to ever get overwhelmed you would remind yourself that they aren’t real. However, that is _only_ our conscience that tells us they’re fake; we could still be affected on a primitive, emotional level. But I don’t think any of the effects would be scarring or long-lasting.

>>9849769
https://blog.openai.com/openai-five/
Bots can train for "180 years" per day

>>9847924
>>Had accuracy and speed gimped to match human (wasn't a big determinant of success)
This also means my predictions on my model of consciousness mean the AI war might not happen.
But damn man I'm the strat guy. Now what.

>>9853027
And yeah, I’ve been thinking about the criminal chase idea for a while. If you commit a crime in real life, there is nothing like the feeling of trying to stay free, an all-stakes version of hide-and-seek. There’s just something beautiful about it. I was thinking, that maybe in Elder Scrolls 6, they could implement this system where the authorities would be pretty lax if you were a minor criminal, but after a certain bounty threshold, a special authoritative character will try to find you and arrest/kill you. Everywhere you go, you’d be susceptible to being found. How different it would be if characters were skilled enough to perform sneak shots on you! Fast travel would be restricted, which means everywhere you go would leave a trail, so you’d have to be careful not to get noticed, or seem too odd from the rest of the crowd.

This is just one specific example, I’m sure there are hundreds of new, immersive possibilities that would put single-player games on the level of online gaming

>>9853110
>This is just one specific example, I’m sure there are hundreds of new, immersive possibilities that would put single-player games on the level of online gaming

Pretty much. In my case, if i were to play an open world game that had unique NPCs that could learn, adapt, and interact with the game world in a human-like way, than even the most mundane things like just walking around the world, holding a conversation with an npc, and following an npc as it goes about its day would be an amazingly fascinating experience. Sucks that such a game is probably decades out.