old news, any no videos are available, so completely useless report, last thing I heard is deep mind has over a billion losses, maybe google is going to sell them just like boston dynamics

also the sc2 one was kinda fraud

can you give me a quick rundown on the sc2 ai thing ? i didn't follow it but see it mentioned a lot when comparing pro gamers to ai

https://esreality.com/post/2932544/google-s-d...pid2932673

yeah, I saw that sometime with that special soldier strategy moving them back and forth all the time in microseconds, attack + flee, attack + flee, humans cant execute this particular task that fast I guess

They use blink stalkers, but they're not anywhere that good, yeah.

DeepMind A.I. now plays on the SC2 ladder using all races. Players can opt-in to be potentially matched agaist AI but those bots play anonymously so players can't say when they match them. There are a couple of videos on YT from those games and it is really impressive how human-like they play. In some games the only way to say it's A.I. is lack of control groups.

It feels like it's not about AI outsmarting people in strategy game but more about crippling computer to be human-like if it's even possible.

it's not about AI outsmarting people in strategy game

well, yeah, it's not because they're not limiting it to human-like mechanics, thus they don't need to outsmart them, they can outmicro them.

That's right. The idea is to run A.I. with human-like limitation to test out its decision making instead of flawless execution. In FPS games like Quake and CS you can easily make AI that will rekt pro team with pure aim but it won't be very interesting. However if you cap A.I. LG accuracy at 35% and 50% RG and you will make it smart to win with rapha then it has to be quite intelligent.

agreed, would love to see deepmind duel with capped accuracy

i'm super skeptical it could ever beat a good player

Yeah, and how impossible is it to create AI that use "human" mechanics?
Building an android that has mechanical limitations?
Doesn't seem like something worth spending time on.

Well, this whole endeavour was about making a bot that outsmarts a human. If you think that's impossible, then you should think their whole goal is not worth time.

However, the limits they had weren't all that difficult to implement, and the ways in which they failed (not limiting APM in enough ways) are easy to implement. They just messed up, purposefully or no.

Who knows, maybe they boosted AI mechanical skills just enough to stomp pro players.
They should work on personal anime waifu holograms instead anyway...

we know. they did

and that the AI will apparently select entities which have 3 pixels visible on the edge of the screen, something a human player could and would not really do.

I don't see how any man-vs-machine match can be fair in a realtime game, as mechanics are a key factor in being successful.




DeepMind has not a great track record so far - they kinda cheated the SC2 matches from earlier, they kinda cheated on their chess match against stockfish.

How did they cheat versus stockfish?

Vast difference in hardware and Stockfish ran on an unsuitable configuration.

In the meantime AI engines like Leela beat conventional engines in fair competitions, but the results are way closer than the bullshit that Deepmind has claimed.

so it seems a patter for deepmind. That's quite lame.

cooller will crush it