|
On September 30 2026 00:55 Warfie wrote:
Agree. If it were more constrained and forced to actually use smarter placements, builds and tactics it could realistically produce stuff that the pros could implement. Kind of what alphazero did to Magnus Carlsen using flank pawns in top level chess.
Gotta keep a couple of things in mind, though...
1. Chess has practically no mechanical component at all for all intents and purposes, so a player can always physically play to the level of an AI, and in chess an AI is far more limited by what it can actually "do". To somehow limit and/or configure the AI to play to the same mechanical level as a human being (as Kraekkling said above) would be a technological feat in and of itself; on the visual intelligence aspect alone, it would be a tremendous breakthrough. Digital cameras and AI-powered image processing have both been around for decades and while certainly they are superior to humans in some areas, they still have not been able to fully mimic the combination of the human eye and brain. Even if we limit the discussion to software alone, how do we mimic stuff like difficulty in reaching all 10 hotkeys on a keyboard, for example? I guess it depends on how* granular you want to get with this, but my point is that even if the AI is somehow neutered to be a general approximation of humanly-possible play, there will inevitably still be things that it does better than us and we do better than it.
2. The time difference between the first primitive chess automaton and the first computer program that could play a full game of chess was 45 years. It would take another 40 years for Deep Blue to beat Kasparov. Stockfish 1.0 came out in 2008 (about 10 years later) and modern Stockfish would clown on 1.0. So while it is clearly true that technology growth is exponential, especially now with reinforcement learning, I can't imagine that implementing/successfully encouraging the changes being described is all that easy. In short, I think it might take a while.
3. All of the aforementioned things (digital cameras, image subject identification, Deep Blue, Stockfish, etc.) were all developed with some combination of: large teams, lots of money, significant profit incentive, significant competition between similar products, global interest and demand, millions of expected long-term customers, etc. There is a pretty significant disparity between that and a BW AI. Don't get me wrong, the foundation is absolutely there and tscmoo did an incredible job; I'm just saying that it might take a while for that reason, too.
4. The AI was released on Remastered without the consent of either the bot author or Blizzard, so there is always that barrier, too...
|
|
|
Netherlands4534 Posts
|
That sucks.
Also, your guardian play is sexy.
|
is this finally what makes us move to shield battery? losing MMR points to AI that blizzard does nothing against?
|
but SB's gonna add these bots too xd, but not on ladder i think, just for customs/melee games
|
I was playing in my current noobie league, C, and faced someone who moved their units in a way I didn't understand, and my mind immediately went "ai?" .... The guy wasn't good and I won and looking at the replay it was obviously a person.
I can't imagine what it is like at the high levels where you face players good enough to not know 100% early on. It is just as cruel psychologically as having a Chess engine make moves versus a grandmaster. The amount of effort it takes to play good is just a bastard thing to abuse.
|
On September 30 2026 06:14 TT1 wrote:
but SB's gonna add these bots too xd, but not on ladder i think, just for customs/melee games
That would be cool. Prolly cant run on super potato pc like mine anyway but still wanna cheese it to death..
|
Ridesky just played a ladder TvT
everything seemed fine, normal game, opponent sort of passive but had good engagements and wraith use
and then all a sudden a bunch of m&m rush in
|
looking at Ret's guardian game on backrooms from ladder, it looks like the human controller is also able to issue orders like some sort of team melee archon mode with the controller+pluto... if you watch the replay on cwal replay viewer there was manual inputs for the medic wall to block the ramp and for the vessal eraser
|
|
|
On September 30 2026 00:30 KwarK wrote:
I think we've more or less exhausted what we can learn from the current iteration of the bot.
On the one hand it is abysmal at strategy and has no real understanding of even basic unit counters. And if it hasn't learned them by now from playing itself then it's not clear that it ever will. Things like building placement to limit damage from vultures etc. just aren't a part of its play.
On the other hand, it is clear that the advantage given by impossible micro is insurmountable in a standard game (people can cheese the bot but that doesn't teach anything). That makes it difficult to teach the bot because it'll win with inferior strategies and draw the wrong conclusions from them. Zealot/dragoon will beat a human hydra bust and what it'll learn is that those units beat hydras whereas what's actually going on is the human micro can't compete with the AI micro.
For it to get better at strategy it needs micro and multitask constraints. That's not knocking the project, I think it's really cool, but it's like a parrot trained to say a few words based on context cues. Once you know how the trick is done you don't need to see it done again. I want to see the next parrot that can hold a conversation.
What is interesting, is that the bot figured out he can counter hydras with zealots and dragoons by playing against it's own AI hydras. AI zvp is still lopsided in zerg favour however.
|
United States44325 Posts
On September 30 2026 07:36 Rainalcar wrote:Show nested quote +On September 30 2026 00:30 KwarK wrote:
I think we've more or less exhausted what we can learn from the current iteration of the bot.
On the one hand it is abysmal at strategy and has no real understanding of even basic unit counters. And if it hasn't learned them by now from playing itself then it's not clear that it ever will. Things like building placement to limit damage from vultures etc. just aren't a part of its play.
On the other hand, it is clear that the advantage given by impossible micro is insurmountable in a standard game (people can cheese the bot but that doesn't teach anything). That makes it difficult to teach the bot because it'll win with inferior strategies and draw the wrong conclusions from them. Zealot/dragoon will beat a human hydra bust and what it'll learn is that those units beat hydras whereas what's actually going on is the human micro can't compete with the AI micro.
For it to get better at strategy it needs micro and multitask constraints. That's not knocking the project, I think it's really cool, but it's like a parrot trained to say a few words based on context cues. Once you know how the trick is done you don't need to see it done again. I want to see the next parrot that can hold a conversation. What is interesting, is that the bot figured out he can counter hydras with zealots and dragoons by playing against it's own AI hydras. AI zvp is still lopsided in zerg favour however.
There's absolutely no evidence that the AI ZvP does a competently executed and timed 973 bust opener.
|
On September 30 2026 00:30 KwarK wrote:
I think we've more or less exhausted what we can learn from the current iteration of the bot.
On the one hand it is abysmal at strategy and has no real understanding of even basic unit counters. And if it hasn't learned them by now from playing itself then it's not clear that it ever will. Things like building placement to limit damage from vultures etc. just aren't a part of its play.
On the other hand, it is clear that the advantage given by impossible micro is insurmountable in a standard game (people can cheese the bot but that doesn't teach anything). That makes it difficult to teach the bot because it'll win with inferior strategies and draw the wrong conclusions from them. Zealot/dragoon will beat a human hydra bust and what it'll learn is that those units beat hydras whereas what's actually going on is the human micro can't compete with the AI micro.
For it to get better at strategy it needs micro and multitask constraints. That's not knocking the project, I think it's really cool, but it's like a parrot trained to say a few words based on context cues. Once you know how the trick is done you don't need to see it done again. I want to see the next parrot that can hold a conversation.
True but it’s still kinda cool to see pros losing against its extreme micro capabilities. It’s an unfair comparison given AI has an extremely high mechanical ceiling but still interesting to see.
|
|
|
|
|
|
EPT![[image loading]](https://i.ibb.co/5mGmTG6/images-4.jpg)
