Thread 720121678

will NPCs finally be smart and actually tricky to play against now that Ai is within everyone's hand?

>>720121678 (OP)
Kill yourself normalfag chasing nigger

no
NPCs are designed to lose, so that gamers feel important and don't kill themselves.

>>720121678 (OP)
Only if you want your the game to look like nethack until your state of the art GPU completely evaporates and you die from inhaling the toxic vapor.

>>720121678 (OP)
>>>/vt/ranny

The progress in Deep Learning/Transformer architecture/LLMs doesn't necessarily indicate progress in video game AIs.

For one thing, it is sometimes the case that making a superhuman AI is already "easy" (it's easier to read player inputs or have the AI know exactly where the player is and shoot them through the walls from the other side of the map, than make basically any other decision procedure), and most commonly it is the case that there's A LOT of low-hanging fruit for improving the AI dramatically within its current architecture but developers don't do that because there's no money in it (as demonstrated by mods made by individuals that do just that) so why would they bother spending more effort than that? And in some sense you don't WANT "good" (as in "competitive with high-level human play") AIs: strategy gamers for example tend to play the games for the story (LARP), and you can't roleplay if you have to play the meta to succeed, although "vastly stronger AI players playing by the same rules, that don't need 500% sort of bonuses to be even remotely competitive" would be preferred over what we tend to have now.

But also, deep learning isn't necessarily even a good approach. For example, in Chess the strongest engine (Stockfish) is basically "good old-fashioned AI" (GOFAI): some years ago there was a big development in using a technique called NNUE (small neural nets that you can run on a CPU) for evaluating balanced positions, but it effectively implements rules like "in certain endgames, migrate King to the edge of the board" that previously were hand-coded, as opposed to inscrutable black box deep neural nets you'd run on a GPU. And anyhow, the cost of training runs for games like Dota (OpenAI Five) and StarCraft 2 (AlphaStar) were in the tens of millions, and once you removed "cheats" like 2000APM micro, they weren't at level of the strongest human players/teams anyway. Of course training would be cheaper now, but still.

>>720121861
>uhm actually you don't want smart NPCs
Says the redditor while I watch some story relevant character getting stuck in a doorway for the 17th time.

>>720124239
Well, that's the sort of AI that was figured out in the 60s, absolutely nothing to do with taking advantage of modern advancements, and everything to do with developers being incompetent hacks that additionally have no incentives to make the game work properly

>>720121678 (OP)
>>>/vt/ranny