Search Results
7/8/2025, 6:57:02 AM
>>714806023
>Actual examples of these models being used to generate actual gameplay?
WHAMM, by Microsoft, is an autoregressive token model trained on Quake II, to recreate Quake II.
It was released back in April of this year. It's barely playable dogshit.
Enemies phase in and out of existence, health jumps around, the 'game' can't remember any of the level's layout so you teleport around,
there's a hard time limit because of the limited memory, ect.
It's also free:
https://copilot.microsoft.com/gaming
To be fair: they cut the time down from 7 years of necessary training data to 1 week of training data from W1.6b to WHAMM which is a massive and impressive improvement for what it is. They went from 7 years to push out a turd, to just days to squish out shit.
>Actual examples of these models being used to generate actual gameplay?
WHAMM, by Microsoft, is an autoregressive token model trained on Quake II, to recreate Quake II.
It was released back in April of this year. It's barely playable dogshit.
Enemies phase in and out of existence, health jumps around, the 'game' can't remember any of the level's layout so you teleport around,
there's a hard time limit because of the limited memory, ect.
It's also free:
https://copilot.microsoft.com/gaming
To be fair: they cut the time down from 7 years of necessary training data to 1 week of training data from W1.6b to WHAMM which is a massive and impressive improvement for what it is. They went from 7 years to push out a turd, to just days to squish out shit.
Page 1