7 comments

[ 0.21 ms ] story [ 29.9 ms ] thread
Obviously the implications of this go far beyond game design... but looking at this makes me so excited for the next few years in gaming.
I wrote up some comments while reading this paper: https://hachyderm.io/@ianbicking/110175433063839064

I found the part on "reflection" was particularly interesting.

The cost of the simulation is also notable, going up to a thousand dollars for 3 days, running on GPT 3.5. That's very high from my experience doing this sort of thing, but not unimaginable when you do things in a loop, and if you are doing 25 queries on a clock tick. But should be very optimizable.