Yeah this is the issue - multimodality would be huge but probably suffer a lot of information loss, plus be a lot more expensive - extrapolating game state requires a looooot of work, you have to map

Yeah this is the issue - multimodality would be huge but probably suffer a lot of information loss, plus be a lot more expensive - extrapolating game state requires a looooot of work, you have to map out all the types with reverse engineering, figure out where they live in memory, then define a structure and make sure you have enough information to make good decisions, as well as dealing with the latency of doing all that expensive work