I think you're talking about totalistic CAs, which are a subset of CAs in general.
Basically, what I'm suggesting is not programming in any sort of physics, and letting the physics itself emerge out of the cellular automata. CGoL didn't have any physics programmed into it, flying and stationary objects emerged naturally.
Sorry about that, I was referring to a totalistic CA...
The problem when you do that is that it's A) hard to determine a good rule and B ) you don't get to pick and choose behavior you want. Many totalistic CAs are quite boring, for instance, because the rules don't result in interesting behavior. If you choose one that does result in interesting behavior, you have to bend whatever game design you want to do around the limitations and features of that rule. It's much better to decide on the sort of behavior you want to see and design the rules, in a broad way, to cater to that behavior. Then you'll get other emergent behavior from the interaction of the rules that is consistent with your original feature set.
There are also ALOT of well-known interesting CAs, CGoL is the most well-known example. There are probably thousands of less well-known rules that also result in interesting behavior, you might need more than 2 states though. (By interesting behaviour, I'm referring to wolfram class-4 automata)
Finally, there are other CAs besides totalistic ones that the hashlife algorithm can be used on... for example, an automata where its state depends on the state of the cells around it at t-2. Or an automata where a cell's state depends on the state of the cells around it, but that are not in contact with it...
Basically, I'm saying that you can "design the rules" of the totalistic/other CA, because there are limitless options.
The main problem I'm thinking of is how to interface a bot with the CA world...