They spent a few minutes discussing determinism - this approach only works if your engine is deterministic. They listed a few techniques to improve determinism: use explicit random sequence algorithms (rather than rand()), use fixed-point instead of floating point, don't use OS audio callbacks but instead just assume when you think the callback would fire (which to me sounds like a bad idea, but ok), use lookup tables instead of transcendental functions, etc.

2/4

They left my biggest issue unaddressed, though, which is: doesn't every modification you make to the game cause all of your recordings to become meaningless? The speaker mentioned that you can, like, have each recording indicate the version of the game it was made with, and then the live game can simulate bugs that were in that previous version... but this approach doesn't sound scalable to me.

3/4

I guess this system becomes more useful near the end of the project, when you aren't messing with the game logic much.

Review: 4/10 IMO, the "record button presses" approach works for smaller games/engines, but falls down in larger games/engines for practicality reasons.

There have been previous GDC presentations that describe a more semantic record/replay system (for testing MMOs), where each subsystem has an API, consisting of the nouns and verbs that the subsystem interacts with.