Reality is merely the cheapest simulator of sufficiently complex models
Conversation
Replying to
Not really a joke. The threshold is much lower than you might think. It’s much cheaper to simply build and try out even fairly simple things than to model them accurately enough in expensive software to try out.
3
2
36
Simulations pay off when you have to do hundreds of refinements and variations for evolution. It’s economies of iteration scale.
2
2
17
Replying to
Trial and error is not simulation. Plus having complete control over the forward model of complex systems is extremely hard/unattainable in the reality "simulator". Peeling the virtual onion is much easier.
1
Replying to
Might this suggest that model-free reinforcement learning is sufficient for complex tasks than trying to directly encode causal models as DAGs? Agents learning causal representation/world models is the obvious compromise. But currently this is a point of debate in AI.
Replying to
if we explored the natural experiment space with even a fraction of the energy with which we explore the reproducible experiment space we would be living in the future
Replying to
Reality is simply the cheapest collection of possible perceptions able to experience complexity.







