Models are by definition a simplification of reality, but having said that, our understanding of reality is a model in itself. Creating a multitude of reality models and run experiments altering one variable in each makes a lot of sense to me and as David alluded will accelerate science.
I'm still in the camp that models are dead, based on data from only the past.
They are by definition not alive -- incapable of receiving new information without starting over again from scratch like an LLM going through full model training. They cannot capture a future influenced by adjacent possibles and the new system dynamics that emerge from evolving interactions in different contexts at different scales.
The compute demand story is becoming increasingly complex. Most people still view AI in terms of "chatbots getting smarter", but world models represent a completely different demand curve.
Rather than generating text, they simulate physics, which is far more computationally intensive and does not plateau in the same way as language models. Having a thousand parallel realities running simultaneously is no longer just a metaphor, but an actual cost on someone's cloud bill.
Models are by definition a simplification of reality, but having said that, our understanding of reality is a model in itself. Creating a multitude of reality models and run experiments altering one variable in each makes a lot of sense to me and as David alluded will accelerate science.
I'm still in the camp that models are dead, based on data from only the past.
They are by definition not alive -- incapable of receiving new information without starting over again from scratch like an LLM going through full model training. They cannot capture a future influenced by adjacent possibles and the new system dynamics that emerge from evolving interactions in different contexts at different scales.
The compute demand story is becoming increasingly complex. Most people still view AI in terms of "chatbots getting smarter", but world models represent a completely different demand curve.
Rather than generating text, they simulate physics, which is far more computationally intensive and does not plateau in the same way as language models. Having a thousand parallel realities running simultaneously is no longer just a metaphor, but an actual cost on someone's cloud bill.