Great post! Is there any workable alternative to sim2real2sim2real2sim2real*? Even when we model things carefully, we must humbly accept that there are unknown failure modes. One partial solution is to carefully instrument our system to detect anomalies and near misses and use those to guide improvements in simulator.
Ben Recht always brings the rigor. The gap between simulation and reality is still the biggest bottleneck in physical AI - excited to see where this goes.
"The map [sim] is not the territory [real]", and all we ever have are maps (A. Korzybski in Science & Sanity and G. Bateson in _Mind and Nature_, who there emphasized that even our senses are 'sim' in a sense, abstracting as they do from What is Going On).
I saw a video of R. Sapolsky conversing with J. Peterson, where Sapolsky explained that, given a finite series of transactions in a game, it always pays off to betray the other player, but that co-operation becomes optimal when the series of projected future transactions is non-finite or of undefined length.
It's interesting to think about that in a context of requiring a non-finite series of transactions between the simulation and the reality.
Can we learn to co-operate with reality when we accept the series to be non-finite?
Great post! Is there any workable alternative to sim2real2sim2real2sim2real*? Even when we model things carefully, we must humbly accept that there are unknown failure modes. One partial solution is to carefully instrument our system to detect anomalies and near misses and use those to guide improvements in simulator.
Ben Recht always brings the rigor. The gap between simulation and reality is still the biggest bottleneck in physical AI - excited to see where this goes.
"The map [sim] is not the territory [real]", and all we ever have are maps (A. Korzybski in Science & Sanity and G. Bateson in _Mind and Nature_, who there emphasized that even our senses are 'sim' in a sense, abstracting as they do from What is Going On).
I saw a video of R. Sapolsky conversing with J. Peterson, where Sapolsky explained that, given a finite series of transactions in a game, it always pays off to betray the other player, but that co-operation becomes optimal when the series of projected future transactions is non-finite or of undefined length.
It's interesting to think about that in a context of requiring a non-finite series of transactions between the simulation and the reality.
Can we learn to co-operate with reality when we accept the series to be non-finite?
Very thought-provoking, thanks Ben.