3 ms·

We used to say that AI development should be "embodied," because only the real world was complex and structured enough to train up an intelligence. Funny how th

We used to say that AI development should be "embodied," because only the real world was complex and structured enough to train up an intelligence. Funny how the tables have turned! My experience matches the author's exactly. You must check your robot against the real world to ensure it works, but simulation is the only feasible way to train and evaluate your system at scale. The challenge is that simulators are kinda sorta almost good enough, but no actually they're not, and you need to solve a bunch of problems around the sim2real gap. The problems range from doable (colors look different under certain lighting conditions), to hard (simulating how other robots and people react to you), to still impossible (the feeling when a USB C plug has snapped into place). Each day we close the gap a little bit, but there is still a ways to go.

ericjang 5y ago
I still think that AI development should be embodied, and am willing to entertain the possibility that you don't need simulation at scale to train it (e.g. some clever unsupervised learning algorithm). Evaluating a system that can do millions of things is simply impossible in reality, and that alone necessitates some kind of software-based evaluation metric. I agree that there is a ways to go - but I think trying to make simulation more like reality is a safer bet than having a low ceiling on iteration speed in real.
hinkley 5y ago
I think it's the siren song of thinking there were going to be more ways for a practical AI to cheat, and it turns out all the corner cases actually matter. Everyone had their fingers crossed that 'the world' was mostly accidental complexity instead of intrinsic. And to be fair I'm not sure that you historically had a choice in this. If you were doing AI research in the 80's or 90's there just weren't enough compute cycles available to model anything complex. Either you believed the gap was narrow and you found another field to work in, or you hoped that it was wide and that you were going to discover just how wide it was. Every mammal has a model of the universe in their head that they work against. Mirror neurons expand that to modelling other mammals. I don't know how you can reason about an environment if you don't have a useful model of it first and foremost.

ericjang 5y ago

I still think that AI development should be embodied, and am willing to entertain the possibility that you don't need simulation at scale to train it (e.g. some clever unsupervised learning algorithm). Evaluating a system that can do millions of things is simply impossible in reality, and that alone necessitates some kind of software-based evaluation metric. I agree that there is a ways to go - but I think trying to make simulation more like reality is a safer bet than having a low ceiling on iteration speed in real.

hinkley 5y ago

I think it's the siren song of thinking there were going to be more ways for a practical AI to cheat, and it turns out all the corner cases actually matter. Everyone had their fingers crossed that 'the world' was mostly accidental complexity instead of intrinsic. And to be fair I'm not sure that you historically had a choice in this. If you were doing AI research in the 80's or 90's there just weren't enough compute cycles available to model anything complex. Either you believed the gap was narrow and you found another field to work in, or you hoped that it was wide and that you were going to discover just how wide it was. Every mammal has a model of the universe in their head that they work against. Mirror neurons expand that to modelling other mammals. I don't know how you can reason about an environment if you don't have a useful model of it first and foremost.