François CholletVerified account

I guess you never heard of Doug Lenat's Cyc initiatlve. https://www.wired.com/2016/03/doug-lenat-artificial-intelligence-common-sense-engine/ … #ai

He didn't start by understanding how humans exhibit common sense, so he went about the task the wrong way. (Obviously wrong because his knowledge system requires more initializing than humans do, and doesn't maintain its knowledge). Better surely to start by understanding the =>

If we only had to encode knowledge once, this might work. But our experience with expert systems showed that the knowledge required continual maintenance (like all software and data-based systems). The cost of ownership of hand-coded knowledge is high

Machine learning has shown potential to reduce the cost of creating a high-performance system. But the cost of ownership of ML systems is also high. I thought it would be just a matter of retraining, but the definition of the ML task itself changes over time.