Re:Peter Norvig critically reviews AlphaCode's code quality
by throw_nbvc1234 on Friday December 16, 2022 @04:53PM
(#34021016)
Coming up with novel content doesn't necessarily mean it can reason (depending on your definition of reason). Take 3 examples:
1) Copying existing bridges
2) Merging concepts from multiple existing bridges in a novel way with much less effort then a human would take to do the same.
3) Understanding the underlying physics and generating novel solutions to building a bridge
The difference between 2 and 3 isn't necessarily the output but how it got to that output; focusing on the output, the lines are blurry. If the AI is able to explain why it came to a solution you can tease out the differences between 2 and 3. And it's probably arguable that for many subject matters (most art?) the difference between 2 and 3 might not matter all that much. But you wouldn't want an AI to design a new bridge unsupervised without knowing if it was following method 2 or method 3.
1) Copying existing bridges 2) Merging concepts from multiple existing bridges in a novel way with much less effort then a human would take to do the same. 3) Understanding the underlying physics and generating novel solutions to building a bridge
The difference between 2 and 3 isn't necessarily the output but how it got to that output; focusing on the output, the lines are blurry. If the AI is able to explain why it came to a solution you can tease out the differences between 2 and 3. And it's probably arguable that for many subject matters (most art?) the difference between 2 and 3 might not matter all that much. But you wouldn't want an AI to design a new bridge unsupervised without knowing if it was following method 2 or method 3.