Retrospring is shutting down on 1st March, 2025 Read more
I'm not sure how to ask this, but can an AI come with something revolutionary, like something beyond what it "learns"?
Yeah, I get what you're saying! Can AI do something original instead of just the task it is trained to do? There are a lot of clear cases where AI can do the task it is trained to do in original ways, but that is still rather passive learning, and I think if we want to talk about something revolutionary we can do better than that.
This is one of the areas that neural networks are weak in since they are a form of supervised learning. They learn from examples which requires that we must have already found a way to evaluate those examples ourselves (even if that method is intuition or perception). To go beyond that requires creativity, to evaluate meaningful things nobody has ever thought of before.
There are several way to approach creativity. The first is unsupervised learning like clustering which sees patterns without having to be told what patterns are there. I've always been fascinated by that, creating a completely new concept from what you see in the world then figuring out how it relates to other things. A second approach is exploration where you search some space to find interesting and maybe unexpected results in it. I'm personally very excited by fitness functions with a lot of structure to them because then there's a lot of different ways things can have value. Things that have value in multiple ways are really intriguing, like how ice cream on a hot summer day is both a good dessert and cold! A third way I really want to work on is something that works on really abstract problems and uses its circumstances to gradually come up with a solution without having to complete prior steps so serendipity and the connections it makes with others play a role. As a matter of practicality, that might really be the best way to be revolutionary.
Retrospring uses Markdown for formatting
*italic text*
for italic text
**bold text**
for bold text
[link](https://example.com)
for link