Develop AI systems to efficiently learn new skills and solve open-ended problems, rather than depend exclusively on AI systems trained with extensive datasets
Collecting some thoughts and I had as we discussed the challenge and how we might go about coming up with a solution.
No doubts about this being a hard challenge. But naiveté may be our greatest strength.
A few things that stick out from the conversation:
Reasoning may not be a sequential process. It's very likely not. Brings up the question about if the brain is quantum mechanical, chemical, or something else. This means that if we're to formalize this as an algorithm, we might not be able to rely on traditional data structures or sequential operations to execute the program.
There's some idea of a set of mini programs or subroutines that we are composing that will be able to be applied to the input to get the output. The reasoning is actually the application and composition of these routines, but the subroutines do exist beforehand.
Seeing new ways of reasoning - as we both solved the same problem, you could see that people reason differently. I know this to be true because as I've worked out different approaches (perspectives) to take when solving the puzzles, the reasoning process also changes. I think this is due to a change of the fundamental subroutines you use to compose your solution to a puzzle. How do we make the reasoning process aware of these subroutines? Probably because the reasoning process uses them to simulate different applications of them, seeing what gets it closer to the output.
This is promising because it means that there are many solutions to the puzzles, and that also the correct choice of subroutines may not be completely essential to the success. As long as the reasoning is sound, the subroutines matter only in reducing the search time to finding the right application of them.
Vision and a wholistic understanding of the puzzles is so important. We make a lot of assumptions and shortcuts just by the sight of the puzzle.
There's some kind of "filtering" that we do at the start of a puzzle that greatly narrows down the ideas we have about how to solve it. But I think we both agreed that this filtering isn't a classical filtering we're use to, where we're looking to see if "condition A" is met. We don't have access to "condition A", because that would imply we have a set of conditions we can look for, which bypasses the reasoning process and puts us in classification mode. Maybe some hybrid of classification but also the boundaries of the filtering are fuzzy to leave other options open.
We talked about some kind of mapping from the input to the output. There's a latent space in between that could give us insight into the operations or subroutines to apply to get us between the two. This is likely to be a new kind of latent space, because it's not mathematical operations we're applying from input to output, but larger operations that might apply to the whole input encoding, like a change of grid dimensions or shifting a whole structure in some way.
The operations we learn in the "latent space" might also need inputs we've learned from prior operations.
Discover other posts like this one