3 ms·

I think a strong argument for 'no' comes from the No Free Lunch (NFL) theorems: https://en.wikipedia.org/wiki/No_free_lunch_theorem https://en.wikipedia.org/wi

by api 11y ago

I think a strong argument for 'no' comes from the No Free Lunch (NFL) theorems: https://en.wikipedia.org/wiki/No_free_lunch_theorem https://en.wikipedia.org/wiki/No_free_lunch_theorem The NFL theorems have been misused by advocates of 'intelligent design' -- they do not factor against evolution since evolution doesn't demand a global maximum, only a local one. (There are other reasons but this is the most fundamental.) But the theorems themselves are incredibly interesting. In short they show that averaged over the space of all possible search landscapes, no search algorithm performs better than any other. Obviously the universe does not provide "all possible search landscapes," since the universe has structure. Some search landscape structures and meta-structures are pathological and rare in nature. But nature does provide a huge diversity of them. The NFL theorem is to me powerful evidence that something as general as human or even animal intelligence must be a superposition of multiple "algorithms" rather than just one. It also, I think, explains why all single-algorithm or single-approach methods of AI end up being domain specific. Obviously they'll be domain specific -- they only work against search landscapes with a certain structure! The NFL theorems are a big reason I am a short term AI skeptic. I don't think human-level or beyond AI is impossible, but I think we are quite far from realizing it. I'll become more optimistic when AI researchers start studying biology more closely and deeply, since biological systems are the only existing examples of truly intelligent systems.

rsaarelm 11y ago
Solomonoff induction seems to be an useful innate bias for the current approaches at general intelligence. It's basically a formalization of Occam's razor that states that instead of being completely random, the surrounding environment is mostly lawful and simple, and therefore most of the processes in it can be described by algorithms with short rather than long source code. Hutter has written a bit about NFL: http://arxiv.org/pdf/1111.3846.pdf http://arxiv.org/pdf/1111.3846.pdf http://arxiv.org/pdf/1105.5721.pdf http://arxiv.org/pdf/1105.5721.pdf
- api 11y ago
  Thanks, those look interesting. Will give them a read when I have time. My naive reaction though is: a lawful universe does not necessarily imply well-behaved fitness landscapes. Many lawful and even very simple processes give rise to chaotic and complex results.

by api 11y ago

rsaarelm 11y ago

Solomonoff induction seems to be an useful innate bias for the current approaches at general intelligence. It's basically a formalization of Occam's razor that states that instead of being completely random, the surrounding environment is mostly lawful and simple, and therefore most of the processes in it can be described by algorithms with short rather than long source code. Hutter has written a bit about NFL: http://arxiv.org/pdf/1111.3846.pdf http://arxiv.org/pdf/1111.3846.pdf http://arxiv.org/pdf/1105.5721.pdf http://arxiv.org/pdf/1105.5721.pdf

api 11y ago
Thanks, those look interesting. Will give them a read when I have time. My naive reaction though is: a lawful universe does not necessarily imply well-behaved fitness landscapes. Many lawful and even very simple processes give rise to chaotic and complex results.