The Key to Real AI | Hacker News Reader

The Key to Real AI

2 pointsmorpheos1372 comments

Postulate: In order to develop real AI you need to simulate the incentives at work in Darwinian evolution. There has never been any intelligence demonstrated in nature outside of the context of the life program to propagate genes in a competitive environment. Intelligence is a tool to solve the replication problem in life. There is no evidence it exists in other areas. Without genetic selection (memory) ML is just fancy combinatorics. Rather than re-inventing the wheel maybe we should be looking at re-wiring existing brains for special purpose tasks. For example re-wiring a mouse brain to factor large numbers rapidly outputting digits serially as toe clicks.

View on HN

Meh. Evolution is a useful tool, no doubt. But it's just a way to search a landscape. There's nothing inherent in evolution that guarantees that it will find all, or even "the best" solutions that can exist in that landscape. I mean, did evolution find turbo-prop engine powered aircraft, or helicopters, or jet aircraft, etc., etc. yadda yadda?

This sounds to me like a classic non-sequitur to me, unless somebody has established that the way biological intelligence works is the only way intelligence could possibly work. And that premise is directly contradicted by all of the intelligent behavior (ignoring the notion of "real intelligence" for the moment since that term is undefined) that has been demonstrated by artificial intelligence programs running on digital computers over the years.

Of course, as the old saying about AI goes "Once it works, it isn't consider artificial intelligence anymore." :-(

>There's nothing inherent in evolution that guarantees that it will find all, or even "the best" solutions that can exist in that landscape.

Yes this is true. But if life persists it demonstrates that a solution that works has been found. Because of time, energy constraints the best solution to an abstract objective function may not not be the optimal real world result. Over successive generations, however evolution will tend toward a better solution to for the problems that life faces. If AI researchers do not simulate the evolutionary process of life then they need to come up with a better heuristic for finding solutions than evolution. So far general intelligence, in humans and animals has only been demonstrated as a result of the evolutionary process and learning using evolved brain structures during life.

As far as I know we have found some good approaches for pattern matching and even problems solving in constrained domains like GO but nobody has yet simulated a general intelligence with a computer model. I argue that is because nobody has modelled a living biological agent yet to a sufficient degree of fidelity.

PS. Jet aircraft are very inefficient from an evolutionary perspective.

Reply on news.ycombinator.com