[resu_nimda]

I’ve been intrigued by Artificial Life for a while; I feel like it’s the way forward for GAI. I think the ML approach leaves out a lot of what is required to develop interesting intelligence.

The main challenges I see are:

1) With any given simulation, it’s going to take a long time before something really noteworthy emerges, and until then how do you know you’re on the right track?

2) What is the fitness function? This has always stumped me. How do you create artificial resources and competition and life and death to drive evolution?

[Enginerrrd]

I've given a lot of thought to this over the years too and come to a very similar conclusion. GAI really probably needs to create itself.

I think that intuition was at least partially confirmed by the successes of GAN's. ...not every problem can be solved that way, per se but it's a good starting place.

The big challenge here though IMO is computational! If you're going to use genetic algorithms to create organisms, you have to then let each of those organisms train themselves. It sounds computationally expensive. And the search space is absolutely massive! It will take some really intelligent effort to bring it down to a level that's feasible.

On that note though, the other thing is that I'm convinced that you also need to include for GAI is fundamental structural arrangements as evolvable parameters. You can argue that this is unnecessary mathematically, but IMO, a little bit of intelligent structure can make the same number of neurons a lot more efficient to train. We see this confirmed in some of the more successful deep learning pipelines. Different components of the whole system get specialized training which gives structure to the whole system.

From an artificial life standpoint, I thought about just offering resources based on correct answers to questions, with exponentially greater resources awarded to harder questions. Questions though would need to be generated in situ of course to avoid memorization.