[T_S_]

People seem to underestimate how much engineering will be required for intelligence to interact with the environment successfully. My dog might not be as "intelligent" as me, but he knows a lot more than I do when it comes to information gained through hearing, taste and smell. He has better engineering in many respects. Almost any life form beats me in "intelligence" if you define the test right.

There are already theories about how to build universal AI algorithms (check out Marcus Hutter's stuff at http://www.hutter1.net/). These theories are quite an accomplishment, no doubt. But they boil the problem down to one involving sequence prediction. Let's say that particular form of the problem is now "solved". OK now what sequences would you like to present to the algorithm for training? Ones from the environment? Great, but those will take time to collect. Simulated ones? It takes a lot of knowledge to generate good ones.

For these reasons, I suspect the Singularity is safe from us for while.

[gwern]

> OK now what sequences would you like to present to the algorithm for training? Ones from the environment? Great, but those will take time to collect. Simulated ones? It takes a lot of knowledge to generate good ones.

With AIXI and variants, since they are universal predictors, it doesn't really matter what sequences you supply so long as they have some connection. Heck, just your algorithm access to an email server or something.

[T_S_]

| Heck, just your algorithm access to an email server or something.

Then my dog will find the bone faster than AIXI. See what I mean?

[gwern]

The computable versions of AIXI have such enormous constant factors that the indirection of email->real-world is as less than a grain of dust in the Ganges.

That indirection's constant factor cost might matter for any feasible AIXI-inspired AI, but since none has yet to be created or even designed, it's hard to say...