[SEQ RERUN] Dreams of AI Design

[MinibearRex]

Today’s post, Dreams of AI Design was originally published on 26 August 2008. A summary (taken from the LW wiki):

It can feel as though you understand how to build an AI, when really, you’re still making all your predictions based on empathy. Your AI design will not work until you figure out a way to reduce the mental to the non-mental.

Discuss the post here (rather than in the comments to the original post).

This post is part of the Rerunning the Sequences series, where we’ll be going through Eliezer Yudkowsky’s old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Three Fallacies of Teleology, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.

Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day’s sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series.

[Luke_A_Somers]

You can say that “lots of little neurons, which are all strictly electrical and chemical with no ontologically basic contemplativeness in them, combine into a complex network that emergently exhibits contemplativeness”. And that is still a fake reduction and you still haven’t looked inside the black box. You still can’t say what a “contemplative” thing will do, using a non-empathic model. You just took a box labeled “lotsa neurons”, and drew an arrow labeled “emergence” to a black box containing your remembered sensation of contemplativeness, which, when you imagine it, tells your brain to empathize with the box by contemplating.

Well, yes. That’s what emergence IS. It constrains the answer space by saying what it emerges from, thereby focusing the question of how it emerges; it isn’t anything close to a complete answer.

[palladias]

Later I will say more upon this subject, but I can go ahead and tell you one of the guiding principles: If you meet someone who says that their AI will do XYZ just like humans, do not give them any venture capital. Say to them rather: “I’m sorry, I’ve never seen a human brain, or any other intelligence, and I have no reason as yet to believe that any such thing can exist. Now please explain to me what your AI does, and why you believe it will do it, without pointing to humans as an example.” Planes would fly just as well, given a fixed design, if birds had never existed; they are not kept aloft by analogies.

It seems like the more productive thing for a venture capitalist to say isn’t “planes are not kept aloft by analogies” but just “Be Specific.” If you’re using humans as shorthand for a black box, that’s unproductive. If you’re saying “A victory condition is that my chat program passes for human about as often as the average human in the following versions of the Turing Test” you have set up a specific goal to shoot for.

There’s still plenty of room to argue about how you’re planning to train your chatbot, or whether that project is particularly useful for developing AI. But working on these small “like-a-human” problems seems like an easier way for most people to notice unnatural or magical category problems. Once you start noticing the failure modes, you might have a better idea of what you’re actually trying to emulate.

Projects like Kismet seem to have taught us that we’re not that interested in making computers that learn like infants. Among other things, it’s hard to check if they’re working as intended. Having AI enthusiasts go down this road—trying to emulate specific human behaviors or categorizations—seems helpful, since they’re diverted from more dangerous AGI research, and the ways they’ll fail will make everyone thing about how to formalize these categories more.