Re: Eliezer's Seed AI concept

Eliezer S. Yudkowsky (
Wed, 02 Jun 1999 23:36:09 -0500

"O'Regan, Emlyn" wrote:
> I've been reading Eliezer's "Coding a Transhuman AI" page (I'm sure most
> people know of it, but here's the URL anyway:
> ). Actually, that URL looks a bit weird - is there a more correct one?

The "correct" one is, which forwards to the one you mentioned.

> Eliezer, I've got a couple of questions:
> 1) Have you started trying to build this beast? If so, how's it going?

I've been thinking about that a lot. (Naturally - it *is* my lifegoal.) It strikes me that while designing is all well and good, I have to make sure there's an Internet capable of running it and a programming community capable of coding it. Right now, the hardware industry is in serious jeopardy because Bill Gates is starting to default on his promise of sucking up all the CPU cycles Intel can give him, and the software industry can't do anything whose complexity is above a certain limit without it starting to look like something out of the Slow Zone.

To build that AI I need three things: Massive hardware backed by Internet-2 and a to pull it together; OpenSource volunteers with the ability to write a new kind of software; and the computer-community will to do it. I'm mostly working on (1) and (2). Which in turn requires that I shove the entire computing community in a certain direction, the direction of scalable software.

Scalable software is software that gets qualitatively better, not just faster, when you give it more computing power. Deep Blue and EURISKO are scalable software, but for scalable software to keep Intel in the money (to pump out the hardware I need), scalable software has to get into word processors and accounting programs, and that means a new kind of software design.

So now I'm working on designing a computer language in which you can write that kind of software; furthermore, a language that can eventually mature into something in which we can write Elisson. The next great transition, the successor to object-oriented programming as OOP replaced structured programs - adaptive code.

> 2) I haven't seen anything about the environment that you'd want to
> "run" such an AI in (maybe I haven't looked hard enough, but its a
> damned big document!). Would you give it an artificial, simplistic
> environment (simulate some simple, maybe 2d space, with some stuff to
> interact with), or are you thinking that full 3d (2.5d?) vision, natural
> language, etc... chuck it in the deep end is the way to go?

AIs do need sensory environments to ground their thoughts, but not artificial environments; they're already native to a sensory environment richer than any of us have ever seen. What is the most elegant, beautiful, complex domain humanity has ever found? Code.

> 3) Do you have some more info on this?

Not yet.

> After decades of effort, and due mainly to the intervention of the late
> David Marr, we now have a relatively good computer model of human
> vision, and specifically of the transition from 2-D pixels to 3-D model.
> Some of the algorithms thus invented were later found to correspond to
> neural computation, so we know we're on the right track. The intervening
> layers are not chaotic, they are not stochastic, and they are not
> high-level semantic; they are what Marr called a "2 1/2-D" world, a
> series of ordered extractions of increasingly high-level features, using
> a lot of computer code.
> (from Coding a Transhuman AI)
> Specifically, I'm interested in David Marr's 2.5d model of vision.

Marr gets major kudos for pre-inventing the concept of domdules, but what we need is not 2.5D vision; we need 2.5D code.

           Eliezer S. Yudkowsky
Running on BeOS           Typing in Dvorak          Programming with Patterns
Voting for Libertarians   Heading for Singularity   There Is A Better Way