Whole Brain Emulation, as a platform for creating safe AGI

Speaker(s): Anna Salamon

Transcriber(s): Ethan Dickinson

Proofreader(s): Carlos Serrano

Anna Salamon: Whole Brain Emulation as a platform for creating safe Artificial General Intelligence. This is joint work with Carl Shulman. He’s here too and will be joining for the panel discussion.

I’m going to go through a bit of background. Why AGI is volatile, why it’s risky, the idea of Whole Brain Emulation, the fact that Whole Brain Emulation is not a stable resting point but may help to build safe AI, and the question of whether we can make Whole Brain Emulation come first as a strategy for long-term AI safety.

To get into the volatility, there’ve been a lot of changes since the Industrial Revolution, but there’s reason to expect that AGI might bring more change, faster. There are at least four major mechanisms that could cause that.

First is that AGIs can be copied. Once you have digital intelligence, it’s hard to design the first one, but you can copy it across the hardware base at the speed of copying software, so that if software is the bottleneck, as many expect, and you have hardware accumulating for a long time before the software hits, you could go rapidly to a state of having billions of copies.

Second major reason why it might bring faster change is the possibility of serial speed-up. Your intelligence is stuck running on your brain. Digital intelligence can be ported between hardware structures. Design something that works at human speed, port it to hardware that works at faster than human speed, if someone develops that. The picture is supposed to be motion-blurred. [laughs]

Third major reason, reduced coordination problems, if you can have lots of copies of a single clone.

Fourth major reason, recursive self-improvement. If AI is developed, it’ll be developed because a set of human researchers is smart enough to design an AI with comparable research abilities. If that AI then turns its abilities on itself, one can plausibly move rapidly outside the human scale.

Why is it risky? Basic reason is that intelligence with goals leads to change. You see some way to rearrange the world that suits your goals better. The smarter you are, the more deeply you can rearrange the world. Our ancestors had chipped, simple stone tools, we have laptops. Smarter intelligence can change things farther.

And unfortunately, most ways of creating change on that scale would kill us, which is especially problematic since values vary starkly, even across the animal kingdom. The space of potential values seems to be much larger than that. Trying to design a powerful AI that would keep our arrangements around seems to be quite difficult. Threading a needle.

By Whole Brain Emulation we just mean copying specific individuals. Making a copy of you that’s so exact that it has your behaviors, your goals, your voice when it speaks in a simulated world. The technical requirements for this have been sketched out in some detail. “Whole Brain Emulation: A Roadmap” by Anders Sandberg and Nick Bostrom is totally worth reading.

It’s attractive because Whole Brain Emulations are a relatively known quantity compared to AIs. They have our values. They do care about your arrangement of atoms, maybe. Although this has its downsides as well. [picture of soldiers marching]

Whole Brain Emulation is not a stable state. Why not? One reason is that, if human researchers can design AGI, simulated humans can design AGI faster. Start off with Whole Brain Emulation, end up at AGI, or at somehow the stable cessation of research.

A Whole Brain Emulation is probably not designed for modular self-improvement. [gestures at head] This looks a whole lot like spaghetti code. Nonetheless, if evolution can try variations, so can we, including both biological variations and variations in educational experiments. And again, using the magic of digitality, copyable software, if you find a version that works, you can copy it instantly across the hardware base.

Other cool tricks you can do when things are digital. Save a copy of yourself at a moment of peak productivity. Feel a little bit differently after a while? Oh well, reboot from disk. Big Brother could do this to keep your goals in a particular arrangement. [laughs] And again, single upload clans could take over using copyable hardware, whether by economic means or by hacking, could thereby use those powers in a [?] ballooning way to perhaps invent powerful AI which could destabilize things further.

If you have a competitive equilibrium for a while, values can also be lost through competition. Nick Bostrom’s essay “The Future of Human Evolution” is great for that, talking about a world where you can engineer job skills separately from the components that we’re used to having go along with it.

To recap the situation. We’re here in the business-as-usual square. Business as usual is not a stable state, you can get from here to totalitarianism, controlled intelligence explosion, uncontrolled intelligence explosion, or human extinction. Those are stable states, go extinct, stay in that square. The question is whether the unstable state of Whole Brain Emulation makes us more or less likely to get to the corners that we care about. Is our shot better or worse with Whole Brain Emulation?

Seems to me it’s probably better. There’s a lot of reasons for that. One is that it can ease coordination. If AI is developed in an arms race, this can incentivize ditching safety for speed. Whole Brain Emulation has the advantage that, since we’re dealing with things more or less like humans, relatively known quantities, you can do more of the safety infrastructure in advance of the relevant technologies.

Imagine an arms race. Imagine that one player is some years or some months ahead of the other. The US was four years ahead of the next runner-up for the nuclear bomb. If that lead player gets to Whole Brain Emulation first, and if there’s in fact a lot of hardware so that those Whole Brain Emulations can run at a hundred, a thousand times speedup, that four-year lead time turns into four thousand years of lead time in which to design more careful safety measures.

Imagine Whole Brain Emulation as a platform for further tech development. It seems to have a number of advantages. This is an artist’s rendering of Foucault's Panopticon, sort of a dystopic vision of how you could stick humans into a structure to enable controlled behavior that’s predictable in a way the individuals aren’t. If you have Whole Brain Emulation and you can carefully test and retest somebody, booting them up from the beginning again, and stick them into very careful roles within an upload system, you could do something much more like that than has ever been possible before.

Again, the safety infrastructure could be designed ahead of time… stable goals... although it’s possible that moral objections would prevent this.

Even without Whole Brain Emulation systems, even if it turns out that moral objections or other factors do prevent this, there are some advantages to Whole Brain Emulations. Some of these are social. Reduced coordination problems, possibly more tendency to take the long view. Digital intelligences are potentially immortal, maybe they’re less risk-prone. It will already have seen radical technological change in their lifetime, it may be more inclined to take further risks seriously.

It can ease the technical challenge as well as social factors. Subjectively slower computers. Eliezer Yudkowsky likes to talk about what he calls “Moore’s Law of Mad Science,” a joke law, whereby every 18 months the minimum IQ necessary to destroy the world drops by one point. [laughs] After a while I’ll be able to do it. Uploads would speed up as the computers sped up, so Moore’s Law of Mad Science would stop for them, which would have certain advantages in terms of keeping track of the AI programs they were experimenting with.

In principle, they could also do cognitive enhancement by messing with the brains, although once you start editing, the brains can become inhuman fast.

If it is safe, can we make it come first? This is the part that I’m most uncertain about. Three major inputs to Whole Brain Emulation, hardware, brain imaging, and computational neuroscience, so you know how to interpret those images. Hardware is hard to budge. Brain imaging can be automated, there’s some successful efforts in this direction. Computational neuroscience seems like the most likely bottleneck, and you could increase research funding there.

But all three of these inputs also speed up brain-like AI. If it turns out that Whole Brain Emulation is a safer platform for AGI development, there’s the question of, when you increase these three arrows, which one does it speed up more? By how much? And is it in fact safer?

Another possibility to consider is the idea of getting Whole Brain Emulation from a partially controlled AGI system. Partial AGI makes Whole Brain Emulation, makes more sophisticated AGI. I’ll be talking more about that in the next talk, actually.

We’re left with the question of what to do. Is it worth trying to accelerate whole brain research as a stepping stone toward AGI? Seems to me it’s at least worth investigating. Again, here’s the business-as-usual situation. The question is, will Whole Brain Emulation make it easier or harder to get to the corners that we care about?

Thanks.

[applause]