Life and expanding steerable consequences

by alexflint3 min read7th May 20212 comments

46

Ω 18

AI TakeoffWorld OptimizationAI
Frontpage
Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

Financial status: This is independent research. I welcome financial support to make further posts like this possible.

Epistemic status: I believe this is a helpful lens through which to view the significance of AI in a way that is not fundamentally about intelligence.


In this world, there are two types of objects: objects whose steerable consequences diminish over time, and objects whose steerable consequences expand over time.

Consider a small rock on a table. Suppose I move that rock a little to the left, and consider the ways that this action might affect the future. The rock might have been holding down some papers, and those papers might now be blown about by a gust of wind. Or someone might walk into the room and, seeing the rock being out of place, walk over and move it back. In fact the rock exerts a gravitational effect on every other object in the universe, and the tiny movement of the rock will have consequences that ripple out for the life of the universe.

But although these consequences are real, the rock cannot be used by us to produce a predictable large-scale effect on the world very far into the future — say, on the timescale of decades. The consequences of moving the rock become too unpredictable for us to reason about. Even if we are allowed to move the rock to any point in the universe, we cannot really use this power to effect any useful control over the future, at least not without involvement from humans. As we consider the causal fallout of moving the rock we quickly hit a wall of foggy uncertainty, and so in this sense rock cannot be used on its own to steer the future.

But consider now the action of introducing a living organism to the surface of Mars. Suppose that some scientists have chosen or engineered a particular kind of mold that will thrive in the environmental conditions present on Mars. Suppose that we move an object of the same size as the rock, only now the object is a mold specimen together with an initial food source, and we move it from some laboratory on Earth to the surface of Mars. Although the physical size of this initial specimen might be quite small, this action could have consequences that eventually affect the entire surface of Mars.

Furthermore, some of these consequences are quite predictable. We can predict that the mold will reproduce. We can predict that the specimen will spread outwards from its initial location. We can predict that after a few decades we might find copies of the mold all over the surface of Mars. Other consequences are fundamentally unpredictable, yet it is clear that there are some predictable large-scale consequences.

Suppose now that we genetically engineer the specimen to grow under some conditions and not others. By picking these conditions precisely, we might cause the mold to spread to only the northern hemisphere of Mars, or to grow only at low altitudes, or only at high altitudes. In each case, the only thing we are transporting to Mars is a single specimen the size of a small rock. We are not ourselves spreading the mold over a mountain range or over the low-altitude parts of the planet, but by tweaking the configuration of atoms within this initial specimen we can choose how and where the mold will spread. In this sense the mold has expanding steerable consequences because a physically small specimen can be altered in a way that predictably steers large-scale effects over a long time horizon.

Another object that has this expanding steerability property is the human being. Transport a small colony of humans together with appropriate resources and an initial life support system to the other side of the universe, and over a few thousands or tens of thousands of years an entire space-faring civilization might spring up, perhaps rearranging the matter and energy in that part of the cosmos at a macroscopic scale.

Which kind of objects have this property of expanding steerability? As of May 2021, there are no non-biological objects on Earth that have this property, without ongoing input from humans. For example, suppose I transported a robot to the surface of Mars. This has been done several times, and it has not had the kind of expanding steerable consequences that transporting a mold specimen to the surface of Mars might have[1]. Furthermore we have not yet built robots that could, without any external help from humans, be used to steer the future, even to the limited extent that a mold specimen might be used to steer the future.

If all biological life on Earth disappeared tomorrow, but all machines built by humans continued operating, the entire ecosystem of machines would quite quickly wind down. Much of the software that runs services on the internet relies on near-constant human oversight, and would cease operating in the absence of humans. But even the most robust pieces of software would cease operating when the power grid decayed to the point of inoperability. And even the most robust machines that humans have ever built, such as some satellites and perhaps some computers located underground with nuclear power sources, will not have the kind of expanding consequences in this neighborhood of the universe that biological life could have.

So in this regard, all the machines that humans have ever built are more like the rock on the table than they are like the mold specimen. Whereas life is winding up, the machines we have built thus far are winding down.

But this may be about to change. Humans appear poised to create machines that could have expanding steerable consequences, independent of biological life. If we succeed at building truly intelligent machines, we might create machines that can collect resources, maintain and upgrade themselves, expand or reproduce themselves, grow their own impact from small to large, and reshape significant patches of the universe. The precise initial configuration of such machines may determine much of what changes they make to their patches of the universe.

All biological life appears to have originated from a single seed organism approximately four billion years ago. This seed organism was almost certainly very small, but its unfolding consequences thus far have been as vast as the Earth, and may yet continue to unfold beyond the Earth. Now, four billion years later, we are about to set in motion a second seed.


  1. Yes, the Mars rovers have had large consequences via the information they have beamed back to Earth, but these consequences have flowed via humans, which are a form of biological life that very much does have the expanding steerability property ↩︎

46

Ω 18

2 comments, sorted by Highlighting new comments since Today at 3:16 PM
New Comment

By picking these conditions precisely, we might cause the mold to spread to only the northern hemisphere of Mars, or to grow only at low altitudes, or only at high altitudes. In each case, the only thing we are transporting to Mars is a single specimen the size of a small rock. We are not ourselves spreading the mold over a mountain range or over the low-altitude parts of the planet, but by tweaking the configuration of atoms within this initial specimen we can choose how and where the mold will spread. In this sense the mold has expanding steerable consequences because a physically small specimen can be altered in a way that predictably steers large-scale effects over a long time horizon.

I actually don't think this is a very good example because once the mold takes root on Mars, Darwinian processes will take over and any mutations with a strong reproductive fitness advantage (such as those that allow the mold to expand to new environments) will be selected for.

But I agree with your general point and think the mold analogy is apt. There's a phrase I heard somewhere in an interview Elon Musk gave where he jokingly said "we're just the biological bootloaders for digital superintelligence." I kind of wonder what that life will look like. It will probably be so unimaginable that it's useless to think about it, but given the types of reinforcement learning systems we have today, I can't help but wonder if the ultimate aim of future digital systems will not be some grand ambition or even reproductive immortality but rather some silly poorly thought-out human goal that is mindlessly pursued until the heat death of the universe.

In some ways, that's the true fear of all AI researchers: that we will not only misalign AI, but that we will do so badly that the digital gods we create will pursue some existentially tragic goal, chewing up all the universe's resources in its light cone.

Yes, that is what I fear for with the future of AI.

I agree re Darwinian selection of the mold. Perhaps a better example would be a deliberately designed reproducing nanofactory with error correction sufficient to prevent viable mutations.