How Can AI Be Used for Space Applications?

[Greg Hullender]

Yeah, probably off topic a bit. But we would have to agree on what could be achieved with AI in order to decide on how it could be used in space. Maybe someday AIs will be discussing how we can be used in space applications. I live in fear of their conclusion.

Nothing based on any existing technology has any hope of ever evolving into something like what you're imagining. An effort to develop human intelligence using current AI technology is like trying to get to the moon using ancient Babylonian technology (i.e. building a tower tall enough to reach the moon). You simply cannot scale it up that far. Nowhere near.

[edzieba]

Yeah, probably off topic a bit. But we would have to agree on what could be achieved with AI in order to decide on how it could be used in space. Maybe someday AIs will be discussing how we can be used in space applications. I live in fear of their conclusion.

Nothing based on any existing technology has any hope of ever evolving into something like what you're imagining. An effort to develop human intelligence using current AI technology is like trying to get to the moon using ancient Babylonian technology (i.e. building a tower tall enough to reach the moon). You simply cannot scale it up that far. Nowhere near.

Until maybe 5-10 years ago I'd have agreed, but the recent 'Deep Learning' boom has shown that techniques long disregarded and woefully inefficient and inelegant can be perfectly adequate at performing real-world tasks if you can throw enough hardware at the problem.
Human consciousness arises from a massively parallel array of incredibly stupid individual neurons working kinda-sorta in tandem and sometimes producing the 'correct' result, so I rather suspect any human-like AI will arise from a similarly messy pile of hacks and excessive use of compute resources rather than some elegant new simulation technique.

[Barley]

Nothing based on any existing technology has any hope of ever evolving into something like what you're imagining. An effort to develop human intelligence using current AI technology is like trying to get to the moon using ancient Babylonian technology (i.e. building a tower tall enough to reach the moon). You simply cannot scale it up that far. Nowhere near.

The Babylonian technology that matters is agriculture.  That lets you expand the population, so you have enough one in a million geniuses (or a million non-geniuses) that you can refine copper, steel, silicon, ... .  Five millennia later you're on the Moon.

Time scales may vary.

[JohnFornaro]

Nothing based on any existing technology has any hope of ever evolving into something like what you're imagining. An effort to develop human intelligence using current AI technology is like trying to get to the moon using ancient Babylonian technology...

The Babylonian technology that matters is agriculture.  That lets you expand the population, so you have enough one in a million geniuses (or a million non-geniuses) that you can refine copper, steel, silicon, ... .  Five millennia later you're on the Moon.

Time scales may vary.

Which is an excellent point. 

It is now thought that mankind can determine its future, and some transhumanists are positing that AI will set us "free".  But what about those Babylonians, back in the day?  How many of them had "visions" of what mankind's future could have been?  Substitute "dreams" for "visions", if desired.

As an aside, I agree that agriculture is crucial to any colonization effort.  I've suggested taking a load of Eastern forest dirt up to the first hab, planting it, and empirically finding out what happens, rather than trying to create an AI model of what lunar agriculture could be.

[sanman]

Whatever the disadvantages, limitations and tradeoffs of AI or machine learning may be, one clear advantage is that it's done with machines - and machines don't automatically need food, water, and oxygen.

Machines may not need the fruits of agriculture, but they can make agriculture possible in space, on some off-world body. And they can do it at scale, and in abundance, for the benefit of humans.

But how can we bootstrap such applications here on Earth, in order to eventually extend them into space?

[JohnFornaro]

Whatever the disadvantages l... machines don't automatically need food, water, and oxygen.
But how can we bootstrap such applications here on Earth, in order to eventually extend them into space?

Pretty sure they "automatically" need electricity.

As to "bootstrapping", that implies sentience.

[Elmar Moelzer]

Things that I am hoping to see machine learning algorithms tackle:
New materials/alloys/ceramics for different applications. E.g. you give it the desired properties for the material and it will come up with a bunch of different solutions. I would like to see what machine learning could do with that.
I could imagine something similar for engine configurations, maybe solar panels, fusion power generator configurations (is already happening to a small extent), etc, etc.
I don't think that we are quite there yet for these things, but given the recent advances it could be sooner rather than later.

[Greg Hullender]

There's lots of good work on using machine learning for alloys. Google machine learning for alloy design.

Basically, if it's a tedious job that humans can do but can't explain precisely how they do it, it's a good candidate for ML.

[Greg Hullender]

. . . An effort to develop human intelligence using current AI technology is like trying to get to the moon using ancient Babylonian technology (i.e. building a tower tall enough to reach the moon). You simply cannot scale it up that far. Nowhere near.

Until maybe 5-10 years ago I'd have agreed, but the recent 'Deep Learning' boom has shown that techniques long disregarded and woefully inefficient and inelegant can be perfectly adequate at performing real-world tasks if you can throw enough hardware at the problem.
Human consciousness arises from a massively parallel array of incredibly stupid individual neurons working kinda-sorta in tandem and sometimes producing the 'correct' result, so I rather suspect any human-like AI will arise from a similarly messy pile of hacks and excessive use of compute resources rather than some elegant new simulation technique.

I've been impressed at deep learning as well--we always had the notion of multi-layer perceptron networks, but no one had a good way to train them. It definitely moves us from mud towers to stone towers. :-)

Where I disagree is with the idea that intelligence will "arise" from a sufficiently large system. I see that as no more likely than new life arising because a kid mixed all the chemicals in his chemistry set together. I know the notion of "emergent behavior" in software systems is popular with a lot of people, but I pretty much reject it outright.

Perhaps the important point is that machine learning has a lot of potential for space applications, but if people think we're on the verge of creating HAL 9000, they're going to miss the things this technology is actually capable of.

[ppnl]

Where I disagree is with the idea that intelligence will "arise" from a sufficiently large system. I see that as no more likely than new life arising because a kid mixed all the chemicals in his chemistry set together. I know the notion of "emergent behavior" in software systems is popular with a lot of people, I pretty much reject it outright.

But do you have an argument for such an opinion? A neuron is just a physical object that obeys physical laws. The Church-Turing thesis suggests that a computer program should be able to emulate the function of a neuron. While emulating a hundred billion neurons with a thousand trillion interconnections is challenging there is no new physics here as far as we can tell. Therefore it seems to be mostly an engineering problem.

The transistor was invented in 1947. It was around the size of a matchbox. In the span of one lifetime we have learned how to put over five billion of them on a little chip of silicone. Before 1947 such a thing would have consumed a large fraction of the electrical output of the entire US and the waste heat would have torched an entire city. I have a dozen of these things operating in my house now. I see no reason that A.I. cannot proceed similarly. A.I. will be embedded in just about every piece of technology we produce. Asking how it can be used in space is like asking how transistors can be used in space.

One idea I like is a full self driving Mars rover that can explore without having to supervise every wheel rotation. In the future a self driving rover that can decide for itself where to go and what samples to collect. Eventually a self driving rover with an effective PhD in geology. I see no hard limit.

[sanman]

Whatever the disadvantages l... machines don't automatically need food, water, and oxygen.
But how can we bootstrap such applications here on Earth, in order to eventually extend them into space?

Pretty sure they "automatically" need electricity.

As to "bootstrapping", that implies sentience.

Sure they need electricity, and there's plenty to be had across much of our solar system (at least the inner parts; for outer parts you need nuclear)

And we sentient humans are the ones to do that bootstrapping of innovation in building machines to carry out activities here on Earth which can be extended to off-world environments.

Maybe we can have semi-autonomous earth-moving machines and construction robots to build things here on Earth, and their core logic can be used in machines that would operate off-world.

[JohnFornaro]

Where I disagree is with the idea that intelligence will "arise" from a sufficiently large system.

Agree.

The process is properly called "Intelligent Design".  [ID]  Design is an iterative process, and the design ends at an arbitrary point when the designer is satisfied with his work.  Even so, there is surely a minimum size of the system, unless one wants to assert that an electron has "intelligence".  We do not yet know how big a "system" can be built either, but your point holds.  A larger system, properly written, would be more likely to generate that intelligent "quality" that is being attempted to replicate.

The intelligence won't "arise" until it has been "created".

[JohnFornaro]

Where I disagree is with the idea that intelligence will "arise" from a sufficiently large system.

But do you have an argument for such an opinion?

He made the argument with the chemistry set analogy.

A neuron is just a physical object that obeys physical laws. The Church-Turing thesis suggests that a computer program should be able to emulate the function of a neuron. While emulating a hundred billion neurons with a thousand trillion interconnections is challenging there is no new physics here as far as we can tell. Therefore it seems to be mostly an engineering problem.

The transistor was invented in 1947. ... I see no reason that A.I. cannot proceed similarly. A.I. will be embedded in just about every piece of technology we produce.

Here, however, you are conflating the size of a physical object with its native or possible intelligence.  No matter how small a transistor you make, it will still be a transistor.

If you want to raise the topic of intelligent matter, here's an interesting essay.

http://www.esalq.usp.br/lepse/imgs/conteudo_thumb/Intelligence-in-Matter.pdf

[JohnFornaro]

Whatever the disadvantages l... machines don't automatically need food, water, and oxygen.
But how can we bootstrap such applications here on Earth, in order to eventually extend them into space?

Pretty sure they "automatically" need electricity.

Sure they need electricity, and there's plenty to be had ...

But wait.  There's also plenty of "food, water, and oxygen" to be had.  What point were you making?

[sanman]

But wait.  There's also plenty of "food, water, and oxygen" to be had.  What point were you making?

I thought we were talking about machines being able to exist and operate independently in space better than fragile human beings can. You responded that machines still need electricity, and I then responded that they can get that electricity out in space (eg. solar). Alright, so there can also be food, water, oxygen out in space away from the Earth, but they're much harder to come by than mere solar flux -- especially the food.

Since machines can be engineered to withstand the rigors of space, while we Earth-coddled human beings can't be similarly re-engineered so easily, then maybe it's more useful to start making more autonomous/semi-autonomous machines for productive use here on Earth, so that they can be transplanted to off-world environment to perform similar functions there.

[Greg Hullender]

But do you have an argument for such an opinion? A neuron is just a physical object that obeys physical laws. The Church-Turing thesis suggests that a computer program should be able to emulate the function of a neuron. While emulating a hundred billion neurons with a thousand trillion interconnections is challenging there is no new physics here as far as we can tell. Therefore it seems to be mostly an engineering problem.

The transistor was invented in 1947. It was around the size of a matchbox. In the span of one lifetime we have learned how to put over five billion of them on a little chip of silicone. Before 1947 such a thing would have consumed a large fraction of the electrical output of the entire US and the waste heat would have torched an entire city. I have a dozen of these things operating in my house now. I see no reason that A.I. cannot proceed similarly. A.I. will be embedded in just about every piece of technology we produce. Asking how it can be used in space is like asking how transistors can be used in space.

I'm not arguing that intelligence is supernatural--just that we don't have the foggiest idea how to engineer such a thing. Nor is it reasonable to suppose that if we just make our computing systems bigger they'll somehow magically become intelligent. I hope you don't think Starship will become intelligent just because it's bigger than Falcon 9! :-)

I think this is a really popular belief among people with vague ideas of how computers work and no idea at all of how artificial intelligence works. Trust me on this--lack of computing power is not what stops current (or foreseeable) AI technology from having human-level intelligence. It's easy to say it's an engineering problem, but it's a problem where it's clear that something critical is missing from our current systems, and we have no idea what that is. That's why I like to make analogy with the ancient Babylonians trying to reach the moon by building higher and higher towers, not realizing that they didn't know enough about the problem to make even a start at it.

Over the rest of the century, AI will make a lot of contributions to society in general and to space travel in specific, but that will be real AI, not some fantasy AI that approximates human intelligence.

[Coastal Ron]

But do you have an argument for such an opinion? A neuron is just a physical object that obeys physical laws. The Church-Turing thesis suggests that a computer program should be able to emulate the function of a neuron. While emulating a hundred billion neurons with a thousand trillion interconnections is challenging there is no new physics here as far as we can tell. Therefore it seems to be mostly an engineering problem.

...

I'm not arguing that intelligence is supernatural--just that we don't have the foggiest idea how to engineer such a thing. Nor is it reasonable to suppose that if we just make our computing systems bigger they'll somehow magically become intelligent.

I remember back in the late 80's a company I was working at, that had LOTS of Phd's, thought that neural networks would solve the A.I. hardware challenge. That was also about the time when "fuzzy logic" was thought to be the next revolution in consumer appliances, for making "smart washers" and such.

Needless to say the hype did not live up to reality, though no doubt we learned more about what we didn't know than what we did know.

Fast forward to fairly recently and we saw a similar boom and bust cycle with A.I. and its various subcategories. Though it looks like the bust cycle is not so bad with A.I., as we have found plenty of applications that can use its limited abilities.

But to your point, it does not yet appear that we understand how to make truly intelligent artificial intelligence, as opposed to smart tools.

So from that standpoint, of course "smart tools" like A.I. can be used for space applications. But I don't think they will be able to solve the challenges we have holding us back from expanding humanity out into space.

[JohnFornaro]

Trust me on this--lack of computing power is not what stops current (or foreseeable) AI technology from having human-level intelligence.

"Trust me on this" -- While you may assert that we [or you?] know the full extent of human intelligence, and the computing power and complexity of the human brain, you and we do not. 

The people who are engineering AI are not, despite their pearl clutching claims to the contrary, trying to create a sentient, reasoning being along the lines of human intelligence.  Rather, they are working more to implement a system of control than anything else.  Musk's recent revelations, and Twitter's instantiation of AI argue for this interpretation.

But your point still holds:

It's easy to say it's an engineering problem, but it's a problem where it's clear that something critical is missing from our current systems, and we have no idea what that is.

[Greg Hullender]

"Trust me on this" -- While you may assert that we [or you?] know the full extent of human intelligence, and the computing power and complexity of the human brain, you and we do not. 

I'm simply saying that scaling up existing systems isn't going to produce human intelligence or anything like it, for the same reasons that scaling up a truck or a rocket isn't going to. People never imagine that a big truck or rocket will suddenly develop its own opinions, because they understand those technologies well enough to see the absurdity of it. But they have almost no understanding of current AI technology. If they did, they'd realize (as all serious people in the field do) that the propositions are equally absurd.

I don't believe you have to know "the full extent of human intelligence" to see this. You just need a broad understanding of the current state of the art in AI.

[edzieba]

I don't believe you have to know "the full extent of human intelligence" to see this. You just need a broad understanding of the current state of the art in AI.

The "current state of the art in AI" is to continue to throw more and more compute at the problem, using techniques (bog standard MLNNs) that were dismissed as dead ends decades ago, and to continue to improve in capability as a result of that increased available compute.

Turing Machines (the majority of current computers) work by being very dumb very quickly, and its turned out that 'AI' can work very well by being very dumb in parallel, sufficient to be just as useful as out UTMs.
Of course, as has happened every single time an AI technique has been adopted into production, it will be dismissed in short order as "not real/true AI", as if the sole goal of AI is to replicate human-level intelligence and ignore the existence and utility of the huge range of other useful intelligences.