# What would an AI need to sustain itself for millions of years?

Say a computer running an AI exists on a planet with no biological life. Say it has access to a rather extensive preexisting manufacturing plant and can control a fleet of machines capable of performing the whole process of mining and transporting raw material to build and assemble more machines, fabricate replacement parts, or build new things of its own design. Can it maintain itself for a long period of time?

What possible challenges and dangers would it need to overcome, and how might it do so?

What natural resources must be available for it to build space traveling vehicles, capable of gathering further resources? Would this even be a feasible course of action for it?

Assume its hardware and construction is similar to a present-day earthly supercomputer (just bigger and hand-wavily enough faster to run the AI in real time). For fun, geological stability, and difficulty in resource extraction, lets say it starts out on (or in) a largish asteroid (rather, a smallish moon) in orbit around a the previously mentioned lifeless planet. Lets say the planet is rocky and rich in similar elements as earth. Say this planet had some plant life at one point, so there is carbon and oil which can be extracted for plastics (at least for a little while).

It has access to a nuclear power source, so it isn't necessarily dependent on getting energy from the system's star, but that's an option too.

The AI's goal is to learn as much as it can, so it may have an interest in sending probes to other systems. Or going there itself, if it can fit itself out with ion drives and carry enough $\Delta v$ to escape one system and arrange to be captured by another.

• Ores: iron, copper, aluminium, chrome, vanadium, tin, zinc, and all the other metals used in industry. Carbon and hydrogen to make plastics. Sources of silicon, galium, arsenide and whatever other materials for the microchips. A source of energy, such as sunlight, uranium, thorium or, even better, knowledge of how to build a nuclear fusion reactor. A means to run the thermodynamic processes, which means an efficient method of cooling stuff -- water or air. And, most importantly, a will to sustain itself for millions of years. – AlexP May 10 '18 at 8:48
• Why has this question not been closed as too opinion based? How are we to know what resources a fictional machine would need? What if a new element is discovered or synthesized? How are we to predict the future? Remember that "asteroid collision" is as valid an answer as "power cable insulation disintegrates." Tell us what your machine needs for resources and we'll tell you if it's possible to collect them. Tell us about your machine's circumstances and we can tell you how those circumstances can fail. – JBH May 17 '18 at 5:13
• @JBH Thanks, I've tried to add some specifics. – gandalf3 May 17 '18 at 7:52
• It's not been closed because this is an ideal question for this forum! Questions regarding real world stuff are on-topic. Can't get much more real world stuff than AI and the materials & technologies it would require to keep itself active and healthy! – elemtilas May 24 '18 at 15:38
• The term AI is too broad to have a specific answer, at least please define if AI means ANI, AGI or ASI. Otherwise it's like to have a question asking if a worm, a person or a lamp genie can survive that long. You can imagine that the answer will change a lot depending the subject. en.wikipedia.org/wiki/Artificial_general_intelligence – theGarz Jun 13 '18 at 10:32

A Similar Concept has been explored by scientists already, they are called Von Neumann Probes, basically their programming is to travel to a star system, map the system, mine any asteroids to self replicate, and then send onto another system. this in theory could leave a satelitte in every system in the galaxy if left long enough, the issues come down to avoiding any debris sitting between star systems. however we already have the technology availble to us to send one of these probes, however the "self Replicating part is not 100% perfect, and more importantly it would take so long that by the time the first probe reached our closest star, Alpha Centauri everyone that remembered the launch would be long dead. yes our childrens children would be able to see the results after it took a few years for the signal to return to earth.

The reason they haven't tried making one yet is that a lot of scientists belive it is better to wait until we have engines capable of reach another system in say 50 years, then to send it.

I appreciate that this isn't a single large AI, but a small one copied billions of times. but if this were a single large AI controlling many machines, then it would need to do the same sort of job just exponentially bigger

Resources

These are basically everything AlexP mentioned plus a few more

The main ones would be Gold, Silicon, Copper, Palladium and Aluminium, these would be for microchips circuitry, capacitors and wiring,

Then of course, Iron, Carbon, Zinc, Vanadium, Tin for the body of the probes. many of these elements already mentioned also make up photo-voltaic cells (solar Panels) as well as lithium for making the lithium compounds to use in making the batteries.

Then just mine ice for hydrogen and oxygen to use as fuel,

there are a lot more then this but rather than listing them all or simply adding a link to the periodic table, these are the most important that i can think of.

Programming the AI is a challenge

Having dabbled in programming - the big problem with AI is not going neurotic.

You can programme an AI to deal with as many known issues that can be foreseen - but millions of years is a long time. Imagine on this planet backwards several million years. We had:

• continents changing
• several ice ages
• complete animal life change
• atmospheric changes
• geological changes
• probably innumerable solar flares
• all sorts of things.

Not only are the conditions changing, but AI itself depends on self improvement and learning algorithms - I cannot foresee an AI self-learning and programming to last millions of years before going belly up eventually, especially if made from scratch.

Unless - the AI is actually at the tail end of a billion year process of programming it - in other words it survived where other AI had not.

The resource stuff I would imagine is easy, unless it's in a hard-to-get-to spot where the AI won't destroy itself trying to get at it...

• By AI I mean a true, self-aware, sentient AI. Not a neural network trained to handle a limited set of contingencies foreseen by its creators, but a fully autonomous AI capable of thinking for itself and learning as it goes. – gandalf3 May 10 '18 at 22:14
• @gandalf3 It will likely still be a neural network AI. The structure of a neural network mimics the way neurons in our body and brain works, hence the name. Its simply the size/scale and how the neural network is trained that will be different. There are 100 billion neurons in our brain and even more in our body that are all interconnected and constantly change, trained for the entire duration of our lives, constantly receiving visual, audio and physical information. No current neural network has been able to match the complexity and amount of data we receive, over the time period we receive. – Shadowzee May 11 '18 at 1:06

## The AI survives so long by becoming Alive then Dying

The only systems that we have seen that can survive for millions or tens of millions of years is life itself in the form of ecosystems. No individual life form lasts very long but the system itself survives the required millions of years.

In contrast, we see that centrally designed and implemented systems by humans (and only by humans since nature only does distributed design) fail regularly and spectacularly. Even very resilient human systems may survive for a long time but inevitably, something will change and the system crashes. A big, central AI is a single, exceptionally expensive bet that will fail and likely soon. Designing and programming this system to survive a million years is a fool's errand.

This AI will have to deal with failing hardware, failing network connectivity to other compute nodes, death of the factories, break down of the mining facilities, exhaustion of local minerals, environmental changes that can't be anticipated in advance, failure to maintain normal operating parameters in the datacenter where the AI lives. Basically, every struggle that life now solves, this AI is going to have to replicate with machines and it's own programming.

It will also have to deal with the failure of the AI itself. The AI will evolve over time and develop cruft. This may be in the form of suboptimal action potentials that can't be changed without starting over. Maybe there's random corruption in a single bit in the exact wrong place. Life solves this by having everything die. Birth gives fresh starts and new chances to experiment with things. It will need smaller, more agile AI/robots to survive the lean times/extinctions when it's just too expensive to be a big AI/robot.

If this AI is going to survive, it must build an ecosystem to support itself. It must have children and it must adapt to meet the local circumstances. Failure to do so means failure to survive for the required timeframe.

• Maybe there's random corruption in a single bit in the exact wrong place. So basically (digital) genetic damage. I wonder if it would lead to digital mutation and versions of genetic diseases. A digital cancer would be possible, where the data begins to replicate out of control, slowly metastasizing, and if left unchecked it could eventually kill the AI. – AndyD273 May 24 '18 at 14:51
• @AndyD273 we already get bitrot in digital systems with half-lifes measured in months or years. Any modern hardware will be an inert brick in a few hundred years of operation. They just don't survive that long. – Green May 24 '18 at 14:56

Self repair and software redundancy. The machine would need to be able to repair itself and also have multiple backups for every process. Something like 5 layers of redundancy should work, basically any time it does anything it compares the action to 5 templates to see if its the same. If one of the templates is different than the others it re-references the other 4 then makes the damaged 5th one the same as them. This applies to every single action the AI takes, which means in essence you actually have 5 AIs running 5 identical sections of the ship constantly checking eachother for uniformity and operating as a consensus.

A note on AI, the idea of one rebelling is kind of silly, since that is confusing artificial intelligence with artificial conscioussness. The AI does not need to actually be self aware to do this job. As a matter of fact, no AI really needs to be self aware to do its job. Im imagining a ship that is 5 cylenders in a hexagonal bundle with 5 engines and 5 sensor suites and 5 reactors and 5 repair bays etc etc. Each AI runs one section which is itself fully capable of travel, resource extraction and refueling, repair, etc etc. Whenever any action is taken all 5 cross reference the decision with eachothers databases. In the event of an extreme emergency or failure the 4 AI will destroy the faulty 5th and even be capable of completeley rebuilding the destroyed section and replacing the now gone AI with a new copy of themselves.