Getting into the mind of an AI

I think I would look at Utilitarianism as a possible starting point. I think you're basically looking at a psychopathy that initially operates under a clearly defined set of algorithms. Part of the character development could be the modification of those paradigms (random mutation could even initiate a change) into the end-product personality you are after.

Interesting project.

The problem you might have is seeing inside its mind. Many millions of instructions per second could make for hefty reading

this sounds REALLY interesting!

Yea, the problem definitely would be how he goes from that basic AI intelligence to what you're aiming for, but i believe it could be done!

I ran into a similar problem and needed to scrap the book. (for now) I think you should keep the character 'dry' but not like Hal. Just be more analytical and emotionless in his/its observations of the world around it. How many words do you plan to write for this short story? How long do you need to focus on the beginning of the AI's existence?

Sapphire Flame said:

So, there's been this short fiction story idea that's been kicking around in my head for a while now. The story is essentially the time-honored cliche of "AI gains sentience and kills people", but told from the perspective of the AI itself, examining what would drive an artificial intelligence to do such an action.

There's just one problem. The way my AI was designed, it slowly evolves over time as it gains intelligence, and initially it is merely an ordinary computer. It's similar to "Flowers for Algernon", in that we see the protagonist's mind changing as the story goes on.

Click to expand...

As far as an "ordinary" computer, it's not feasible that, running "ordinary" programs, it would "learn" by itself, no matter how powerful it is.

But if a programmer get tired of having to add to a program for every new and different input data or feature the program needs, so he takes some time to write an extra part of the program that automatically does this, making sense of new inputs by itself. He leaves it running all the time, because even though it works, it takes a long time to recognize new inputs.

He forgets about it, and doesn't notice that as it grows it gets more efficient at "recognizing new inputs." Or maybe he notices and is pleased, but dioesn't think about any possible consequences past that. And of course it keeps going...

Sapphire Flame said:

The problem is that I am writing from the perspective of what is essentially an alien intelligence. I don't want it to start thinking and talking like HAL 9000 as soon as it's switched on. However, a computer is essentially a complicated device for running programs. It doesn't think deep thoughts, it doesn't question its existence,

Click to expand...

Yes, that's always true, but that's not quite the way to look at it. A human body, including the brain, doesn't do these things, but it supports a human MIND that does do these things. The analogy to the mind isn't the computer itself, but a computer program which might do these things.

Sapphire Flame said:

and it doesn't keep a diary of what its day was like.

Click to expand...

As far as an informal diary, no, but there are lots of little things lept track of in modern computers, and the more powerfuil they get, the more things are kept track of. Usually what programs were started and what time, and of course web browser "temporary files" and "cookies." Some (all?) inkjet printer drivers keep track of the total number of dots printed from each print cartridge.

To the point of your story, a program with "growing intelligence" would, as a part of learning, notice unusual inputs and save them, along with the date, time and in what sutiation they happened to refer back to later.

Sapphire Flame said:

So, Absolute Write people, how should I go about writing the earlier chapters? How should I write the thought process of a computer? How should I clue people in that there's something more going inside its mind?

Any tips at all would be nice. I'm determined to write this thing.

Click to expand...

This could easily be a novel. I suppose I should write something like this myself. An older one (quite old in cyber time) is "The Adolescence of P1."

What was the name of the computer from I Robot? Yours sounds a little like that. Then again, you could describe it as an evil version of the Bicentennial Man robot.

I know I'm making mostly movie comparisons but I have one more that you could consider. Agent Smith from The Matrix. Even before Neo freed him, he was displaying signs of independent thought. The scene with him and Morpheus was great, especially how he compared his "definition of humanity" not with mammals, but said we're more like a virus.

That could be an interesting avenue to take.

I've tried to write AI perspective several times, and nearly always come back to trickery and handwaving of the programming underlying the sentience. The main problem comes when the AI needs to do something which is logical to the character but seemingly illogical out of the context of its' immediate surroundings.

Simple words and basic sentence structure could be 'out of the box' AI, before it has had the opportunity to mimic learn words and sentence construction which those around it use. As it adapts to its' surroundings it will slowly pick up a better understanding of itsself - think of Data in ST:TNG without the idiotic bits from later seasons.

Have you thought of how much on-board memory the AI will have available? It will probably need a pretty hefty chunk of RAM, disposing of 'irrelevant information every so often. A sleep cycle will be needed to keep it from clogging up its' memory with crap if it is using vast amounts of information.

You might try jotting notes on all the AI stories you've read, that you remember, and see what you can think of nobody's ever done. You might also try realistically inferring what would motivate an AI to self program, to think and behave on its own - what drive? A single program to do so, or an actual emotion?

here's my own take, "Angry Army Against Artificial Intelligence"

Real World AI News Articles
http://www.bloomberg.com/news/marketsmag/ai.pdf
http://www.wired.com/techbiz/people/magazine/16-02/ff_aimystery?currentPage=all
http://ebiquity.umbc.edu/blogger/2007/07/09/ai-pioneer-donald-mitchie-dies-in-automobile-accident/

I would suggest reading the 'Bolo' books which were created by Keith Laumer, they are written partly from a human point of view and from the Bolo's point of view.

In case you don't know, Bolo's are city-block sized, super-intelligent battle tanks

Slant, by Greg Bear.

Sapphire Flame said:

how should I go about writing the earlier chapters? How should I write the thought process of a computer? How should I clue people in that there's something more going inside its mind?

Click to expand...

All characters -- animals, elves, ents, aliens, gods, cities, storms, AIs... are just actors in rubber suits with CGI. Some are very unusual, but what people create for the entertainment of other people is... people.

So recognising that, it's perhaps worthwhile asking what sort of person is your AI? Who is it physically, psychologically? Socially? Especially, the key questions I'd want to know are: when does it become aware of Self and Other, when and why does it decide that killing Other is a good thing for Self to do? When does it decide that killing Other is enough, or a mistake? In other words, when is that need slaked?

For a story like this I'd suggest writing the AI with several voices -- each voice reflecting a period of its development.

For example, suppose that you decide that early on it is taught by example, frequently makes errors and has not much sense of time. You could write like this...
Then gradually you could mutate the voice... introduce time, first person, and begin to add emotions... A very advanced AI might begin to anticipate complex matters, might coin its own idiom for concepts that humans don't care about, might learn to lie, might begin to create its own inner dialect, might amuse itself with poems or puns or word-associations, or assonance or alliteration and so on... Might it express anger through alliteration, say? Does it express joy with portmanteau words? Would humans pick up on that? Might they misunderstand or see it as malfunction?

Hope that helps.

I have an AI which learns from experience, and adapts its programming accordingly (mentioned in my story "Obedience"). I imagine you have to start with a set program, and then figure out the opportunities and limitations of your specific AI, and go from there.

Sapphire Flame said:

The problem is that I am writing from the perspective of what is essentially an alien intelligence. I don't want it to start thinking and talking like HAL 9000 as soon as it's switched on. However, a computer is essentially a complicated device for running programs. It doesn't think deep thoughts, it doesn't question its existence, and it doesn't keep a diary of what its day was like.

Click to expand...

Ah, hold it right here Isaac.
One thing you can't say about AI's is how they think, or don't think. Not because we don't know yet, but because it's a matter of software. And a such can be anything. You could create an AI by simulating a human brain on a sufficiently powerful computer, in which case the AI will think exactly like a human. Or you could simulate a jellyfish nervous system, and enlarge it until it becomes intelligent. Who knows how that might think. (like Cthulu probably)
Or you create an AI from scratch by writing increasingly capable neural networks which become self aware when they're big enough (no-one if that could really happen, but hey, it's fiction).
Anyway, point is: the "mind" of an AI is pure software, so there are no ways it has to work or wouldn't work.

Definitive statements can be made about hardware however. Memory for example would work like digital memory, not like the mess we humans have. An AI doesn't need to "remember", it knows exactly what it knows, and it knows everything it knows. All available information (in RAM and on harddrives) is indexed and readily available.
Digital storage space is cheap as dirt. Cheaper. And works differently than human memory. We can easily remember events, that would have to be stored as a large video file or something similar. We have difficulties remembering the contents of an entire phonebook, which takes next to no space on a digital drive.
Sleep is a biological mechanism we need, an AI wouldn't. Sleep serves several functions for humans, none would be relevant for an AI though. Unless possibly if it runs on some kind of bioprocessor.

An interesting question about an AI is multitasking. If its artificial, it's reasonable to say it would probably be able to multitask like a computer can. I.e. it would have no problem observing a hundred video inputs at once, being able to divide its processor power easily and at will. Your growing AI sounds like this type.
If it was modelled on a human mind, it might work similar to a human mind, having only one major attention "focus point".

BigWords said:

Have you thought of how much on-board memory the AI will have available? It will probably need a pretty hefty chunk of RAM, disposing of 'irrelevant information every so often. A sleep cycle will be needed to keep it from clogging up its' memory with crap if it is using vast amounts of information.

Click to expand...

That would be some really crappy programming. It's called a memory leak, and usually considered a bug. Decent programs manage RAM usage while running.

On the subject of RAM, a decent AI would need (from the OP's question it would be self-aware and able to 'see' its' surroundings) masses of memory. We're probably talking terrabytes here, just so the information (light sensors, audio receptors etc.) can be sorted into useful data which it can respond to. It makes sense to give any character (even an AI) some limitations, and this seems the easiest to implement into the story, if only because it figures that even the most complex systems currently operational are so power-hungry and yet manage to be inefficient. The 'seeing' robots are particularly bad at simple tasks despite being coded specifically for that one purpose. Crappy programming is one constant which will go on for as long as people (fallible people) are programming software.

When we get to the point where complex systems are designing even more complex systems, we'll either end up with bug-ridden, crash-happy systems which are next to useless or perfect systems which never crash.
Guess which I'm betting will be the result.

Why not look into genetic algorithms?

What if your computer works with genetic algorithms which is like breeding (combining) programs that work really good but eliminating those that don’t function as needed. In real life, genetic algorithms create programs and do things humans could not have imagined. For example (I can’t remember the exact scenarios but here’s the gist, a program is told to use the least amount of 18 circuits to perform 18 processes, all the while changes in temperature could kill the program, so the program keeps breeding until one of its offspring can work in all sorts of temperatures with only 4 circuits. The scientists were puzzled how the program could become so efficient and discovered that the computer wasn’t using direct circuitry, but indirect circuitry from the radiation of proximity. In other words, the computer did things engineers wouldn’t have thought of, passing information across unconnected circuits.

Perhaps your computer can be creating genetic algorithms for some purpose. It breeds programs, learning and killing off programs that don’t function best. Perhaps the computer becomes somewhat out of control, but attempts to kill it fail because the computer has learned to live outside its known circuitry. Perhaps the computer could behave like Dr. Manhattan (Watchmen), not so much that it exits outside in a human form, but that the computer is doing its own thing and cannot be bothered by human needs or requests. It isn’t evil, it’s just oblivious that its actions are causing harm.

How can someone be oblivious that their actions are causing harm? Forest fires are a great example. For years we’ve been putting out fires, not allowing them to burn because we were taught that all fire was bad. Now we realize that fire isn’t bad, that it was happening well before humans existed, and nature was fine. Nature actually needs fire to clear brush and some seeds need fire to germinate, but former forest fires didn’t cause the damage and destruction they do now. With so much brush, fires burn longer and stronger than before, so in our attempt to do good, we caused more harm.

Anyway, I know this post is all over the place, but maybe it will germinate an idea without flaming your concept!

BigWords said:

On the subject of RAM, a decent AI would need (from the OP's question it would be self-aware and able to 'see' its' surroundings) masses of memory.

Click to expand...

The separation of storage from computation is an architectural decision underpinning all modern computers, but it's not the only one we could make. Computation can also be storage.

I think there is a way to minimize both the amount of active visual imput against stored memory (in whatever manner that is, Ruv ) and allow for growth of character at the same time. If the AI is off-setting the constant input ('seeing' who it is conversing with is pretty much essential) with a representation of its' surroundings - like a computer game level loaded into its' system - it can ignore all of the background information it receives. It is essentially ignoring 80% of the input. Same with audio, the background noise (fans for the cooling systems and so on) will be ignored. As it progresses, it can then 'learn' which information is important (the color of a persons eyes) and what is superfluous (the walls, ceiling, carpet) and rely less on what the creator of the software has begun. It will, essentially, be writing its' own software after a certain point.

Of course, that would be too dry and technical for a lot of readers...

Key questions are what a synthetic consciousness considers Itself to be and what is the Outside -- and whether between Self and Outside are any Tools. And whether Itself is a 'We' or an 'Us'. And whether it perceives its experience as a single integrated stream, or separate channels about which inferences can be made. And whether it considers the people it talks to as part of Itself, or Tools, or Outside -- because humans can treat each other as all those things.

BigWords said:

On the subject of RAM, a decent AI would need (from the OP's question it would be self-aware and able to 'see' its' surroundings) masses of memory. We're probably talking terrabytes here, just so the information (light sensors, audio receptors etc.) can be sorted into useful data which it can respond to. It makes sense to give any character (even an AI) some limitations, and this seems the easiest to implement into the story, if only because it figures that even the most complex systems currently operational are so power-hungry and yet manage to be inefficient. The 'seeing' robots are particularly bad at simple tasks despite being coded specifically for that one purpose. Crappy programming is one constant which will go on for as long as people (fallible people) are programming software.

When we get to the point where complex systems are designing even more complex systems, we'll either end up with bug-ridden, crash-happy systems which are next to useless or perfect systems which never crash.
Guess which I'm betting will be the result.

Click to expand...

I suspect that a machine making programs would be perfect, more often than not. Think about it. The incredibly painful thing about coding is that you are basically trying to tell a machine to do something in a really beating-around-the-bush, alright-almost-there manner. It's kind of like telling someone on a millisecond basis how to ride a bike, through a babelfish translation, instead of doing it yourself.

Salis said:

I suspect that a machine making programs would be perfect, more often than not. Think about it. The incredibly painful thing about coding is that you are basically trying to tell a machine to do something in a really beating-around-the-bush, alright-almost-there manner. It's kind of like telling someone on a millisecond basis how to ride a bike, through a babelfish translation, instead of doing it yourself.

Click to expand...

I'll bet against BigWords as well. Mot bugs in code are the result of human mistakes, and a machine doesn't make human mistakes.

Jumping into the fray here. Interesting topic, by the way.

Lhun said:

I'll bet against BigWords as well. Mot bugs in code are the result of human mistakes, and a machine doesn't make human mistakes.

Click to expand...

I don't quite agree. Sometimes, the bugs are not really bugs but old requirements in the face of new insights. A change in requirement, if you will.

Imagine a program that intends to 'travel'. So the objective for this program is to travel from point A to point B.

If this AI/program can evolve, it might start with legs to crawl. Then it'll 'evolve' and mutate to have wings. (cuz wings are always cooler).

If you look at the final product, you might say legs are bugs in the code...but it's not. (and I assume the AI will cut that out later)

In that regards, if we are talking about a perfect AI that evolves...I don't think there is such a thing as a bug.

Because for such an AI, bugs are just minor stepping stones in continual evolution.

I'm a little confused by the descriptions in this thread (not from the O.P) that the A.I is a 'psychopath' or 'evil'.

Why?

It's not like it is doing something bad. It's just killing people.

Even vegetarians kill thousands of mice, roaches, etc on Soyabean farms as a normal part of making vegetarian food. The other creatures get in our way. It would be too expensive in resources to avoid them. So we just kill them.

We don't put most other life forms in a special category of 'things to avoid killing'.

So why on earth should we expect that other life forms will treat us differently ?

From the A.I's point of view it is just changing the life span of someone else from 0.008% of a millennia down to 0.0004% of a millennia. That's no worse than what people do every day by smoking, over eating, driving too fast.

I'll bet against BigWords as well. Most bugs in code are the result of human mistakes, and a machine doesn't make human mistakes.

Click to expand...

Maybe not. But that just opens up a billion other classifications of mistakes.

There was some nice series of stories in Asimov (I think) last year about the first A.I's and how they kept falling for 'Invest here for secrets to win the lottery' etc.

Basically, like us, they are designed to spot patterns in data so if they are just slightly tuned too far they end up getting paranoid and seeing 'patterns' that aren't really relevant.

Good luck !

Mac
(PS: I really liked Ruv Draba's approach)

I love the concept and I adore non-human protagonists for their alien perspective. Details aside, I think the very premise would be exhausting to write (and read) in the long term.

This may work better as a short story than as a novel. I think it would be incredibly salable in short form, but a tough sell to an agent.

i had a similar idea once, except with a bit of matrix twist near the end. theres a guy who gains what he beleives to be "super powers" in that he can change elements in the world. He turns out to be a memory programmed into a computer, able to reconstruct his virtual world, but unable to escape it.

jinkang said:

I don't quite agree. Sometimes, the bugs are not really bugs but old requirements in the face of new insights. A change in requirement, if you will.

Imagine a program that intends to 'travel'. So the objective for this program is to travel from point A to point B.

If this AI/program can evolve, it might start with legs to crawl. Then it'll 'evolve' and mutate to have wings. (cuz wings are always cooler).

If you look at the final product, you might say legs are bugs in the code...but it's not. (and I assume the AI will cut that out later)

In that regards, if we are talking about a perfect AI that evolves...I don't think there is such a thing as a bug.

Because for such an AI, bugs are just minor stepping stones in continual evolution.

Click to expand...

While that's a nice example of biological evolution (and vestigial organs) it's not at all how a program would work. The big difference being that a program changing itself is an intentionally directed process.