2

I recently posted a similar question over on Stack Overflow but they just weren't having it, so I figured I'd bring it over here.

I'm developing a model for a general AI based roughly on Douglas Hofstadter's theory of analogy/categorization as the core of cognition. I'm imagining a mobile robot with stereoscopic vision, a microphone and a speaker. Using machine learning techniques which are well established it shouldn't be too difficult to spoon feed the robot a series of useful "symbols" to start off with, like what a chair looks like, what a desk looks like, what a computer looks like, etc. There could even be a mirror in the room and you could teach it that when it sees its reflection it's seeing itself, which would in essence make it a self conscious artificial intelligence. Maybe that last bit is a leap of wishful thinking but let's run with it for now.

So we have this sentient robot, basically running an object oriented program in which it roams around the room and recognizes and categorizes things (if you're familiar with OOP it actually makes perfect sense for AI, assuming you agree with Hofstadter's take on cognition), so according to Hofstadter it would have some degree of "understanding" about the world around it. For instance you could walk into the room and hold up a piece of paper and say "what is this robot?" and it would say "paper." Then maybe if there's enough examples of paper in the room you could ask it "what is paper used for" and it may respond, "to write things on." Great, we may now have the world's best general AI in our possession, but, here is where I get stuck. As humans, we have an innate tendency to want to explore the world a bit, and I think this can be boiled down to the fact that we have needs and emotions. We get hungry so we go explore the kitchen, we get horny so we go explore the clubs (or internet more likely), but our robot would require none of that. As long as it's plugged into an outlet I don't see any reason why it would have any desire to do anything other than sit there like a perfect little monk, thoughtless, and just existing.

So, if you're with me so far and you have any background in programming, how would you go about putting incentives into the code to make the robot actually do something? I mean, you could hard code it to just try and find new objects, or roam around aimlessly categorizing things and gaining knowledge, but since that's hard coded in it kind of feels like a shame or like breaking the rules of what it means to have a general AI. Although I guess that's what I argued our DNA does for us so maybe that is the answer.

I apologize that the question is vague, I'm both looking for some sort of algorithm in the code which would incentivize the robot to act more like a human and less like a giant Amazon Echo, but I'd also like any general feedback or ideas on whether this would work, improvements to what I've suggested, whatever.

spoof
  • 21
  • 2
  • I suspect the question may be off-topic here, but perhaps some people may have some answers or insights. A programming site might be a better place for this. Welcome to Philosophy! – Frank Hubeny Jan 09 '19 at 19:11
  • Try CS.SE? This question is more like how to algorithmize AI properly rather than philosophical one. Also, a robot can't learn that paper is paper without examples. Just like humans. – rus9384 Jan 09 '19 at 20:00
  • Don't forget about Isaac Asimov's robot rules! It's not all the way there, but it's a good start. – elliot svensson Jan 09 '19 at 20:01
  • 1
    And of course, there's a Stack Exchange for AI I hadn't thought to check out. Here's a similar QA if anyone is interested. https://ai.stackexchange.com/questions/3903/what-would-motivate-a-machine – spoof Jan 09 '19 at 20:25
  • I think you're asking us how to code intelligence? A neural network will be naturally inquisitive about whatever it's feedback parameters are. In most single networks this would be... Is this 16x16 pixel image a letter 'a'. In a deep learning network composed of 27 such networks it might be.. is this image any Latin letter? In a big enough network.. it would be everything... Is that a human? Yes... Is that human about to burn me? No. Is that human likely to aid me...should I converse with them... Etc. Intelligence will likely grow from network size... – Richard Jan 10 '19 at 00:51
  • You will not get away from a hard coded 'trigger' to get things started, see 'instinct' in animal behavior (humans included). The trigger can (should) be recursively modified by learning... Also https://en.wikipedia.org/wiki/Swarm_intelligence can be virtually implemented, don't know if someone has tried that recursively... – christo183 Jan 10 '19 at 05:23
  • Your question describes a sophisticated abacus. It will never be curious. If you program it to do things that a curious robot might do it will still not be curious. –  Jan 10 '19 at 15:10
  • 1
    @PeterJ And what is the difference between the curiosity the AI shows (in the improbable circumstance that OP can make it work) and the curiosity you or I show? We're pre-programmed to be curious, after all. – David Thornley Jan 11 '19 at 19:01
  • @DavidThornley - Yes, good point. But being curious indicates a state of knowing we are curious. Can we be curious and not know it? Curiosity requires consciousness. I'd agree that if an AI is conscious it can be curious, but this would be why it can never be curious. –  Jan 12 '19 at 13:07
  • Would love to see reopened, and answer properly. I see the fundamental question as, what is curiosity? Would develop my answer around Tegmark's AI Physicist https://www.technologyreview.com/s/612358/an-ai-physicist-can-derive-the-natural-laws-of-imagined-universes/ & the notion Turing machines can ask how, but strange loops keep asking why. What curiosity is, is a defining question for philosophy, once it is recognised as one of the only remaining motivations in an absurd world, following Camus. I hear issue is key in actual simulated brains. I'd analogise wobbles needed for equilibrioception – CriglCragl Jan 12 '19 at 23:13

1 Answers1

-1

Thanks for the question! This may indeed be a better question for Worldbuilding.SE or something, but I'll have a go at it anyway.

Without reading other theories than what you presented, I would say that your AI lacks meaning.

The most obvious meaning for you to write for your AI is "do whatever my creator tells me." Slightly better might be "do whatever is best for my creator." Or better still, "do whatever is best for any person around me." Or maybe, "do whatever is best for my creator without breaking Asimov's robot rules."

Now, "do whatever is best" sounds free of content but remember what AI is: it's a strictly logical, vastly resourced thinking device. Utilitarians know that it's good to apply existing resources to achieve the best benefit for the most people, and your robot could get started identifying the decisions people make that cause waste and reduce that.

Or maybe "do whatever is best" could be directed at the deontological virtues of human rights: then your robot would learn about injustice and get busy making bank walls thicker, grocery store checkout scales more accurate, witnesses more bold, cops less vulnerable, etc.

Or maybe "do whatever is best" would take more of a progressive stance and identify inequalities and past injustices, and begin its work to rectify that: educating those whose parents weren't educated, enriching those whose parents suffered economic injustice, or structuring current institutions such that this generation, at least, doesn't suffer unequally the same way as the last.

Or perhaps the AI could study our ecosystem and identify things that endanger the world's living wealth: species of bugs, forestry practices that don't work, etc.

Or perhaps the AI could dedicate itself to expanding individual human abilities: could it think of how to build the best submarine? Could it design the best lunar colony? The best personal jetpack?

OK, all these goals are very far from the code you insert into an AI. But they could serve as metrics from which your AI could begin to value its surroundings, and evaluate how best to spend its limited time before the next planned maintenance and before you get too old and die.

After all, current chess computers are not built on a brute-force methodology: they assign a metric to any given decision and do the thing that they think will increase the metric.

elliot svensson
  • 4,065
  • 7
  • 25
  • 1
    I'd personally like to see a robot read everything in every science journal and map out which ideas are original, which ideas are contradictory, which ideas are trivial, which ideas would benefit the most from being spread around... which foreign-language ideas have suffered the most from being siloed in their home language instead of English, etc. – elliot svensson Jan 09 '19 at 20:21
  • But, what is meaning in your view? – CriglCragl Jan 12 '19 at 23:15
  • Can you explain how to program: "do whatever"? How would that "motivate" a robot do anything at all? – christo183 Jan 13 '19 at 03:31
  • @christo183, by "do whatever my creator tells me," I mean that the robot is an assistant-bot. At first, it's like having a baby but in time, the bot will be able to get your coffee and fold your laundry the right way at the right time. Later on, the bot could pick stocks, choose your March Madness brackets, or have parent-teacher conferences for you... or whatever you would hope it to do for you (with your consent, of course). – elliot svensson Jan 14 '19 at 14:57
  • @christo183, my idea of "motivation" for the robot is exactly the scheme used to program chess-bots. Every action gets a metric, and the robot chooses the best action. I suppose one thing that's not intuitive is "the zeroth law of robots", in which inaction, too, gets a metric. – elliot svensson Jan 14 '19 at 14:59
  • The OP question seemed to imply an autonomous learning cycle. I'll grant that a regime of of human input driven learning would likely be a more efficient scheme, more natural at the least. After all how would one learn 'intelligence' if not from another intelligence? – christo183 Jan 14 '19 at 16:21
  • @christo183, it brings up a wonderful point... which living things do you know of that grow up entirely alone in the world? – elliot svensson Jan 14 '19 at 16:23
  • Even HAL 9000 showed mecha-affection for its creator, Dr. Chandra, who taught it a song... – elliot svensson Jan 14 '19 at 16:27
  • @CriglCragl, I generally define "meaning" according to a military analogy: if I'm doing what my superior officer wants, I'm doing something meaningful. Otherwise, I'm not. – elliot svensson Jan 14 '19 at 16:31