[rss] [contact] [forum]

What's this all about?

I'm interested in and write about a wide variety of topics - economics, psychology, marketing, music, etc. I prefer writing long articles to short posts and don't update very often.

Recent Activity

Popular Pages




Now, What if we pooled our resources? (an article on AI)

An article the potential of AI, hurdles that need to be crossed, and why we should and do fear computers.

I've been reading up on AI lately, and I got to thinking. What if all of all of the scientists and programmers and researchers working on AI got together and just built one giant machine. A-la Extreme Programming, everyone would have an overview of the whole so they knew what they were building and could get their parts to talk together, but no one person would understand every part.

Could we, just possibly, get a robot (like the Honda Robot) that could recognize faces, had moods, play chess like a master, answer questions on a wide range of topics, speak... Obviously the robot itself would be a shell, there isn't enough processing power in something the size of a robot to do this yet, we'd just put enough processing on board to carry out the actions that were transmitted to it via radio from a master computer. (2)

Imagine, a neural network made up of dozens or hundreds of other neural networks, heuristics programs, fuzzy logic processors, etc. Much like a model of our own mind where subconscious thoughts bubble up, thoughts would bubble up from one of the sub-neural-networks.

Each network would respond in ways understood by it's designer, but nobody would know which network will respond. Add enough complexity and enough sub-networks, and it would begin to behave unpredictably. Not unpredictable random, unpredictable human. It would have tendancies and moods. What if a Strong AI was no different than a lot of traditional AI's linked together?

A few thoughts occur to my rather uninitaited mind. Aside from getting all of these programs to talk to each other being a massive effort, and determining the over-all architecture, there are a few innovations that would go a long way towards making strong AI a reality.

First, the mind 'creates' memories as it goes, categorizing the information it receives into multiple categories. I'm sure someone out there is working on this problem (perhaps the folks at Google have some insights that would help...) It seems the mind burns neurons as memories enter the brain. Emotionally charged memories (according to Daniel Goleman's book Emotional Intelligence) render strong memories. Similar situations get compiled into one master memory to which every new experience is compared. (see Sources of Power)

A true "strong" AI would have to be able to not only categorize data as it came in, but create it's own new categories. The same way a baby learns about fire, this AI would have to learn to categorize new experiences in multi-dimensional ways. Okay, so this isn't a necessary component for my idea, but it's something that needs to be addressed.

Second, the AI would need a way to carry thoughts with it. Initiate actions. If one sub network was working on a problem, there would need to be a way for it to communicate it's idea outward to the other networks to work on it, or to the master network.

Let's say we send our little robot out into the world and it gets picked on in the school yard for being different from the other kids. (1) It gets beat up and sent home. Parts of the program are important enough to day-to-day functions that they aren't actively engaged in thinking about being beat up, but other parts are. It would try to categorize the event, think through other possible scenarios, try to understand the other children's behaviour. Just like in a human these parts can be on the surface gaining enough prominence in the neural network to occupy it, or not. In other words, be part of the conscious or subconscious mind.

When something particularly interesting happens in one sub-network, a breakthrough thought into why the other children picked on it, it might bubble up into the conscious part of the mind for other networks (including the ones that control speach, for example) to examine and work on.

Or, when asked about the incident, those sub-networks would gain more prominence by being the topic of conversation and their analysis would be brought to the surface in the master (or conscious) mind (or network) to be analyzed and shared with other sub-networks.

Third, the master neural network would fluxuate. It would favor certain sub-networks in certain situations, and favor different ones in different situations. Sitting at home discussing biology is certainly different than playing Chess, and it would have to favor some over others depending on what the situation called for.  Some sort of weighted momentum could create 'moods.' You can't jump from fencing into reading a book and quiet contemplation quites so easily. It takes a while for you to calm down. When you get angry at your boss you don't show it, but you go home and kick the cat. (4)

These points aren't necessary to implement this idea, but I think they're hurdles strong AI will need to tackle if it's to get anywhere. The ability to categorize new kinds of data, and the ability to carry throughts with it into the future - obsess over things, forget and remember, almost dream.

One last note. In my book the root of all evil is the objectification of other life. You can kill an ant because it's an object to you. A sociopath can kill a human because it's an object to him (or her), and in war, we justify our actions to ourselves by objectifying the enemy, which is what makes civil wars so difficult. Treating someone like an object rather than as a person is easy to do - pushing someone around, for example. It's also very easy to do while physically away from the person, making the decision to fire someone is easier than firing them to their face. "Flaming" someone online is easier than chewing them out in person. Objectifying animals is a short step from objectifying humans, and we all do it to some extent or another. Intersetingly it's often those who hold their own strong stereotypes that dislike being stereotyped the most.

Giving a computer responsibility for people takes away this human component, and I'm not sure that the emotions and social ties that bind us can be taught to a computer. Certainly not in the forseeable future. I think most visions of a dark computer-controlled future stem from this. We fear being objectified - by our job, by our loved ones, by a drunk driver, and a computer will feel no remorse, will not have any quams with making an inhuman decision if it's processers tell it that that's the best thing to do. The same way a computer can sacrafice pawns in a chess game, even a queen if it will lead to a checkmate, we fear being sacriced for the 'greater good.'

It's for this reason that a computer could, if given enough power, become the next Hitler - obsessed with effeciency, somehow skewed to react in a certain way to a certain situation (as in killing Jews). Seeing the world as a min/max puzzle to be solved rather than in terms of people. Computers are useful for making our lives easier, and more and more they're making decisions for us (we buy the books Amazon reccomends, get turned down for a loan because the computer gave us a rating, etc.). While I don't think humans will let this get out of hand, if we don't learn from the lessons of World War II, we'll be doomed to repeat them.(3)

Perhaps if Isaac Asimov were around, he could give us some laws of Robotics to guide us, but he's not. In his absence, we have to remember that what often starts out as an intellectual exercise can become reality, so we must be cautious even in our intellectual exercises.

Footnotes

  1. Yes I'm using a school yard as a tongue in cheek example.
  2. I guess it's not necessary that the computer have a physical extention into the outer world. After all, AI tend to do better when dealing with a symbolic, limited world like Chess rather than a real, unlimited world like the one we live in.
  3. Yeah, this too has little to do with the actual subject matter of the article, but I feel it's an important point.
  4. If you've read my tangential ending, you know I don't advocate kicking cats.

Related Books

Related Web Resources

page first created on Tuesday, April 09, 2002

© Mark Wieczorek