Ruminations on matters of some importance. Also programming and AI.

Wednesday, June 16, 2010

Digital immortality, true AI and the destruction of mankind

Around the world a few people and companies are working towards the goal of strong AI or artificial general intelligence, which is more or less the same thing. Too few, if you ask me.

Today I was reminded of a strategy towards AGI that I have sometimes dreamed of. It is not the strategy I believe most in, but I think it is interesting nonetheless.

What if you (assuming you are a competent programmer), from now on, tried to automate as many of your computer tasks as possible. Instead of doing something, try making the computer do it. Even if it takes you ten or a hundred times as long, your time investment will hopefully pay dividends in the future. Starting out, many things will be out of reach, but you will slowly build a knowledge base and an algorithm base that can mimic your preferences and skills. this will enable you to take on harder tasks and so on. You are building a digital assistant from the ground up.

Maybe this undertaking is too ambitious for one person, especially if they actually wanted to get something done besides building a digital assistant. In that case, my proposal stands, but instead use a small team (Google and Microsoft, I know you have a few talented guys to spare for a grand project) that tries to automate the computer tasks of one guy.

Take email, for example. Propose automatic actions on incoming mail, including replies, forwards and adding stuff to the calendar. Initially, very few emails will be understandable, but gradually I expect the algorithm to get better at parsing language and to get a better model of the user and the world. Perhaps one part of the knowledge base is building a Bayesian network that models the user's preferences. The important thing is: solve the emails one at a time, using as general rules as you can get away with, but as specialized rules as you practically have to.

Want to look something up on Wikipedia? Make travel plans online? Make a purchasing decision? Solve it in code, as general as you can. When the AI is further advanced, you start to write documents and code collaboratively, and so forth. One way of developing the AI is to let it observe your digital life and all your actions. Ultimately, what you end up with is a digital model of yourself, that gets more and more like the original. It answers mail, reads news and maybe comments on it in tweets and blogs. In effect, you achieve digital immortality.

Obviously, the weakness here is that I have not proposed what algorithms should be used for this digital alter ego. I do, however, feel that the task of general AI will benefit from both clever general algorithms and clever specialized algorithms and specialized knowledge. An organic hodge-podge of hacks and patches, very much like the brain itself.

When a mathematician approaches the problem of AGI, they want a clean solution. One algorithm to bind them all, like the Gödel Machine of Jürgen Schmidhuber and Marcus Hutter. When engineers approach the same problem, they tend to engineer grand designs, like Ben Goertzel's Novamente and OpenCog. I can certainly see the charm of both approaches and I hope that they succeed, but maybe the most practical way forward is just to tackle one small real-world task at a time - the "guided hodge-podge" approach.

About the destruction of mankind? No, I don't think we will have any of that, but some smart people do. Like Michael Anissimov: "Why is AI dangerous?". Still, a title is better when it involves destruction, don't you think?

4 comments:

  1. I'm a part of one of those companies that's working on an AGI - the project is RIA (http://www.qtara.com/) and the beta looks just like the website. It can help you write email, use Skype, research, read the news - all via a natural language spoken interface.

    An open source equivalent by the very smart John is: http://code.google.com/p/open-allure-ds/

    I'm building up a set of like-minded folk for the A.I.Cookbook, mostly using Python to solve useful problems. Some of the active projects are documented here: http://blog.aicookbook.com/ with a budding discussion group: http://groups.google.com/group/aicookbook

    Cheers,
    Ian.

    ReplyDelete
  2. "solve the emails one at a time, using as general rules as you can get away with, but as specialized rules as you practically have to".
    From what I learned doing my PhD in NLP, the more emails you "solve" (i.e. parse the syntax, disambiguate the 67 possible solutions, analyze the sense of each word in its context, combine these senses, disambiguate the 670 possible interpretations), the more your general rules/ specific rules ratio will decrease.
    I can't prove this, but it's based on my observation of several rule-based systems in NLP.

    ReplyDelete
  3. The problem with most strong AI advocates like yourself is that you think that all it takes to be intelligent is to behave intelligently. This is false. In psychology, can you truly determine someones intent from their behaviour alone? Take Steven Hawking for example, he can communicate only through technology. If he didn't have that technology, he would still be highly intelligent, but not according to you since he can't behave intelligently. There is far more to intelligence than behaviour: understanding is what needs to be modelled.

    I suggest you read the book On Intelligence by Jeff Hawkins.

    ReplyDelete
  4. Why don't you think there's a risk from AI destroying us? Just context-insensitive optimism, or what? I suggest you check out Stephen Omohundro's paper, "Basic AI Drives", and also Eliezer Yudkowsky's "AI as a Positive and Negative Factor in Global Risk.

    ReplyDelete