Future Current

Peter Voss started his career as an entrepreneur, inventor, engineer and scientist at age 16. After a few years of experience in electronics engineering, at age 25 he started a company to provide advanced custom software development and information-technology services. Seven years later the company employed several hundred people and was listed on the Johannesburg Stock Exchange. Having recently taken his artificial intelligence company Adaptive A.I. Inc. (a2i2) out of stealth mode, he presented at the BIL unconference in Long Beach, California in February on the prospects of creating artificial general intelligence, or “Real AI,” in less than a decade.

The following transcript of Peter Voss’s BIL unconference presentation “Toward Real AI” has not been approved by the speaker. Video and audio are also available.

Toward Real AI

Thanks everyone for coming to this talk.  After seven years I am very happy to be able to talk about some of my work.  I am not going to be able to talk in much detail, but we have been in stealth mode for seven years and we have just come out of it.

I have been instructed to make this a firehose presentation, that you guys are all very smart, so I am going to give you a lot of bullet points.  I am going to be covering a lot of ground on my personal experience working on AI for so many years now.  As some of you might know, we launched a commercial venture recently to automate call centers.  These things are also called IVRs: interactive voice response systems.    We have basically been using our AI brain to do that.

What is artificial general intelligence, the field that we are working in?  It can also be called “Real AI.”  It is basically getting back to the roots of AI.  What AI used to be fifty years ago when people started it was to build computers that could think and learn like humans.  That dream had been abandoned with the AI winter, but artificial general intelligence is getting back to that vision to create machines that can think and learn.

If we look at intelligence as knowledge and skills, then AGI is about building machines where acquiring the knowledge is more important than having it.  The system’s ability to learn is the important part, not some programmer sitting down and coding information.  General knowledge relates to general AI, as opposed to domain-specific knowledge in narrow AI.  Things are learned in the abstract, where possible, and the specific context will guide the application of those skills. That is the idea.

Another detail is that it is self-directed, ongoing learning as opposed to a programmer or user deciding, “Hey, we need some new feature,” and then externally deciding to build that feature in.  AGI is about this ongoing, interactive learning process.  It is a system that has to interact with the world.  It has to have input and output, dynamically learning from interacting with the world.

Talking about the longer-term, more ambitious goals of AGI, people often think in terms of human-level intelligence and beyond.  What in my view would qualify for AGI?  How would I recognize it if I saw it?  We might build something only for people to say it is just technology, not really AGI.

My view on that is AGI will not solve all possible problems.  This is a theoretical view of AGI that some people have—something with infinite computing power that can solve any kind of problem you can throw at it…  I think that is nonsense.  To me, if AGI can solve problems that humans can solve, that is perfectly adequate.  To make it simpler, let’s exclude the problems in arts and music and let’s just talk about technical problems.  That to me would be the domain of AGI.

If you have a machine where you can tell it what the problem is, or it can read up about the problem, think about it, or do research, that to me would be AGI.  It does not need to have all human senses.  It does not need to have the sense acuity and the dexterity of a human.  Those are actually very difficult to engineer and computationally expensive.  I think they are actually not necessary.  You guys are all smart, so just think “Helen Hawking.”

Switching over to intelligence augmentation, there is a lot of confusion as to what intelligence augmentation entails.  I put it into three different categories.  One is the use of technology.  Wearing glasses is intelligence augmentation by that definition.  Using a calculator or the internet, to me that is mundane and does not qualify for intelligence augmentation in the realm of AGI and future enhancement.

The second category is the real thing, upgrading our wetware: whether through drugs, implants or some kind of engineering, really upgrading our brain capacity.  That to me is real intelligence augmentation, but it is really, really difficult.  This is why I believe AGI is going to happen way sooner than this kind of intelligence augmentation.  The third category then is uploading and upgrading—placing our brains into another substrate.  That is an even more advanced step and would give us more cognitive power, but I see that as being way, way more difficult.  I believe in fact it is going to require advanced AGI to achieve that.  There are people who disagree with the view, but this is to give you my view on that.

When do I think full-blown AGI will happen?  Some of you may have heard my talk some 18 months ago.  At the time, I was pretty sure the timeframe was less than ten years and as little as five.  Today, I think it is going to be around eight years.  I know that is way optimistic and a shorter period than pretty much anyone else gives. Perhaps other people are just pessimistic on this.

I will give you three different angles.  The one is, practically no one is working on AGI.  Is that cause or is that effect?  Do we think it is not going to happen in the next fifty years because nobody is working on it, or do people really have a good reason not to be working on it?  I will argue that people are actually blindsided in many ways.  The second thing that I like to remind people is that things really have changed since the last attempt at Real AI, which was probably thirty years ago.  Lastly, I would like to talk about the personal experience from the work that we have been doing over the last twelve years and give you some perspective on why I am more optimistic about the time table.

Getting back to the first point, 80% of the people working in AI do not believe in general intelligence.  They do not believe there is such a thing as “g,” in psychology… I think they’re wrong. Out of those remaining, many do not believe that human-level AI is going to be possible in the next hundred years, if ever.  Therefore, they do not attempt it.

Out of the remaining people, 80% are actually working in narrow AI.  Maybe they would like to work in general AI, but they actually work in narrow AI. This is because commercially it is much more viable, you get short-term results, while AGI is very long-term. There is a very strong bias to be working on narrow AI. In academia there is very little support, little funding and it gets no respect. We are only now having the second AGI conference, and it is a tiny subfield. It is really hard to get any support to work on AGI, so people work on narrow AI.

Out of the remaining people, 80% of them do not have a good theory.  They might want to work on AGI, but they really do not have an idea of how to solve the problem.  That does not leave a lot of people working on AGI. That is both a cause and an effect of estimating how soon Real AI can come about.  It was something that really troubled me when I entered the field some twelve years ago.  I was puzzled: “Why is nobody working on this?” Then I actually did spend a significant amount of time analyzing that.

As regards things being better, I have been around long enough in the tech industry to give a personal perspective on this.  On the hardware side, 25 years ago when there was the last real attempt at AGI, machines were a million times slower than they are now.  A one-second response time to test my computer today took twelve days to get that same answer.  If you can imagine experiments where you are testing thousands of hypotheses, which is what we do now routinely, it makes a real difference.

We also spent an awful amount of time writing programs in 64k—that is “k” as in kilobytes.  In fact, we were writing in 48k often.  It wasted an enormous amount of time.  The software side is also very significant in providing advances that shrink down time.  Over the last seven years the chances of being able to find a program on the internet, some tool that previously would have taken months or years to develop that moves us on to the next step, are really substantially better.

This is a quick timeline before switching over to my personal view of working on this.  I spent about five years learning, educating myself and experimenting.  We then started the company a2i2, and for about three years we did pretty much pure research.   We tried to put a lot of these things into practice to find out what worked and what did not work. We then got to a point where we believed we understood it well enough to focus on commercial development, and spent almost a year putting together a business plan, making presentations, raising money and all of that.

It took a lot of time putting together demos and presentations over the last seven years. It has definitely bitten into the AGI work significantly.  We then spent about two years with about fifteen people doing some solid AGI development with some good funding.  Over the last year we have really focused on our commercial track.  We are still doing about a third of our work in pure AGI, but a lot of our work is now getting our company off the ground.  You know, we have to pay the bills.  Fortunately, our AGI system is now getting smart enough to earn a living.  That’s going to help.

What are the other lessons that I have learned?  One needs a clear theory.  To try and do what our brains can do is very intimidating.  All of these issues of what is consciousness, free will and the complexity of vision, learning and epigenisis, these are really complex.  You really need to simplify the problem as much as you can so that you can focus on it.  Having that clear goal and vision is absolutely key.

My view is that all the pieces of the puzzle are out there.  No fundamental technology needs to be developed to achieve full blown AGI.  Over the years that we have improved the intelligence of our system and developed it, we have hit no brick walls that we could not break through in a period of six months or a year.  I do not anticipate that changing.  The problems are hard, but I believe they can be solved by putting effort, thought and good engineering into it and focusing on the issue.

Another thing I have found is that it is very hard to find AI psychologists.  This is a profession we invented.  It is basically the skill of having a good software designer and cognitive psychologist in one person.  Seeing both perspectives is really hard.  Most psychologists hate computers and most computer geeks really do not care for psychology.  To find people who have the skills to think of building a system that thinks, learns and interacts with the environment and has layered learning is really difficult.  It has been hard but we have put together a good team of people to do that.

This is why I also believe that a large company throwing hundreds of millions of dollars at this problem may not make any progress at all.  That is because their mindset will be a software engineering mindset.  Unless you can get the right leadership and get into the right groove, I think you are actually not going to make much progress.

The next three problems that I want to talk about deal with an open-ended general AI.  As the system improves in capability it is difficult to find the right level of problems that you pose.   If you make things too hard, you cannot succeed.  If you make things too simple, it becomes trivial and you do not make progress.  Getting the data into the system, kids and animals can run around in the wild and get all of this feedback, but with a simple AI it is really hard to get it the right training and testing data.

Funding is obviously very difficult.  AGI is a very long-term thing and gets no respect.  What helps us is that we have found a way of having a parallel path of AGI development and a highly viable business model that we could actually marry synergetically.  There are hazards and benefits to marrying commercial goals with AGI.

What will it take for ongoing progress in AGI?  It will take a workable design theory to cut through all the clutter and make this huge problem more manageable.  It takes a workable AGI platform, which is also not easy.  There are not tools out there that allow you to experiment and develop AGI stuff.  There are very few people working on that.  When I started on the project, we spent several years developing this platform.  This is important to have for ongoing development.   Money, focus and real-world feedback to serve as a reality check to your AGI design are very important.  It is very easy to subconsciously dodge some of the problems that real-world constraints impose on you.

We have an IVR system with an AGI brain.  The brain improves the capability of the IVR and the IVR gives us both practical feedback and hopefully lots of money.  This allows for improving the AGI.  There are human operators with different levels of skills in call centers.  The very lowest level, touchtone IVR—press two for one option, press seven for another—is very simple technology.  Voice activated IVR is slightly higher up in the food chain, and we are aiming at the upper end of that.  As the brain gets better, we move up the food chain with our hopefully virtuous cycle of AGI and business model interaction.

Speaking to the risks of commercializing, if we take narrow AI such as a chess-playing program, over time it becomes slightly more capable, but all it will do is play chess.  It will not direct traffic, do medical diagnosis or anything else.  Up until now, IVR has been a narrow AI application.  AGI, as it gets smarter, encompasses a wider and wider range of applicatons.  That is basically the huge difference.  It is much harder to get to some useful application, but once you get that intelligence, it solves a whole bunch of problems.  It’s general.

If you focus on one particular area, you run a real risk of putting all of your resources into solving one problem.  The risk is that in order to get some sales, you put all your efforts into making this IVR the best possible.  That is a risk we run as well.  We have got ourselves against it by having lots of applications that we are pursuing in parallel.  Our business model and ongoing AGI development aims at lots of applications.  We are just choosing one of them at the moment to commercialize because it is an existing market that is lucrative, but one does run the risk.  One needs to mitigate that to really have the effect of being able to continue to fulfill the AGI’s potential.

What is the inflection point?  How do we know when AGI will take off?  It is really hard to say.  With hindsight analysis has been done to see what was the inflection point for making companies like Amazon successful.  In AGI, what will it be?  I believe it is very likely to be the first real successful commercialization of an AGI engine that will have that virtuous cycle.  It may be ours or someone else’s, but I believe that will quite possibly be the inflection point for AGI.

Switching gears back to AI versus intelligence augmentation, since there is a question of which one is likely to happen sooner.  It is really much harder to upgrade wetware than it is to deal with hardware.  The FDA approval alone is probably worth thirty years of development.  Another thing is that with AGI the design information is available to the designer.  With wetware, we don’t have any blueprints.  AGI development, even though it is complex and emergent in many ways, we still at least have some design information.  Once AGI becomes smart enough, it can use that design information itself.  That is Seed AI potential.

The singularity is another controversial topic.  That is when AIs become smarter than humans.  I am quite happy to work on making the dumb AI that we have right now less dumb, for them to earn a living and for us to move the technology forward so that ultimately these AGIs can help us solve the problems that I think most of us want to see solved, whether they are disease, aging, over-population, resources or pollution.  I believe we need better intelligence than we have in humans to be able to solve these problems, and I think AGI will help to give us that.

Will it be a hard take-off?  I do not believe so—not in the sense that it will wake up in some garage or some basement somewhere and overnight take over the world.  There are many physical limits that have to be dealt with, physical processes that work in real time.  Even though your computer could be solving problems in minutes or seconds, there are things that have to be tested in the real world, and then you are hitting the time limitations and dynamics of real-world problems.  For example, if we wanted a computer to upgrade our wetware, it takes a long time for anything in our brains that we implant to heal or to grow. That timeframe I believe will make it a firm take-off rather than a hard take-off.

This entry was posted on Tuesday, March 3rd, 2009 at 2:34 pm and is filed under BIL, Peter Voss, Artificial Intelligence. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.