[agi] An idea for promoting AI development.

[Alan Grimes]

om

In this posting I will try to walk you through a fairly complex and, in some places tenuous, chain of reasoning related to my own personal attempt to develop an AI. My motovation in writing this post is to seek out people who might be interested in this venture. And, ofcourse, this will serve as a sanity check for myself.

I know that people skim longer postings however I do have some very important things to say, skim to the second 'om' in the posting for the skinny...

===

For the last month I have been trying to put togeather a box for research. The USPS lost the ram and its going to be a while before I will be able to afford it again =(. In the mean time I am trying to sketch out my research agenda.

To start with I need an OS and develment environment. It can't be linux because I am building the machine with my old 850MB HD (I spent all my money on the Mobo). That's not a real reason, ofcourse, because I could put any HD I want in it, if I had the money. The real issue is that Linux is not a real-time operating system. Its simply the wrong solution for this job. I'm not going to use windows, for obvious reasons.

My immediate plan is to put BeOS on the machine cuz I have the disk on hand. Unfortunately Microsoft killed Be Inc, so I will probably switch to QNX or some other real-time platform at some point. Other options include using DOS to boot the machine then run the AI as a self-supporting system providing its own OS functionality.


These platform issues seem to be minor but I will demonstrate their importance presently.

I have chosen a cybernetic approach to AI because it makes the most sense to me. I don't think I could work any other approach. A cybernetic approach is one that is based on a cybernetic loop between actor and object. Where the AI is designed in the context of a problem to be solved.

To develop an AI based on this approach one first sets about to construct a problem domain for the AI to work on. For general AI one requires an open problem space such as a box of Legos. Today's computers already come with a broad selection of software that would be suitable for this purpose. The problem then becomes how to make it so that the AI can see and use these applications just as the human user does? One would like to create a virtual desktop which is mirrored to the physical console and then have the AI take its input from what it displays. On this desktop the AI can run games and other instructional software as well as communicate with the human operator.

The development of such a system is still very chalenging as nobody writes software to run in the configuration I just described. This problem is not insurmountable, it will only require a great deal of money. Here is where I make my make my first logical step. I have concluded that to get the money and resources I need to develop the AI training software I will need to launch a commercial venture either under an existing company or organization or as a new business.

So I start to sketch out a business plan. My idea would be to produce AI development tools. These would include software adaptors to let the AI use Mozilla or sumpfin as well as two lines of robotic systems. The first line would feature laboratory rigs that include white-boxes for controlling stimulii. In that system, the AI would be given Fischer Price toys to play with. It would have a camera and a dexterous high-feedback robotic arm. Basically a pre-fab version of the type of laboratory that has been in use for years...

The higher line would be a mobile platform with a big honkin on-board computer and a dexterous manipulator.

So how do I market this? Could I match expenses? The market for the lab equipment would be fairly narrow... Maybe a few dozen universities would buy it. Probably not enough to achieve an economy of scale and make it affordable to a broader market of hobyists and entheuseasts. In other words the hardware angle of the venture would not do much to advance the cause of AI develment.

The software, however could be produced for practically nothing. So what kind of software would be able to make it big enough to justify a price-drop into hobyist teritory? The most advanced AI available to consumers today are in games such as Creatures and "Black and White".
A game engine architected to support general AI actors would be close to ideal.

The Sims Online is only the latest example. That kind of game would be great for teaching an AI to be good citizen. I, however, have a great bias against virtual communities beacuse it makes me nervous that it might one day gain a status above the real world, creating a situation inimical to someone, such as myself, who has an interest in real-world things. I used to have a "reality engine" project on my website but I withdrew it after reading Egan's Diaspora. So my first choice would be to go with a mobile platform or implement some in-house application for the job. No matter how attractive that approach is, it is still most problematic.

om

Software is too hard to develop. The curent state of software is that there are high barriers to entering the circle of developers. Furthermore once one has made a comittment to devel software you are faced with C++ which I have been told requires 7 years of daily experience to truly master or with the nightmare of getting a better language to work with an operating system which is literally built to support C and C++. Switching to a better OS makes things worse because once you get out of the mainstream you can't find support and you find yourself having to write/port more of your applications. While the OS you choose may be much much better than linux (which isn't hard at all)
such as BeOS your costs will go up because you have to do alot more from scratch.

[damn, I'm having trouble writing streight today; not enough sleap cuz I've been downloading 5.5GB of leenooks over 56k...]

YES, it is technically possible to nuckle under and work linux. When one thinks of an application such as the internet one looks at the work and what it takes to do the work. In all cases you want to maximize the ratio of work to overhead.

Lets think of software as a pyramid. The simplest softwarez are at the bottom, and AI is the little point at the very top...

^ < AI.
/_\ < applications.
/___\ < operating systems.
/_____\ < languages
/_______\ < raw computation.

The ease of developing AI is a function of how well the lower levels support the next level. This isn't strictly the case but given my state of tiredness, it makes perfect sense. ;)

To give you a better feel for what exactly is going on here let me try, in my over fatigued state, to relate the core of what is going on here through an anictdote that happened to me a week or so ago.

I have a long-time pen-pal, David G. Shreeves ( [EMAIL PROTECTED] ). He has been a Slackerware (linux) user for many years and reccomends it highly. On his advice, six months ago, I had a friend of mine (who has DSL) burn me the discs. He, most kindly, provided me with all three disks. My HD had reciently been slaughtered by an asshole after 6 years of uninterupted operation and I was in need of a new OS. When I tried to install the OS, it didn't work, it simply couldn't handle the optimal configuration of my HD. I asked around and they told me that it was a bug in the kernel and the work-around was to change a setting in the BIOS to a lower setting and later to recompile the kernel with the correct driver.

I use DOS as a benchmark for OS quality. Since this problem _NEVER_ happens in DOS, linux is clearly inferior.

So I was talking with Shreeves about it a few days ago and he knew about the bug too and suggested that its just a hacker's OS and that these inconveniences are minor.

He, and many like him overlook a critical fact. That being that the _USER_, and in this case AI researcher, was forced to take time and mental resources AWAY from his work on AI to handle these linux "quirks".

While it is theoreticaly possible for someone to be such a strong programmer that he can put up with linux and still have some time left over on the weekends to solve the AI problem, I hope the people receiving this message will see that to be a real problem. Humans are finite creatures. You can't keep piling stuff on them and expect them to be just as fast.

The burdeons of using linux have gotten to the point that its development has been thouroughly arrested by all the cruft that has built up around it.

Our basic goal here is to make computing much easier so that AI researchers can jump right in and start working the real problems of AI rather than spending all their time and mental resources getting their flaming hard drives working. By making AI research dramaticly easier to accomplish we can dramaticly improve the chances that someone out there will "get it right", and that's a Good Thing (tm).

The easier computing systems that I am talking about already exist. One of the most shining examples of these is the Squeak environment. ( www.squeak.org ).

The question now is, what is the killer app that can bring Squeak into the mainstream and encourage the open source community to give it the functionality that it needs to really succede without having to make a massive initial investment in that same functionality?

I think the answer is this:

http://www.business2.com/articles/web/0,1653,45665,FF.html

A game is not like traditional open source. Very few people will work on the real problems in the linux system, the core of GNU software that it is based on, because it is difficult and because, as humans, they don't have sufficient emotional motovation to work on such projects.

A MMORPG based on Squeak can solve that problem. While the details of the virtual world I propose are beyond the scope of what I want to say here. I feel that I have an excelant chance at being extremely sucessful in this venture. I will just have to set asside my fears and go whole hog on this because these on-line games have such an enormous potential for driving development of the cultural and technical foundations of AI development.

The economic value of virtual communities is only now beginning to be realized. The on-line RPGs we see today are merely experaments to find the right set of concepts that will make it big. I can project from the recient trends in the field and I am very confidient in my ability to design a virtual world with great potential.

I know how to do it.

I do face a significant challenge in getting it off the ground. While it has a clear mission, its qualities as a piece of software are not readily apparent. Investors have been mistaking complexity for sophistication for quite a while now and I don't see that changing any time soon. They will look at the Squeak package and won't see its true value because it looks too simple to them. ;)

The Virtual world I propose, the first generation at least, is quite obnoxious in that it will be a simple 2-D system. It is 2-D for very good reasons but investors will look at it and balk. =(

For an AI focused institution they will look at it and say "What value does this bring us? It has nothing to do with AI!" While, on the surface, they would be right, they will fail to see that promoting the virtual world I here propose, is what I call an "indispensible luxury". Its something that you don't really really need, but is a real pain in the butt to do without. Let me try to sumarize the payoff here. Squeak today is a wonderful system but it is immature and not yet sufficient for major projects. Today's operating systems have a massive ammount of inertia behind them and that won't change unless major effort is put into displacing them.

Through the virtual world squeak can become a major factor in the future of computing. In five to seven years, my virtual world system can be in a position to begin to displace these entrenched monopolies. The payoff comes on the day when system administrators are fired because the computers are simple enough for even the boss to configure. ;) The payoff comes when all the hours spent maintaining a windows or linux installation and doing day-to-day chores can be put into pushing the frontiers of the art. While all this liberated effort won't be going into AI directly, it will be put into software that AI researchers will use in their daily work and thereby form the foundation on which AI will be built.

It is apparent to me that AI is, and has been for some time, strictly a software problem. The time it will take to solve this software problem will depend on how much overhead and complexity the AI researcher has to deal wtith _BEFORE_ beginning his work.

WE CAN MOVE THE SINGULARITY FORWARD BY YEARS JUST BY FIXING THE PROBLEMS WITH TODAY'S SOFTWARE.

om

In conclusion, I am about to go against my reservations and fears and plunge head-long into a virtual world project because I beleive, by this convoluted logic, that it is vital to advancing the development of AI.

Bottom line: I need $5,000,000 to start, and another $10,000,000 to go on-line.

-------
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]