Ben Goertzel, author of The Hidden Pattern and Artificial General Intelligence (both books highly recommended), talks about the Singularity, artificial intelligence, how to be human after transcending and a lot of other wild topics. [via boingboing]

Links:

• Interview at Neofiles
• Ben Goertzel
• Novamente

Given is an arbitrary computational problem whose solution may require interaction with a possibly reactive environment. For example, the goal may be to maximize the future expected reward of a robot. While executing its initial (possibly sub-optimal) problem solving strategy, the Gödel machine simultaneously runs a proof searcher which systematically and repeatedly tests proof techniques. Proof techniques are programs that may read any part of the Gödel machine's storage, and write on a reserved part which may be reset for each new proof technique test. In our example Gödel machine this writable storage includes the variables proof and switchprog, where switchprog holds a potentially unrestricted program whose execution could completely rewrite any part of the Gödel machine's current software. Normally the current switchprog is not executed. However, proof techniques may invoke a special subroutine check() which tests whether proof currently holds a proof showing that the utility of stopping the systematic proof searcher and transferring control to the current switchprog at a precisely defined point in the near future exceeds the utility of continuing the search until some alternative switchprog is found. Such proofs are derivable from the proof searcher's axiom scheme which formally describes the utility function to be maximized (typically the expected future reward in the expected remaining lifetime of the Gödel machine), the computational costs of hardware instructions (from which all programs are composed), and the effects of hardware instructions on the Gödel machine's state. The axiom scheme also formalizes known probabilistic properties of the environment, and also the initial Gödel machine state and software, which includes the axiom scheme itself (no circular argument here). Thus proof techniques can reason about expected costs and results of all programs including the proof searcher.

(thanks Calypso!)

ITConversations has a recording of an interesting panel discussion at Accelerating Change 2005 on The Prospects for AI with Neil Jacobstein, Patrick Lincoln, Peter Norvig and Bruno Olshausen.

Stefan Lehmke, author and maintainer of the supercool TeXPower presentation system, recommended the Mozart Programming System and the programming language Oz as an ideal platform for building agent based systems (like my conceptual workbench, which is of course pure Common Lisp).

The Mozart Programming System is an advanced development platform for intelligent, distributed applications. The system is the result of a decade of research in programming language design and implementation, constraint-based inference, distributed computing, and human-computer interfaces. As a result, Mozart is unequaled in expressive power and functionality. Mozart has an interactive incremental development environment and a production-quality implementation for Unix and Windows platforms.

Mozart is based on the Oz language, which supports declarative programming, object-oriented programming, constraint programming, and concurrency as part of a coherent whole. For distribution, Mozart provides a true network transparent implementation with support for network awareness, openness, and fault tolerance. Security is upcoming.

Back to the reading list:

• The Truth Machine by James L. Halperin
• The Paleo Diet by Loren Cordain
• Richard Feynman by John and Mary Gribbin
• Bayesian Artificial Intelligence by Korb and Nicholson
• The Confusion by Neal Stephenson (again)
• On Lisp - Advanced Techniques for Common Lisp by Paul Graham (again)

Over the last couple of weeks I collected numerous papers in a Read This! folder. This morning I read a paper written by Hans Moravec in 1975 titled: The Role of Raw Power in Intelligence. Thirty years after its publication it's still quite to the point:

The enormous shortage of ability to compute is distorting our work, creating problems where there are none, making others impossibly difficult, and generally causing effort to be misdirected. Shouldn't this view be more widespread, if it is as obvious as I claim?

In the early days of AI the thought that existing machines might be much too small was widespread, but there was hope that clever mathematics and advancing computer technology could soon make up the difference. Since then computers have improved by a factor of ten every five years, but, in spite of reasonably diligent work by a reasonable number of people, the results have been embarrassingly sparse. The realization that available compute power might still be vastly inadequate has since been swept under the rug, due to wishful thinking and a feeling that there was nothing to be done about it anyway and that voicing such an opinion could cause AI to be considered impractical, resulting in reduced funding.

There is also an element of scientific snobbery. Many of the most influential names in the field seem to feel that AI should be like the theoretical side of physics, the essential problem being to find the laws of universe relating to intelligence. Once these are known, the thinking goes, construction of efficient intelligent machines will be trivial. Suggestions that the problems are essentially engineering ones of scale and complexity, and can be solved by incremental improvements and occasional insights into sub-problems, are treated with disdain.

This attitude is a variant of the philosophical notion that all truth can be arrived at by pure thought, and is unfounded and harmful. One wonders what state space travel would be in if the Goddards and von Brauns had spent their time trying to find the universal laws of rocket construction before trying to build space ships. AI needs a stronger experimental base. Like other branches of endeavor (notably physics, aeronautics and meteorology), we should realize our desperate need for more computing, and do things about it.

In hindsight, we had made several major mistakes - mistakes that seem to be repeated again and again throughout the software industry.

Part of the problem was how arrogant we were. We believed that we could spend a couple of days watching trained lawyers perform a highly-skilled job, talk briefly to them, and then make their jobs completely obsolete.

Worse, we made the job completely non-fudgable. In any human process there's always a degree to which the outcome can be fudged by the person performing the task. Even when the rules are simple or well-understood, there are always cases when someone will have a compelling reason to do things differently. In this case we didn't even know all the rules, and discovered to our horror that there were many more edge-cases than we'd imagined.

Via Emery Berger in gmane.comp.lang.lightweight comes a recommendation for this research page by Umut A. Acar. Some of tasty papers there:

• Selective Memoization
• Adaptive Memoization
• Stability, Adaptivity, and Dynamic Algorithms

And here's a photo of the OSPNL development team. And here's Intels introduction to the PNL.