Why AGI May Be Near
From The Transhumanist Wiki
The following are some of the interrelated reasons why general AI may be accomplished, with effective, diligent effort, within 10 to 30 years. (This page is still rough.) Please consider adding your own, critiquing these, etc. Best, -- Anand
- Both high end computersTop500 List, and commodity components are improving at about 75% per year. At that rate sufficiently powerful hardware (estimated at 100 to 100,000 TFlops) will be available at reasonable cost (estimated at $100K to $3M for hardware) in the 2010-2025 time frame. Assuming sofware development/AI training will take an additional 5-10 years after adequate hardware is affordable, then general AI may be developed in the 2015-2035 time frame.
- Research and development on general AI is presently being pioneered by four organizations: Adaptive AI, the Artificial General Intelligence Research Institute, Cycorp, and the Singularity Institute. (These projects, and others, will be featured in Advances in Artificial General Intelligence (forthcoming in ?03).) An advanced conceptual understanding of intelligence has recently been developed by the Singularity Institute?s principal researcher, Eliezer Yudkowsky, in ?Levels of Organization in General Intelligence,? which details the theory of deliberative General Intelligence that will be the basis of the Institute?s general AI design. Ben Goertzel and Cassio Pennachin, lead researchers at the AGI Research Institute, are presently at work on a ~1200 page manuscript that will provide a detailed mathematical and conceptual overview of their Institute?s general AI design, the NovamenteAIEngine. And Peter Voss, in his ?Essentials of General Intelligence: The Direct Path to AGI,? has provided an overview of Adaptive AI?s work-in-progress general AI design. The diligent effort of these three organizations exemplifies the kind of work that is continually needed in order to achieve general AI within 10 to 30 years. For more, see the Artificial General Intelligence Resources.
- During the past three decades we have made exponential progress in reverse engineering the human brain. Presently, the speed of brain scanning devices is doubling every 26 months; the resolution of noninvasive brain scanning devices, e.g., positron emission tomography (PET) and functional magnetic resonance imaging (fMRI), is doubling in improvement every 36 months; and the image reconstruction time of brain scanning devices is reducing in half every 18 months. (Those three figures are based on chart data in Ray Kurzweil's "The Law of Accelerating Returns.") An important individual example of brain reverse-engineering work is Lloyd Watts?s, who has successfully reproduced, within a computing system, many of the cochlea?s auditory processes. Another important development is the Brain Tissue Scanner, created by Texas A&M University?s Brain Networks Laboratory. The scanner is a microscopic digital camera that uses a diamond knife to precisely slice a mouse?s brain tissue. Once sliced, the tissue is illumniated by a laser light, recorded as an image by a camera, stored on a computer as data, and finally reconstructed as a detailed three-dimensional model. (Presently, one brain can be scanned within a month, the camera can image details smaller than one-hundred-thousandth of an inch, and the team intends to begin a human brain scan in 2003.) These examples of brain reverse-engineering are evidence for the near term feasibility of a Human Cognome Project: a nationwide, government-funded effort to map the areas, regions, and columns of the brain's two hemispheres; the brain's neurons, dendrites, synapses, and neurotransmitters; and the functional aspects of the mind. This project will have profound pragmatic value, as the better we understand ourselves the better we will affect the world according to what we value, and will reduce the difficulty of general AI development, as the brain and mind will be a guide for said development.
- Technical progress in one area often interacts synergenetically with progress in other areas. These interactions create positive feedback effects in technical development, and occur more frequently as time passes. The following technologies interact (in many of these, exponential progress occurs): biotechnology; computer software; computer power, data storage, and memory; Internet bandwidth, connectivity, and transmission speed; microelectrical mechanical systems (MEMS); nanotechnology; neurotechnology; wireless data services; and the WWW. Synergetic interactions in technical progress will continually reduce general AI?s development time.
- A fundamental imperative of an economic environment is to design, develop, commercialize, and distribute technology that displays better ?intelligence? than competing technology. A humanitarian imperative also exists to develop technology that positively effects how well we model, predict, and modify environments so as to alleviate nonvoluntary suffering. These two imperatives are behind the use of specialized AI in areas where only human endeavor could previously accomplish intelligent results. They will also be behind the development of general AI because of its clear economic, humanitarian, medical, and scientific value.
- Problems of software complexity (brittleness, legacy, sedimentation, unwieldiness) may be overcome by innovative improvements to software development and programming languages. The ongoing work on supercompilers is one example of innovative research and development that may substantially improve how we develop software. When compiling into executable code, regular compilers will make a few improvements to the source code. A supercompiler, however, will be able to mathematically model preexisting code, and to create new code that implements, with greater efficiency, the original?s function. By allowing programmers to solely focus on code elegance and maintenance, rather than on efficiency, supercompilers may achieve a breakthrough in our present methodology of software optimization. And one example of an innovative programming language is the Flare Programming Language, which is presently in early development. The language has been explicitly designed to handle issues of software complexity; to take the next forward step beyond the present method of object-oriented programming by becoming the first annotative programming language.
- We have made slow though still exponential progress in developing better software, e.g., better algorithms, speech recognition, algorithmic techniques (genetic algorithms), chat bots (ALICE), chess programs (Deep Fritz). For more, see the Artificial Specialized Intelligence Resources.
- Research and development on general AI will acquire greater support as its R&D becomes more fashionable-as its benefits, commercial viability, and near term feasibility are better realized by academia, businesses, and the government.
- As the benefits of specialized AI become clearer, so will the benefits of general AI. For example, in March 2002, researchers at the University of Maryland Greenebaum Cancer Center in Baltimore created specialized AI software to distinguish, when diagnosing patients, between two types of cancer (common colon and inflammatory bowel disease-related) that individuals with Crohn?s disease or ulcerative colitis have an increased risk of developing. The software can be used to assist doctors in identifying the two forms of cancer, which, if identified early, will not require debilating surgery to remove. In a related, and even more profound development, an extremely accurate blood test has been created for early detection of ovarian cancer, a disease that 23,000 women acquire each year, and that 14,000 die from. The average five-year survival rate of stage I ovarian cancer is 90 perecent, with stage III being only 20 percent. Since early symptoms are not usually present, nine of 10 women don?t receive their diagnosis until after entering the cancer?s late stages. This is why a recently developed test, assisted by specialized AI software, is being considered a medical breakthrough. So far, 100 blood samples have been examined in a blinded study, that is, a study where the researchers didn?t know whether a sample was from a woman with or without cancer. The results showed a 100% accuracy detection rate in the 50 cases of ovarian cancer. A larger study is now being planned to verify the results. Dr. David Fishman of Northwestern University, commenting on this development, said, ?I am completely overjoyed at the potential this technology represents in being able to save lives.? I believe this is a sentiment often to be repeated in future applications of specialized AI. For more, see Neural net programs diagnose colon tumors (source: Kurzweil AI.Net), and New blood test detects early ovarian cancer (source: MSNBC Online).
- Computing systems have many design and functional advantages over our brain. See Computing System Advantages.
- General AI is likely to have many cognitive advantages over us, which will reduce the difficulty of its development. See AGI Cognitive Advantages.
To whoever keeps changing the lines that say "within 10 to 30 years":
Please stop editing other people's text. If you want to add a comment to this page, mentioning that you disagree with the time frame, then please do so using your own name, rather than editing text to which someone else has already signed their own name.
If you disagree with this policy, then you can discuss it on the Wiki Etiquette page.
If you actually are Anand, finally returned to this wiki after over 3 years of absence, then I apologize for reverting your edits.
--observer
