Shane Legg: Machine Super-Intelligence
September 10th, 2008 –
Dr. Shane Legg has published his dissertation, Machine Super-Intelligence. Continuing in the line of research of Professors Marcus Hutter and Jürgen Schmidhuber, Legg explores theoretical models for artificial general intelligence. He also discusses the practical possibilities for achieving AGI and its ethical implications, with a reference to SIAI’s work in the field.
Dr. Legg recently received a fellowship from SIAI-Canada.
I like the fact that this paper starts to look broadly at all of the definitions of intelligence.
It is important to note that there are scores of these and they often vary in scope or focus and they often conflict or overlap with each other.
Related to this, and regarding the quote from Johnson (1992) on page 167:
“…we need a defition of intelligence that is applicable to machines as well as humans or even dogs.”…
The only way researchers will arrive at this ’single definition’ is to leave the contraints of intelligence in the frame of the human mind/brain, animal mind/brain, machine mind/brain. global mind/brain—and to leave disciplinary constraints like mathematics, biology, psychology, computer science, etc.–and to look instead at the overall PROCESS of intelligence.
Generally speaking, this process is how intelligence is created in the individual intellect, flows into the social knowledge base, is accepted or rejected by society, and is extracted from the social knowledge base by individual learners.
That single process is common to all disciplines, humans, and machines and it is the key to everything AI and AGI researchers want to accomplish.
All of these other foci are like trying to visualize the forest by looking in detail at a tree. It is absolutely impossible to solve this problem of AI or AGI outside of the framework of this interdisciplinary process.
Why should societal acceptance have anything to do with intelligence? Better to think of intelligence as a characteristic rather than a process. Then we can look at the processes within the boundaries of an “intelligent” entity to measure the characteristic. The primary process would be Problem Solving. How well an entity performs this process could be measured to determine the degree of “intelligence” exhibited. Creation of knowledge useful to society would be only one outcome. Other outcomes would include providing information/instructions for other entities, performing machine operations (building things, putting out fires, surgery), identifying needs for information and predicting probabilities of events (although these could also be seen as useful to society in the largest sense).
Knowledge can be created in the individual mind and shared or not shared with society…it is the individual’s choice as to whether or not to do this.
But if the individual shares new knowledge with society and that society does not accept that knowledge, the social intelligence cannot grow or advance. Individual intelligence grows through knowledge creation and social intelligence grows through this process of social acceptance.
New knowledge itself is an outcome; the creation of new knowledge is a process. Problem solving is a form of knowledge creation that works by the same process, as does innovation and creativity. Many different terms are used to describe this one process.
The ability to create knowledge and solve problems is genius, not intelligence. These two are routinely confused in scholarship. If one measures problem solving capability, they are measuring genius, not intelligence. People tend to use genius to mean ultra-high intelligence, but this is an erroneous view of the term.
Intelligence is simply the volume of knowledge stored that can be recalled from the individual or social brain. Genius is our ability to create knowledge, solve problems, innovate, and create.
Doing physical tasks, e.g. robotics, is performance, not knowledge working. Once again, these two are also routinely confused, especially in business. Knowledge working is an enabler to performance.
Predicting probabilities of events is an example of decision-making which again is distinct from both knowledge working and performance.
In general, what we have are silos, self-defeating, or conflicting definitions of terms that don’t work together as an integral system. As such, modern scholarship has created extensive confusion in the mind and computer sciences.
Bruce, Shane Legg is lucky you were not in the approval cycle for his dissertation on machine super intelligence. He gave considerable effort to defining intelligence. He posited a simple agent-environment interaction model to support his working definition: “Intelligence measures an agent’s ability to achieve goals in a wide range of environments.” A performance based understanding of intelligence is more useful than a knowledge model for comprehending the road to the Singularity, and its effects.
Bradley,
I just enjoy all this and don’t expect anyone will be jumping to have me disagree with them….but that’s ok.
But anyway, I do respectfully disagree with Shane’s definition and the usefulness of this to comprehending the road to Singularity. The way I see this, it’s impossible to measure ‘the ability to perform,’…you’re really measuring performance itself.
Bruce,
I just now saw these comments.
Your definitions of intelligence and genius, while I have seen some related ideas in the literature, are quite non-standard in their emphasis. Also, if you look at my mathematical model, you will see that the creation of knowledge, prediction etc. must all be working well together in order for the system to have much intelligence.
As for the usefulness of my mathematical definition? For that, you’ll have to wait.