Re: Intelligence, IE, and SI - Part 1

hal@rain.org
Wed, 27 Jan 1999 14:37:32 -0800

Billy Brown, <bbrown@conemsco.com>, writes:
> Since there seems to be a fair amount of interest in the topic, I've decided
> to go ahead and post what I've got so far (in sections, of course). The
> final version of all of this will (hopefully) end up as a paper-length
> treatment of the topic, but it has a ways to go before its ready for that.

This looks like a great start.

> Now, intelligence in this sense is not a unitary entity. It is perfectly
> possible to be good at solving one class of problems, and do a poor job of
> solving others. There is a very large (possibly infinite) number of
> different problem domains in which an intelligent entity
might have some
> ability. Some of these domains are related to each other, such that a high
> level of ability in one domain can be used to solve problems in the related
> domains. Other problem domains are independent of each other, and require
> completely different problem-solving approaches.

The nature of intelligence is somewhat controversial, tied in to the dispute over whether intelligence is heritable, and whether different races have differences in intelligence. As you say, there are a number of different skills which can be identified as part of overall "intelligence". One disputed point is whether there is correlation among these abilities. Do people who are better at memorizing lists, say, also tend to have better logical deduction skills? Are better readers also better at math?

Apparently the statistical data is ambiguous enough that controversy remains, although it is possible that the high political stakes are enough to explain it. The hypothetical correlation factor is called "g", the general intelligence. My impression is that most studies show that there is such a factor, that broadly speaking people who are better at some mental skills tend to be better at others, but there is definitely considerable variation in abilities.

> The answer to this apparent contradiction is also the reason why so many
> people think of intelligence as a single ability. Picture a graph with
> different cognitive abilities along the X-axis, and increasing ability along
> the Y-axis, so that the graph depicts varying ability levels in different
> problem domains. What we commonly call 'intelligence' is the area under
> this curve, with various distortions based on our own ideas about which
> abilities are important. Different humans may have different levels of
> ability in each problem domain, but we expect the total area under the curve
> to fall within a given range.

This should be a bar (column) graph; I think the various cognitive skills would be discrete rather than continuous. (Although this may be in part an artifact of the factor analysis technique, which inherently identifies discrete basis vectors.)

The difficult issue as you try to generalize this to super-intelligence is how to scale the Y axis. With humans, we can test people and come up with a range of scores which we can normalize with a desired mean and standard deviation. However the nature of tests is such that there may be no unambiguous way to extrapolate them beyond the range of abilities seen. As a simple example, you can't do better than 100% on a test. A test may not be able to produce meaningful data outside of a given range.

Even with open-ended tests, like how quickly something can be done, or how many numbers you can remember, it is not clear how to measure an ability which is beyond the human level, in quantitative terms. How much better do you have to do on a test to have an IQ of 400 rather than 350? What would such values mean?

An interesting point this suggests is that a person, today, who is allowed to "cheat" on an IQ test by having access to computers and other helpful devices, could probably score very highly on some of the sub-tests. He could "remember" sequences of virtually any length, for example. (All he really needs is pencil and paper for that.) He could come up with synonyms, and might be able to handle geometric problems more easily as well. The standard formulas assemble the sub-tests into an overall score, and his massively increased ability in some areas would increase his apparent IQ, possibly significantly.

This leads to a couple of conclusions. First, such an "augmented" person is effectivelly more intelligent than an ordinary human. This is unsurprising; it is one reason our culture has advanced from caveman days.

But second, it shows that the tests don't always produce meaningful results when taken outside their intended range. A test of memory is more relevant when measuring brain capacity than paper capacity. Someone who can memorize hundreds of digits on hearing them has an amazing mental ability, and this would show up in his IQ score; but someone who can write down hundreds of digits and read them back is just an ordinary person, and he doesn't deserve to have his IQ score boosted by a formula created with the intention of testing a purely mental ability.

> We can therefore say that an entity has human-equivalent intelligence if it
> meets the same criterion: its total ability in the relevant problem domains
> must fall within the same range as that of humans. A transhuman entity
> would be one whose total ability falls well beyond the human range, and an
> SI would be an entity with an astronomically large total ability.

What is an SI? A Super-Intelligence?

> <to be continued>

Looking forward to seeing more...

Hal