Quus, Redux
[updated 3/20/2022 to add footnote 2]
Philip Goff explores Kripke's quss function, defined as:
Goff then claims:
This statement rests on some unstated assumptions. The calculator is a finite state machine. For simplicity, suppose the calculator has 10 digits, a function key (labelled "?" for mystery function) and "=" for result. There is a three character screen, so that any three digit numbers can be "quadded". The calculator can then produce 1000x1000 different results. A larger finite state machine can query the calculator for all one million possible inputs then collect and analyze the results. Given the definition of quss, the analyzer can then feed all one million possible inputs to quss, and show that the output of quss matches the output of the calculator.
Goff then tries to extend this result by making N larger than what the calculator can "handle". But this attempt fails, because if the calculator cannot handle bigN, then the conditionals (< a bigN) and (< b bigN) cannot be expressed, so the calculator can't implement quss on bigN. Since the function cannot even be implemented with bigN, it's pointless to ask what it's doing. Questions can only be asked about what the actual implementation is doing; not what an imagined unimplementable implementation is doing.
Goff then tries to apply this to brains and this is where the sleight of hand occurs. The supposed dichotomy between brains and calculators is that brains can know they are adding or quadding with numbers that are too big for the brain to handle. Therefore, brains are not calculators.
The sleight of hand is that our brains can work with the descriptions of the behavior, while most calculators are built with only the behavior. With calculators, and much software, the descriptions are stripped away so that only the behavior remains. But there exists software that works with descriptions to generate behavior. This technique is known as "symbolic computation". Programs such as Maxima, Mathematica, and Maple can know that they are adding or quadding because they can work from the symbolic description of the behavior. Like humans, they deal with short descriptions of large things1. We can't write out all of the digits in the number 10^120. But because we can manipulate short descriptions of big things, we can answer what quss would do if bigN were 10^80. 10^80 is less than 10^120, so quss would return 5. Symbolic computation would give the same answer. But if we tried to do that with the actual numbers, we couldn't. When the thing described doesn't fit, it can't be worked on. Or, if the attempt is made, the old programming adage, Garbage In - Garbage Out, applies to humans and machines alike.
[1] We deal with infinity via short descriptions, e.g. "10 goto 10". We don't actually try to evaluate this, because we know we would get stuck if we did. We tag it "don't evaluate". If we actually need a result with these kinds of objects, we get rid of the infinity by various finite techniques.
[2] This post title refers to a prior brief mention of quss here. In that post, it suggested looking at the wiring of a device to determine what it does. In this post, we look at the behavior of the device across all of its inputs to determine what it does. But we only do that because we don't embed a rich description of each behavior in most devices. If we did, we could simply ask the device what it is doing. Then, just as with people, we'd have to correlate their behavior with their description of their behavior to see if they are acting as advertised.
Philip Goff explores Kripke's quss function, defined as:
In English, if the two inputs are both less than 100, the inputs are added, otherwise the result is 5.
(defparameter N 100)
(defun quss (a b)
(if (and (< a N) (< b N))
(+ a b)
5))
Goff then claims:
Rather, it’s indeterminate whether it’s adding orquadding.
This statement rests on some unstated assumptions. The calculator is a finite state machine. For simplicity, suppose the calculator has 10 digits, a function key (labelled "?" for mystery function) and "=" for result. There is a three character screen, so that any three digit numbers can be "quadded". The calculator can then produce 1000x1000 different results. A larger finite state machine can query the calculator for all one million possible inputs then collect and analyze the results. Given the definition of quss, the analyzer can then feed all one million possible inputs to quss, and show that the output of quss matches the output of the calculator.
Goff then tries to extend this result by making N larger than what the calculator can "handle". But this attempt fails, because if the calculator cannot handle bigN, then the conditionals (< a bigN) and (< b bigN) cannot be expressed, so the calculator can't implement quss on bigN. Since the function cannot even be implemented with bigN, it's pointless to ask what it's doing. Questions can only be asked about what the actual implementation is doing; not what an imagined unimplementable implementation is doing.
Goff then tries to apply this to brains and this is where the sleight of hand occurs. The supposed dichotomy between brains and calculators is that brains can know they are adding or quadding with numbers that are too big for the brain to handle. Therefore, brains are not calculators.
The sleight of hand is that our brains can work with the descriptions of the behavior, while most calculators are built with only the behavior. With calculators, and much software, the descriptions are stripped away so that only the behavior remains. But there exists software that works with descriptions to generate behavior. This technique is known as "symbolic computation". Programs such as Maxima, Mathematica, and Maple can know that they are adding or quadding because they can work from the symbolic description of the behavior. Like humans, they deal with short descriptions of large things1. We can't write out all of the digits in the number 10^120. But because we can manipulate short descriptions of big things, we can answer what quss would do if bigN were 10^80. 10^80 is less than 10^120, so quss would return 5. Symbolic computation would give the same answer. But if we tried to do that with the actual numbers, we couldn't. When the thing described doesn't fit, it can't be worked on. Or, if the attempt is made, the old programming adage, Garbage In - Garbage Out, applies to humans and machines alike.
[1] We deal with infinity via short descriptions, e.g. "10 goto 10". We don't actually try to evaluate this, because we know we would get stuck if we did. We tag it "don't evaluate". If we actually need a result with these kinds of objects, we get rid of the infinity by various finite techniques.
[2] This post title refers to a prior brief mention of quss here. In that post, it suggested looking at the wiring of a device to determine what it does. In this post, we look at the behavior of the device across all of its inputs to determine what it does. But we only do that because we don't embed a rich description of each behavior in most devices. If we did, we could simply ask the device what it is doing. Then, just as with people, we'd have to correlate their behavior with their description of their behavior to see if they are acting as advertised.
blog comments powered by Disqus