> The most efficient brain will have the highest information-processing rate I, and the lowest mass M, hence the highest ratio I/M. Since very large exponents are involved, for the convenience we define the Sentience Quotient or SQ as the logarithm of I/M, that is, its order of magnitude.
Doesn't dividing by mass give small things too much of an advantage? If we have a construct of 3 atoms that can process some small non-zero amount of information, wouldn't it be likely that this construct would be many orders of magnitude more intelligent than a human, based on the very low mass?
Further, the Apple II and Cray are rated at +5 and +9. It seems like it'd be quite straightforward to cut the mass in half (maybe multiple times) for both of those machines and radically improve their score. Is that rating based on the computer with case and cooling, or just the chip and package, or the silicon without packaging, or just the transistors on the silicon? Each of those answers would, under the current formulation, radically change the sentience quotient, right?
I suppose I'm saying I'm concerned about the mass having too great of a role in this equation. This either indicates some kind of misunderstanding on my part, or perhaps it'd be improved by using something like sqrt(M)?
Although it might be possible to modify the mass of a computer by a factor of 2 this is really not going to affect its score so much, given the presence of a logarithm. You're definitely not going to be able to adjust its mass by an order of magnitude after all. A similar problem exists for humans: should you include the whole body or just the brain? What about some kind of power source? Luckily all of these factors are only going to add/subtract a few points (doubling/halving mass is +-0.3).
A bigger source of error is whether the ability to process information can be put to good use. Since this was written computing power has increased by at least 6 orders of magnitude, putting computers ahead of humans. Yet we are no closer to being able to have an intelligent conversation with a computer than we are to having an intelligent conversation with a plough or any other machine that can be used to lever our minds.
Do you really think that "we are no closer to being able to have an intelligent conversation with a computer" ?
I mean, we aren't there yet, but we're definitely closer than before, the gap is narrowing in part due to the 6 order of magnitude increase. And do take note that at the time when we'd be 95% close, then it would still feel like having an intelligent conversation with a dog; the conversation would only 'feel right' if the capabilities matched 100% or more.
That's a good point, I guess with something like that you'll never know how close you are until you get there.
Nonetheless, once we get there (and I see no reason why we shouldn't eventually - even if it's via and exact simulation of a human brain) I'll be interested to look back and see how much of what we were missing was knowledge and how much was computational capacity.
Incorporating mass makes sense in considering whether the intelligence type would be useful for an organism.
For example, a little sea creature with the capability of an Apple II (not the programs, but the capacity to process data and respond to stimuli) might be competitive if the neural tissue were compact and light, but not if it had to be the mass of an Apple II. So its capacity to process 10^5 bits per second per kg of weight is not that useful. Though apparently more efficient than plants' processing methods.
This points the way to an interesting new metric which would be something like average decisions per kilogram per kilowatt input power. To one sig fig (if that much) human bodies at a constant weight around 100 Kg run a constant power level "around" 100 watts and we seem capable of making a decision per second or so. So 3600 decisions at 100 Kg per 0.1 KW gives you about 4 billion, so our figure of merit is about four billion decision kilos per kilowatt/hr. How much energy it takes to control your hands and how capable your hands are and how agile your hands are. Octopi have more "hands" to control and they're lighter so they can accomplish less.
There's probably a very human tendency to fudge the equations to put us on top. Although given the peculiarity of random evolution there's no reason ants or dolphins couldn't numerically beat us, although selection pressures in their environments means they've never as a group accomplished even a tiny fraction we have. This "slacker coefficient" cannot be underestimated, I'm quite sure a theoretical SQ=50 entity from a slacker civilization could make the human idiocracy movie look brilliant. In that way I think the SQ impedance barrier in the article is much more impermeable than the article implies. Idiotic celebrity news in SQ=50 land is probably about the same as idiotic celebrity news in SQ=10 land, although the pace and individual stories might vary. Humans have entire subcultures which seem to exist solely to drag down the rest of civilization, and that's probably not unique to humanity. The tired analogy of crabs dragging any crab back into the pot that tries to escape come to mind. Some factor of civilization signal to noise ratio is required, not just brute signal level.
Wow, I read the article, and ended up falling in a Wikipedia rabbithole, reading up about Aristotelian logic, Peano arithmetic, communication in sea mammals, and now I need to go eat dinner before it's 3 am. :)
That said, the part about (computational) capabilities seems unwilling to contemplate any but the most rigorously-established facts, leading to completely irrelevant estimates. It's fairly obvious that the limit on human intelligence is not Goedel's incompleteness theorem; it should also be fairly obvious that humans do not process 10 ^ 11 * 1000 = 10 ^ 14 bits = 100 Tb per second (this combines Wikipedia's number of neurons in the brain and the article's bits/second/neuron), not in any meaningful way and definitely not in the "10 ^ 14 yes/no decisions per second" that the article gives as a definition/explanation for "bit". In fact, the article picks efficiency, rather than efficiency times mass, as a measure of intelligence; it's not at all obvious why this should be so, in a simple computational model - bigger computers do tend to be faster, after all. (On the other hand, animal intelligence seems to be better predicted by brain mass per body mass.)
Unfortunately, I'm not aware of any better analysis that is still fairly rigorous; does anyone have suggestions?
This article was written in 1984, 29 years ago. Some of the information used in it is a bit out of date.
1) It refers to the triune brain model that has been debunked already. See this Scientific American blog entry: http://blogs.scientificamerican.com/guest-blog/2012/09/07/re... So most of the speculation regarding reptilian, limbic and neocortical intelligences isn't really valid. Though this doesn't affect the chordate vs. ganglionic argument.
2) The article states that "consciousness is an emergent of neuronal sentience" which isn't a necessary assumption. Recent research[1] has hinted at intelligent behavior as being a thermodynamic process occurring when a physical system acts such as to maximize its number of possible future states. Quick video summary: https://www.youtube.com/watch?v=rZB8TNaG-ik If true, a typical neuronal structure (i.e. brain) wouldn't necessarily be required for intelligent or conscious behavior, maybe not even a typical computational structure. One could imagine a cloud of gas undergoing chemistry such that it maximized its entropy production and could potentially possess superhuman intelligence or consciousness.
3) The article discusses the theoretical limits of intelligence as being a system possessing 10^50 bits of information per kilogram. But the maximum limit of a computational system should be given by the Bekenstein bound[2], which limits the information the system can process to a factor proportional to the surface area of the enclosed volume (for a 1 cubic cm sphere this comes to about 10^66 bits), the speed of light limiting the speed at which these bits can be processed, and the particular structure of how they are processed (e.g. the flow of logic gates). One could imagine a (quantum?) computer with the maximum encloseable surface area of the universe which transmitted information optically between different gates in the system. There may be other limits preventing that from being remotely feasible (e.g. heat dissipation, energy requirements, entropy production leading to the heat death of the universe, etc.). The main point is that comparing the Apple II at +5 SQ on a scale of 0 to 50 is kind of silly, when one could imagine an intelligence the size of galaxies, or even the known universe, operating according to a principal of maximum entropy production similar as referenced in the previous paragraph.
Not to knock the article; my main point is that there have been a lot more theoretical advances in the almost-30-years since it was written.
A bit of nitpick on point 3) - it's a logarithmic scale. The difference between +10 and +5 is 148 times and between +5 and +50 it's about 3.5e+19 which should be quite doable by something as large as a galaxy...
Reading this article reminded me a bit of Carl Sagan's "The Dragons of Eden"[1] (1977). It's subtitle: 'Speculations on the Evolution of Human Intelligence' explains the content pretty well. Mr. Sagan explains quite clearly in the introduction that he is speculating, but it is very entertaining speculation.
13 comments
[ 2.8 ms ] story [ 37.7 ms ] threadDoesn't dividing by mass give small things too much of an advantage? If we have a construct of 3 atoms that can process some small non-zero amount of information, wouldn't it be likely that this construct would be many orders of magnitude more intelligent than a human, based on the very low mass?
Further, the Apple II and Cray are rated at +5 and +9. It seems like it'd be quite straightforward to cut the mass in half (maybe multiple times) for both of those machines and radically improve their score. Is that rating based on the computer with case and cooling, or just the chip and package, or the silicon without packaging, or just the transistors on the silicon? Each of those answers would, under the current formulation, radically change the sentience quotient, right?
I suppose I'm saying I'm concerned about the mass having too great of a role in this equation. This either indicates some kind of misunderstanding on my part, or perhaps it'd be improved by using something like sqrt(M)?
A bigger source of error is whether the ability to process information can be put to good use. Since this was written computing power has increased by at least 6 orders of magnitude, putting computers ahead of humans. Yet we are no closer to being able to have an intelligent conversation with a computer than we are to having an intelligent conversation with a plough or any other machine that can be used to lever our minds.
I mean, we aren't there yet, but we're definitely closer than before, the gap is narrowing in part due to the 6 order of magnitude increase. And do take note that at the time when we'd be 95% close, then it would still feel like having an intelligent conversation with a dog; the conversation would only 'feel right' if the capabilities matched 100% or more.
Nonetheless, once we get there (and I see no reason why we shouldn't eventually - even if it's via and exact simulation of a human brain) I'll be interested to look back and see how much of what we were missing was knowledge and how much was computational capacity.
For example, a little sea creature with the capability of an Apple II (not the programs, but the capacity to process data and respond to stimuli) might be competitive if the neural tissue were compact and light, but not if it had to be the mass of an Apple II. So its capacity to process 10^5 bits per second per kg of weight is not that useful. Though apparently more efficient than plants' processing methods.
This points the way to an interesting new metric which would be something like average decisions per kilogram per kilowatt input power. To one sig fig (if that much) human bodies at a constant weight around 100 Kg run a constant power level "around" 100 watts and we seem capable of making a decision per second or so. So 3600 decisions at 100 Kg per 0.1 KW gives you about 4 billion, so our figure of merit is about four billion decision kilos per kilowatt/hr. How much energy it takes to control your hands and how capable your hands are and how agile your hands are. Octopi have more "hands" to control and they're lighter so they can accomplish less.
There's probably a very human tendency to fudge the equations to put us on top. Although given the peculiarity of random evolution there's no reason ants or dolphins couldn't numerically beat us, although selection pressures in their environments means they've never as a group accomplished even a tiny fraction we have. This "slacker coefficient" cannot be underestimated, I'm quite sure a theoretical SQ=50 entity from a slacker civilization could make the human idiocracy movie look brilliant. In that way I think the SQ impedance barrier in the article is much more impermeable than the article implies. Idiotic celebrity news in SQ=50 land is probably about the same as idiotic celebrity news in SQ=10 land, although the pace and individual stories might vary. Humans have entire subcultures which seem to exist solely to drag down the rest of civilization, and that's probably not unique to humanity. The tired analogy of crabs dragging any crab back into the pot that tries to escape come to mind. Some factor of civilization signal to noise ratio is required, not just brute signal level.
Thanks for posting this amazing article.
That said, the part about (computational) capabilities seems unwilling to contemplate any but the most rigorously-established facts, leading to completely irrelevant estimates. It's fairly obvious that the limit on human intelligence is not Goedel's incompleteness theorem; it should also be fairly obvious that humans do not process 10 ^ 11 * 1000 = 10 ^ 14 bits = 100 Tb per second (this combines Wikipedia's number of neurons in the brain and the article's bits/second/neuron), not in any meaningful way and definitely not in the "10 ^ 14 yes/no decisions per second" that the article gives as a definition/explanation for "bit". In fact, the article picks efficiency, rather than efficiency times mass, as a measure of intelligence; it's not at all obvious why this should be so, in a simple computational model - bigger computers do tend to be faster, after all. (On the other hand, animal intelligence seems to be better predicted by brain mass per body mass.)
Unfortunately, I'm not aware of any better analysis that is still fairly rigorous; does anyone have suggestions?
1) It refers to the triune brain model that has been debunked already. See this Scientific American blog entry: http://blogs.scientificamerican.com/guest-blog/2012/09/07/re... So most of the speculation regarding reptilian, limbic and neocortical intelligences isn't really valid. Though this doesn't affect the chordate vs. ganglionic argument.
2) The article states that "consciousness is an emergent of neuronal sentience" which isn't a necessary assumption. Recent research[1] has hinted at intelligent behavior as being a thermodynamic process occurring when a physical system acts such as to maximize its number of possible future states. Quick video summary: https://www.youtube.com/watch?v=rZB8TNaG-ik If true, a typical neuronal structure (i.e. brain) wouldn't necessarily be required for intelligent or conscious behavior, maybe not even a typical computational structure. One could imagine a cloud of gas undergoing chemistry such that it maximized its entropy production and could potentially possess superhuman intelligence or consciousness.
3) The article discusses the theoretical limits of intelligence as being a system possessing 10^50 bits of information per kilogram. But the maximum limit of a computational system should be given by the Bekenstein bound[2], which limits the information the system can process to a factor proportional to the surface area of the enclosed volume (for a 1 cubic cm sphere this comes to about 10^66 bits), the speed of light limiting the speed at which these bits can be processed, and the particular structure of how they are processed (e.g. the flow of logic gates). One could imagine a (quantum?) computer with the maximum encloseable surface area of the universe which transmitted information optically between different gates in the system. There may be other limits preventing that from being remotely feasible (e.g. heat dissipation, energy requirements, entropy production leading to the heat death of the universe, etc.). The main point is that comparing the Apple II at +5 SQ on a scale of 0 to 50 is kind of silly, when one could imagine an intelligence the size of galaxies, or even the known universe, operating according to a principal of maximum entropy production similar as referenced in the previous paragraph.
Not to knock the article; my main point is that there have been a lot more theoretical advances in the almost-30-years since it was written.
[1] http://math.mit.edu/~freer/papers/PhysRevLett_110-168702.pdf
[2] http://www.scholarpedia.org/article/Bekenstein_bound
[1] http://en.wikipedia.org/wiki/The_Dragons_of_Eden