Headlock
Whenever I see an article written about AI, it either falls into the category of very learned and backed up by research and experience, or it’s unmitigated tripe fuelled by romanticism and fiction. ‘“What are the odds?”, by Mitchell How’ falls very much into the latter. It presents two scenarios based on a recent wager, which, when you get right down to it, is a sucker’s bet. It may have a time limit which prevents it from being a true sucker’s bet, but the hallmarks are there.
It’s hard to know where to start refuting such an article filled with a plethora of misconceptions and fallacies. The first place would really be the Turing Test, one of the most well known ways of “proving” an AI. The first thing any AI course worth its salt will teach you is not so much that the Turing Test is worthless, just that it doesn’t necessarily prove that something is intelligent. It makes the mistake that if a human is fooled into thinking the machine is human, that it has demonstrated thought or intelligence when in actuality, a lot of humans wouldn’t “pass” the Turing test. There is an immense body of work dedicated to the Turing test and whether it should be just one of a suite of tests or whether the test itself fallacious or whether tests are even necessary.
Cutting to the heart of this matter is the definition of intelligence. One cannot give a definition because one does and can not exist, a measure of intelligence is built in a variety of ways but there is no cold-clinical definition that would allow a strict separation of intelligence and stupidity, light and dark, pass or fail. Just like someone cannot prove they are alive, they cannot prove they are intelligent because one definition does not match up with another. “I am alive because I am cognisant of my surroundings,” “I am intelligent because I am not yet dead”.
The primary part of what I see bandied about as discussion of AI is the ridiculous idea of “true” AI; as if we could somehow imbue a system with life and intelligence through lightning bolts and Tesla-coils and that a shining paradigm of intelligence would be born. Computers are deterministic; we get out of them exactly what we put in. When it comes to AI there is the situation of “emergent behaviour”, whereby we get something out that we didn’t necessarily expect; this is not to say that the outcome could not have been predicted or modelled, just that it wasn’t expected. So by that definition, to create some kind of “true” AI, we would have to have someone, either one or a group of people, who would be as smart as the system they created; intelligence could not just “fall out” of a vastly complex system. The article goes on to speak about cyclic improvement as if it was the simplest thing in the world: by virtue of being a machine, the AI could of course improve itself whereby we have failed. Just like how if we had two AIs, they could of course “merge” to create a better amalgam of the two. The entire concept of “true” AI is so outlandish and completely outside the bounds of rational and logical thinking it gobsmacks me.
The comments to the article range from the informed to the abstract but cover a wider gamut of reasoning than the article itself. I think what amuses me most about the concept presented in the article and in numerous other bits of futurism and fiction is that fundamentally, we (as human-kind) do not understand how our brain works. We see the brain as the nexus of our “intelligence”, how a dense cluster of nerves which coagulated over generations of evolution gives rise to “consciousness” and “thought”. If we could model that or at least the “important” parts of the brain, we would birth something at least as intelligent as us. To me, this falls into that category of magic and gods, whereby we don’t understand something so find ways to explain it away. Maybe what we think of as cognitive thought is simply smoke and mirrors, emergent behaviour we can’t understand simply because we’re part of the process, and trying to fool ourselves into believing we created something intelligent is just an Escher painting of madness.
I digress, really this isn’t a discussion on the philosophical aspects and impacts of AI and its fundamentals, but simply to indicate that we are merely scratching the surface of a vast and nebulous concept. The article makes no attempt to be researched or intelligent in the matter and simply comes across as sensationalist and shallow.
10th July 2007
Tom McCabe“The first thing any AI course worth its salt will teach you is not so much that the Turing Test is worthless, just that it doesn’t necessarily prove that something is intelligent.”
If an AGI can do anything a human can do- if it can perform as a human so well that we can’t detect the difference- then it must be intelligent. If an AGI could do anything a human could do, and yet we didn’t call it “intelligent”, that just proves that we’re a perverse species, since we’re obviously defining “intelligence” using something other than behavior.
“when in actuality, a lot of humans wouldn’t “pass” the Turing test.”
Obviously, if you pit a human against another human in a Turing test, one of them is going to lose. This doesn’t say anything about humans or intelligence; it’s just an artifact of the way the test is designed. Testing humans this way is as absurd as trying to prove that a plane can fly by comparing it to itself.
“but there is no cold-clinical definition that would allow a strict separation of intelligence and stupidity, light and dark, pass or fail.”
This is true; however, that does not mean that intelligence is imaginary. Heat obviously exists; this does not mean we can classify the world into things “with heat” and things “without heat”.
“Just like someone cannot prove they are alive, ”
Sure they can. Unlike “intelligence”, “life” does have a practical definition, as something that can reproduce. You can show you are alive by demonstrating that you have a mechanism to reproduce.
“as if we could somehow imbue a system with life and intelligence through lightning bolts and Tesla-coils and that a shining paradigm of intelligence would be born.”
This smacks of Hollywood Frankenstein movies. Movies, as in fiction, as in *has no bearing on reality whatsoever*. See http://www.acceleratingfuture.com/tom/?p=12.
“Computers are deterministic; we get out of them exactly what we put in. ”
http://www.acceleratingfuture.com/tom/?p=8
“this is not to say that the outcome could not have been predicted or modelled, just that it wasn’t expected.”
Rice’s Theorem states that it is mathematically impossible to build a computer which can say *anything whatsoever* about an arbitrary AI. See http://en.wikipedia.org/wiki/Rice’s_Theorem.
“Just like how if we had two AIs, they could of course “merge” to create a better amalgam of the two”
This claim has nothing to do whatsoever with your previous claim.
“To me, this falls into that category of magic and gods, whereby we don’t understand something so find ways to explain it away.”
We don’t understand *everything* about the brain, but to claim that “the brain” is just a magical explanation for intelligence like “phlogiston” is a magical explanation for fire is just silly. We’ve chopped down the brain into its components, we’ve figured out what each component does, and we’ve figured out how the components communicate. Anyone who claims that the brain is a total mystery should be slapped upside the head with MITECS. All one thousand ninety-six pages of it.
“The article makes no attempt to be researched or intelligent in the matter and simply comes across as sensationalist and shallow.”
If you want to have a deep, technical understanding of AGI, you have to do a lot of studying. A *lot* of studying. See http://www.singinst.org/reading/corereading for starters.
11th July 2007
ChaosTangent“If an AGI can do anything a human can do- if it can perform as a human so well that we can’t detect the difference- then it must be intelligent.”
No, it is just means that it can emulate a human, it has nothing to do with intelligence. Look at the world of robotics, we already have machines that can emulate complex human interactions, and there is nothing “intelligent” about them. This is exactly the point about why the Turing test shouldn’t be used as a measure of intelligence because it favours smoke-and-mirros and deception versus measuring ACTUAL intelligence.
“This is true; however, that does not mean that intelligence is imaginary.”
I never said it was imaginary (an absurd suggestion), but the Turing test has a binary outcome rather than an analogue one. If a machine passes the test, it’s intelligent. This is the fundamental flaw of the test. Even from what you’re saying you seem to agree that intelligence is a measure, it has scale, which the Turing test doesn’t measure.
“Sure they can. Unlike “intelligence”, “life” does have a practical definition, as something that can reproduce. You can show you are alive by demonstrating that you have a mechanism to reproduce.”
What of people who are impotent? Or sterile? Does losing the ability reproduce automatically make them not alive? Machines can already produce themselves, does this mean the factory worker machines are alive? This isn’t to do with the point at hand but discussing away a very core philosophical point smacks of arrogance.
“This smacks of Hollywood Frankenstein movies. Movies, as in fiction, as in *has no bearing on reality whatsoever*.”
And finally, the crux of the argument. The entire point of my post was to point out that the article seems to have absolutely no grounding in reality whatsoever. It seems that your “theories” and and their words are not based in the here-and-now workings of AI and are so far removed that it seems to indicate that AGI or “True AI” will just sprout out of holes in the ground. It’s a ridiculous, romantic idea that is posed and doesn’t help current, grass-roots AI research at all; it’s the posturing of a self-defined intellectual.
“http://www.acceleratingfuture.com/tom/?p=8”
This is not a refutation of the fact that computers are deterministic. It’s not a case of free will, computers work on the precept of determinism as a practicality rather than a philosophical concept.
“http://en.wikipedia.org/wiki/Rice’s_Theorem”
This is fine and dandy, but again has nothing to do with reality. Computers, modern or otherwise are not fully-fledged Turing machines.
“but to claim that “the brain” is just a magical explanation for intelligence” - “Anyone who claims that the brain is a total mystery”
Again, I never said or implied either of these. Just because we’ve dissected the brain doesn’t mean we understand it, it’s a case of looking at the component parts and not seeing how they fit together. We know they do, just not necessarily how. My point to this was that by not understanding the brain, the undisputed centre of our “intelligence”, how can we create a system which models our intelligence if we don’t even understand the physical aspects of how it works? It would be like trying to build a car without having a plan or knowing how the gearbox connects.
“If you want to have a deep, technical understanding of AGI, you have to do a lot of studying. A *lot* of studying.”
I never said that you hadn’t done a lot of studying, evidently you have on your given subject. My point is that the article didn’t show this; as I started my article: “unmitigated tripe fueled by romanticism and fiction”. So far you have done nothing to prove otherwise as your “technical” understanding of the subject at hand seems to focus on theorems with no practical implementations. I have no doubt you have a deep “understanding” of your subject area, but technical it is not.
11th July 2007
Tom McCabeAll of your points have been addressed in the original thread back on acceleratingfuture.com — I invite you to join the discussion there.
9th August 2007
ChaosTangentPoint 6 of the following article is highly relevant to this discussion: http://discovermagazine.com/2007/aug/unsolved-brain-mysteries/article_view?b_start:int=1&-C=