More Massimo Pigliucci Saturday, Nov 14 2009
AI 11:56 am
There is more discussion on Massimo Pigliucci’s critique of Chalmers’ Singularity talk at Scientific Blogging, including numerous comments. What I find most unusual about Pigliucci’s arguments is his assertion that human thought is non-algorithmic. It reminds me of vitalism. His idea seems to be that human thoughts are somehow fundamentally inaccessible to definition or reproduction, just how we once thought biology was. Why do people say that when we have already delineated numerous algorithms from how the mind works? There are books and books of this stuff — it would take years to read it all. See, for instance, The Bayesian Brain: Probabilistic Approaches to Neural Coding from MIT Press.
Could it that be Pigliucci is not aware of probabilistic algorithms, or algorithms that deal well with uncertainty? Or perhaps there is another magical threshold where he is imagining “things the human mind can do and algorithms can’t”. Whatever that thing one is imagining is, if it’s necessary for intelligence, it will eventually be analyzed and distilled into an algorithm. AI researchers have done this thousands of times already.
I would appreciate seeing a nice long debate between an intelligent person like Pigliucci who believes human thought is somehow irredeemably inscrutable, and someone like Chalmers who has a more sensible view. I’d actually prefer Pigliucci vs Eliezer, but that would probably be too much to hope for. There seems to be two camps that think they are more or less absolutely right and the other side is a little bit crazy, which generally means both sides (or at least one side) need to hear more of the other’s ideas. Of course, I think the SIAI et al are right about their position on this, but I bet there are a lot of people who haven’t heard it explained in a calm rational way addressing all objections and counterarguments.
Or could it be that you are reading too much into what I wrote? I simply raised the possibility that AI’s definition of “algorithm” is a bit too broad to be useful. And if you read any of my blog posts you will see that I am as much of a materialist as you can find them. But statements like “if it’s necessary for intelligence, it will eventually be analyzed and distilled into an algorithm. AI researchers have done this thousands of times already” are just silly. What gives you that cocky assurance?
What is the AI definition of algorithm? As far as I know, the definition is broad enough to encompass all aspects of intelligence.
You think it’s silly that science progressively understands more about intelligence? I haven’t read your earlier blog posts, no. But if you are “as much of a materialist as you can find them” (doubtful), then you would acknowledge the tens of thousands of pages of work done by computational neuroscientists towards distilling neural operation into discrete algorithms. We disagree so intensely here, that one of us has to be blatantly right and the other blatantly wrong, and I think that you’re the wrong one.
Pigliucci never claimed that human thoughts are “irredeemably inscrutable” or that they are “fundamentally inaccessible to definition or reproduction.” Those are just strawmen, and the crack about vitalism is a shameful low blow.
What I think he was saying (correct me if I’m wrong, Massimo) is that since we don’t yet know whether human-quality intelligence can be replicated in silicon, to blithely assume that it’s a given betrays an unscientific approach to problem-solving.
I find it disturbing that whenever anyone raises the point that higher intelligence might be an emergent property of biological brains and might prove difficult to instantiate in silicon, they are instantly labeled by this community as heretical and derided as “blatantly wrong” if not stupid. And then you wonder why singularitarianism is sometimes equated with religious belief?
Mike T,
Are you saying that there are some special atomic properties of carbon related to thought?
I don’t see why there would be (there’s certainly no evidence to suggest that there are)…but on the off chance that there were, there’s no reason to limit yourself to silicon. The important thing is that intelligence CAN be replicated (“instantiated”) from relative scratch; it happens every few seconds.
Yes, it will prove difficult to create human-level AI. However, I would say it is a difficult *engineering* problem. I have a feeling that you would say it’s something more. This implies a dualistic, non-materialist, philosophy; are you willing to admit that you’re a dualist?
Are you willing to admit that you’re a dualist?
That question is like asking how long ago did you stop beating your wife. Of course I’m not a dualist, and you are reading all sorts of things into my objectively straightforward statements that simply aren’t there.
Mike, when someone says that human thought is “non-algorithmic”, they mean that there is no algorithm that can describe how it works. That means inscrutable. How is that a straw man?
You are sort of feeding him a (vaguely) more reasonable-sounding argument, but he had hundreds of words in which to make that argument himself, and I didn’t see it.
Massimo didn’t say that intelligence might be an emergent property of biology alone, he said that human intelligence was non-algorithmic. There is a huge difference.
I have never said that precise belief you mention is blatantly wrong. (Maybe other commenters here have, but not me.) Still, trying to delineate sharp boundaries in thought is not unique to singularitarianism. Similar phenomena occur in every area of science, and egos clash in much the same way. Your accusation that singularitarianism is “religious” is mirrored by accusations all throughout science that the group that holds position X is just fanatical. People don’t like being told that they’re blatantly wrong — some consider it rude.
And yeah, I agree that “Are you willing to admit that you’re a dualist” is a silly line.
But if physicalism is true, and physics is computably approximable (AFAIK, very few physicists take seriously the possibility that it’s not), there’s very good reason to think it can be in principle and in practice. (The philosophical questions about the consciousness and personal identity of an AI/upload are less clear, but are independent of the arguments about its ability to act intelligently.)
I think we disagree on how far we can trust theoretical arguments, such as I linked, while empirical testing remains far off. Singularitarians generally don’t think that doing so is “unscientific” and/or don’t think that a narrow construal of “science” encompasses all rational thought, which lets these arguments seem sufficient to justify confidence in AI. They don’t necessarily intuitively understand that others place less trust in non-empirical reasoning, and take this distrust as indicating inability to follow the arguments, hence accusations of stupidity. Homeopathy obviously can’t work on theoretical grounds, and it wouldn’t have been reasonable for someone to say “homeopathy might work” if we possessed a good knowledge of physics and chemistry but homeopathy had never been experimentally tested; but this is roughly how “AI might be impossible” sounds to someone who buys the theoretical arguments for AI.
I really wish these arguments would more often go into the real, meta-level disagreements involved, like how much humans can be trusted to argue non-empirically, how reasonable it can be to hold a confident opinion in the face of expert uncertainty, or how strong superficially religous/pseudoscientific behavior is as evidence for irrationality and how defeasible it is as evidence.
(Also, much less importantly, I have no idea what you mean by “emergent property” or why intelligence being “emergent” would make it hard/impossible to implement.)
Just because something is in principle materialistic and able to be reduced to algorithms that can be run on silicon doesn’t mean in practice we can do so. For example, quantum chromodynamics is exceedingly difficult to simulate in a computer, and it isn’t clear that any classical computer will ever be able to model to high accuracy a uranium nucleus.
That said, it seems like everyone here is a reductionist, and the discussion should move on to the more pertinent question: are the algorithms for human intelligence *tractable* on hardware to become available in the next few decades? For this I recommend the Whole Brain Emulation Roadmap as good starting material:
http://www.philosophy.ox.ac.uk/__data/assets/pdf_file/0019/3853/brain-emulation-roadmap-report.pdf
Mike,
“That question is like asking how long ago did you stop beating your wife.”
Beating your wife is bad…being a dualist, not necessarily so.
“Of course I’m not a dualist, and you are reading all sorts of things into my objectively straightforward statements that simply aren’t there.”
I think that to straighten this conflict out would require a lot of time and effort.
Michael,
“And yeah, I agree that “Are you willing to admit that you’re a dualist” is a silly line.”
How can you reconcile not being able to instantiate intelligence with being a monist? I was simply trying to get Mike to commit to a consistent position. Perhaps I could have phrased it more delicately, but I didn’t think people would react so emotionally to it.
My position in a nutshell is this: all phenomena, whether rule-bound, random, or a combination, should be able to be instantiated given the right programing. I feel the statement “instantiating x might be hard” is vague, unless you define what you mean by “hard”. If, by “hard”, you mean impossible, then you are implicitly saying that the phenomenon in question is neither rule-bound nor random — a statement, that, in my mind, means that the phenomenon is outside the realm of the physical universe.
People don’t always realize what their beliefs imply.
As my above comment talks about, skepticism of the argument that one implies the other.