= 27d567bc2d48d9f40e9ccc20ea370b15)
lukeprog comments on Less Wrong Rationality and Mainstream Philosophy - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (328)
I'll await your next post, but in retrospect you should have started with the big concrete example of mainstream philosophy doing an LW-style dissolution-to-algorithm not already covered on LW, and then told us that the moral was that we shouldn't ignore mainstream philosophy.
I did the whole sequence on QM to make the final point that people shouldn't trust physicists to get elementary Bayesian problems right. I didn't just walk in and tell them that physicists were untrustworthy.
If you want to make a point about medicine, you start by showing people a Bayesian problem that doctors get wrong; you don't start by telling them that doctors are untrustworthy.
If you want me to believe that philosophy isn't a terribly sick field, devoted to arguing instead of facing real-world tests and admiring problems instead of solving them and moving on, whose poison a novice should avoid in favor of eating healthy fields like settled physics (not string theory) or mainstream AI (not AGI), you're probably better off starting with the specific example first. "I disagree with your decision not to cover terminal vs. instrumental in CEV" doesn't cover it, and neither does "Quineans agree the world is made of atoms". Show me this field's power!
Eliezer,
When I wrote the post I didn't know that what you meant by "reductionist-grade naturalistic cognitive philosophy" was only the very narrow thing of dissolving philosophical problems to cognitive algorithms. After all, most of the useful philosophy you've done on Less Wrong is not specifically related to that very particular thing... which again supports my point that mainstream philosophy has more to offer than dissolution-to-algorithm. (Unless you think most of your philosophical writing on Less Wrong is useless.)
Also, I don't disagree with your decision not to cover means and ends in CEV.
Anyway. Here are some useful contributions of mainstream philosophy:
Of course that's far from all there is, but it's a start.
...also, you occasionally stumble across some neato quotes, like Dennett saying "AI makes philosophy honest." :)
Note that useful insights come from unexpected places. Rawls was not a Quinean naturalist, but his concept of reflective equilibrium plays a central role in your plan for Friendly AI to save the world.
P.S. Predicate logic was removed from the original list for these reasons.
It seems a shame to leave this list with several useful cites as a comment, where it is likely to be missed. Not sure what to suggest - maybe append it to the main article?
I added a link to this list to the end of the original post.
Saying this may count as staking an exciting position in philosophy, already right there; but merely saying this doesn't shape my expectations about how people think, or tell me how to build an AI, or how to expect or do anything concrete that I couldn't do before, so from an LW perspective this isn't yet a move on the gameboard. At best it introduces a move on the gameboard.
I know Tarski as a mathematician and have acknowledged my debt to him as a mathematician. Perhaps you can learn about him in philosophy, but that doesn't imply people should study philosophy if they will also run into Tarski by doing mathematics.
...was great for introducing mainstream academia to Good, but if you compare it to http://wiki.lesswrong.com/wiki/The_Hanson-Yudkowsky_AI-Foom_Debate then you'll see that most of the issues raised didn't fit into Chalmers's decomposition at all. Not suggesting that he should've done it differently in a first paper, but still, Chalmers's formalization doesn't yet represent most of the debates that have been done in this community. It's more an illustration of how far you have to simplify things down for the sake of getting published in the mainstream, than an argument that you ought to be learning this sort of thing from the mainstream.
Acknowledged and credited. Like Drescher, Dennett is one of the known exceptions.
Appears as a citation only in AIMA 2nd edition, described as a philosopher who approves of GOFAI. "Not all philosophers are critical of GOFAI, however; some are, in fact, ardent advocates and even practitioners... Michael Bratman has applied his "belief-desire-intention" model of human psychology (Bratman, 1987) to AI research on planning (Bratman, 1992)." This is the only mention in the 2nd edition. Perhaps by the time they wrote the third edition they read more Bratman and figured that he could be used to describe work they had already done? Not exactly a "major inspiration", if so...
This comes under the heading of "things that rather a lot of computer programmers, though not all of them, can see as immediately obvious even if philosophers argue it afterward". I really don't think that computer programmers would be at a loss to understand that different systems can implement the same algorithm if not for Putnam and Lewis.
Same comment as for Quine: This might introduce interesting work, but while saying just this may count as an exciting philosophical position, it's not a move on the LW gameboard until you get to specifics. Then it's not a very impressive move unless it involves doing nonobvious reductionism, not just "Bias X might make philosophers want to believe in position Y". You are not being held to a special standard as Luke here; a friend named Kip Werking once did some work arguing that we have lots of cognitive biases pushing us to believe in libertarian free will that I thought made a nice illustration of the difference between LW-style decomposition of a cognitive algorithm and treating biases as an argument in the war of surface intuitions.
Mathematician and AI researcher. He may have mentioned the philosophical literature in his book. It's what academics do. He may even have read the philosophers before he worked out the answer for himself. He may even have found that reading philosophers getting it wrong helped spur him to think about the problem and deduce the right answer by contrast - I've done some of that over the course of my career, though more in the early phases than the later phases. Can you really describe Pearl's work as "building" on philosophy, when IIRC, most of the philosophers were claiming at this point that causality was a mere illusion of correlation? Has Pearl named a previous philosopher, who was not a mathematician, who Pearl thought was getting it right?
Previously named by me as good philosophy, as done by an AI researcher coming in from outside for some odd reason. Not exactly a good sign for philosophy when you think about it.
For a change I actually did read about this before forming my own AI theories. I can't recall ever actually using it, though. It's for helping people who are confused in a way that I wasn't confused to begin with. Dennett is in any case a widely known and named exception.
A friend and colleague who was part of the transhumanist community and a founder of the World Transhumanist Association long before he was the Director of the Oxford Future of Humanity Institute, and who's done a great deal to precisionize transhumanist ideas about global catastrophic risks and inform academia about them, as well as excellent original work on anthropic reasoning and the simulation argument. Bostrom is familiar with Less Wrong and has even tried to bring some of the work done here into mainstream academia, such as Pascal's Mugging, which was invented right here on Less Wrong by none other than yours truly - although of course, owing to the constraints of academia and their prior unfamiliarity with elementary probability theory and decision theory, Bostrom was unable to convey the most exciting part of Pascal's Mugging in his academic writeup, namely the idea that Solomonoff-induction-style reasoning will explode the size of remote possibilities much faster than their Kolmogorov complexity diminishes their probability.
Reading Bostrom is a triumph of the rule "Read the most famous transhumanists" not "Read the most famous philosophers".
The doomsday argument, which was not invented by Bostrom, is a rare case of genuinely interesting work done in mainstream philosophy - anthropic issues are genuinely not obvious, genuinely worth arguing about and philosophers have done genuinely interesting work on it. Similarly, although LW has gotten further, there has been genuinely interesting work in philosophy on the genuinely interesting problems of Newcomblike dilemmas. There are people in the field who can do good work on the rather rare occasions when there is something worth arguing about that is still classed as "philosophy" rather than as a separate science, although they cannot actually solve those problems (as very clearly illustrated by the Newcomblike case) and the field as a whole is not capable of distinguishing good work from bad work on even the genuinely interesting subjects.
Argued it on Less Wrong before he wrote the mainstream paper. The LW discussion got further, IMO. (And AFAIK, since I don't know if there was any academic debate or if the paper just dropped into the void.)
Is not useful for anything in real life / AI. This is instantly obvious to any sufficiently competent AI researcher. See e.g. http://norvig.com/design-patterns/img070.htm, a mention that turned up in passing back when I was doing my own search for prior work on Friendly AI.
...I'll stop there, but do want to note, even if it's out-of-order, that the work you glowingly cite on statistical prediction rules is familiar to me from having read the famous edited volume "Judgment Under Uncertainty: Heuristics and Biases" where it appears as a lovely chapter by Robyn Dawes on "The robust beauty of improper linear models", which quite stuck in my mind (citation from memory). You may have learned about this from philosophy, and I can see how you would credit that as a use of reading philosophy, but it's not work done in philosophy and, well, I didn't learn about it there so this particular citation feels a bit odd to me.
That this isn't at all the case should be obvious even if the only thing you've read on the subject is Pearl's book. The entire counterfactual approach is due to Lewis and Stalnaker. Salmon's theory isn't about correlation either. Also, see James Woodward who has done very similar work to Pearl but from a philosophy department. Pearl cites all of them if I recall.
Stalnaker's name sounds familiar from Pearl, so I'll take your word for this and concede the point.
As I pointed out before, the same is true for me of Quine. I don't know if lukeprog means to include Mathematical Logic when he keeps saying not to read Quine, but that book was effectively my introduction to the subject, and I still hold it in high regard. It's an elegant system with some important innovations, and features a particularly nice treatment of Gödel's incompleteness theorem (one of his main objectives in writing the book). I don't know if it's the best book on mathematical logic there is (I doubt it), but it appeals to a certain kind of personality, and I would certainly recommend it to a young high-schooler over reading Principia Mathematica, for example.
Cool. Let me know when you've finished your comment here and I'll respond.
Done.
Quine's naturalized epistemology: agreed.
Tarski: But I thought you said you were not only influenced by Tarski's mathematics but also his philosophical work on truth?
Chalmers' paper: Yeah, it's mostly useful as an overview. I should have clarified that I meant that Chalmers' paper makes a more organized and compelling case for Good's intelligence explosion than anybody at SIAI has in one place. Obviously, your work (and your debate with Robin) goes far beyond Chalmers' introductory paper, but it's scattered all over the place and takes a lot of reading to track down and understand.
And this would be the main reason to learn something from the mainstream: If it takes way less time than tracking down the same arguments and answers through hundreds of Less Wrong posts and other articles, and does a better job of pointing you to other discussions of the relevant ideas.
But we could have the best of both worlds if SIAI spent some time writing well-referenced survey articles on their work, in the professional style instead of telling people to read hundreds of pages of blog posts (that mostly lack references) in order to figure out what you're talking about.
Bratman: I don't know his influence first hand, either - it's just that I've seen his 1987 book mentioned in several books on AI and cognitive science.
Pearl: Jack beat me to the punch on this.
Talbot: I guess I'll have to read more about what you mean by dissolution to cognitive algorithm. I thought the point was that even if you can solve the problem, there's that lingering wonder about why people believe in free will, and once you explain why it is that humans believe in free will, not even a hint of the problem remains. The difference being that your dissolution of free will to cognitive algorithm didn't (as I recall) cite any of the relevant science, whereas Talbot's (and others') dissolutions to cognitive algorithms do cite the relevant science.
Is there somewhere where you explain the difference between what Talbot, and also Kip Werking, have done versus what you think is so special and important about LW-style philosophy?
As for the others: Yeah, we seem to agree that useful work does sometimes come from philosophy, but that it mostly doesn't, and people are better off reading statistics and AI and cognitive science, like I said. So I'm not sure there's anything left to argue.
The one major thing I'd like clarification on (if you can find the time) is the difference between what experimental philosophers are doing (or what Joshua Greene is doing) and the dissolution-to-algorithm that you consider so central to LW-style philosophy.
I'd like to emphasize, to no one in particular, that the evaluation that seems to be going on here is about whether or not reading these philosophers is useful for building a Friendly recursively self-improving artificial intelligence. While thats a good criteria for whether or not Eliezer should read them, failure to meet this criteria doesn't render the work of the philosopher valueless (really! it doesn't!). The question "is philosophy helpful for researching AI" is not the same as the question "is philosophy helpful for a rational person trying to better understand the world".
Tarski did philosophical work on truth? Apart from his mathematical logic work on truth? Haven't read it if so.
What does Talbot say about a cognitive algorithm generating the appearance of free will? Is it one of the cognitive algorithms referenced in the LW dissolution or a different one? Does Talbot talk about labeling possibilities as reachable? About causal models with separate nodes for self and physics? Can you please take a moment to be specific about this?
Okay, now you're just drawing lines around what you don't like and calling everything in that box philosophy.
Should we just hold a draft? With the first pick the philosophers select... Judea Pearl! What? whats that? The mathematicians have just grabbed Alfred Tarski from right under the noses the of the philosophers!
To philosophers, Tarski's work on truth is considered one of the triumphs of 20th century philosophy. But that sort of thing is typical of analytic and especially naturalistic philosophy (including your own philosophy): the lines between mathematics and science and philosophy are pretty fuzzy.
Talbot's paper isn't about free will (though others in experimental philosophy are); it's about the cognitive mechanisms that produce intuitions in general. But anyway this is the post I'm drafting right now, so I'll be happy to pick up the conversation once I've posted it. I might do a post on experimental philosophy and free will, too.
Yet to Wikipedia, Tarski is a mathematician. Period. Philosophy is not mentioned.
It is true that mathematical logic can be considered as a joint construction by philosophers and mathematicians. Frege, Russell, and Godel are all listed in Wikipedia as both mathematicians and philosophers. So are a couple of modern contributors to logic - Dana Scott and Per Martin-Lof. But just about everyone else who made major contributions to mathematical logic - Peano, Cantor, Hilbert, Zermelo, Skolem, von Neumann, Gentzen, Church, Turing, Komolgorov, Kleene, Robinson, Curry, Cohen, Lawvere, and Girard are listed as mathematicians, not philosophers. To my knowledge, the only pure philosopher who has made a contribution to logic at the level of these people is Kripke, and I'm not sure that should count (because the bulk of his contribution was done before he got to college and picked philosophy as a major. :)
Quine, incidentally, made a minor contribution to mathematical logic with his idea of 'stratified' formulas in his 'New Foundations' version of set theory. Unfortunately, Quine's theory was found to be inconsistent. But a few decades later, a fix was discovered and today some of the most interesting Computer Science work on higher-order logic uses a variant of Quine's idea to avoid Girard's paradox.
Wikipedia is not authoritative (and recognizes this explicitly - hence the need to give citations). Here is a quote from Tarski himself:
This sort of thing is less a fact about the world and more an artifact of the epistemological bias in English Wikipedia's wording and application of its verifiability rules. en:wp's way of thinking started at computer technology - as far as I can tell, the first field in which Wikipedia was the most useful encyclopedia - and went in concentric circles out from there (comp sci, maths, physics, the other sciences); work in the humanities less than a hundred or so years old gets screwed over regularly. This is because the verifiability rules have to more or less compress a degree's worth of training in sifting through human-generated evidence into a few quickly-comprehensible paragraphs, which are then overly misapplied by teenage science geek rulebots who have an "ugh" reaction to fuzzy subjects.
This is admittedly a bit of an overgeneralisation, but this sort of thing is actually a serious problem with Wikipedia's coverage of the humanities. (Which I'm currently researching with the assistance of upset academics in the area in order to make a suitable amount of targeted fuss about.)
tl;dr: that's stronger evidence of how Wikipedia works than of how the world works.
I believe Carnap is also primarily listed as a philosopher in wikipiedia, and he certainly counts as a major contributor to modern logic (although, of course, much of his work relates to mathamatics as well).
Quine's set theory NF has not been shown to be inconsistent. Neither has it been proven consistent, even relative to large cardinals. This is actually a famous open problem (by the standards of set theory...)
However, NFU (New Foundations with Urelements) is consistent relative to ZF.
I thought Chalmers was a newbie to all this - and showed it quite a bit. However, a definite step forward from zombies. Next, see if Penrose or Searle can be recruited.
No, it's more than that, but only things of that level are useful philosophy. Other things are not philosophy or more like background intros.
Amy just arrived and I've got to start book-writing, but I'll take one example from this list, the first one, so that I'm not picking and choosing; later if I've got a moment I'll do some others, in the order listed.
Funny you should mention that.
There is this incredibly toxic view of predicate logic that I first encountered in Good Old-Fashioned AI. And then this entirely different, highly useful and precise view of the uses and bounds of logic that I encountered when I started studying mathematical logic and learned about things like model theory.
Now considering that philosophers of the sort I inveighed against in "against modal logic" seem to talk and think like the GOFAI people and not like the model-theoretic people, I'm guessing that the GOFAI people made the terrible, horrible, no good, very bad mistake of getting their views of logic from the descendants of Bertrand Russell who still called themselves "philosophers" instead of those descendants who considered themselves part of the thriving edifice of mathematics.
Anyway. If you and I agree that philosophy is an extremely sick field, that there is no standardized repository of the good stuff, that it would be a desperate and terrible mistake for anyone to start their life studying philosophy before they had learned a lot of cognitive science and math and AI algorithms and plain old material science as explained by non-philosophers, and that it's not worth my time to read through philosophy to pick out the good stuff even if there are a few small nuggets of goodness or competent people buried here and there, then I'm not sure we disagree on much - except this post sort of did seem to suggest that people ought to run out and read philosophy-qua-philosophy as written by professional philosophers, rather than this being a terrible mistake.
Will try to get to some of the other items, in order, later.
You may enjoy the following exchange between two philosophers and one mathematician.
Bertrand Russell, speaking of Godel's incompleteness theorem, wrote:
Wittgenstein dismissed the theorem as trickery:
Godel replied:
According to Gleick (in The Information), the only person who understood Godel's theorem when Godel first presented it was another mathematician, Neumann Janos, who moved to the USA and began presenting it wherever he went, by then calling himself John von Neumann.
The soundtrack for Godel's incompleteness theorem should be, I think, the last couple minutes of 'Ludus' from Tabula Rasa by Arvo Part.
I've been wondering why von Neumann didn't do much work in the foundations of mathematics. (It seems like something he should have been very interested in.) Your comment made me do some searching. It turns out:
ETA: Am I the only one who fantasizes about cloning a few dozen individuals from von Neumann's DNA, teaching them rationality, and setting them to work on FAI? There must be some Everett branches where that is being done, right?
We'd need to inoculate the clones against vanity, it appears.
Interesting story. Thanks for sharing your findings.
Well spoken! :)
Of course, since this is a community blog, we can have it both ways. Those of us interested in philosophy can go out and read (and/or write) lots of it, and we'll chuck the good stuff this way. No need for anyone to miss out.
Exactly. Like I did with my statistical prediction rules post.
Yeah, we don't disagree much on all those points.
I didn't say in my original post that people should run out and start reading mainstream philosophy. If that's what people got from it, then I'll add some clarifications to my original post.
Instead, I said that mainstream philosophy has some useful things to offer, and shouldn't be ignored. Which I think you agree with if you've benefited from the work of Bostrom and Dennett (including, via Drescher) and so on. But maybe you still disagree with it, for reasons that are forthcoming in your response to my other examples of mainstream philosophy contributions useful to Less Wrong.
But yeah, don't let me keep you from your book!
As for predicate logic, I'll have to take your word on that. I'll 'downgrade it' in my list above.
FWIW, what I got from your original post was not "LW readers should all go out and start reading mainstream philosophy," but rather "LW is part of a mainstream philosophical lineage, whether its members want to acknowledge that or not."
Thanks for sharing. That too. :)
I'm part of Roger Bacon's lineage too, and not ashamed of it either, but time passes and things improve and then there's not much point in looking back.
Meh. Historical context can help put things in perspective. You've done that plenty of times in your own posts on Less Wrong. Again, you seem to be holding my post to a different standard of usefulness than your own posts. But like I said, I don't recommend anybody actually read Quine.
Oftentimes you simply can't understand what some theorem or experiment was for without at least knowing about its historical context. Take something as basic as calculus: if you've never heard the slightest thing about classical mechanics, what possible meaning could a derivative, integral, or differential equation have to you?
Does human nature improve, too?
What's "human nature"?
Something that probably hasn't changed much over the history of philosophy.
I'd be curious to know what that "toxic view" was. My GOFAI academic advisor back in grad school swore by predicate logic. The only argument against that I ever heard was that proving or disproving something is undecidable (in theory) and frequently intractible (in practice).
Model theory as opposed to proof theory? What is it you think is great about model theory?
I have no idea what you are saying here. That "Against Modal Logic" posting, and some of your commentary following it strike me as one of your most bizarre and incomprehensible pieces of writing at OB. Looking at the karma and comments suggests that I am not alone in this assessment.
Somehow, you have picked up a very strange notion of what modal logic is all about. The whole field of hardware and software verification is based on modal logics. Modal logics largely solve the undecidability and intractability problems the bedeviled GOFAI approaches to these problems using predicate logic. Temporal logics are modal. Epistemic and game-theoretic logics are modal.
Or maybe it is just the philosophical approaches to modal logic that offended you. The classical modal logic of necessity and possibility. The puzzles over the Barcan formulas when you try to combine modality and quantification. Or maybe something bizarre involving zombies or Goedel/Anselm ontological proofs.
Whatever it was that poisoned your mind against modal logic, I hope it isn't contagious. Modal logic is something that everyone should be exposed to, if they are exposed to logic at all. A classic introductory text: Robert Goldblatt: Logics of Time and Computation (pdf) is now available free online. I just got the current standard text from the library. It - Blackburn et al.: Modal Logic (textbook) - is also very good. And the standard reference work - Blackburn et al.: Handbook of Modal Logic - is outstanding (and available for less than $150 as Borders continues to go out of business :)
Reading Plantinga could poison almost anybody's opinion of modal logic. :)
That is entirely possible. A five star review at the Amazon link you provided calls this "The classic work on the metaphysics of modality". Another review there says:
Yet among the literally thousands of references in the three books I linked, Platinga is not even mentioned. A fact which pretty much demonstrates that modal logic has left mainstream philosophy behind. Modal logic (in the sense I am promoting) is a branch of logic, not a branch of metaphysics.
I'd very much like to see a post explaining that.
I'm not sure what "of that level" (of dissolving-to-algorithm) means, but I think I've demonstrated that quite a lot of useful stuff comes from mainstream philosophy, and indeed that a lot of mainstream philosophy is already being used by yourself and Less Wrong.