Followup toThe Planning Fallacy

Plato's Phaedo:

    "The state of sleep is opposed to the state of waking; and out of sleeping, waking is generated; and out of waking, sleeping; and the process of generation is in the one case falling asleep, and in the other waking up.  Do you agree?"
    "Quite."
    "Then suppose that you analyze life and death to me in the same manner.  Is not death opposed to life?"
    "Yes."
    "And they are generated one from the other?"
    "Yes."
    "What is generated from life?"
    "Death."
    "And what from death?"
    "I can only say in answer - life."
    "Then the living, whether things or persons, Cebes, are generated from the dead?"
    "That is clear."
    "Then our souls exist in the house of Hades."
    "It seems so."

Now suppose that the foil in the dialogue had objected a bit more strongly, and also that Plato himself had known about the standard research on the Inside View vs. Outside View...

(As I disapprove of Plato's use of Socrates as his character mouthpiece, I shall let one of the characters be Plato; and the other... let's call him "Phaecrinon".)

Plato:  "The state of sleep is opposed to the state of waking; and out of sleeping, waking is generated; and out of waking, sleeping; and the process of generation is in the one case falling asleep, and in the other waking up...  Then suppose that you analyze life and death to me in the same manner."

Phaecrinon:  "Why should I?  They are different things."

Plato:  "Oh, Phaecrinon, have you not heard what researchers have shown, that the outside view is a better predictor than the inside view? You come to me and point out the differences between life-death and wake-sleep, all so that you can avoid making the obvious generalization that you prefer to deny.  Yet if we allow such reasoning as this, will not software project managers say, 'My project is different from yours, because I have better programmers'?  And will not textbook authors say, "We are wiser than those other textbook authors, and therefore we will finish sooner'?  Therefore you can see that to point out the similarities between things is superior, and to point out the differences between them, inferior."

Phaecrinon:  "You say that your reasoning is like to the reasoning of Daniel Kahneman, yet it seems to me that they are importantly different.  For Daniel Kahneman dealt with generalization over things that are almost quite as similar to each other, as different flippings of the same coin.  Yet you deal with wholly different processes with different internal mechanisms, and try to generalize across one to the other."

Plato:  "But Phaecrinon, now you only compound your error; for you have pointed out the difference between myself and Kahneman, where I have pointed out the similarity.  And this is again inferior, by reason of the inferiority of the Inside View over the Outside View.  You have only given me one more special reason why the Outside View should not apply to your particular case - all so that you can deny that our souls exist in the house of Hades."

Phaecrinon:  "Yet Plato, if you propose indiscriminately to apply the Outside View to all things, how do you explain the ability of engineers to construct a new bridge that is unlike any other bridge, and calculate its properties in advance by computer simulation?  How can the Wright Flyer fly, when all previous human-made flying machines had failed?  How indeed can anything at all happen for the first time?"

Plato:  "Perhaps sometimes things do happen for the first time, but this does not mean we can predict them."

Phaecrinon:  "Ah, Plato, you do too little justice to engineers. Out of all the possible structures of metal and tubes and explosive fuel, very few such structures constitute a spaceship that will land on the Moon.  To land on the Moon for the first time, then, human engineers must have known, in advance, which of many designs would have the exceedingly rare property of landing upon the Moon.  And is this not the very activity that engineers perform - calculating questions in detail?  Do not engineers take the Inside View with great success?"

Plato:  "But they assume that each screw and part will behave just like it does on all the other times observed."

Phaecrinon:  "That is so.  Yet nonetheless they construct detailed internal models of exceeding complexity, and do not only collect the statistics of whole cases.  This is the Inside View if anything is the Inside View; no one claims that Inside Views are generated purely from nothingness."

Plato:  "Then I answer that when engineers have shown many times their ability to perform detailed calculations with success, we trust on future occasions that they will succeed similarly.  We trust the Inside View when the Outside View tells us to trust the Inside View.  But if this is not so, and there is not a past record of success with detailed calculations, then the Outside View is all that is left to us; and we should foresake all attempts at internal modeling, for they will only lead us astray."

Phaecrinon:  "But now you have admitted that the notion of 'trust Outside View, distrust Inside View' has a limited domain of applicability, and we may as well restrict that domain further.  Just as you try to seal off the successes of engineers from the Outside View, so too, I wish to seal off the failures of Greek philosophers from the Outside View.  Specifically, the record of Greek philosophers does not inspire in me any confidence that the Outside View can be applied across processes with greatly different internal causal structures, like life-and-death versus sleeping-and-waking.  Daniel Kahneman and his fellows, writing a textbook, encountered a challenge drawn from a structurally similar causal generator as many other cases of textbook-writing; subject to just the same sort of unforeseen delays.  Likewise the students who failed to predict when they would finish their Christmas shopping; the task of Christmas shopping does not change so much from one Christmas to another.  It would be another matter entirely to say, 'Each year I have finished my Christmas shopping one day before Christmas - therefore I expect to finish my textbook one day before my deadline.'"

Plato:  "But this only sounds foolish, because we know from the Outside View that textbooks are delayed far longer than this. Perhaps if you had never written a textbook before, and neither had anyone else, 'one day before deadline' would be the most reasonable estimate."

Phaecrinon:  "You would not allow me to predict in advance that textbook writing is more difficult than Christmas shopping?"

Plato:  "No.  For you have chosen this particular special plea, using your hindsight of the correct answer.  If you had truly needed to write a textbook for the first time in history, you would have pled, 'No one can foresee driving delays and crowds in the store, but the work I do to write a textbook is all under my own control - therefore, I will finish more than one day before deadline.'"

Phaecrinon:  "But even you admit that to draw analogies across wider and wider differences is to make those analogies less and less reliable.  If you see many different humans sleeping, you can conclude that a newly observed human will probably sleep for eight hours; but if you see a cat sleeping, you must be less confident; and if you wish to draw an analogy to life and death, that is a greater distance still."

Plato:  "If I allow that, will not software project managers say, 'My software project is as unlike to all other software projects as is a cat to a human?'"

Phaecrinon:  "Then they are fools and nothing can be done about it.  Surely you do not think that the prediction from many humans to one cat is just as strong as the prediction from many humans to one human?  Insensitivity to the reliability of predictors is also a standard bias."

Plato:  "That is true.  Yet an Outside View may not be a good estimate, and yet still be the best estimate.  If we have only seen the sleep cycles of many humans, then the Outside View on the whole group may be the best estimate for a newly observed cat, if you have no other data.  Even likewise with our guesses as to life and death."

Phaecrinon:  "And one sign of when the Outside View might not provide a good estimate, is when there are many different reference classes to which you might compare your new thing.  A candle burns low, and exhausts itself and extinguishes, and does not light again the next day. How do you know that life and death is not analogous to a candle which burns and fails?  Why not generalize over the similarity to a candle, rather than the similarity to sleep cycles?"

Plato:  "Oh, but Phaecrinon, if we allow arguments over reference classes, we may as well toss the notion of an Outside View out the window.  For then software project managers will say that the proper reference class for their project is the class of projects that delivered on time, or the class of projects with managers as wise as themselves.  As for your analogy of the candle, it is self-evident that life is similar to sleeping and waking, not to candles.  When a man is born, he is weak, but he grows to adulthood and is strong, and then with age he grows weaker again.  In this he is like the Sun, that is weak when it rises, and strong at its apex, and then sinks below the horizon; in this a man is like the sleepy riser, who becomes sleepy again at the end of the day.  It is self-evident that life and death belongs to the class of cyclical processes, and not the class of irreversible processes."

Phaecrinon:  "What is self-evident to you does not seem so self-evident to me, Plato; and just because you call several widely different things 'cyclical processes', it does not follow that they were all random samples drawn from a great Barrel of Cyclical Processes, and that the next thing you choose to call a 'cyclical process' will have the same distribution of properties."

Plato:  "Again you compound your mistake by pleading special exceptions.  Will we let the software manager plead that his project is not drawn from the same barrel as the others?"

Phaecrinon:  "Again you extend the Outside View beyond its domain of applicability.  In engineering where all internal parts are precisely understood, but the whole is not quite similar to anything else that has been built before, then the Inside View is superior to the Outside View.  And the sign of this is that results are routinely predicted in advance with great precision."

Plato:  "This is just to say that when the Outside View tells us to use the Inside View, we should use it.  But surely not otherwise, Phaecrinon!"

Phaecrinon:  "When many different people try to accomplish the same task, and the internal details cannot be precisely calculated, and yet people have a tendency to optimism and to not visualize incidental catastrophes, then the Outside View is superior to the Inside View.  And the sign of this is that the same kind of task - with the same sort of internal structure, the same difficulties and challenges - has been done many times, and the result cannot be predicted with precision; yet people's predictions are usually biased optimistically."

Plato:  "This is the triumph of the Outside View!"

Phaecrinon:  "But when you deal with attempted analogies across structually different processes, perhaps unique or poorly understood, then things which are similar in some surface respects are often different in other respects.  And the sign of this domain is that when people try to reason by similarity, it is not at all clear what is similar to what, or which surface resemblances they should focus upon as opposed to others."

Plato:  "I think the resemblance of life-death to sleep-waking is perfectly clear.  But what do you assert we should do in such a case, if it is not taking the Outside View?"

Phaecrinon:  "Perhaps there is nothing to be done at all, with either the Inside View or the Outside View.  Not all problems are solvable; and it may be that the best we can do is avoid the overconfidence from asserting that analogies are much stronger than they are.  But it seems to me that in those cases where we know something of the internal structure, then we can sometimes produce predictions by imagining the internals, even though the whole thing is not similar to any other whole thing in our experience."

Plato:  "Now I challenge you to consider how well such thoughts have done, historically."

Phaecrinon:  "What I have just described is the way that engineers build the first prototype of anything.  But that, I admit, is when they understand very precisely the parts they use.  If the internals are not well-understood, then the whole will in most cases be even less well-understood.  It is only your idea that the Outside View can yield better predictions, that I am protesting against.  It seems to me that the result of taking the Outside View of things poorly understood or structurally dissimilar to other things in the purported reference class, is only to create great disputes about definitional boundaries, and clashing analogies, and arguments over which surface similarities are important.  When all along the new process may not be similar to anything that already exists."

Plato:  "But there is no alternative to the Outside View."

Phaecrinon:  "Yes, there is; you can try to imagine the internal process, if you know anything at all about it.  At least then two people can focus on the internal structure and argue about what happens and their dispute will be commensurable.  But if two people both say 'I am taking the Outside View' and then form different 'self-evident' reference classes, what do they do from there?  How can they resolve their dispute about which surface characteristics are important?  At least if you make predictions about internal causal processes, the results are finally testable if the dispute is empirical at all.  How do you test the assertion that life is more importantly similar to sleep and waking, than to a candle?  Perhaps life is simply like neither.  Something must happen internally when a human thinks and reasons, but there does not have to exist any other process in nature similar enough that we could predict the characteristics of human thought by looking at it."

Plato:  "What it boils down to, is that you are constructing a detailed excuse not to use the Outside View in your own case."

Phaecrinon:  "And if each of two people with different Outside Views says to the other, 'You are a fool, for disregarding the Outside View!' then they will make no progress on their disagreement at all.  This is the danger of proposing an absolute mandate for philosophers encountering new and structurally different phenomena, because you want to prevent software project managers from making special excuses for their software project.  Reversed stupidity is not intelligence, and there is no language in which it is difficult to write bad computer programs, and in the art of rationality it is never difficult to shoot off your own foot if you desire to do so.  The standard Outside View relies on your seeing the common-sense difference between textbook writing and Christmas shopping, so that you don't try to lump them into the same reference class.  I am similarly hoping that you can see by common sense that the Outside View works rather better to predict Christmas shopping times, than what you are arguing is the analogous 'Outside View' technique in philosophy."

Plato:  "And you believe you can do better with the Inside View."

Phaecrinon:  "Reasoning about the internals of things whose output is not yet observed, is fraught with difficulty.  One must be constantly aware of what one can and cannot reasonably guess, based on the strength of your knowledge.  The uncertainties of such an Inside View, end up being much greater than the uncertainties of the Outside View on Christmas shopping.  Only when the Inside View support appears extremely lopsided can you dare to come to even a tentative conclusion!  But I do think that sometimes the Inside View support can be extremely lopsided - though it is a strain on your rationality even to correctly distinguish such cases."

Plato:  "The evidence shows that people cannot successfully use the Inside View at all."

Phaecrinon:  "No, the evidence shows that the Outside View yields better answers than the Inside View for problems like writing a textbook.  But even an Inside View of writing a textbook would tell you that the project was unlikely to destroy the Earth.  Taking the Inside View of a new and strange process is a Difficult Problem, where taking the Outside View on textbook composition is a Straightforward Problem.  But to try and argue like alchemists from surface resemblances is a Hopeless Problem.  Then there cannot even be any meeting of minds, if you start with different assumptions about which similarities are important.  An answer need not exist even in principle, for there may be nothing else that is enough like this new thing to yield successful predictions by analogy."

Plato:  "So you have said that it is easier for two people to conduct their dispute if they both take the Inside View and argue about internal causal processes.  But from this it does not follow that the Outside View based on surface resemblances is inferior.  Perhaps you are only coming to agreement on folly, and either of two conflicting Outside Views would be more reliable than the best Inside View."

New to LessWrong?

New Comment
14 comments, sorted by Click to highlight new comments since: Today at 7:44 AM

The question of whether trying to consistently adopt meta-reasoning position A will raise the percentage of time you're correct, compared with meta-reasoning position B, is often a difficult one.

When someone uses a disliked heuristic to produce a wrong result, the temptation is to pronounce the heuristic "toxic". When someone uses a favored heuristic to produce a wrong result, the temptation is to shrug and say "there is no safe harbor for a rationalist" or "such a person is biased, stupid, and beyond help; he would have gotten to the wrong conclusion anyway, no matter what his meta-reasoning position was. The idiot reasoner, rather than my beautiful heuristic, has to be discarded." In the absence of hard data, consensus seems difficult; the problem is exacerbated when a novel meta-reasoning argument is brought up in the middle of a debate on a separate disagreement, in which case the opposing sides have even more temptation to "dig in" to separate meta-reasoning positions.

Phaecrinon: But even an Inside View of writing a textbook would tell you that the project was unlikely to destroy the Earth.

Eric Drexler might have something to say about that, along with one or two twentieth century physicists.

Good post nonetheless :)

The implied disagreement here between the "inside view" of "outside views" (i.e. a limited domain) and the "outside view" of "outside views" (i.e. something that applies in general) is the same as Eliezer's disagreement with Robin about the meaning of Aumann.

If Robin is right, then Eliezer is against overcoming bias in principle, since this would be taking an outside view (according to Robin's understanding). Of course, if Eliezer is right, it just means that Robin is biased against inside views. Each of these consequences is very strange; if Robin is right, Eliezer is in favor of bias despite posting on a blog on overcoming bias, while if Eliezer is right, Robin is biased against his own positions, among other things.

Unknown: If Robin is right, then Eliezer is against overcoming bias in principle, since this would be taking an outside view (according to Robin's understanding)

I thought that overcoming bias was about reaching true beliefs, not adhering to some particular ritual like using outside views.

Bayes' Theorem says that P(H|DI) is proportional to p(H|I) * p(D|HI), where D is the data, H is the hypothesis, and I is our background information. A common mistaken intuition is that P(H|DI) = p(D|HI). The "outside view" seems to be that p(H|DI) = p(H|I). This is just plain wrong. As Eli states above, by varying your choice of D' and I' such that D'I' = DI, you can make p(H|D'I') equal to all sorts of things by using the outside view heuristic. Getting good predictions from the outside view requires you to have good intuitions about what to use as D and what to use as I.

Having useful approximations is great. Turning a useful approximation into your new definition of rationality is a bad, bad idea. You say that Robin thinks that overcoming bias means taking outside views. I think Robin knows better than that.

Peter de Blanc: see http://www.overcomingbias.com/2007/07/beware-the-insi.html, posted by Robin Hanson. In particular : "Most, perhaps all, ways to overcome bias seem like this. In the language of Kahneman and Lovallo's classic '93 paper, we allow an outside view to overrule an inside view... If overcoming bias comes down to having an outside view overrule an inside view, then our questions become: what are valid outside views, and what will motivate us to apply them?"

What do you think this means, if not that overcoming bias means taking outside views?

There are a variety of things we can take the inside view or outside view with regards to AI.

  • How powerful AI can be.
  • How much progress an individual or small team of humans will make towards making strong AI.

The second question seems appropriate for the outside view, the first less so in my view. The outside answer to the second question seems to be very little.

Inside and outside views are different levels of description of the problem, and reasoning always happens on many levels, with analogies drawn between levels and with feedback between levels. The failure of rationality occurs with particular use of representation where you have your inbuilt overconfidence pull in the wrong direction. By shifting to a different level of description that doesn't include the faulty parts, you restore rationality of the conclusion. In other cases, you don't need to do that, and the success depends primarily on having sufficient information about the domain and correct estimates of similarities of the parts of the problems, on all considered levels of description.

Are there no social scientists left reading this blog anymore, to comment on the implicit accusation that analyses of social transitions are just "surface" analogies no more trustworthy that Plato's analogy of death to sleep?

If we are trading in analogies, then I'd liken Robin's argument to noticing that the shoe size of the last four Miss World winners decreased each year, and then predicting the shoe size of next year's winner on that basis. Also it turns out that his measure of shoe size is based on measuring the shoe size of other members of the participant's country - and that he skipped over the 2006 competition for some reason.

Are the judges really preferring smaller feet? Maybe, but the evidence seems rather tenuous.

Robin, questioning the analogy from biology to postbiology, and questioning the extension of a trend in interest rates over the leap to transistor-based thought, is not the same as dismissing the whole of economics!

Eliezer, you describe the economics concepts I use to compare past and future as "surface analogies" and not "deep causes" and give that as your reason for not thinking much of the analysis. That is surely some flavor of dismissal of those concepts.

In accordance with the conjunction rule of probability theory, a deep principle times a surface analogy equals a surface analogy. The principles you've described are deep for analyzing human economies, it's the analogy over to the posthuman side that I have trouble with.

I'll see if I can write a post addressing this specific topic later today, though it's going to be way out of order in the sequence.

May I suggest that Plato's words carry some different and non-obvious sensibility, that has little to do with the outside vs inside, if we take the original text and the circumstances into account? For, in that age, people had fewer reasons to believe the physicality of the individual. They saw dead people remain dead, but that's pretty much all of it. And they had more motives to believe in the soul, because there's no scientific transhumanism, and religion was their only hope of personal immortality. So the introspecting self may feel that just as it has slept and awaken, and remained itself, so it is possible to survive an abscence of consciousness, and death must be temporary. Not because they superficially looked and sounded alike, but because of the common factor of lack of consciousness, which seemed much less distinguishable than they do now in light of neuroscience. It's not outside vs inside, it might as well have been Phaecrinon thinking "the two pairs are structually different" and Plato thinking "they are equivalent and symmetrical".

Plato has dismissed his share of strawmen opponents, and I have no problem with adapting his words, but I feel confused by this post's focus on the principles of thinking when this more obvious reaction to the analogy comes to mind. How about choosing a purer example next time?