To what extent can we predict the future? With Robin Hanson and David Deutsch

00:00:00.000 This is Daniel Martin.
00:00:00.920 I'm here with Robin Hanson and David Deutsch.
00:00:03.120 Yesterday I was asked if I'd like to help moderate a conversation.
00:00:05.240 So here we are, have fun.
00:00:06.960 And Robin, do you have a question for David?
00:00:09.800 David, it's great to talk to you.
00:00:11.840 I like you.
00:00:12.680 I'm a big fan.
00:00:14.000 Thank you.
00:00:14.840 I've spent a lot of my career doing various kinds
00:00:17.440 of future forecasting.
00:00:19.400 I hear you're not entirely a fan of all of that.
00:00:22.520 So I wanted to ask more specifically
00:00:24.560 which kinds of forecasting you do or don't like
00:00:27.240 and just to start with very concrete examples.
00:00:30.360 So for example, solar power costs have been falling steadily
00:00:35.120 and there's a standard power law relationship
00:00:37.240 that predicts that fall of costs.
00:00:39.640 And in fact surveys of technology show
00:00:42.120 that more than half of technologies do follow
00:00:44.160 that power law of cost as a function of the quantity produced.
00:00:48.080 I presume you're okay with that kind of forecasting.
00:00:50.960 Not entirely okay, actually,
00:00:53.160 because you seem to be, well, when you say half of all,
00:00:59.720 whatever you said, innovations or whatever.
00:01:01.920 Technology.
00:01:02.760 It's all over the forecast.
00:01:05.200 You're kind of implicitly assuming that's a frequency,
00:01:07.880 half of them.
00:01:09.280 And you're kind of assuming that that's equal to a probability.
00:01:12.480 Namely, it's going to be, there's a probability of one half
00:01:19.640 that in the particular innovation
00:01:22.760 we're considering like solar power,
00:01:25.400 solar power or whatever will follow that law.
00:01:28.960 And that's the kind of inference I think isn't valid.
00:01:33.400 However, I think part of the reason,
00:01:37.440 I think it may not be as bad as it sounds
00:01:40.720 because part of the reason why people,
00:01:45.720 for example, confuse frequency and probability
00:01:48.640 in their language
00:01:51.600 and why they express forecasts in terms of probability
00:01:57.840 and then use frequency to judge that
00:02:00.400 is that there's an unspoken assumption there,
00:02:05.040 which is also about the technology.
00:02:07.680 Like you know that, for example,
00:02:12.280 the maximum distance that a human can travel from Earth
00:02:17.440 as a function of time
00:02:19.160 is may have followed a certain law
00:02:23.640 until we reach the edge of the atmosphere
00:02:25.920 and then followed a completely different law
00:02:28.920 between the Earth and the Moon
00:02:30.800 and then we'll follow a completely different law again
00:02:33.640 from the Moon to Mars.
00:02:36.160 And when you talk about solar power,
00:02:39.880 you are implicitly referring not only to the statistics
00:02:44.920 but to a mental model that you have,
00:02:46.760 or theory, or explanatory theory that you have
00:02:49.960 about that specific technology.
00:02:52.720 And although the answer you get
00:02:57.320 is entirely derived from that explanatory theory
00:03:02.160 and not from the statistics,
00:03:05.280 it doesn't matter if they match well.
00:03:09.000 And for example, in the case of solar power,
00:03:13.680 I think it's correct,
00:03:14.640 although that power law will eventually stop.
00:03:18.520 And it may stop before the laws of physics say so
00:03:22.320 if there are contingent factors that stop it.
00:03:28.160 So, I mean, I think you are right
00:03:30.600 that we'd like to talk in terms of frequencies
00:03:33.440 because it's a safer thing to say.
00:03:35.120 And we are implicitly trying to get someone
00:03:37.640 to think in terms of probabilities
00:03:40.680 as the thing we're trying to get in their head.
00:03:42.840 But because it's harder to justify a claim
00:03:45.360 about a probability than a frequency,
00:03:46.960 we speak in terms of frequencies.
00:03:49.720 And of course, whenever we see a trend
00:03:52.040 like the solar power trend,
00:03:53.320 we can't be overwhelmingly confident
00:03:56.320 that it will continue forever.
00:03:57.800 And we should have a substantial expectation
00:04:01.160 that it might deviate soon.
00:04:04.200 But nevertheless, we might say,
00:04:05.960 well, look, if the standard here is just better than chance,
00:04:09.480 if we're just trying to understand the future,
00:04:11.240 surely we have to do things like this,
00:04:13.160 like find the kind of trends that technologies follow
00:04:15.680 and pick the one that's most common.
00:04:17.920 And if it seems to fit the trajectory so far,
00:04:19.880 surely we'd want to do something like projecting that forward.
00:04:22.640 We want to have some degree of uncertainty
00:04:24.600 about that projection,
00:04:25.840 but isn't it a reasonable thing to do to take our best fit
00:04:29.440 to test data and then try to use it
00:04:31.600 to project future data?
00:04:33.000 I mean, I think it's a reasonable thing to do
00:04:36.680 if our best explanations imply that.
00:04:41.400 So I think the futureologists of the 1960s
00:04:47.200 had a surprise free forecast.
00:04:54.200 And a surprise free forecast is them saying,
00:04:57.680 don't blame us if we turn out,
00:04:59.280 if this turns out to be false,
00:05:00.840 because we will always be able to look back and say,
00:05:02.960 ah, this thing happened and it wasn't in our predictions,
00:05:07.240 but we're surprised by it.
00:05:08.960 Well, a surprise, again,
00:05:12.280 they're referring to explanations there.
00:05:14.560 A surprise free, a surprising outcome
00:05:18.560 is one that your best explanation didn't foretell.
00:05:23.560 So a surprise free one is what will happen
00:05:27.800 if your current explanation is true.
00:05:30.520 But then everything hangs on your current explanation.
00:05:36.480 But nevertheless, surely,
00:05:38.160 I mean, we do look at say solar power
00:05:40.160 and it does look similar to many other kinds of technologies.
00:05:43.360 And we are implicitly using our sense
00:05:45.160 of what the nature of solar power is compared
00:05:47.680 to other kinds of say electronic technologies
00:05:49.680 to compare it to say electronic technologies
00:05:52.160 or materials technologies.
00:05:53.920 And then when we, and then we are looking at sort of rates
00:05:57.040 of patterns that we've seen in those,
00:05:58.800 and we are using many things we know
00:06:01.280 to make this judgment that say solar power
00:06:03.440 would probably follow this power law,
00:06:05.080 at least for a while.
00:06:06.320 Nevertheless, isn't it a reasonable thing
00:06:08.040 to do all of that taking everything we know
00:06:10.400 and using that to make predictions
00:06:12.240 about the future solar power?
00:06:14.160 I, I mean, I don't know that much
00:06:17.320 about the details of solar power,
00:06:19.120 but I, I'm assuming that that is a good assumption.
00:06:23.000 But you're, you're making an assumption
00:06:25.680 there about a very complex and detailed thing.
00:06:30.400 When you say this is like other technologies,
00:06:33.320 you mean not only that the graphs look the same,
00:06:37.960 you mean that you expect solar power technology
00:06:43.400 to be dependent on making factories,
00:06:46.480 which use materials of a certain kind,
00:06:51.000 which aren't going to be throttled
00:06:53.160 by a hostile foreign power and so on.
00:06:57.560 And it's not going to be something that you haven't foreseen,
00:07:02.960 like for example, if you were,
00:07:04.600 if you were predicting the future of power stations in 1900,
00:07:09.160 you wouldn't be expecting nuclear power stations.
00:07:12.840 So your forecast based on the best existing knowledge
00:07:18.600 would not be just out by a factor.
00:07:22.800 It would be out by and having the wrong explanation,
00:07:28.040 which would make the real things behave in a different way,
00:07:32.800 not just be quantitatively different.
00:07:37.080 And then as I, as I've, I've, I've written somewhere,
00:07:40.840 I can't, was it in the beginning of infinity?
00:07:43.360 I can't remember, but after a radioactive activity
00:07:46.520 was discovered, you could have then made the prediction
00:07:49.840 that, okay, technology will harness that.
00:07:53.320 And then you wouldn't have been able to predict
00:07:55.480 the environmental movement, which will sabotage that.
00:07:59.320 So you would have been completely wrong again about,
00:08:03.240 but yeah, I mean, if the standard here is,
00:08:06.640 can you ever predict things perfectly with certainty?
00:08:09.200 Of course we're going to agree, no.
00:08:11.840 But if the opposite standard is,
00:08:13.600 you can't ever predict anything whatsoever
00:08:15.760 compared to some absolute chance, that's wrong too.
00:08:19.080 So isn't obviously the question,
00:08:21.360 what can you predict how well not can you predict the future?
00:08:25.080 And then we want to go through particular topics
00:08:27.200 and just point out the various degrees to which
00:08:29.680 we might be more or less confident and accurate
00:08:32.960 in any one prediction.
00:08:33.800 So isn't it come down to those details?
00:08:37.280 It comes down to the explanatory details.
00:08:40.360 So like, you know, I like to say when,
00:08:44.240 when people ask me about say Bitcoin,
00:08:47.400 I've talked in terms of whether someone's going to win
00:08:51.240 the fields medal or not for cracking it.
00:08:56.120 And that makes it clear that part of the answer
00:08:59.440 to this question is unknowable.
00:09:02.120 I can't tell whether someone's going to tomorrow
00:09:06.960 discover a piece of mathematics that wins the field medal
00:09:10.280 and invalidates Bitcoin.
00:09:12.800 So if we think about predicting a longer timescale,
00:09:15.840 of course, all else equal, that should be harder.
00:09:18.400 But then if we go to the most solid things we know,
00:09:21.800 the things we're the most confident in,
00:09:23.680 it would seem like that's where we have the best chance
00:09:26.040 of making a longer term predictions.
00:09:28.360 And so since I tend to make longer term predictions
00:09:30.560 and you might be especially wary of those,
00:09:33.720 then it would seem like to focus on these most solid predictions.
00:09:36.680 So one of the most solid predictions
00:09:38.280 I would based on is say the speed of light
00:09:42.160 as a limitation on causal influence.
00:09:45.560 I might say we're limited to our future light code.
00:09:49.240 And that has consequences, substantial consequences
00:09:52.400 for the future.
00:09:53.240 Now, I'm going to admit that of course,
00:09:54.840 yes, that might be wrong.
00:09:55.920 But surely it's reasonable to tentatively guess
00:09:59.000 that maybe that is a fundamental limit
00:10:01.120 and explore those limits,
00:10:02.800 making them explicitly conditional saying,
00:10:04.480 yes, this could be wrong and here's how,
00:10:06.200 but we're pretty sure of the speed of light limits.
00:10:08.680 So let's go with that and analyze that.
00:10:10.840 So for example, I've done some work
00:10:13.760 on predicting where aliens are in space and time.
00:10:16.440 And that's based on this assumption
00:10:18.440 of the light cone limitation.
00:10:20.120 If the light cone limitation weren't there,
00:10:21.640 then basically any alien anywhere in the universe
00:10:24.240 could be here right now because there would be no limits
00:10:27.640 on where the how fast it could get places.
00:10:29.560 And therefore, if there's no aliens,
00:10:31.280 now we would have to predict there's no aliens anywhere
00:10:33.640 because or anywhere in who want to get here
00:10:35.640 because anyone could get here.
00:10:36.760 But with the light cone limit,
00:10:38.360 then we could say, uh-huh,
00:10:39.840 only the aliens in our backward light cone could be here.
00:10:42.800 And therefore, then we can make many more concrete predictions
00:10:47.200 about, you know, the nature of the advanced development
00:10:50.720 of advanced life based on that
00:10:52.120 and that's something that I have done recently.
00:10:54.480 And it seems like again,
00:10:56.240 a reasonable tentative thing to do is to assume
00:10:58.520 that light cone is limited.
00:11:00.400 It's also another basic implication is that, you know,
00:11:03.520 it seems like exponential growth in of things we care about
00:11:07.440 is hard to continue in a light-conlimited world
00:11:11.080 because the light cone only expands the volume
00:11:13.480 at a cube rate and exponential growth requires much faster growth.
00:11:17.200 So isn't it reasonable to say,
00:11:18.960 well, it looks like exponential growth can't go on forever
00:11:22.600 if we're limited to the future light cone.
00:11:24.800 So I think that those kind of speculations
00:11:29.640 or predictions are very legitimate.
00:11:31.760 In fact, we ought to be making them
00:11:35.480 because it is a way of exploring the consequences,
00:11:40.480 the cast iron consequences of our existing best knowledge.
00:11:46.240 And there, I think it makes very little difference
00:11:49.000 to say whether you make the mistake of saying,
00:11:53.520 we're doing this because we're relatively very sure
00:11:56.920 of the speed of light limitation and so on.
00:11:59.520 Or whether you say this prediction is subject
00:12:05.040 to no Nobel Prize being one invalidating it.
00:12:09.720 That those are pretty much the same thing,
00:12:12.640 although I think it's just wrong to say
00:12:14.560 that we're sure of something.
00:12:16.120 We should not believe things.
00:12:19.680 But for some purposes, like in this case, it doesn't matter.
00:12:24.160 But although it is very legitimate
00:12:28.400 to make speculations based on the fineness
00:12:30.800 of the speed of light, I think going on to say
00:12:34.280 that will then have consequences
00:12:37.920 for what we say about our future.
00:12:40.480 That's a bit more, a bit more rickety
00:12:42.960 because we don't know that the limiting factor
00:12:47.320 in our future will be accessibility
00:12:51.160 of stars basically or galaxies in the longer run
00:12:56.160 because we might expand downwards.
00:13:01.520 Like Feynman said, there's plenty of room at the bottom.
00:13:04.600 And there's not an infinite amount of room at the bottom.
00:13:07.760 So then I would invoke thermodynamics
00:13:09.720 and the fact that nagentropy seems to be finite
00:13:12.320 and to say that's going to set a finite limit
00:13:14.880 on the amount of room at the bottom.
00:13:16.840 And that also sets the limit on our future growth
00:13:19.800 in terms of nagentropy because nagentropy is bounded
00:13:22.520 by a finite spatial density of nagentropy
00:13:25.320 and a finite volume by the speed of light.
00:13:29.320 Now you're talking about things that are not as secure
00:13:34.920 as the fineness of the speed of light.
00:13:37.080 For example, there isn't an infinite amount
00:13:39.160 of room at the bottom.
00:13:40.560 Is something that crazy theories of quantum gravity
00:13:43.760 have contradicted?
00:13:45.680 They may not be our best theories,
00:13:47.440 but they're not completely to be true.
00:13:50.520 But they don't predict infinite amounts of nagentropy
00:13:53.200 at the bottom.
00:13:54.040 They predict infinite amounts of potential structure
00:13:56.240 if you add a nagentropy.
00:13:57.440 That is, you could add enough energy to a system
00:14:00.000 then you could create lots of small scale structure.
00:14:02.360 But we do have even within quantum gravity,
00:14:06.040 standard theories of the fineness of nagentropy,
00:14:08.720 say within the volume or within a black hole, et cetera.
00:14:12.640 And then I could appeal to those.
00:14:13.720 But I'm happy to take your point
00:14:15.880 that anyone analysis could have more ways it could be wrong.
00:14:20.880 We could sort of count the ways thing could be wrong.
00:14:23.480 But we'd kind of like a way to weigh all the different ways
00:14:26.320 something could be wrong in some sort of overall measure.
00:14:28.360 And that's what people usually use probability for.
00:14:30.680 So you don't want to use probability
00:14:32.160 that you need some other way to count.
00:14:34.320 But we do roughly want to see for anyone
00:14:37.280 how much it depends on things we might be wrong about
00:14:40.760 and make a judgment about focusing on the things
00:14:43.440 we are going to be the most confident of.
00:14:45.080 And it may not affect the bottom line that much in many cases.
00:14:51.600 But I think it is simply wrong to use probability
00:14:56.400 in that way.
00:14:57.480 And it would be better to make the assumptions explicit
00:15:02.200 rather than say, well, well, it's not going to be,
00:15:07.320 we're not going to supersede the second or the thermodynamics
00:15:11.960 or something.
00:15:12.520 By the way, thermodynamics, at least in one version of it,
00:15:18.480 depends on information theory.
00:15:20.560 Information theory has been overturned.
00:15:25.440 I mean, Shannon information theory
00:15:29.320 has been overturned by quantum information theory.
00:15:33.360 So for example, supplanted, perhaps, is a better word.
00:15:36.640 I mean, still most of the results of information theory
00:15:41.040 are still true in quantum information theory.
00:15:42.920 Most of the usual claims about information
00:15:45.200 still continue there.
00:15:46.600 So the classic entropy remains finite
00:15:50.320 in quantum information theory.
00:15:52.880 It does.
00:15:53.600 But the fact that entropy is governed
00:15:57.520 by slightly different laws.
00:16:00.280 And that these allow exponentially more information
00:16:05.720 to apparently exist in some situations
00:16:09.160 and not other means that we have to change our conception
00:16:11.840 of information.
00:16:13.000 And then our conception of information
00:16:16.840 may or may not feed into the practical conclusions
00:16:21.760 that we want this information theory to inform.
00:16:25.960 But the amount of negative entropy in this glass
00:16:28.800 is not substantially changed by quantum information theory.
00:16:32.360 We could still say, if that's all you have,
00:16:35.360 then you can only grow at a certain rate,
00:16:37.400 then that means our future economy descendants
00:16:40.480 can only grow at a certain rate.
00:16:42.000 I mean, that's still a robust conclusion,
00:16:43.920 even if we switch to quantum information theory.
00:16:46.160 It happens to be so for quantum information theory.
00:16:51.000 But quantum information theory allows information channels
00:16:55.680 to transmit twice as much information as classical channels.
00:17:00.720 So it would have been, you know,
00:17:02.960 if we've been having this conversation 40 years ago
00:17:05.360 or whatever it is, whenever channel capacity doubling
00:17:08.920 was invented, we would have said, you know,
00:17:12.040 we know for sure that if an object can be in two
00:17:17.880 distinguishable states and used for information transfer,
00:17:21.240 then the rate at which it transfers information
00:17:24.520 is also one bit per unit time or whatever it is.
00:17:28.840 But it's actually two.
00:17:30.720 So now we couldn't have predicted
00:17:33.040 that it won't be 10 or 50 or infinite.
00:17:37.520 Say you were with me on a project
00:17:39.160 to develop a nuclear reactor for thorium,
00:17:42.080 you know, you might have some qualifications for that, right?
00:17:44.600 And say I'm a part of this project
00:17:46.440 and I'm trying to, you know,
00:17:47.960 work out a tentative design for our thorium reactor
00:17:50.560 and some tentative ways to test this.
00:17:52.320 And I had to make some, you know, suggested claims to you
00:17:55.280 that this particular design could have this rate
00:17:57.520 and this density and then this heat conductivity, et cetera.
00:18:01.120 And then for any tentative design I offered you,
00:18:03.480 you could play the same game you're playing now
00:18:05.200 and say, well, yeah, but you're using standard information theory
00:18:07.800 and what about potential future information theories
00:18:09.880 and you're assuming standard gravity
00:18:11.880 and what about potential future gravities?
00:18:13.920 And I might say, well, you're not helping very much
00:18:16.000 to design this reactor if you're going to
00:18:17.400 to point out all the logical ways it could be wrong.
00:18:19.920 Don't you want to like make some tentative assumptions
00:18:22.720 that we get in here in the game, right?
00:18:24.080 Let's help design something.
00:18:26.040 And that's the way I want to ask about the future.
00:18:27.760 I want to say don't we want to think about the future?
00:18:29.480 Yes, of course we could be wrong.
00:18:30.680 Well, shouldn't we like take our best understanding
00:18:33.480 and make some tentative guesses here?
00:18:35.960 We could be wrong is a bad objection
00:18:40.080 because you could always make it.
00:18:43.320 And so you have to reject on principle all objections
00:18:52.320 to a project that are of the form we could be wrong.
00:18:57.280 Our best explanations could be wrong.
00:19:01.280 So I'm not objecting to that part
00:19:04.560 and I wouldn't be making those objections to the thorium reactor
00:19:08.000 and I don't make those objections to your analysis
00:19:12.120 of future aliens occupying galactic clusters and so on.
00:19:20.120 It's only due to my stuff.
00:19:22.200 Well, then let's talk more specifically
00:19:24.080 about some other ways we could guess about the future.
00:19:26.480 And then we could say specifically like what assumptions
00:19:28.720 were worried about being wrong that such might
00:19:30.640 derail the analysis.
00:19:31.680 But let's just talk about particular things
00:19:34.040 we could guess about the future, right?
00:19:35.960 So for example, we might say future stock prices
00:19:39.440 will follow a random walk just like current stock prices do
00:19:42.320 because our theory of why stock prices follow a random walk
00:19:45.720 is very solid theory about information basically.
00:19:48.440 It's an information based theory about the random walk
00:19:50.720 and you say as long as the information theory holds,
00:19:53.560 then future prices should follow a random walk.
00:19:55.720 So that sounds like a reasonable long term claim.
00:19:59.200 I go out a million years and say in a million years
00:20:02.040 in the future stock prices will follow a random walk
00:20:04.160 because unless information theory is dramatically changed,
00:20:06.840 that'll just be true.
00:20:08.520 Yes, so I agree with the conclusion,
00:20:11.760 but I would be explicit about the explanatory assumption there.
00:20:19.200 To me, the explanatory assumption there is basically
00:20:23.000 that if that were not so then speculator
00:20:27.360 could make money out of it not being so
00:20:29.360 and eventually it would become so.
00:20:31.760 And therefore it will be so.
00:20:34.000 So I think that's an epistemological argument
00:20:37.280 fundamentally.
00:20:39.040 And therefore I would say we can expect that to be true
00:20:45.560 until somebody wins a Nobel Prize in economics
00:20:49.240 for showing that the epistemological isn't what it is.
00:20:52.600 But for most purposes in the world,
00:20:54.240 it's perfectly reasonable to take what we think we know now
00:20:57.040 and apply it even if a Nobel Prize might be one someday
00:21:00.400 for showing that it's wrong, right?
00:21:02.200 Yes.
00:21:03.040 And so for example, there's a whole literature
00:21:05.240 on evolutionary investment, which talks about
00:21:08.440 if there's an evolutionary competition
00:21:10.400 between different investment strategies,
00:21:11.960 what would be the investment strategies
00:21:13.560 that win in the long run?
00:21:15.240 And that's based on something called the Kelly Rule.
00:21:17.600 And there's a theorem basically that the maximum investment
00:21:21.840 would have to invest in proportion
00:21:23.440 to the future final value of different investment categories.
00:21:27.080 And that seems pretty solid to me,
00:21:29.440 at least given the assumptions
00:21:30.640 that there'd be some evolutionary competition
00:21:32.760 among investors, such that the one who invest more successfully
00:21:36.960 reproduce and become more common.
00:21:38.920 Yes, and underneath the assumption
00:21:42.040 that there might be investors who have certain objectives,
00:21:46.120 there again, epistemological assumptions
00:21:48.560 about how success in achieving things can come about
00:21:55.360 when people cooperate.
00:21:57.040 And I think those theories are true.
00:21:59.160 And therefore, as I keep saying,
00:22:03.960 I think in many cases the bottom line is the same,
00:22:08.080 whether we are explicit about the explanatory basis
00:22:12.600 of the prediction or not.
00:22:14.800 But in some cases, you know, you can wander off,
00:22:17.800 I think you just did it just now
00:22:19.680 where the speed of light being constant
00:22:22.080 is a different assumption,
00:22:23.480 different species of assumption
00:22:26.360 from the assumption that the amount of information
00:22:31.280 carrying capacity at small scales is not infinite.
00:22:36.040 That those are two different kinds of things
00:22:39.000 because one of them is kind of saying,
00:22:41.600 well, we don't know about it.
00:22:42.840 We, you know, therefore it's probably finite
00:22:45.400 because most things are.
00:22:47.080 That is a different kind of assumption
00:22:48.920 from one that says, here is a physical limit
00:22:53.360 and there is no rival theory on the horizon.
00:22:57.880 So I mean, as you know, say, beckonstein's mound
00:23:00.360 does give a finite entropy within a volume.
00:23:03.360 And of course, our tip, you know,
00:23:05.160 when we actually look at physical systems,
00:23:06.640 we do tend to measure their entropy
00:23:08.440 the temperature, for example.
00:23:10.160 And we measure sort of total entropy
00:23:12.400 by the integral over what would happen
00:23:14.520 if temperature went to its limits.
00:23:16.960 And so we do have physical estimates
00:23:19.200 of the total amount of entropy in physical systems
00:23:21.440 that we actually deal with based on the physical properties
00:23:23.800 and those quantities are finite, right?
00:23:26.320 That's a very standard thing to find
00:23:28.640 in actual physical systems we interact with.
00:23:31.520 And so it's perfectly reasonable
00:23:34.160 to at least tentatively say these look like
00:23:36.840 finite entropy systems, you know, the glass of water here
00:23:39.680 is this is finite.
00:23:42.520 The beckonstein bound is a good example
00:23:44.560 because actually beckonstein had several bounds.
00:23:47.560 And I think one of them is solid, the one you mentioned,
00:23:54.480 but he had other bounds.
00:23:56.760 One of them turned out to be wrong
00:23:58.400 and I wrote paper showing it was wrong,
00:24:00.720 which purported to limit the maximum speed of a computer.
00:24:06.680 And that was just wrong.
00:24:07.960 It was now, it looked right
00:24:10.360 because the arguments were very similar
00:24:12.640 to the arguments about the entropy being
00:24:16.680 that equal to the surface area and that kind of thing,
00:24:19.120 but it wasn't right.
00:24:20.480 And I have actually co-authored a paper recently
00:24:23.560 where I corrected, I and co-authors,
00:24:26.080 corrected other people on mistaken judgments
00:24:28.480 about quantum, quantum, that quantum,
00:24:32.640 to just entropy of computation bounds.
00:24:35.080 So I think there's a difference.
00:24:37.600 And we could have known there was,
00:24:39.080 even before we discovered those errors,
00:24:41.280 we could have known there was a difference
00:24:42.960 between a beckonstein bound and the speed of light.
00:24:46.920 And that is, is rests on ultimately on assumptions
00:24:54.360 that aren't the direct result of well-tested,
00:24:57.920 well-explained physical theories,
00:24:59.680 whereas the speed of light is.
00:25:01.640 But you and I are right now in the context
00:25:03.520 of a larger intellectual world,
00:25:05.480 which accepts most of the same basic theories
00:25:08.120 that you and I are now accepting,
00:25:09.600 but which mostly puts a mental block on an analyzing
00:25:13.280 of the future.
00:25:14.120 They basically go, that's too speculative.
00:25:15.720 We're just not gonna do it.
00:25:16.880 And so, I mean, right.
00:25:18.880 And so you and I can at least share this approval of,
00:25:22.440 let's take our best theories and do the best
00:25:24.400 we can to project them forward.
00:25:26.280 And so again, I'm interested in like going with you there
00:25:29.520 to sort of think about.
00:25:30.640 So I just gave you the sort of the evolutionary analysis
00:25:33.720 or I pointed to the evolution analysis
00:25:35.560 of investment returns.
00:25:36.840 But even a more basic one is the idea
00:25:40.920 that we have a discount rate today.
00:25:43.040 So today, people discount the future in the market
00:25:46.720 and in the private actions.
00:25:47.880 And that is a big issue with respect
00:25:49.760 to say global warming or something
00:25:51.040 because the market value of distant future consequences
00:25:54.080 is very low.
00:25:54.920 And according to that usual discount rates,
00:25:56.920 we shouldn't bother to do much to prevent future harm.
00:26:00.520 And people often think that's wrong.
00:26:03.320 And we have a evolutionary analysis that gives a plausible
00:26:08.480 explanation for why we evolved such a high discount rate,
00:26:11.680 which is to say that the analysis suggests that
00:26:15.680 we often have a choice between investing in ourselves
00:26:19.040 or investing in our future generations.
00:26:22.480 And so that gives, because our descendants
00:26:24.720 have typically a factor of two difference
00:26:26.600 in their genes related to us,
00:26:29.480 that we then have a factor of two per generation discount rate,
00:26:32.480 which does roughly match the typical market discount rates.
00:26:36.480 And so that's an explanation.
00:26:37.720 And it says, well, we're failing sort of to collectively
00:26:40.960 care about the future because we each care
00:26:42.760 about the future generation half as much as ourselves.
00:26:46.200 And asexually reproducing organisms
00:26:48.840 wouldn't have this discount rate
00:26:50.240 according to standard theory.
00:26:52.560 And so these theories of sort of self reproducing investment
00:26:55.920 funds, they would be asexually reproducing.
00:26:58.160 And so they wouldn't have a discount rate.
00:27:00.560 And so the prediction is that in the long run,
00:27:02.960 something will arise that doesn't have this long-term discount
00:27:06.680 rate, and it will spread because it will
00:27:08.840 care more about the future.
00:27:09.840 And so we'll have more future consequences.
00:27:11.520 And so my prediction based on this analysis
00:27:13.720 is long-user coming.
00:27:15.960 Even though today, we don't care much about the future
00:27:18.720 because of this inherited biological tendency
00:27:21.480 to discount the future, a factor of two per generation,
00:27:24.920 future more asexually reproducing organizations
00:27:27.800 than units would care more about the future.
00:27:30.680 And therefore, they will eventually
00:27:32.560 care about the future.
00:27:33.360 So eventually, people will care about the future,
00:27:35.720 even if not soon.
00:27:37.160 That's a prediction I will make about the future.
00:27:39.440 Our descendants will care about their future.
00:27:42.680 So I see, you can point out the potential errors
00:27:45.360 in the potential ways the analysis could go wrong.
00:27:47.240 And we should do that.
00:27:48.160 But still, it's worth having projections like that,
00:27:51.560 at least to tentatively put them up and say, does that make sense?
00:27:55.160 Well, the idea that decision theory
00:28:00.000 should be analyzed in terms of how much
00:28:02.240 we care about our future selves or about our descendants.
00:28:06.640 I'm very skeptical of that theory
00:28:09.880 because it's a subjective psychological theory
00:28:14.520 of really an objective thing, which
00:28:16.960 is the growth of knowledge.
00:28:19.560 So I haven't actually heard before this theory
00:28:23.240 of where the offsprings matter.
00:28:30.160 I mean, there's also the, I don't know if you've
00:28:33.600 studied the case as well, there's all sooner or later
00:28:36.640 we're going to go immortal.
00:28:38.520 And then we won't have offspring.
00:28:40.560 And how will we value our future then?
00:28:43.120 Well, I think this can't be deduced from biology.
00:28:47.160 It can't be deduced from psychology.
00:28:49.520 But there will be objective factors.
00:28:54.800 So evolution hasn't encoded preferences as an us
00:28:58.400 for what happens when we get immortal, because that never happened.
00:29:01.400 So at the best, it would project the preferences
00:29:04.600 we have across our lifetime into the future.
00:29:06.720 But that does seem to be roughly this factor
00:29:08.760 of two per generation that we, at the moment,
00:29:11.440 are discounting our own personal immortality with.
00:29:14.080 That is, if we ask how much is it worth
00:29:16.160 to us to prevent ourselves from dying in a century,
00:29:19.040 we don't actually willing to put very many resources
00:29:21.840 into it because of this factor of two per generation.
00:29:24.000 But over time, that will change.
00:29:27.160 And then I might use an evolutionary analysis
00:29:29.400 to talk about how that would change.
00:29:30.640 I might say, well, when there are immortal creatures
00:29:33.200 and there's selection, variation, selection,
00:29:35.120 then what kind of preferences will be selected for?
00:29:38.560 That seems to me the way I want to do the analysis.
00:29:40.840 But that seems, sounds like you're
00:29:42.280 somewhat skeptical of analyzing preferences for time using
00:29:48.120 evolution.
00:29:49.400 Well, you're using really biologically evolution,
00:29:53.120 because it could be that this number, the discount rate,
00:29:59.440 is dominated by, say, meme evolution.
00:30:02.920 So suppose that some idea spreads, which
00:30:06.360 say that it ought to be some, or implied
00:30:09.120 that it ought to be some value.
00:30:10.800 Then you will say, ah, but then there's group selection,
00:30:14.120 because some other subset of the people
00:30:17.520 will not have that meme or religion or whatever it will be.
00:30:23.360 And if they can produce weapons faster,
00:30:27.560 then they will, in the long run, win.
00:30:32.240 So that's group selection.
00:30:34.040 And as you know, group selection is an extremely problematic
00:30:39.920 concept.
00:30:40.880 And I'm inclined to go with Dawkins and co
00:30:48.880 to say that group selection basically never happens,
00:30:52.440 even in the biological case.
00:30:54.600 And when you then take it to memes and societies
00:31:00.200 and civilizations, it can go wrong in all sorts of other ways
00:31:04.720 as well.
00:31:06.280 And there just aren't enough subgroups of society,
00:31:11.520 because there will be subgroups within those subgroups who
00:31:15.120 think that what they're doing is wrong.
00:31:17.280 Our whole society at the moment, the entire world,
00:31:20.680 is dominated by ideas which say that we shouldn't
00:31:26.760 make progress on principle, either in particular cases
00:31:31.840 or we shouldn't go too fast or whatever.
00:31:34.960 The idea that we shouldn't make progress as fast as possible
00:31:38.440 such that our point of view will become dominant in the long run
00:31:43.520 is almost held by nobody.
00:31:46.480 And you're assuming that in the long run,
00:31:51.520 that view will be superseded by some other view
00:31:54.080 that you think is more viable.
00:31:57.120 But it need not be so.
00:31:59.200 It's very rickety, it seems to me.
00:32:02.000 So David, I think you'd screw with Robin and just in the nature
00:32:05.160 of what he's proposing as well.
00:32:06.640 So for example, Robin has also said that proposed that we humans
00:32:10.680 have explicit goals in the future, we will evolve such
00:32:13.400 that we may have explicit goal to have more descendants.
00:32:16.200 That being an idea in our mind, it's
00:32:17.560 like a freely floating propensity, which our future generations
00:32:21.040 will have more than we do.
00:32:23.440 And the idea of this discount rate, my model of view,
00:32:25.400 is that you see thoughts as highly iterated and things
00:32:30.040 like culture and memes now dominate.
00:32:31.800 So how would this really floating number
00:32:34.240 about say one half or two to get into?
00:32:36.520 So I agree, Daniel, that the key issue here
00:32:38.680 is to think about what are our best robust models
00:32:41.520 of evolution that includes different potential units
00:32:44.920 like groups and includes culture as well as physical genes.
00:32:48.040 That is, if we thought, say, a simple model of physical gene
00:32:52.000 evolution was old hat, not really applicable,
00:32:56.560 and that the conclusions I was talking about were
00:32:58.600 dependent on that particular set of assumptions,
00:33:00.680 and we might say, well, that's just the wrong set of assumptions.
00:33:03.320 It's not relevant.
00:33:05.360 I think the conclusions that I've talked about
00:33:07.520 are relatively robust to the different size units of selection
00:33:11.560 and to culture being part of the co-evolutionary package.
00:33:15.240 And so I'm certainly happy to grant an interesting thing
00:33:18.360 to think about that at the moment our genetic induced
00:33:23.280 and culturally induced packages of behavior
00:33:24.960 are far from equilibrium.
00:33:26.440 That is, they're far from the thing
00:33:27.520 that would be selected for in the long run.
00:33:29.280 And something has gone very wrong in the past to induce that.
00:33:31.960 And I've thought a lot about what that could be.
00:33:34.200 But nevertheless, when I make the long-term predictions,
00:33:36.920 I set aside current behavior and just ask what in theory
00:33:40.320 would be selected for for not just genes,
00:33:42.600 but gene culture, co-evolutionary packages.
00:33:45.520 And most of the analysis I know is robust to which kind
00:33:48.920 of packages those are.
00:33:49.840 It doesn't that much whether behaviors encoded in textbooks
00:33:54.280 that people teach or in genes that influence in individual
00:33:59.360 construction, wherever it is, there's
00:34:02.200 evolution analysis of what sort of behavior
00:34:04.400 would went out in the long run.
00:34:06.040 Are you think things will equilibrate rather than staying
00:34:08.320 out of equilibrium, what culture we're going away from?
00:34:10.480 Yeah, yeah, that is the robust prediction
00:34:12.240 that evolution would say.
00:34:13.080 Yes, if some people have behavior
00:34:15.520 that produces much fewer descendants
00:34:17.520 or some people culture packages do,
00:34:19.560 then those packages will lose out.
00:34:21.760 They won't reproduce as much.
00:34:22.960 And in the future, there'll be more of the packages
00:34:25.040 that do reproduce more.
00:34:27.240 If the universe were infinite and there
00:34:30.080 were infinitely many cultures and civilizations and so on.
00:34:36.160 And if they were uncorrelated,
00:34:39.640 so that there was no particular reason
00:34:41.480 why a particular type of culture might take over
00:34:44.880 even though it's disadvantageous, then
00:34:47.920 you could make a robust evolutionary theory of it.
00:34:51.360 But I don't see how it can be robust in predicting
00:34:56.520 what errors people will make, for example.
00:35:00.400 Like I said, in 1990, they could have predicted
00:35:04.520 the environmental movement or nuclear power.
00:35:07.320 And all the ideas that have informed the use of those things.
00:35:13.760 Now, yes, if there were a million Earths, well, actually,
00:35:17.960 I think no, it's not yet.
00:35:19.680 Even if there were a million Earths or a trillion or whatever,
00:35:23.440 there is no way of using evolutionary theory
00:35:27.320 to predict what kind of mistakes of this kind
00:35:30.640 will be made by most of them.
00:35:32.200 There might be a theory of cultures that
00:35:35.600 says that certain types of mistakes will be made.
00:35:39.720 Some people say all cultures destroy themselves
00:35:42.640 because they become weak and so.
00:35:47.480 And I don't believe that theory.
00:35:48.920 But I'm just using it to illustrate the fact
00:35:50.800 that there is a substantive assumption there
00:35:54.640 that underlies the assumption that evolution theory alone
00:36:01.440 can determine the outcome.
00:36:04.680 So as you know, our standard theory of thermodynamics
00:36:07.640 both has equilibrium versions.
00:36:09.880 It has fluctuation versions.
00:36:11.280 The fluctuation versions predict how fluctuations
00:36:15.280 change over time and decay to equilibrium.
00:36:18.000 And of course, we have similar very alternatives
00:36:20.600 in evolution that is evolution isn't just
00:36:22.600 a theory of a final stable state.
00:36:24.880 It's also a theory of how initial deviations
00:36:27.480 will decay or move toward the equilibrium.
00:36:30.640 And I think those parameters suggest
00:36:32.880 that Earth is plenty big and the future is plenty long
00:36:36.000 for us to move from where we are now
00:36:38.240 to closer to an equilibrium.
00:36:39.480 Isn't that, I mean, do you have time estimates
00:36:42.280 that say it would take a trillion years
00:36:44.600 for us to equilibrate now?
00:36:47.000 I think there's no reason to think
00:36:49.520 that we're going to equilibrate because it
00:36:51.920 depends on the creation of knowledge,
00:36:54.120 which is not a relaxation process.
00:36:56.720 It's an error correcting process.
00:37:00.760 And the error correcting processes can go wrong.
00:37:03.640 They're not infallible.
00:37:05.680 They, in fact, they go wrong all the time.
00:37:08.000 And what you're proposing is that there's
00:37:11.720 an overarching reason why things will come to equilibrium.
00:37:16.960 In which case, you know, if you'd been around four
00:37:22.400 billion years ago, you would have predicted
00:37:24.200 that humans will never exist and that nothing from Earth
00:37:28.720 will ever get to the moon.
00:37:31.000 Let's just take concrete example.
00:37:32.800 Today, religious people are more
00:37:35.000 are having higher fertility than nonreligious people.
00:37:37.760 So the straightforward prediction
00:37:39.400 is that over the next generation or two,
00:37:41.400 there will be more religious people, right?
00:37:43.640 Now, at least people who descended from religious people,
00:37:46.480 now you might ask, well, will the children of religious people
00:37:49.080 be as religious as their parents?
00:37:50.400 And of course, that could be a thing
00:37:52.080 that then means that religion doesn't actually increase.
00:37:54.760 But you know, we can certainly expect a few generations now.
00:37:58.000 More people will have descended from religious people,
00:38:01.280 than will have descended from nonreligious people
00:38:03.240 because that's the selection argument.
00:38:05.160 And you could say, well, some new,
00:38:06.720 some we might discover a new thing that suddenly
00:38:08.880 helps the nonreligious people, you know,
00:38:10.680 have a big advantage.
00:38:11.520 And yes, we might.
00:38:12.280 But just like with the stock market,
00:38:13.880 is there any particular reason to expect
00:38:16.160 that the correlation will be that way,
00:38:17.800 if we don't have a particular reason
00:38:19.120 that, again, our expectation is that, on average,
00:38:22.800 for the deal.
00:38:23.640 Again, that's an explanatory assumption.
00:38:26.920 And that explanatory assumption is a very kind of
00:38:31.920 arbitrary assumption compared with the things
00:38:36.360 that it's kind of purports to be about,
00:38:39.480 which is very rigid connections
00:38:41.880 between the population at one time
00:38:43.840 and the population at a generation later,
00:38:47.200 some things about that relationship are rigid
00:38:51.040 or, you know, as they're rigid in the sense
00:38:53.680 that the explanatory theory underneath it
00:38:56.680 doesn't have any rivals, but any viable rivals.
00:39:00.640 But there are other things which aren't rigid.
00:39:03.360 And, you know, people's attitude towards
00:39:08.240 the number of children they have changes
00:39:10.960 on a time scale of a generation at the moment.
00:39:14.240 It's, you know, it's that fast.
00:39:16.640 So, you know, it might settle down
00:39:20.800 and it might, there might be an asymptotic form of it.
00:39:25.000 But things that, you know, things that contain
00:39:32.080 touring complete subsystems
00:39:34.920 don't have asymptotic forms.
00:39:38.040 That's, you know, you can almost define a touring complete thing
00:39:44.160 as a thing that doesn't have an asymptotic state
00:39:47.240 if it runs forever.
00:39:49.320 So therefore, I think it's a mistake
00:39:53.800 to make biology type assumptions about human growth of knowledge
00:39:59.240 just as it's a mistake to make chemistry type assumptions
00:40:02.520 about the growth of life.
00:40:04.000 So as an analogy that perhaps on a faster time scale,
00:40:09.960 we have somewhat of an evolutionary theory of growth
00:40:13.080 of the economy in the sense that we say that there are many firms
00:40:16.720 in an industry and some of them are more productive than others
00:40:19.520 and the more productive ones consistently grow relative
00:40:22.040 to the others.
00:40:22.880 And so we predict in some ways that the,
00:40:24.840 that each industry will be more productive in the future
00:40:27.280 based on an evolutionary selection of argument
00:40:29.640 among the firms in the industry.
00:40:31.400 It sounds like you'd say, no, we can't predict
00:40:33.600 that industries will get more productive in the future
00:40:35.400 because that's just making arbitrary assumptions
00:40:37.520 about correlations between culture and productivity of firms
00:40:40.720 which we have no reason to believe.
00:40:42.160 And then you do not believe that industries
00:40:44.480 will get more productive in the future
00:40:45.720 basically by the selection of the more productive firms winning.
00:40:49.120 As I said, I resist believing things,
00:40:52.280 but I think it's a good theory.
00:40:55.120 And you know, if I want to plan my pension or whatever,
00:40:58.520 then I would assume that there will be economic growth
00:41:03.520 in the future.
00:41:04.840 But the reason, and it's a selection argument,
00:41:07.520 it's an evolutionary selection argument
00:41:08.880 which is one of the primary reasons we believe
00:41:11.040 there is economic growth.
00:41:12.280 Well, the, I don't know about primary.
00:41:16.040 You see, that argument depends on things like
00:41:21.960 civilization continuing, you know.
00:41:24.680 Well, yeah.
00:41:25.680 If, if, again, in 19.
00:41:28.360 No asteroid destroys everything, yeah, of course.
00:41:30.800 Yeah, well, but you know, no world war
00:41:33.400 destroys everything, no crazy economic theory adopted
00:41:38.200 by the majority destroys everything.
00:41:40.040 And again, if you're choosing your pension,
00:41:41.560 you're going to make those assumptions, aren't you?
00:41:44.960 Well, I'm, I think I'm going to hedge against
00:41:48.080 those events to some degree,
00:41:51.040 but you'll still mostly bet that that's the most
00:41:53.360 like the outcome.
00:41:54.240 Well, I, I have no choice.
00:41:57.640 So it's, it's more like, you know,
00:42:00.800 they've got this Bitcoin thing.
00:42:01.920 If you really think everything's going to hell,
00:42:03.440 you could just buy a lot of Bitcoin, right?
00:42:05.840 Yeah, sure, you can find some people willing to sell it to you.
00:42:08.400 People say that and it's the reason is
00:42:11.640 that they are not being explicit about their assumptions.
00:42:14.760 Again, and I would rather say nothing is secure.
00:42:21.040 Therefore, I'm going to, I'm going to make my investments
00:42:24.680 on the basis that if this investment goes wrong,
00:42:29.160 I will have bigger worries than that.
00:42:32.320 For example, it may be that democracy is broken down
00:42:37.320 or that there's a world war or something like that.
00:42:40.560 So then then I will, I will have been wrong
00:42:43.760 about my investment, you know, I should have put all my money
00:42:47.480 into rubidium because that's the basis of the future weapons.
00:42:52.680 But I was not to know that.
00:42:54.760 So, so when I try to do future analysis,
00:42:58.000 one of the biggest contrary assumptions or scenarios
00:43:01.560 that I focus on is what if we end up creating
00:43:05.080 a strong world government that strongly regulates investments
00:43:08.360 and reproduction and other sorts of things
00:43:10.600 and then prevents the sort of competitive environment
00:43:13.400 in which the evolutionary analysis applies.
00:43:16.360 And I'm very concerned about this scenario.
00:43:18.800 That is my best judgment of our biggest long-term risk
00:43:22.160 in terms of, you know, the thing we should focus attention on
00:43:24.520 is this creation of a strong civilization-wide government,
00:43:29.000 say solar system-wide government that is going to be
00:43:31.800 wary of competition and wary of allowing independent choices
00:43:36.600 and probably wary of allowing interstellar colonization.
00:43:39.760 That is this vast expansion into the universe
00:43:42.720 could well be prevented and stopped by that.
00:43:45.720 So I'm very into sort of a baseline set of assumptions
00:43:49.840 and the conclusions and then asking what's the most likely
00:43:52.360 or the biggest thing that could go wrong there.
00:43:54.160 And that would be my guess for the biggest thing
00:43:56.240 that could go wrong.
00:43:58.200 It seems more useful to have a concrete scenario like that
00:44:01.080 that I can get my hands on and think about
00:44:02.800 than just the general idea that, hey, who knows,
00:44:05.520 you might be wrong.
00:44:06.800 I agree.
00:44:07.640 And that the concreteness is what I would call
00:44:10.360 the explanatoryness.
00:44:11.760 I mean, that is explaining how things could go wrong.
00:44:16.080 There's also the fact that though that things could go wrong
00:44:20.360 in a way that our existing explanations don't predict.
00:44:24.440 By the way, you haven't yet spoken of a type of prediction
00:44:30.400 that I don't think you should be making.
00:44:32.640 Well, all the specific examples you've given are,
00:44:37.520 I would just, if you like, it's a matter of terminology
00:44:42.520 but it isn't a matter of terminology.
00:44:44.160 It's a matter of epistemology and fundamental theory.
00:44:47.640 But the bottom line is you haven't yet given an example
00:44:51.320 where I would disagree with you that something is worth
00:44:54.440 investigating.
00:44:56.680 But I also think that if, if let's say there's
00:45:01.760 going to be a conditional on the our civilization
00:45:08.680 or our species not surviving for the next century,
00:45:14.280 I think it is overwhelmingly likely that the reason
00:45:18.240 will be one that we haven't thought of yet.
00:45:22.280 We should work on thinking of more things.
00:45:24.640 Yes, absolutely.
00:45:26.840 That's that's my basic conclusion.
00:45:29.760 That we know over that I'm into generating new scenarios
00:45:33.960 for things to go wrong that other people haven't
00:45:35.560 been talking about.
00:45:36.360 I have been generating such scenarios.
00:45:39.000 Great, great, but this is just a special case
00:45:43.040 of the fact that what we really should be doing
00:45:45.440 is creating more knowledge, more general purpose knowledge.
00:45:50.680 Knowledge that is potentially applicable
00:45:53.440 to applications that we don't yet know.
00:45:56.440 So one of them is the one you've just mentioned,
00:45:59.640 but there's also just doing pure mathematics
00:46:02.240 or working on things that don't even have names yet.
00:46:09.080 I agree we should have some broad just exploration
00:46:11.520 of all possible areas of knowledge, which isn't terribly
00:46:14.240 directed by any particular goals.
00:46:15.760 But then we should also focus our attention
00:46:17.640 on particular things that seem promising for particular reasons.
00:46:20.560 So for example, I have this book, The Age of M, as you may know,
00:46:24.320 takes a very particular technology scenario,
00:46:26.520 which some people might find out unlikely and just tries
00:46:29.080 to work through in great detail all the consequences of that.
00:46:32.520 And my justification would be, even if you think it only
00:46:35.720 has a 1% chance, if it's worth having 100 books on the future,
00:46:38.760 then it's worth having a book like this
00:46:40.360 because with this and the other 99 books,
00:46:42.920 we could lay out a picture of a wide range of options, right?
00:46:46.480 And so that seems to me that we should just take particular scenarios
00:46:51.320 that we can think are not crazy and work out their consequences
00:46:55.000 to think about what would happen if.
00:46:56.600 We should, but just as a side remark,
00:46:59.560 you've caught me in an illegitimate use of probability
00:47:03.360 when I said that I think it's overwhelming.
00:47:05.400 Given that the only assumption that the civilization
00:47:12.960 or species will be destroyed, then it's overwhelmingly likely
00:47:16.320 that it will be, I shouldn't have said that.
00:47:19.640 And this just shows how deeply this mistaken notion
00:47:25.880 of probability, of ideas having probabilities
00:47:30.800 has permeated our culture, even though I hate it,
00:47:35.040 I can't help using it.
00:47:37.240 So as you know, there are betting markets out there.
00:47:40.440 And you can interpret betting markets
00:47:42.400 as representing probabilities.
00:47:43.840 And in fact, there's standard theorems and finance
00:47:45.960 how you can represent basically all prices
00:47:48.080 through some set of probabilities.
00:47:50.560 And I've worked on mechanisms to create
00:47:52.920 sort of conditional probabilities or combinatorial probabilities.
00:47:56.320 We can elaborate pretty large complicated combinatorial sets.
00:47:59.440 And we have a lot of literature that at least says these
00:48:02.520 probability estimates are well calibrated.
00:48:04.960 That is what, when it says 20%, 20% of the time that happens
00:48:08.360 and that they are relatively accurate in the sense
00:48:10.440 that compared to other ways that make these forecasts,
00:48:13.600 they are giving you the truth more often.
00:48:17.880 I mean, this seems to me pretty useful to me
00:48:19.600 to have these systems that produce calibrated accurate forecasts
00:48:22.680 according to a probabilistic measure.
00:48:24.760 You don't approve of them, but you probably still
00:48:27.080 want to use them.
00:48:27.960 I mean, when you get a weather forecast that says 70% chance
00:48:31.160 of rain, then that weighs on you more than at 10% chance
00:48:34.200 of rain, don't you listen to these things, aren't they useful?
00:48:37.040 Yes.
00:48:37.880 And I admit that I think the connection between what
00:48:43.600 you might call risk and probability, that
00:48:47.080 is the reason why risks can be approximated by probabilities.
00:48:52.520 And also the reason why frequencies in certain situations
00:48:57.160 can be approximated by probabilities is an unsolved problem.
00:49:01.880 And I think it's a very important problem.
00:49:05.080 And if I wasn't working on other things,
00:49:06.880 I would be working on that.
00:49:08.280 I mean, I think-
00:49:09.720 OK, so until you figure that out, it's not too crazy
00:49:13.040 for the rest of us to continue to use them, right?
00:49:15.400 I mean, since they seem to be useful so far,
00:49:18.000 this appearance of use isn't just an illusion, right?
00:49:20.040 I mean, we are actually getting substantial value
00:49:22.480 out of weather forecasts and financial prices.
00:49:25.760 Assuming that there are no asteroids out there
00:49:29.640 heading on a collision course is also
00:49:31.760 a very useful assumption in practice.
00:49:35.120 But-
00:49:35.800 Because the chance is low, most people would say.
00:49:38.480 Yes, but they're wrong.
00:49:40.040 We don't know what the chance is.
00:49:41.920 And we need this idea that the chance is low is a break
00:49:48.480 on the urgency of working out what the chance actually is.
00:49:54.440 So we need to work on that.
00:49:57.000 And the fact that life continues productively
00:50:02.480 on the opposite assumption isn't an argument.
00:50:06.000 We need knowledge about asteroids.
00:50:08.880 We need that if there is an asteroid,
00:50:13.600 there either is or isn't an asteroid out there heading
00:50:16.800 towards us in the next 100 years or in the next year.
00:50:20.800 And that's a fact.
00:50:22.520 It has nothing to do with probability.
00:50:24.400 Probabilities can't help us.
00:50:26.920 Once it's in the sky, it's going to be no good with this.
00:50:30.520 We can't sue anybody and say the probability of that was very low.
00:50:34.640 I mean, God will just say, no, it wasn't.
00:50:37.040 It was one.
00:50:37.760 It always was one.
00:50:38.920 And no, but when you have the right theory of asteroids,
00:50:41.280 then it can make a stochastic prediction and you have chance.
00:50:44.600 And you know what it is.
00:50:45.520 So you need a theory of asteroids with a theory
00:50:48.840 that predicts their distribution.
00:50:52.000 They're distributing their trajectories.
00:50:54.000 We do actually have such theories.
00:50:55.440 I mean, people have actually calculated the rate of asteroids
00:50:58.480 with stochastic models like that.
00:51:00.760 That's a common thing.
00:51:02.760 So our best theories of planetary formations
00:51:05.800 say that a gas giants ought to form
00:51:11.160 in the outer stellar system.
00:51:14.400 Like, and indeed, they have in our solar system.
00:51:18.440 But all the other solar systems we look at have them
00:51:21.280 be formed in the near to their star.
00:51:26.240 And nobody knows why.
00:51:28.200 And therefore, I don't hold much with stochastic theories
00:51:32.360 of asteroids either.
00:51:34.240 You might remember that the first steam engines
00:51:36.960 didn't have thermodynamics as a basis.
00:51:39.200 Thermodynamics was generated after the steam engines, right?
00:51:42.320 So once you have a thing that seems to work,
00:51:45.000 we might be justified in just using it.
00:51:46.920 And if you say it doesn't have a good theoretical foundation,
00:51:49.120 we might say, well, then go find it.
00:51:51.200 But we're still going to use it until you prove theoretically
00:51:54.720 that it's badly founded that we're misjudged.
00:51:57.840 I agree that not only can we use it, but we should use it.
00:52:02.880 So as I said, everything you're doing
00:52:06.000 is not just legitimate, but at least things you've told me,
00:52:11.040 not just legitimate, but morally required.
00:52:15.200 And it's a bit of a scandal that more people aren't doing it.
00:52:19.760 But I think the same is true of all fundamental theories,
00:52:26.720 all fundamental branches in the village.
00:52:30.320 We don't know how to make quantum mechanics
00:52:32.880 and gravity coherent or generality coherent.
00:52:35.680 Nevertheless, we're completely justified in using each of them
00:52:38.400 in their typical current areas of application.
00:52:41.640 Well, using them for predicting things like satellite
00:52:47.880 or bits, but not for predicting things like how much information
00:52:52.280 you can get into a finite volume.
00:52:54.840 Right.
00:52:55.680 So there, the theory itself says that it doesn't know.
00:52:59.440 And that's exactly what we should.
00:53:01.400 So if you could tell me where exactly
00:53:03.720 is probability most problematic, then I
00:53:05.720 will try to avoid it there, because you've
00:53:07.800 thought more about where problematic is problematic.
00:53:09.840 But if you don't mention a particular area
00:53:12.120 is especially problematic, then I'm probably just
00:53:13.920 going to keep using it, because it's interesting.
00:53:17.040 Just probably a part.
00:53:18.040 Yes.
00:53:20.200 In short, the place where it's dangerous
00:53:23.480 is where the thing you're predicting
00:53:26.000 depends on the future growth of knowledge.
00:53:29.000 So you've given some, it still works even then,
00:53:34.000 like when you're saying that the stock price of something
00:53:39.560 will be a random walk.
00:53:41.640 Even when knowledge accumulates, the stock price
00:53:44.120 follows a random walk.
00:53:44.960 Yes, exactly.
00:53:46.480 So there are cases, but there are cases
00:53:49.320 where it's very misleading.
00:53:52.280 And so for example, when you say that all long-lived civilizations
00:54:00.920 in the past have failed within, well, they have failed.
00:54:07.640 All long-lived civilizations, say,
00:54:10.600 all civilizations that have lasted longer than 400 years.
00:54:14.160 If you count ours as having lasted for 300 years,
00:54:17.600 all civilizations have lasted for more than 400 years
00:54:21.360 have already failed.
00:54:23.320 And therefore ours will, that's illegitimate.
00:54:27.080 That's because it's making an assumption
00:54:31.840 that the frequency is the probability.
00:54:34.400 And here, I have a substantive theory
00:54:37.960 that says why it isn't namely that our civilization
00:54:42.640 is different from all the others in the relevant way.
00:54:45.520 But it doesn't matter.
00:54:46.280 Even if I didn't know that theory,
00:54:48.120 I would still say it was illegitimate
00:54:50.320 to extrapolate the future of our theory
00:54:53.560 based on our civilization, based on past civilizations,
00:54:57.800 because all of them depended on the future growth of knowledge.
00:55:01.680 And if you look in detail, actually, at how they failed,
00:55:05.880 they will fail in different ways.
00:55:08.600 But one thing you can say about it
00:55:10.200 is that in all cases, more knowledge would have saved them.
00:55:14.920 But the growth of knowledge was, in fact,
00:55:18.040 the major reason for growth for all past history.
00:55:22.680 That is, even biology, the major reason for innovation
00:55:25.160 biology was the growth of knowledge.
00:55:26.560 Forager humans, slowly over 2 million years,
00:55:29.920 they slowly grew by accumulating culture,
00:55:32.520 the growth of knowledge, of farming culture,
00:55:34.880 accumulated doubling roughly every 1,000 years
00:55:37.040 through the growth of knowledge, primarily
00:55:38.720 via different kinds of plants and animals
00:55:40.600 were domesticated.
00:55:42.040 The growth of knowledge has always
00:55:43.240 been the dominant factor in growth for all of history
00:55:46.960 that's still true today.
00:55:48.120 So that fact doesn't distinguish us from our ancestors.
00:55:52.160 Maybe the way we grow knowledge or the rate at which we
00:55:54.600 can grow is the distinction.
00:55:55.920 But the fact that knowledge was growing
00:55:58.240 and was the main long-term force has not changed.
00:56:01.880 Yeah, it's the way we grow knowledge.
00:56:06.000 And then we might find that our descendants
00:56:07.840 will grow knowledge even faster.
00:56:10.000 And I've thought a lot about the ways
00:56:12.840 that our descendants might grow knowledge faster
00:56:15.120 than we do and what are the obstacles
00:56:17.560 to our growing knowledge now.
00:56:19.400 Yeah, I'll say to difference, not just
00:56:21.400 the difference of scale or rate.
00:56:24.240 Yes, it's qualitative.
00:56:30.480 Again, just like the life as a chemical process
00:56:35.880 is different from all other chemical processes.
00:56:39.040 And thinking as a information processing process
00:56:46.320 is different from all other information processing processes.
00:56:50.160 Therefore, when you try to predict one
00:56:53.200 in terms of frequencies derived from the other,
00:56:55.920 you're going to make mistakes.
00:56:57.160 You're going to make bad mistakes.
00:56:58.880 Always, by the way, in the direction of pessimism.
00:57:02.040 Because it's not just that we are unlike previous civilizations.
00:57:06.160 We're unlike them in a systematic way.
00:57:11.920 It's a big topic.
00:57:13.280 Go ahead, Daniel.
00:57:14.280 It's a big topic.
00:57:14.800 And people use probability for more than,
00:57:17.280 so just matching to frequencies.
00:57:19.280 And Robin has papers on priors and stuff.
00:57:21.240 And there's more there, but it's a big topic.
00:57:24.600 It sounds like we agreed to talk for about an hour.
00:57:26.880 We've talked for about an hour.
00:57:28.000 We feel like we're at a sort of a resting
00:57:31.200 or not in the middle of some dispute or something like that.
00:57:34.440 Yeah, well, very interesting.
00:57:36.080 I mean, I think we should publish this.
00:57:40.080 And see what we'll say.
00:57:41.560 Absolutely.
00:57:42.560 If they like it, maybe we can begin.
00:57:44.720 And we'll make them take sides, who's the winner here, right?
00:57:47.160 I'm not sure I don't think we need to win.
00:57:51.080 I don't think we need to win.
00:57:52.920 I was just teasing you about the previous debate.
00:57:54.920 Yeah.
00:57:55.920 OK, it was a nice conversation.
00:57:58.120 It's a nice to see you.
00:57:58.920 All right.
00:57:59.440 Thank you for hosting, Daniel.
00:58:01.280 Welcome.
00:58:02.920 Bye, bye.