Thoughts on Robin Hanson and David Deutsch on predicting the future

David Deutsch has an influential book that contains statements like the following:

The future of civilization is unknowable, because the knowledge that is going to affect it has yet to be created.

Unfortunately, the closest he comes to explaining this claim is:

The ability of scientific theories to predict the future depends on the reach of their explanations, but no explanation has enough reach to predict the content of its own successors — or their effects, or those of other ideas that have not yet been thought of. Just as no one in 1900 could have foreseen the consequences of innovations made during the twentieth century — including whole new fields such as nuclear physics, computer science and biotechnology — so our own future will be shaped by knowledge that we do not yet have.

In the same book Deutsch writes:

The philosopher Roger Bacon (1214–94) […] foresaw the invention of microscopes, telescopes, self-powered vehicles and flying machines — and that mathematics would be a key to future scientific discoveries.

Later, he predicts:

Illness and old age are going to be cured soon — certainly within the next few lifetimes — and technology will also be able to prevent deaths through homicide or accidents by creating backups of the states of brains, which could be uploaded into new, blank brains in identical bodies if a person should die. Once that technology exists, people will consider it considerably more foolish not to make frequent backups of themselves than they do today in regard to their computers. If nothing else, evolution alone will ensure that, because those who do not back themselves up will gradually die out. So there can be only one outcome: effective immortality for the whole human population, with the present generation being one of the last that will have short lives.

So—it seems like we should take this unknowability claim with a big pinch of salt. A more plausible slogan would be:

The future of civilization is very hard to know, partly because the knowledge that is going to affect it has yet to be created.

Anyway, with this context I expected that Deutsch would be quite critical of Robin Hanson’s approach to prediction. In fact, during their recent hour-long discussion, Deutsch did not identify any particular cases where he thought Hanson’s approach was widlly off. After discussing several cases (including the cost of solar power; demographic projections; the grabby aliens model), he said:

DD: You haven’t yet given an example of something where I would disagree with you that it’s worth investigating.

And then, towards the end of the discussion:

DD: Everything you are doing is legitimate and indeed morally required, and it’s a bit of a scandal that more people aren’t doing it. But I think the same is true of all fundamental theories, branches of knowledge.

In closing, Deutsch summarised his position:

DD: In short, the place where [probability] is dangerous is where the thing you are predicting depends on the future growth of knowledge. You’ve given examples where it still works even then, e.g. you’ve mentioned the idea where stock prices will be a random walk [even when knowledge accumulates].

He continues:

DD: But there are cases where it’s very misleading. So for example, [if one says] that all long-lived civilisations in the past have failed, and therefore ours will—that’s illegitimate. Because it’s making an assumption that the frequency is the probability. And here I have a substantive theory that says why it isn’t. Namely that our civilisation is different from all the others.

But then:

But it doesn’t matter—even if I didn’t know that theory, I would still say it was illegitimate to extrapolate the future of our civilisation based on past civilisations. Because all of them depended on the future growth of knowledge. And if you look in detail about how they failed, they all failed in different ways, but one thing you can say about it is that in all cases, more knowledge would have saved them.

He says he would say its illegitimate to extrapolate even if he didn’t have a theory as to why—but does not explain why that’s the case, instead just restates the theory he actually has.

Notably, his central principle of optimism” (“all evils are caused by lack of knowledge) is a claim about the future of our civilisation based on… extrapolation from past civilisations.

My suspicion is that Deutsch doesn’t have a crisp way to distinguish (legitimate) prediction from (illegitimate) prophecy, and that often he just labels as prophecy” the predictions that he does not want to seriously engage. At least, this is what I think I’ve repeatedly seen going on during discussions of existential risk that threaten his principle of optimism.

At the level of theory, Deutsch concedes:

DD: I admit that I think the connection between risk and what you might call probability, the reason why risks can be approximated by probabilities, and also the reason why frequencies, in certain situations, can be approximated by probabilities, is an unsolved problem. And I think it’s a very important problem and if I wasn’t working on other things I would be working on that.

Deutsch is right to emphasise is that if you’re going to extrapolate trends or make reference class comparisons, it’s worth trying to articulate the grounds on which you’ve selected your reference classes, and reasons the trend might not continue. In his language: predictions are always underwritten by explanations, implicit or otherwise, and if you make them explicit you may be able to improve on them. That said, it’s surprising how often a prior in favour of simple extrapolations can peform better than more complicated models (see e.g. COVID-19, spring 2020).

Deutsch is also right to emphasise that probability estimates can easily be thrown wildly off due to unknown unknowns or other kinds of model error. This is not controversial, but it is often forgotten, and this mistake can be expensive.

DD: I think it’s reasonable if your best explanations imply that.

DD: You mean not only that the graphs are the same, you mean that you expect solar power technology to be dependent on making factories which use materials of a certain kind which aren’t going to be throttled by a hostile foreign power and so on.

DD: It is simply wrong to use probability in this way and it would be better to make the assumptions explicit.

I note that Deutsch does not offer an argument for the it is simply wrong to use probability in this way” claim, in much the way that Mervyn King and John Kay fail to in their book. In particular, he does not discuss or argue against the Bayesian notion of assigning subjective probabilities to beliefs.

In a talk titled Knowledge Creation and It’s Risks, Deutsch says:

Outcomes can’t be analysed in terms of probability unless we have specific explanatory models that predict that something is or can be approximated as a random process, and predicts the probabilities. Otherwise one is fooling oneself, picking arbitrary numbers as probabilities and arbitrary numbers as utilities and then claiming authority for the result by misdirection, away from the baseless assumptions.

For example, when we were building the Hadron collider, should we not switch it on just in case it destroys the universe? Well either the theory that it will destroy the universe is true, or the theory that it’s safe is true. The theories don’t have probabilities. The real probability is zero or one, it’s just unknown. And the issue must be decided by explanation, not game theory. And the explanation that it was more dangerous to use the collider than to scrap it, and forgo the resulting knowledge, was a bad explanation, because it could be applied to any fundamental research.

He’s right that—objectively—either the collidor will destroy the universe or it won’t. But it seems fine for the Bayesian to say: in cases like this, with evidence like this, I expect my beliefs about the objective world to be correct X% of the time.

Towards the end of the dialog, Deutsch accidentally states a prediction in rather Bayesian-sounding terms:

DD: Conditional on our species not surviving this century, I think it is overwhelmingly likely that the reason is one we have not thought of yet.

He notices the mistake, and corrects himself:

DD: As a side remark, you’ve caught me in an illegitimate use of probability. When I said that conditional on our species being destroyed it’s overwhelmingly likely to be [due to a reason we have not thought of]… I shouldn’t have said that. This just shows how deeply this mistaken notion of ideas having probability has permeated our culture. Even though I hate it, I can’t help using it.

At the end of the dialog, I was still left waiting for Deutsch’s explanation for why assigning subjective probabilities is not a sensible thing to do1. Perhaps that will not be forthcoming. I would gladly settle, instead, for his object-level discussion of the Vulnerable World Hypothesis 2.

Edit (2021-12-29): Tyler Cowen comments on his interview with David Deutsch:

Deutsch convinced me there’s often a lot of hot air behind Popper. He didn’t argue very well on Popper’s behalf. Deutsch is way smarter than I am but he seemed to me in some fundamental ways a dogmatist, and not really able to defend Popper very well. He’s made up his mind and you get a particular kind of emphatic statement, but I thought that at the philosophical level his defences were weak. […] There’s this odd feature of Popperianism. It somehow attracts a lot of dogmatists. I don’t know why.

  1. Edit (2021-12-27): since writing this, Joesph Walker pointed me to a 2016 paper by Deutsch, which contains an extended discussion of the Popperian conception of scientific explanation which Deutsch favours, and contrasts it to Bayesian conception. I’ve not yet had a chance to give it a proper read. I will update this footnote/post when I do.↩︎

  2. For what it’s worth: my current take is that Deutsch is right to worry about widespread pessimism about technology, but his reaction is too sweeping. I worry that the slogan he presents as the principle of optimism is used by technologists to wave away legitimate concerns about catastrophic and existential risk. In many areas (e.g. nuclear power)—and probably on average—I guess Deutsch is right that we need more techno-optimism. But in some areas, I suspect we need discouragement and regulation, with a view to pulling off some strategy of differential technological development.↩︎

writing robin hanson david deutsch futurism bayesianism radical uncertainty