Data Republican believes EA should be banned from AI safety, policy-making, and governance. That's a pretty extreme position, but she makes a solid case for both the mild and extreme versions.
Fundamentally, EA is a maximizing philosophy: it always tries to price future value. And extinction is the “negative infinity” number in the ledger. Once you are determining outcome by probabilities, just having even a tiny chance of extinction dominates every other calculation.
One such scenario that many conservatives might find sympathetic is: do you do gain-of-function research for a 70% chance that you will prevent the next swine flu, or is the 1% risk that the virus escapes containment in Wuhan and kills 10 million people acceptable? Many of us, having lived through COVID-19, would put our foot down on allowing any kind of gain-of-function research again within our lifetimes. That’s one scenario where conservatives can potentially reason their way to the same answer as EAs.
EAs run that same logic in every calculation. What this means, inside their heads: an extinction level event scenario is the greatest threat vector.
Scott Alexander, like others, is obsessed with AI, alignment, and its effects. He is very much a supporter of Effective Altruism in general, though not slavishly so. Still, I imagine he would object to such a blanket disallowing of EA from the discussion (or at least, from the decision-making). There are so many discussions and counter discussions over at Astral Codex Ten that I won't even attempt to choose for you. As his commenters are also quite brilliant - I consider it one of the few sites to compete with this one for average IQ - you can either dabble or go hard against them, trying to steelman your own arguments. This would be the major leagues. And yet...
I have felt uncomfortable with Effective Altruism since I first heard of it, though it took me a while to articulate why. Frankly, it probably took reading others who had thought about it more rigorously to be able to articulate it. One of the consequences of becoming lazier, and thus less intelligent, is not being able to provide some of the service I was designed for. I have excuses. Not good excuses, but at least some things I can say out loud with some earnestness.
The best arguments against EA are in Data Republican's essay. Pricing future value is a great example. It makes entire sense, but somehow leads to insane conclusions. If a future life is worth the same as a life now, then the sheer volume of future humans and their needs with forever dwarf what is good for us. More interestingly, perhaps, is how that completely devalues our ancestors. There is the (bogus in reality but worth valuing as a myth) Native American belief in Seven Generations, that we owe something to the three generations before us and the three after us. GK Chesterton wrote similarly that tradition is the only way of giving a vote to the dead. CS Lewis, describing the Tao in The Abolition of Man shows that the wise in every tradition teach that we owe something to both our heritage and our posterity. But EA very quickly reveals that it owes nothing to heritage and everything to posterity. The dead have made their contribution and their suffering is now a sunk cost.
There is also comparative suffering among species, leading to calculations that the horrible suffering of three chickens or thirty-two shrimp are equivalent to the suffering of one human. The actual numbers are immaterial. Thirty or three-hundred chickens would be the same. The principle of comparing it to any human suffering is something new in the world. It may be a logically consistent and defensible philosophical and moral position.
But it is not a human position, despite being favored by "humanists." It gives away the game, recognising that humanists are in fact subhumanists, valuing some bare minimum of the species as fungible with any other human. Dismiss it as a mere prejudice on my part if you will, but the idea of perpetuating humans who are only biologically and not culturally related to me is not just disquieting, it's silly.
Humanists are actually Subhumanists. I should probably think about that and post on it. Yet now that I have said that, I recall that CS Lewis already covered that in the last debate section of Out of the Silent Planet.
4 comments:
Infinities generally have to be canceled out when they crop up, because (a) the status of the infinite is a deep philosophical problem to which we do not know the answer, though most of our thinkers think that somehow there are at least two classes of infinite things and they are of different 'sizes' (I dissent from this view, but it is the standard one and mathematicians will tell you that it is settled and proven); and (b) they otherwise break the math that we rely on to calculate how to do things in a universe that possibly only admits of potential rather than actual infinities anyway.
Those two points are tightly related.
I'm not Terrance Tao, but I can give infinity a try.
In analysis one of those things they emphasized was that infinity wasn't so much a thing as a process. E.g. start with 1, add 1, add 1, and keep doing that.
It's the "not stopping" part that turns the algorithm into an infinite process.
It turns out that pretty much any algorithmic process (Do A, do B, do C, ...) that goes on forever is equivalent to any other. Adding 1 each time and never stopping doesn't get you any "bigger" than adding 2 each time, or 1/n each time, or looking at n^n^n...^n n times each time. You can map them all on each other with nothing left out. Famously, the positive integers map to the rationals: 1->1, 2-> 1/2, 3-> 2/1, 4->1/3, 5->2/2. 6-> 3/1, etc. FWIW, these mappings are also algorithmic processes, so things get a bit messier later.
On the other hand, you can define representations for numbers, e.g. the decimal representation. When you do you find that there are more representations of distinct numbers than you can have algorithmic descriptions for. So algorithmic infinite processes don't exhaust the possible ways to have infinite processes.
From there you can go to functions, and can imagine more ways to have functions than there are representations for numbers -- which seems pretty obvious, but it's harder to prove because algorithmic mappings quit working an infinity ago.
Aristotle in the Physics worries that the additive sort of infinity is obviously false because, however you define 'a thing,' there are only so many things in the universe. He is genuinely worried about an analogous problem to the one you are raising (which is allegedly the proof of different sizes of infinities), that if you try to assign numbers to things, eventually you'll run out of things and the pairing process will have to stop.
What he determines is a workable answer to that is a version of your "1/n": since we can at least conceptually divide any thing into 1/2 and 1/2, and then the 1/2 into two 1/4ths, etc., we can get potential infinities in the real world after all. You can't actually divide things forever, though: we have even more reason to doubt that now than he did. If you want to divide a hydrogen atom in half conceptually, you'd get half each of a proton and an electron; but they don't really exist physically as halves of themselves, and a very short jump from that it isn't clear what you're dividing, or if it's sensible to continue to speak of division.
Or, to borrow from Feynman, how much time does it take for a photon to travel from the surface of the sun to here? In a sense, we can fudge the answer by saying that it doesn't take any time at all, zero time; but we're really doing something improper by saying that. There's actually no traveling 'from the photon's point of view,' because the photon has no point of view and no mass and, thus sort of, no time. Any sort of infinity of magnitude that crops up -- and they do seem to crop up -- doesn't really work.
Post a Comment