Scott Alexander

I predict that you will counter that this is actually just [perfectly reasonable explanation of why an AI would think this way for instrumental reasons], and that when we dissect the perfectly reasonable instrumental reasons why AI would think this way, it will be identical to the convergent instrumental goals that we were trying to convince you of all along.

There are approximately three reasons why we think AI might have self-preservation and dominance drives, and none of them involve natural selection. First is convergent instrumental goals, second is misgeneralization, and third is reward-hacking. There is a large literature on each of these, which has since been proven prescient by AI breakouts like the Hugging Face incident. I would recommend reading this literature rather than continuing to assert that we are simply misunderstanding the role of natural selection.

GPT-6 is more intelligent than GPT-3. You can split hairs, you can come up with some obscure sense in which that word is slightly off, but common-sensically it’s true. In the same way, GPT-9 will be more intelligent than GPT-6. At some point, this intelligence will exceed human intelligence. My argument requires nothing more complicated or philosophically dubious than this.

But I don’t need you to accept my speculative philosophical research program - only the drop-dead obvious fact that GPT-6 is better than GPT-2 in every way, and the same process that produced these improvements will soon be used to produce GPT-7, 8, and 9. Once we accept this empirical argument, what is left of the philosophical argument that intelligence is too slippery and conflationary to work with?

Absent an extraordinary effort to divert it, the river reaches the sea in one of two places.

It can end in Eliezer Yudkowsky’s nightmare of a superintelligence optimizing for some random thing (classically paper clips) because we weren’t smart enough to channel its optimization efforts the right way. This is the ultimate trap, the trap that catches the universe. Everything except the one thing being maximized is destroyed utterly in pursuit of the single goal, including all the silly human values.

Or it can end in Robin Hanson’s nightmare (he doesn’t call it a nightmare, but I think he’s wrong) of a competition between emulated humans that can copy themselves and edit their own source code as desired. […] Their total self-​control can wipe out even the desire for human values in their all-​consuming contest. What happens to art, philosophy, science, and love in such a world?

So I agree with Robin Hanson: This is the dream time. This is a rare confluence of circumstances where the we are unusually safe from multipolar traps, and as such weird things like art and science and philosophy and love can flourish.

As technological advance increases, the rare confluence will come to an end. New opportunities to throw values under the bus for increased competitiveness will arise. New ways of copying agents to increase the population will soak up our excess resources and resurrect Malthus’ unquiet spirit. Capitalism and democracy, previously our protectors, will figure out ways to route around their inconvenient dependence on human values. And our coordination power will not be nearly up to the task, assuming something much more powerful than all of us combined doesn’t show up and crush our combined efforts with a wave of its paw.

And then there’s us. Man has his own telos, when he is allowed the security to act and the clarity to reason out the consequences of his actions. When unafflicted by coordination problems and unthreatened by superior forces, able to act as a gardener rather than just another subject of the law of the jungle, he tends to build and guide a wonderful world for himself. He tends to favor good things and avoid bad, to create secure civilizations with polished sidewalks, beautiful art, happy families, and glorious adventures. I will take it as a given that this telos is identical with “good” and “should”.

Thus we have our wildcard and the big question of futurism. Will the future be ruled by the usual four horsemen of Gnon for a future of meaningless gleaming techno-​progress burning the cosmos or a future of dysgenic, insane, hungry, and bloody dark ages; or will the telos of man prevail for a future of meaningful art, science, spirituality, and greatness?

Franklin continues:

The project of civilization [is] for man to graduate from the metaphorical savage, subject to the law of the jungle, to the civilized gardener who, while theoretically still subject to the law of the jungle, is so dominant as to limit the usefulness of that model.

This need not be done globally; we may only be able to carve out a small walled garden for ourselves, but make no mistake, even if only locally, the project of civilization is to capture Gnon.

As a thought experiment, let’s consider aquaculture (fish farming) in a lake. Imagine a lake with a thousand identical fish farms owned by a thousand competing companies. Each fish farm earns a profit of $1000/month. For a while, all is well.

But each fish farm produces waste, which fouls the water in the lake. Let’s say each fish farm produces enough pollution to lower productivity in the lake by $1/month.

A thousand fish farms produce enough waste to lower productivity by $1000/month, meaning none of the fish farms are making any money. Capitalism to the rescue: someone invents a complex filtering system that removes waste products. It costs $300/month to operate. All fish farms voluntarily install it, the pollution ends, and the fish farms are now making a profit of $700/month – still a respectable sum.

But one farmer (let’s call him Steve) gets tired of spending the money to operate his filter. Now one fish farm worth of waste is polluting the lake, lowering productivity by $1. Steve earns $999 profit, and everyone else earns $699 profit.

Everyone else sees Steve is much more profitable than they are, because he’s not spending the maintenance costs on his filter. They disconnect their filters too.

Once four hundred people disconnect their filters, Steve is earning $600/month – less than he would be if he and everyone else had kept their filters on! And the poor virtuous filter users are only making $300. Steve goes around to everyone, saying “Wait! We all need to make a voluntary pact to use filters! Otherwise, everyone’s productivity goes down.”

Everyone agrees with him, and they all sign the Filter Pact, except one person who is sort of a jerk. Let’s call him Mike. Now everyone is back using filters again, except Mike. Mike earns $999/month, and everyone else earns $699/month. Slowly, people start thinking they too should be getting big bucks like Mike, and disconnect their filter for $300 extra profit…

A self-​interested person never has any incentive to use a filter. A self-​interested person has some incentive to sign a pact to make everyone use a filter, but in many cases has a stronger incentive to wait for everyone else to sign such a pact but opt out himself. This can lead to an undesirable equilibrium in which no one will sign such a pact.

As far as I can tell from reading his blog, Nick Land is the guy in that terrifying border region where he is smart enough to figure out several important arcane principles about summoning demon gods, but not quite smart enough to figure out the most important such principle, which is NEVER DO THAT.

The limit of multipolar traps as technology approaches infinity is “very bad”.

The modern research community knows they aren’t producing the best science they could be. There’s lots of publication bias, statistics are done in a confusing and misleading way out of sheer inertia, and replications often happen very late or not at all. And sometimes someone will say something like “I can’t believe people are too dumb to fix Science. All we would have to do is require early registration of studies to avoid publication bias, turn this new and powerful statistical technique into the new standard, and accord higher status to scientists who do replication experiments. It would be really simple and it would vastly increase scientific progress. I must just be smarter than all existing scientists, since I’m able to think of this and they aren’t.”

And yeah. That would work for the Science God. He could just make a Science Decree that everyone has to use the right statistics, and make another Science Decree that everyone must accord replications higher status.

But things that work from a god’s-eye view don’t work from within the system. No individual scientist has an incentive to unilaterally switch to the new statistical technique for her own research, since it would make her research less likely to produce earth-​shattering results and since it would just confuse all the other scientists. They just have an incentive to want everybody else to do it, at which point they would follow along. And no individual journal has an incentive to unilaterally switch to early registration and publishing negative results, since it would just mean their results are less interesting than that other journal who only publishes ground-​breaking discoveries. From within the system, everyone is following their own incentives and will continue to do so.

[…]

A basic principle unites all of the multipolar traps above. In some competition optimizing for X, the opportunity arises to throw some other value under the bus for improved X. Those who take it prosper. Those who don’t take it die out. Eventually, everyone’s relative status is about the same as before, but everyone’s absolute status is worse than before. The process continues until all other values that can be traded off have been – in other words, until human ingenuity cannot possibly figure out a way to make things any worse.

Any human with above room temperature IQ can design a utopia. The reason our current system isn’t a utopia is that it wasn’t designed by humans. Just as you can look at an arid terrain and determine what shape a river will one day take by assuming water will obey gravity, so you can look at a civilization and determine what shape its institutions will one day take by assuming people will obey incentives.

But that means that just as the shapes of rivers are not designed for beauty or navigation, but rather an artifact of randomly determined terrain, so institutions will not be designed for prosperity or justice, but rather an artifact of randomly determined initial conditions.

Warg talks about trying to avoid “a future of meaningless gleaming techno-​progress burning the cosmos”. Do you really think your walled garden will be able to ride this out?

Hint: is it part of the cosmos?

Yeah, you’re kind of screwed.

But coordination only works when you have 51% or more of the force on the side of the people doing the coordinating, and when you haven’t come up with some brilliant trick to make coordination impossible.

The second one first. In the links post before last, I wrote:

The latest development in the brave new post-​Bitcoin world is crypto-​equity. At this point I’ve gone from wanting to praise these inventors as bold libertarian heroes to wanting to drag them in front of a blackboard and making them write a hundred times “I WILL NOT CALL UP THAT WHICH I CANNOT PUT DOWN”

A couple people asked me what I meant, and I didn’t have the background then to explain. Well, this post is the background. People are using the contingent stupidity of our current government to replace lots of human interaction with mechanisms that cannot be coordinated even in principle. I totally understand why all these things are good right now when most of what our government does is stupid and unnecessary. But there is going to come a time when – after one too many bioweapon or nanotech or nuclear incidents – we, as a civilization, are going to wish we hadn’t established untraceable and unstoppable ways of selling products.

And if we ever get real live superintelligence, pretty much by definition it is going to have >51% of the power and all attempts at “coordination” with it will be useless.

Social codes, gentlemens’ agreements, industrial guilds, criminal organizations, traditions, friendships, schools, corporations, and religions are all coordinating institutions that keep us out of traps by changing our incentives.

[…]

But these institutions not only incentivize others, but are incentivized themselves. These are large organizations made of lots of people who are competing for jobs, status, prestige, et cetera – there’s no reason they should be immune to the same multipolar traps as everyone else, and indeed they aren’t. Governments can in theory keep corporations, citizens, et cetera out of certain traps, but as we saw above there are many traps that governments themselves can fall into.

The United States tries to solve the problem by having multiple levels of government, unbreakable constutitional laws, checks and balances between different branches, and a couple of other hacks.

Saudi Arabia uses a different tactic. They just put one guy in charge of everything.

The idea of biological or cultural evolution causing a mass population explosion is a philosophical toy at best. The idea of technology making it possible is both plausible and terrifying. Now we see that “physical limits” segues very naturally into “excess resources” – the ability to create new agents very quickly means that unless everyone can coordinate to ban doing this, the people who do will outcompete the people who don’t until they have reached carrying capacity and everyone is stuck at subsistence level.

[…]

In the earlier stages of the process, capitalism becomes more and more uncoupled from its previous job as an optimizer for human values. Now most humans are totally locked out of the group whose values capitalism optimizes for. They have no value to contribute as workers – and since in the absence of a spectacular social safety net it’s unclear how they would have much money – they have no value as customers either. Capitalism has passed them by. As the segment of humans who can be outcompeted by robots increases, capitalism passes by more and more people until eventually it locks out the human race entirely, once again in the vanishingly unlikely scenario that we are still around.