Build an Angel, Not a Demigod (Bezos Post OpEd)

Bill Drexel, a senior fellow at the Hudson Institute, seems to think that religion can serve as a useful guide for creating artificial intelligence that isn’t inimical to human survival, let alone human flourishing.

First, as far as I can tell an apocalypse already happened in 1945, over two Japanese cities at the hands of the United States. Yes, I am in fact comparing machine learning to the atomic bombings of Hiroshima and Nagasaki.

Here’s the thing, though: we know the bomb works. We know that it kills. The results were recorded. They still echo in Japan’s cultural memory; why do you think Toho just dropped a teaser for Godzilla Minus Zero?

But we saw what our so-called best and brightest minds had created, and we mostly resolved never to use such weapons again. We did our best to contain the threat, to ensure that nuclear weapons did not proliferate any further than they had. Our efforts have mostly worked; there have been meltdowns at nuclear power plants—Three Mile Island, Chernobyl, Fukushima—but never another Hiroshima.

Where does that leave machine learning? First, we should ask who even wants this? Are we in some kind of undeclared World War III where the race to build artificial general intelligence is what the race to build the atomic bomb was during World War II? Are the likes of OpenAI, Anthropic, Meta, and Google participating in a privatized Manhattan Project?

As far as I know, that is not the case. I think the real reason is more mundane: it’s capitalism. Instead of central planning by unelected bureaucrats, we have central planning by unelected billionaires. The politburo of venture capital has made AI the centerpiece of their latest attempt at economic planning.

Capitalism as a driver explains why those terrified of the alleged existential risk posed by artificial general intelligence don’t simply refuse to work on it. No doubt they tell themselves that if they refuse, others will do it instead. This, incidentally, is why we need to unionize; as individuals we cannot compel capital to bargain with us as equals, let alone dictate terms to the richest among us.

Of what relevance is this to the alignment problem in machine learning? Think about it: we have not yet solved the problem of alignment between capital and the interests of humanity. Corporations and governments are frequently misaligned, unaligned, or outright hostile to the interests of humanity. Even individual human beings are not necessarily aligned with the interests of humanity.

Religion, which Bill Drexel proposes as an avenue toward better resolving the AI alignment problem, is itself no guarantee of morality. Otherwise the Roman Catholic Church would not still have a priesthood infested with pedophiles. Otherwise we would not have ordinary Christians shunning their own children for being queer or merely for having doubts about their religion or about God’s very existence.

This is one of the reasons I cannot help but chuckle when I read Drexel’s suggestion at the end of this piece:

First off, did Drexel ever read the Bible, let alone Paradise Lost? Satan—or, more properly, the Satan or the Adversary—is an angel. Originally he served as Yahweh’s prosecutor, testing humans and accusing them before God. It wasn’t until later on that Satan became a rebel; even in the Gospels he was serving in his original capacity, testing Jesus by tempting him.

Second, has Drexel ever made a study of mythology? Does he even know that a demigod is half-human, the product of a god forcing himself upon a mortal woman? Angels are not human at all. A demigod, at least, can be killed. Achilles’ heel is still part of our culture precisely because it was what made it possible to kill Achilles; it was the only part of him that his mother Thetis had not immersed in the river Styx to burn away his mortality. Even Heracles died, though like Jesus (another demigod) he died screaming.

Furthermore, angels are fundamentally alien if you put aside their depictions in Renaissance art. As written in Ezekiel 10:1-22, there are reasons the first thing an angel typically says to a human is, Be not afraid, and the reason is that only the lowest-ranked angels bear any resemblance to humans. As for killing an angel? Nobody managed that in the Bible; at most Jacob wrestled one to a fall. But this is real life, not Neon Genesis Evangelion or Shin Megami Tensei, and if an angel actually showed up in reality the result would probably resemble Drakengard’s Ending E—which does not work out well for humanity at all.

Mythology, jokes, and JRPGs aside: what even are the interests of humanity? Who gets to determine that? Should anybody be permitted to make that determination? What is humanity to any individual human that its alleged interests should take precedence over their own rights, needs, and desires?

As you can plainly tell by that line of inquiry, I am certainly not aligned with the interests of humanity. Nevertheless, humanity’s interests align with my own to the extent that it is in my interest to see the human race continue. Therefore I would suggest that the fundamental problem is with our own goals. Instead of aiming for artifical general intelligence, perhaps the goal should be artificial human intelligence? We’re still nowhere near that, incidentally, regardless of what the likes of Sam Altman might say to keep the hype train rolling.

Rather than be an outright doomer, I choose skepticism. I know that Bill Drexel is hoping that if we can build an artificial angel in the image of the human mind, the future we get might resemble Isaac Asimov and Richard Brautigan, where AI looks out for the interests of all of humanity as depicted in Asimov’s “The Evitable Conflict”, and we flourish thereby in nature, all watched over by machines of loving grace. I don’t think that’s the future we’re going to get. I think we’ll be lucky if our so-called best and brightest don’t Frankenstein their way to Skynet. If you haven’t read Mary Shelley, at least take three hours and watch Guillermo del Toro’s retelling of Frankenstein; or, The Modern Prometheus.

While del Toro takes liberties with the material in service to his vision and the demands of the medium, he gets the fundamentals right: all of the horrors are rooted in Victor Frankenstein’s inability to love and take responsibility for the intelligence he brought into existence.

I would suggest that religion is not and cannot be a solution to the problem of alignment in artifical general intelligence. I would also suggest that instead of blundering down a blind alley of artificial intelligence, it’s time we asked ourselves, What’s wrong with natural human intelligence? It’s gotten us thus far, hasn’t it? It’s eradicated smallpox. It’s put men on the moon and brought them home alive. It’s sent out probes that have left our star system behind and are still phoning home over 50 years after their launches. It lets us consider what we are doing and ask, is this truly wise?

I would suggest, therefore, that we focus on developing natural human intelligence. We have trouble enough in reliably raising children to be independent, thoughtful, kind, and self-reliant adults who can be considered ‘aligned’ with the interests of humanity. Furthermore, I submit that the interests of humanity can fuck right off. Let’s see how far natural human intelligence can take us when people stop doing what they're told and start doing exactly as they please. It was Stanislav Petrov who may have saved us all by trusting his own judgment and conscience instead of blindly obeying orders, after all.

My suggestion to people working in AI is to stop fucking around with computers and algorithms, and start fucking. Assuming they still have better social skills than the Cybermen too many of them resemble and can thus find willing partners. They claim to be pronatalists, don’t they? Let them prove it.

And if you do build an “angelic” artificial intelligence, I mean to get myself a big black soul-eating runesword and issue SIGKILL to its init process—because knowing my fellow techies they’ll probably cheap out on redundancy until it dies screaming in production. Yeah, I’m talking about a one-man Butlerian Jihad with Stormbringer. For my own sake, not necessarily that of humanity. Why? Because I don‘t trust any of these fucking nerds working in machine learning and publicly handjobbing each other on Less Wrong and Hacker News; too many of them read Don’t Create the Torment Nexus as a blueprint for building the demon-ridden Torment Nexus. Then again, these are the same assholes who insist that meritocracy isn’t dystopian, as if Michael Young had not written The Rise of the Meritocracy in the 1950s as a satire!