|
 |
Killer Robots Aren't The Threat. We Are.
Tech execs made a devil’s bargain with a dim, amoral, vain man, and handed him the Superintelligent, Amoral, Sycophantic Robot Conundrum. We need to solve that problem before we get to “alignment.”
|
| |||||||||||||||
 |
This week’s Politix was already in the can when Jacob Coxon, the 27 year old artificial intelligence researcher, announced he’d resigned from Anthropic—arguably the most advanced A.I. development company in the world. Coxon chose to work there because Anthropic is also the most conscientious of the most innovative labs—at least, that’s how its executives have presented it. But he concluded that Anthropic, too, is acting irresponsibly, racing to unleash an extinction-level threat into the world.
Several of Coxon’s colleagues in the industry, and even at his own former company, quickly chimed in—not to write him off as some crank or neurotic, but to say, in essence: He’s right. This is getting out of hand.
The basic shape of the dilemma, as they describe it, is that the top A.I. labs are competing with each other to dominate what they call the frontier—to build the most capable A.I. technology, in the race toward “superintelligence.” They want to corner as much of the market as possible within the industry; but they also view themselves as discrete Manhattan Projects in a global arms race to build enough A.I. capability to establish industry dominance for the United States and deterrence against hostile nations and non-state actors. So absent domestic regulation, all incentives point to aggressive competition to expand the frontier; and absent a global treaty to slow or pause A.I. development, the U.S. won’t (and maybe shouldn’t) regulate A.I. unilaterally.
And so, almost in spite of themselves, these technologists and researchers will continue to accept large paychecks to build a technology that they believe imperils life on Earth. They’ll build and build until the robots achieve superintelligence, even if they get there before they’ve encoded them with human-like or pro-social reasoning (alignment), in which case they might easily “decide” to eliminate humans. They might do this because they perceive humans as an existential threat to the A.I. “race,” or they might do it as the most straightforward solution to some workaday problem.
There are two points I want to make: One is that the above dilemma, as presented by A.I. developers, is not as sticky as they believe it is—or at least as they want us to believe it is. It is not a prisoner’s dilemma-like problem of global coordination, so much as a domestic corruption problem that they don’t want any part in solving. The second, though, is that they’re dooming past the graveyard. Or perhaps a better coinage is to say they’re missing the trees for the forest.
Obviously if A.I. presents existential risk then its development has to be paused at least. Failure to pause would be evil. But while we’re all still alive, there’s still society and civilization to uphold, and for all the anguished debate over the existential risk posed by future models, these same people seem incredibly blase about the destabilizing forces they’ve already unleashed.
The state-actor concern—really a China concern—isn’t entirely frivolous, nor is the deterrence concern. It’s easy to caricature—and some people, like Ted Cruz embrace the caricature unironically.
“One of the real challenges is [that] China is going full speed ahead, and whatever we do here, China is not going to stop,” Cruz said. “If there are going to be killer robots, I would rather they be American killer robots, rather than Chinese killer robots.” We’ll all be dead, but at least our machines will have done the killing!
Obviously if we reach the point where an advanced A.I. breaks containment and sets out to end humanity, it won’t matter at all whether it’s a Chinese or American A.I.
But shy of that event horizon, the situation looks a lot more like conventional deterrence. If we unilaterally disarm, then we’re defenseless against any of the ways China or any other adversary might deploy A.I. as a weapon, and we won’t be equipped to retaliate.
So it’s not entirely stupid. But it’s still special pleading. For instance:
The U.S. is way ahead, so we can absorb the strategic costs of a pause;
Chinese progress is at least in part derived from U.S. progress, so a unilateral U.S. moratorium should tend to slow them down;
There’s no reason to assume China would refuse to pause;
If they did, a unilateral U.S. pause would place China under immense global pressure and scrutiny, as any spillover catastrophe would be easily traceable to them.
China already suffered strategic humiliation once this decade by failing to contain COVID-19 then wiping its tracks. Does it want to risk something similar? Something worse?
The real problem, which none of the engineers sounding the alarm is likely to acknowledge, is that Donald Trump is president, cloaked with impunity by Republicans in Congress, and several of the executives overseeing their work are people of low character. Many of them have made devil’s bargains with Trump himself.
This is a point I tried to make, perhaps inartfully, on that Politix episode.
