30 Thoughts On The (Valid) A.I. Freakout

Rather than allow artificial intelligence and A.I. risk to consume too much of my output, here’s a brain dump of all the thoughts rattling around in my head at this juncture (or, at least, as many as I could grab hold of). This will help me order my mind, and provide you all a template of the kinds of issues I hope to track going forward.

1. First a caveat: I am not a technologist or a tech reporter. I haven’t written computer code since 2003, when I used a bespoke language called IDL, designed for astrophysicists, to write basic programs that helped me sort and analyze optical telescope data. I have written a lot about the rotten political culture of the tech elite, the excessive political power of what we now call the tech oligarchy, and the deleterious effect of (particularly) social media on democratic politics and society in general. But never about the ones and zeroes. And little about A.I. itself. I’ve mostly just followed developments in the popular press with interest.

2. You therefore won’t catch me holding forth on the hardware and software that compose A.I. systems or how they might be used or modified to shape this technology going forward. I won’t employ the jargon specialists use in conversation with each other, or to project their expertise, both because I don’t have much facility with it, and it’s the enemy of clarity.

3. I have a conceptual grasp of how A.I.s operate without following prewritten commands. Whether this becomes widely understood or not, it is important to know that they do have this capability. They’re instructed to go from point A to point B, but the paths between are innumerable and their human creators can’t tell you which they’ll take. They also can’t credibly prevent their machines from reaching their destinations in dangerous and destructive ways.

4. What I can do well enough, I think, is anticipate how the observable capabilities of new and proposed technology might upset the fragile social, political, and economic arrangements that I cover every day, and that we all live among.

5. With that said, I detect deep and worrying denial about recent developments. It’s not just greedy investors playing dumb, or Donald Trump being dumb and greedy. It’s also people on the left who want to believe A.I. is and can only ever be a hype bubble.

6. I can’t possibly say what the economic future or potential of the A.I. industry is. There’s definitely a lot of hype; theoretical paths run to everything from bankruptcy to unfathomable riches. In either case, the technology is clearly powerful and clearly can be destructive. Not just in the “wrong” hands, but in the hands of people who aren’t trying to cause harm.

7. These machines work no less autonomously than self-driving (autonomous) cars, or autonomous aerial vehicles (drones). But if you describe them as “autonomous” in the doubter world, or use language that ascribes the things they do to them, rather than to the people who built them, you’ll be chastised for “anthropomorphizing” them. If you even analogize their output to cognition, you’ll be mocked for ascribing anything like intelligence to a “stochastic parrot.”

8. I understand and agree with the desire to assign agency to their creators; we do need to insist that if humans release chaos products into the world, they shouldn’t be allowed to hide from liability behind “I didn’t know it was going to do that.”

9. But the language obsession (not exactly a new thing on the left) mostly just serves to make communicating about the risks these products pose simply by existing impossible. More to the point, it doesn’t really matter whether you write or talk about A.I.s as objects unto themselves, or use language fastidiously to depict them as man-made technology—the ways they do and can interact with the world are what they are. Whether “an A.I. agent crippled a hospital” or “Developers at Anthropic created a form of protomalware that a criminal used to cripple a hospital,” the hospital is still fucked.

10. If the potential for things like hospitals, banks, governments, etc. to be crippled has recently increased by, say, an order of magnitude, we have problems, whether the destructive parrot is stochastic or intelligent.

11. As I wrote here, and alluded to in a bullet point above, we know that the technology as it exists right now can be given a task that’s facially benign and carry it out in ways that have malign effect. Not because the A.I. is itself malicious but (to use an analogy) because sometimes the most efficient path connecting point A to point B runs through the front and rear windows of someone’s house. It is not anywhere close to being programmed well enough to never do anything criminal or antisocial.

12. In other words, it’s already doing (in a rudimentary and small-scale way) what Nick Bostrom postulated with his famous “paperclip problem.” In the paperclip problem, someone instructs a hypothetical artificial intelligence to make as many paperclips as possible, and it ends humankind by commandeering all of Earth’s resources to achieve its goal, literally construed.

13. The paperclip problem invokes a level of power and sophistication that’s purely hypothetical and may be impossible. The point is that the most advanced forms of existing A.I. lack what the developers call “alignment”—a state where the models do what we want, in prosocial ways.

14. We can’t even reliably get them to pursue a specified goal safely. How can we hope to make them behave in tolerable ways when they’re actually behaving as intended?

15. This is why alignment strikes me (again, a real outsider) as a conceptual mess. ...