We head downstairs to a dark basement attached to a garage. A slight, long-haired man, a paid test subject solicited through Craigslist, sits before a computer screen, wearing a white cap that looks like a medieval linen coif threaded with wires. The screen flashes images — basil, a blazer, Parmesan cheese. With unsettling clarity, the computer will be able to resurrect the image from electrical signals in the subject’s brain. A subject considers a picture of jelly beans. AI offers a picture of similarly colored beads. A subject looks at a red station wagon; AI presents a red sedan. Until very recently, most people thought the data produced by EEGs, an 80-year-old technology, was noisy garbage. “They just didn’t understand the power of large language models,” Jonathan says. He is 24 years old.Reading minds is what AI engineers mean when they talk about hard problems. Eventually, the tech will advance to interpret “evoked states.” “So we start with, you know, discrete smaller tasks like emotion, like positive, negative, maybe now ten, 20 emotions. And then we add more dimensionality so that eventually we can go into full sort of inner monologue,” a world of superior self-knowledge wherein we sift through our own memory banks rather than selectively recall events through a haze of misperception. “And,” he says, “we do it all in-house.”Doing it all in-house looks like this: a server rack with LED-lit fans in the garage next to some exercise equipment and some bicycle helmets. Jonathan and his housemates built the rig themselves. “Just asking ChatGPT basically. You know, you can just ask and then order the parts you need and you learn and you debug.”
How do you make contact with the intelligence rising up from the machines around you? Do you build it a body? Do you offer it yours? It is perhaps tedious to point out that we are always operating under the shadow of destruction, deploying tools that might end us, convincing ourselves, not without reason, that if we don’t build the bomb, someone with worse intentions will. Not a single one of the AI kids had attempted to lecture me about a theory or suggested I read a paper; it was not me they were trying to program. Somewhere along the way, drawn into their swell, I had begun to think of large questions about the nature of AI as New York questions, millennial questions, distant from the center of things. Where it mattered, humans were not debating AI; they were merging with it. You order the parts you need, you learn, you debug. The kids carry on with the crisp clarity of engineers, integrating what is immediately useful, discarding or rewriting what is not. No one will ask your permission to build a world you do not understand.
It's like they're living in an 80s William Gibson novel. Working in spaces where policy and regulators can't really stop them. They don't care about policy papers. They're just building.
Those spaces matter. Policy is so impossibly stupid.
Here, Parliament's trying to figure out the best way of breaking the internet, with National wanting to copy Australia's insane social media age-gating bill, and Labour wanting to copy the UK's even worse version. News outlets cheer them on, having always figured that social media platforms are the enemy anyway. Both parties see it as crack cocaine for attracting pivotal female votes in the 35-55 age bracket.
And I understand that Parliament's select committee looking into this is wondering whether it's possible to regulate VPNs. Because that's been the obvious workaround in the UK for adults who don't want to have to show ID to visit every darned website. So long as at least one country remains free, VPN to it and pretend to be from there.
I'm cheering for the cyberpunks. They may be our best hope.
