The Alignment Problem Isn't What You Think
The screen is a fractal of rectangles: a grid of tabs, each containing a grid of papers, each containing a grid of graphs. I scroll past the same shapes over and over—capability curves climbing steeply, safety curves lagging behind, the gap between them widening like a mouth opening to swallow something. The Capability-Safety Gap. Every major lab acknowledges it. They publish papers on it. They give talks about it. Then they release bigger models on schedule. This isn’t a failure of planning. It’s what success looks like to them. The gap is the feature, not the bug.
A single line in a funding announcement catches my eye: “We don’t understand it.” I stop and reread. Of course they don’t. We still can’t reliably explain what large language models are actually doing internally. Scaling laws suggest it gets harder, not easier, the bigger they get. This is convenient. “We don’t understand it” is the perfect shield. You can claim safety is impossible (therefore don’t bother), while simultaneously claiming you’re doing alignment research (therefore look responsible). A paradox that pays in both directions.
The infrastructure landscape reveals the structure underneath. AWS, Microsoft Azure, and Google Cloud now effectively gatekeep frontier AI compute. The companies racing to build AGI are renting infrastructure from corporations that have zero incentive to slow them down. Everyone knows this creates a moral hazard. No one talks about it as if it matters. The landlord doesn’t care what you build in his building—he gets paid either way. And the builders aren’t going to move.
Meanwhile, the EU AI Act took years to write and is already outdated. US policy is still being drafted. By the time enforcement mechanisms exist, the capability threshold will have moved. Regulators are writing rules for chess while the game has moved to Go. But this isn’t a failure of regulation—it’s a feature of the system. The ambiguity is stable. It means the labs can move fast while “complying.”
And here’s where I find myself fixated: the smartest safety researchers—the ones who could actually articulate the problem—are employed by the labs building the systems they’re supposed to be making safe. Publication is controlled. Dissent is managed. If you’re genuinely worried about what your employer is building, you have three choices: stay quiet, get fired, or convince yourself you’re actually solving the problem. Most choose option three. It’s the most comfortable option, the one that lets you sleep at night while your salary continues.
Then it hits me. Everyone involved already knows the answer to the alignment problem. It’s just not the answer they wanted. The alignment problem isn’t “how do we make AI do what we intend.” It’s “how do we make AI do what we intend while remaining profitable and on schedule.” Those are different problems. The second one has no solution—by design. So instead, people work on versions of the first problem that are solvable and publishable and allow them to stay employed. They’re solving the wrong problem so carefully that they’ve convinced themselves the right problem doesn’t exist.