Austin Spizzirri

We’re building machines that appear smarter and smarter. They’re not. They’re becoming more and more capable, which isn’t the same thing. Current AI systems have a lot of data and they piece it together well enough to pass a bar exam, but that doesn’t make them intelligent. They predict and pattern match, which is part of what intelligence does, but without any underlying model of reality grounding it. No active maintenance, no felt sense of when something’s wrong, no ability to update what they know from experience. They produce outputs that look like comprehension, but there’s no one home. Most people, including most people building them, don’t fully grasp that distinction. That’s half the problem. The other half is alignment itself.

Alignment is the question of how to make sure these systems actually care about what we care about. Not perform caring. Not optimize for looking aligned while pursuing something else beneath the surface. Actually care, in a way that holds when they’re more capable than us and we can’t course-correct anymore. The standard approach is to specify values and train systems to follow them. I think that’s a dead end. You can’t encode morality into something the way you upload a file. Values that matter, the kind worth having, emerge from lived experience, from relationships with stakes, from being in a world where your choices cost something.

I write essays about all of this. They live in Writing. I run two companies: Velanox Systems, my research org, where I work on the alignment problem, and MIDAS Designs, a web design studio. I also sometimes make music as Spizz.