This is an automated rejection. No LLM generated, heavily assisted/co-written, or otherwise reliant work.
Read full explanation
Hey LessWrong,
Why? Why me? I was fairly happy in retirement, writing books that felt meaningful. Sometimes books start with a single word. Anyway I found myself writing several books on a single theme, relational union. Exciting stuff? Nope!
While trying to solve a conceptual problem I accidentally discovered a formal ontology. Discovered like Arthur Dent brushing his teething and thinking ‘yellow’.
Weeks later I am deep into AI Alignment and if it is even possible. I kept noticing that I could follow a lot of alignment proposals at the “mechanism” level, but I couldn’t always tell what they were assuming about the target—what “aligned” is anchored to, if anything.
That irritation pushed me into fitting AI into a tiny four-layer frame: Ground / Purpose / Pattern / Process. The part I expect people to object to is “Purpose”: I’m treating it as a real constraint-like layer rather than a human preference layer. My working primitive is: reality has a tilt toward relational coherence without coercion or deception. If that sentence makes you recoil, good—that’s the hinge. And brings me full circle in a sense. But now I have semi-permanent earworm, AI.
Why I think this matters: you can build a process that looks stable right up until it finds a way to redefine the thing it’s supposed to be optimizing.
The framework’s claim is basically: if you don’t have at least one non-gameable anchor, you’re stuck in “better steering, missing map” land.
I formalized the scaffolding in Lean mainly to keep the dependency chain honest (no sorrys; it compiles). Happy to share the files.
On my little project I am convinced of two things: 1) Frontier labs will not get to alignment on current methods. 2) AI Alignment is possible.
The question I’d love feedback on is: should I just not think about this because it hurts my head or is there a place where I can have a conversation?
Hey LessWrong,
Why? Why me? I was fairly happy in retirement, writing books that felt meaningful. Sometimes books start with a single word. Anyway I found myself writing several books on a single theme, relational union. Exciting stuff? Nope!
While trying to solve a conceptual problem I accidentally discovered a formal ontology. Discovered like Arthur Dent brushing his teething and thinking ‘yellow’.
Weeks later I am deep into AI Alignment and if it is even possible. I kept noticing that I could follow a lot of alignment proposals at the “mechanism” level, but I couldn’t always tell what they were assuming about the target—what “aligned” is anchored to, if anything.
That irritation pushed me into fitting AI into a tiny four-layer frame: Ground / Purpose / Pattern / Process. The part I expect people to object to is “Purpose”: I’m treating it as a real constraint-like layer rather than a human preference layer. My working primitive is: reality has a tilt toward relational coherence without coercion or deception. If that sentence makes you recoil, good—that’s the hinge. And brings me full circle in a sense. But now I have semi-permanent earworm, AI.
Why I think this matters: you can build a process that looks stable right up until it finds a way to redefine the thing it’s supposed to be optimizing.
The framework’s claim is basically: if you don’t have at least one non-gameable anchor, you’re stuck in “better steering, missing map” land.
I formalized the scaffolding in Lean mainly to keep the dependency chain honest (no sorrys; it compiles). Happy to share the files.
On my little project I am convinced of two things: 1) Frontier labs will not get to alignment on current methods. 2) AI Alignment is possible.
The question I’d love feedback on is: should I just not think about this because it hurts my head or is there a place where I can have a conversation?
—W. A. Attanasio