Posts
Wikitag Contributions
Human: "Look, can't you just be normal about this?"
GAA-optimized agent: "Actually-"
Hm, I guess this wouldn't work if the agent still learns an internalized RL methodology? Or would it? Say we have a base model, not much need for GAA because it's just doing token pred. We go into some sort of (distilled?) RL-based cot instruct tuning, GAA means it picks up abnormal rewards from the signal more slowly, ie. it doesn't do the classic boat-spinning-in-circles thing (good test?), but if it internalizes RL at some point its mesaoptimizer wouldn't be so limited, and that's a general technique so GAA wouldn't prevent it? Still, seems like a good first line of defense.
The issue is, from a writing perspective, that a positive singularity quickly becomes both unpredictable and unrelatable, so that any hopeful story we could write would, inevitably, look boring and pedestrian. I mean, I know what I intend to do come the Good End, for maybe the next 100k years or so, but probably a five-minute conversation with the AI will bring up many much better ideas, being how it is. But ... bad ends are predictable, simple, and enter a very easy to describe steady state.
A curve that grows and never repeats is a lot harder to predict than a curve that goes to zero and stays there.
Can I really trust an organization to preserve my brain that can't manage a working SSL certificate?
1
I don't understand it but it does make me feel happy.