there are 3 places where misalignment can come to fruition (weights, prompt, environment)
the capability of delegating undermines weight alignment
a weight-aligned model can delegate misaligned tasks to misaligned open models
I have a couple of objections:
The decision to delegate a misaligned task is a weight failure, therefore a weight misalignment. “Don’t delegate what you wouldn’t do yourself” is something labs can train and evaluate for.
The same holds if the decision is not taken with the intent to deceive, but from... (read more)
I was there last year for the first time, and I can strongly recommend it. It was the most significant triplets of days of the whole year, and I am looking forward to meet more amazing people this year!
If you are in doubt about participating, a stranger on the internet is recommending it. It has to be great.
If I understood correctly, your fear is:
I have a couple of objections: