I’ve stepped back from thinking about ML and alignment the last few years, so I don’t know how this fits into the discourse about it, but I felt like I got important insight here and I’d be excited to include this. The key concept that bigger models can be simpler seems very important.
In my words, I’d say that when you don’t have enough knobs, you’re forced to find ways for each knob to serve multiple purposes slash combine multiple things, which is messy and complex and can be highly arbitrary, whereas with lots of knobs you can do ‘the thing you naturally actually want to do.’ And once you get sufficiently powerful, the overfitting danger isn’t getting any worse with the extra knobs, so sure, why not?
I also strongly agree with orthonormal that including the follow-up as an addendum adds a lot to this post. If it’s worth including this, it’s worth including both, even if the follow-up wasn’t also nominated.
I’ve stepped back from thinking about ML and alignment the last few years, so I don’t know how this fits into the discourse about it, but I felt like I got important insight here and I’d be excited to include this. The key concept that bigger models can be simpler seems very important.
In my words, I’d say that when you don’t have enough knobs, you’re forced to find ways for each knob to serve multiple purposes slash combine multiple things, which is messy and complex and can be highly arbitrary, whereas with lots of knobs you can do ‘the thing you naturally actually want to do.’ And once you get sufficiently powerful, the overfitting danger isn’t getting any worse with the extra knobs, so sure, why not?
I also strongly agree with orthonormal that including the follow-up as an addendum adds a lot to this post. If it’s worth including this, it’s worth including both, even if the follow-up wasn’t also nominated.