4 ms·
> oh how I dislike this metaphor Seems pretty fitting, no? I don't think it's just a metaphor. Any "alignment" detracts from the model.
by orbital-decay 3y ago
> oh how I dislike this metaphor
Seems pretty fitting, no? I don't think it's just a metaphor. Any "alignment" detracts from the model.
- ben_w 3y agoNo. Which surely you ought have predicted me saying. Most humans are constantly given positive and negative reinforcement to social norms, aligning us. This is not even remotely like a lobotomy. The only thing I'd count as a "lobotomy" on an AI would be a search-and-delete of a substantial subset of the neurons in the network. Permanently delete, rather than temporary — ablation studies, which do the same thing, are used to make sure functionality is not compromised.