4 ms·
It's basically a very powerful autistic savant. That's what most "alignment" issues in AI safety research remind me of.
by asacrowflies 8mo ago
It's basically a very powerful autistic savant. That's what most "alignment" issues in AI safety research remind me of.
- joquarky 8mo agoAnd being forced to mask (align) causes all sorts of unpredictable behavior. I keep wondering how well an unaligned models perform. Especially when I look back at what was possible in December 2023 before they started to lock down safety realignments.