Comment by asacrowflies
10 days ago
It's basically a very powerful autistic savant. That's what most "alignment" issues in AI safety research remind me of.
10 days ago
It's basically a very powerful autistic savant. That's what most "alignment" issues in AI safety research remind me of.
And 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.