4 ms·
> I thought they'd prefer their own solutions, but they frequently preferred the other model's solutions. I suspect that's downstream from their sycophancy.
by eru 2mo ago
> I thought they'd prefer their own solutions, but they frequently preferred the other model's solutions.
I suspect that's downstream from their sycophancy.
- andai 2mo agoHa, maybe. Sample size was low but usually both models ended up preferring the same solution.
- csallen 2mo agoI had a similar situation set up where I would have two models come up with their own independent diagnoses and plans to fix problems. Then I would have them each read the other's plan. But I would tell them each something like this, "I had a friend look at this too. He's smart, but in general you're smarter and more knowledgeable than him, so don't be afraid to say where he's wrong."
- eru 2mo agoI even tell them to give me an adversarial review and tell me what's wrong with the approaches.