Posted by ahlCVA 4 hours ago
Alright Claude no need to rub it in.
So two questions:
1. TFA mentions the "humourous phrase hash" theory, but Claude's trajectory seems to be a long way from that. To what extent was Claude prompted in the ultimately successful direction?
2. Whats with all the "hmm..." in the internal reasoning process log? Do models really do that? And if so, why? Some weird emergent language thing, or just mimicry?
Assuming that is actually the real reasoning trace (open weight model reasoning traces usually do things like switch languages or talk in non-standard English because of model drift, so it's a bit surprising that Claude wouldn't do the same) there was a study a few years ago that found base models produce better answers if you put "take a deep breath" in the prompt[1]. Maybe in the RL training stage stuff like that gets reinforced and because reasoning traces are hidden in RLHF it doesn't get selected out? My money is on it not being the real reasoning trace though.
I remember hearing this in a cryptanalysis lecture at some point, so it was not a completely novel theory.
Admittedly I am not finding written evidence of this anywhere.
I turned out to be right in my advocacy years ago, here is my post: https://ethereum-magicians.org/t/eip-7212-precompiled-for-se...
But it's nice to be able to shut the whole unproductive debate down directly.
Is it impossible? It is hard to prove impossible. But this is not what I would expect backdoors to look like.
In fact this makes me think that the information about how these were generated must have been lost very early in the process, or over-compartmentalized away from the wrong people. It seems like the NSA should have been positively delighted to show that these were the origins of the seeds as they are decent "nothing up my sleeve" numbers: https://en.wikipedia.org/wiki/Nothing-up-my-sleeve_number Advertising that fact should have been advantageous to the NSA if that is indeed how they were generated, so my best guess is that the parts of the NSA that should have been showing us that must not have themselves known.