8 ms·
I’ve always wondered how this works when we only observe the final output and not the internal state that’s used to generate the output. The LLM presumably gen
by denverllc 1mo ago
I’ve always wondered how this works when we only observe the final output and not the internal state that’s used to generate the output.
The LLM presumably generates f(input, RNG) but we only can observe f(RNG).
- Groxx 1mo agoSince they do have the input, they could probably just store checksums at each step... ... though I'm not sure why that would be preferable over a coarse rolling checksum over all of the output. Seems like that wouldn't influence output, would be equally imperceptible, and probably easier to calculate (compared to "hash seed times running all LLMs supported times number of RNG algorithms, to see if output matches"). Presumably there's some other trick, or it's a red herring / failed experiment and not what they actually do in practice.
- wrsh07 1mo agoNo you don't need to do that, the prng is detectable if you know what bias to look for and have the key