LLMs (and most computer applications) don’t operate with true random number generators, mostly pseudorandom generators of varying degrees of strength of randomness.
Physical dice are more likely to be actually random than most computer stuff. Unless they’re loaded or otherwise biased.
LLM output is also not an unbiased PRNG; it will give you a result based on the probabilities of the next word according to the model, which will likely be heavily biased towards more common values.
Heavily biased towards more common values of a particular subculture of a particular culture, in fact.
Want to have some fun? Ask a culturally weighted question (say about gun ownership) in English. Keep the question as neutral as possible. Watch the pro-gun stance take hold, nine times out of ten. Now ask the same question in a new session translated to idiomatic German. Again make sure the question is as neutral as possible. Compare the answers.
LLMs (and most computer applications) don’t operate with true random number generators, mostly pseudorandom generators of varying degrees of strength of randomness.
Physical dice are more likely to be actually random than most computer stuff. Unless they’re loaded or otherwise biased.
LLM output is also not an unbiased PRNG; it will give you a result based on the probabilities of the next word according to the model, which will likely be heavily biased towards more common values.
Heavily biased towards more common values of a particular subculture of a particular culture, in fact.
Want to have some fun? Ask a culturally weighted question (say about gun ownership) in English. Keep the question as neutral as possible. Watch the pro-gun stance take hold, nine times out of ten. Now ask the same question in a new session translated to idiomatic German. Again make sure the question is as neutral as possible. Compare the answers.
Granted they don’t, but can you run a 2x10^500 die? I wouldn’t trust it with my login but choice of dinner tonight…