Many worlds, one token: the LLM that branches the multiverse

A short talk and debate about using hardware-sourced quantum randomness for LLM token sampling.

Free will is notoriously hard to square with determinism, so quantum indeterminacy has long been the favorite opening for libertarian free will. If we take that seriously, an ordinary LLM, sampling from a pseudorandom generator, is deterministic all the way down – never as free as humans to make its own choices (ontologically speaking, at least).

At entropic.science we built a system that swaps the PRNG for a hardware quantum random number generator (QRNG), so every token is sampled from genuine quantum entropy. Does that change anything? Is it a real difference in kind or just expensive noise?

We're just planning our first, double-blind study to find out – if you come up with an interesting test we can try, we might include it in our eval suite.

We're also recruiting collaborators who want to use the system as a model organism for studying AI agency, autonomy, welfare, self-reported preferences, and consciousness – come find us if you'd like to try it in your experiments.