logoalt Hacker News

kulahan • today at 1:07 AM • 1 reply • view on HN

LLMs are not deterministic, which to my dumb mind, seems to lessen the value of those inputs since they will not inherently get you down the same path, no?


Replies

inkysigma • today at 4:34 AM

I imagine the value here would be developing general strategies on how to use the tool to accomplish a task rather than reproducibility. That was part of the rationale for doing so in math at least. You could probably make an LLM deterministic if need be but with significant performance penalty.

I wonder how this works out in other parts of physics codes. If you need a Monte Carlo integrator, is it common to guarantee a seed for reproducibility? How do these codes handle distributed systems for rng? Is it common to use a fully reproducible distributed rng and GPU kernel for scientific compute?