In case Copilot was too much work
In case Copilot was too much work
cross-posted from: https://lemmy.zip/post/27030131
The full repo: https://github.com/vongaisberg/gpt3_macro
In case Copilot was too much work
cross-posted from: https://lemmy.zip/post/27030131
The full repo: https://github.com/vongaisberg/gpt3_macro
Is this the freaking antithesis of reproducible builds‽ Sheesh, just thinking of the implications in the build pipeline/supply chain makes me shudder
Just set the temperature to zero, duh
When your CPU is at 0 degrees Kelvin, nothing unpredictable can happen.
Looking at the source they thankfully already use a temp of zero, but max tokens is 320. That doesn't seem like much for code especially since most symbols are a whole token.
Just hash the binary and include it with the build. When somebody else compiles they can check the hash and just recompile until it is the same. Deterministic outcome in presumambly finite time. Untill the weights of the model change then all bets are off.
this is how we end up with lost tech a few decades later
You'd have to consider it somewhat of a black box, which is what people already do.
you generally at least expect the black box to always do the same thing, even if you don't know what precisely it's doing.
someone post this to the guix mailinglist 😄