3 ms·
If you can prove the LLM was not trained on the code it is reproducing, and has never seen the code as part of a spec to follow, I don't see a problem. Proving
by ori_b 7mo ago
If you can prove the LLM was not trained on the code it is reproducing, and has never seen the code as part of a spec to follow, I don't see a problem.
Proving this is going to be hard with current "open source" models.
- eschaton 7mo agoIndeed, you have to prove that the LLM is generating code from a specification. Right now they don’t do that; what they do is regurgitate portions of their training data based on correlations with input tokens. Put the programmer’s reference for the Digital Equipment DEQNA QBus Ethernet adapter in your favorite slop tool and tell it to make a C or C++ implementation for an emulator, and you know what you get? Code from SIMH. That’s not “generating,” that’s “copying.”