I'm not (at least not yet) persuaded by the argument that the output of an LLM is a derivative of all the material it was trained on. The problem with the argument is that it also applies to me: if I see some cool code you wrote, you bet I'm going to make a mental note of it. The trip hazards here are patents and non-free licenses, but they also apply equally to people.
There are other problems with LLMs, but I'm not convinced by that one specifically.
@mariusor Exactly! There are patents and licenses to prevent people from using other people's code, which requires measures like clean room if you want to implement the same functionality. Others which are designed to facilitate it. I don't see what makes LLMs unique in this.
@mattb to me it's pretty clear that you can't trust what an LLM produces exactly because it's not clear room, because it's a derivative of all the code it was trained on.
@mattb I doubt that you have seen the same amount of code as an LLM, or that you "internalized" it in a similar way that it "has". Also the difference between an LLM and a person is that one can understand the context in which it programs and can make a decision if a "clean code" approach is required or not.
@mattb in the end I guess the difference is that with a person there's a non-zero, but small, chance that they produce copyright questionable source code, but with the LLM that chance is much higher(*) at a level that lawyers get uncomfortable with.
depending how close aligned the problem is with the training material.
Add comment