You’re still correct. The thing about LLMs is that they’re statistical models that output one of the most likely responses, from the list of most likely responses. It still has some randomness though. You can tune this, but no randomness is shit, and too much randomness sometimes generates stupid garbage. With a large enough output, you’re statistically likely, with any randomness, to generate some garbage. A compiler is sufficiently large and complex that it’s going to end up creating garbage somewhere, even if it’s trained on these compilers specifically.
that’s a great point but wouldn’t the output for a solved problem like “make a working C compiler in rust” work better if the temperature/randomness were zero? or am i fundamentally misunderstanding?
Probably. At that point you might as well just copy/paste the existing compiler though. The temperature is basically the thing that makes it seem intelligent, because it gives different responses each time, so it seems like it’s thinking. But yeah, having it just always give the most likely response would probably be better, but also probably wouldn’t play well with copyright laws when you have the exact same code as an existing compiler.
This is so fucked up. The AI company has the perfect answer and yet it rolls the die to recreate the same thing by chance. What are they expecting, really?
any llm must have several C compilers in its training data, so it would be a reasonably competent almost-clone of gcc/clang/msvc anyway, right?
is what i would have said if you didn’t put that last part
You’re still correct. The thing about LLMs is that they’re statistical models that output one of the most likely responses, from the list of most likely responses. It still has some randomness though. You can tune this, but no randomness is shit, and too much randomness sometimes generates stupid garbage. With a large enough output, you’re statistically likely, with any randomness, to generate some garbage. A compiler is sufficiently large and complex that it’s going to end up creating garbage somewhere, even if it’s trained on these compilers specifically.
that’s a great point but wouldn’t the output for a solved problem like “make a working C compiler in rust” work better if the temperature/randomness were zero? or am i fundamentally misunderstanding?
Probably. At that point you might as well just copy/paste the existing compiler though. The temperature is basically the thing that makes it seem intelligent, because it gives different responses each time, so it seems like it’s thinking. But yeah, having it just always give the most likely response would probably be better, but also probably wouldn’t play well with copyright laws when you have the exact same code as an existing compiler.
This is so fucked up. The AI company has the perfect answer and yet it rolls the die to recreate the same thing by chance. What are they expecting, really?
they don’t care. they expect to be able to say “our AI agent made a C compiler that passed 99% of our tests”