4 ms·

I find it dubious that an ML model would outperform existing compilers (AOT or JIT) for Python and JS, both of which exist and have many engineer years invested

by duped 2y ago

I find it dubious that an ML model would outperform existing compilers (AOT or JIT) for Python and JS, both of which exist and have many engineer years invested in their design and testing. I find it even more dubious that someone would want something that could hallucinate generating machine code. The difficulty of optimizing compiler passes is not in writing code that appears to be "better" or "faster" but guaranteeing that it is correct in all possible contexts.

ActorNightly 2y ago
This would be a much different training task than LLMs. The reference to it being possible with large LLMs is just a proof that it can be done. The reason its different is because you are working with a finite set of token sequences, and you will be training the model on every value of that set, because its fairly small. So hallucination won't be a problem. Even without ML, its a lengthy but P hard task to really build a python to C translator. Once you unroll things like classes, list comprehensions, generators, e.t.c, you end up with basically the same rough structure of code minus memory allocation. And for the latter, its a process of semantic analysis to figure out how to allocate memory, very deterministic. Then you have your C compiler code as it exists. Put the two together, and you basically have a much faster python without any dynamic memory handling. The advantage of doing it through ML is that once you do the initial setup of the training set, and set up the pipeline to train the compiler, to integrate any pattern recognition into the compiler would be very trivial.
- duped 2y ago
  > Once you unroll things like classes, list comprehensions, generators, e.t.c, you end up with basically the same rough structure of code minus memory allocation. No, you don't, and that's why there are many engineer years invested into designing AoT and JIT compilers for JS and Python. If you write C like Python you get Python but slower. > The advantage of doing it through ML is that once you do the initial setup of the training set, and set up the pipeline to train the compiler, to integrate any pattern recognition into the compiler would be very trivial. Except this has already been done, so what advantage does ML bring? Other than doing it again, but worse, and possibly incorrectly?
  - ActorNightly 2y ago
    AoT/JIT work would literally be the training tool to train the model, with additional optimizations along the way. The issue is that manual removal of all the runtime stuff from the generated code is just cumbersome right now, but with ML, its just a matter of having enough examples. The advantage is that you get native optimized code like someone wrote it in C directly, and ability to automatically generate code to be offloaded to GPU as people start doing expanded training with higher level pattern recognition. The incorectness part I adressed already, stochastic output doesnt matter when your domain is finite.

by duped 2y ago

ActorNightly 2y ago

This would be a much different training task than LLMs. The reference to it being possible with large LLMs is just a proof that it can be done. The reason its different is because you are working with a finite set of token sequences, and you will be training the model on every value of that set, because its fairly small. So hallucination won't be a problem. Even without ML, its a lengthy but P hard task to really build a python to C translator. Once you unroll things like classes, list comprehensions, generators, e.t.c, you end up with basically the same rough structure of code minus memory allocation. And for the latter, its a process of semantic analysis to figure out how to allocate memory, very deterministic. Then you have your C compiler code as it exists. Put the two together, and you basically have a much faster python without any dynamic memory handling. The advantage of doing it through ML is that once you do the initial setup of the training set, and set up the pipeline to train the compiler, to integrate any pattern recognition into the compiler would be very trivial.

duped 2y ago
> Once you unroll things like classes, list comprehensions, generators, e.t.c, you end up with basically the same rough structure of code minus memory allocation. No, you don't, and that's why there are many engineer years invested into designing AoT and JIT compilers for JS and Python. If you write C like Python you get Python but slower. > The advantage of doing it through ML is that once you do the initial setup of the training set, and set up the pipeline to train the compiler, to integrate any pattern recognition into the compiler would be very trivial. Except this has already been done, so what advantage does ML bring? Other than doing it again, but worse, and possibly incorrectly?
- ActorNightly 2y ago
  AoT/JIT work would literally be the training tool to train the model, with additional optimizations along the way. The issue is that manual removal of all the runtime stuff from the generated code is just cumbersome right now, but with ML, its just a matter of having enough examples. The advantage is that you get native optimized code like someone wrote it in C directly, and ability to automatically generate code to be offloaded to GPU as people start doing expanded training with higher level pattern recognition. The incorectness part I adressed already, stochastic output doesnt matter when your domain is finite.