anthropics / anthropics/claudes-c-compiler

Use a different problem for the next capability trial?

Ouverte
#239 3 commentaires 0 réactions 0 personnes assignées Voir sur GitHub
Langage dominant
Rust
Étoiles
2.8k
Forks
247
Métriques de merge des PR
Aucune PR mergée en 30 j

Description

This project seems to be an artifact of trying the capabilities of next generation of LLMs.

While the compiler works, the code quality is inferior, poorly maintainable (article mentioned attempts to add features often triggered emergence of bugs, so LLM struggels and probably no human would want to touch the codebase as well) and it is likely full of various and yet undiscovered bugs, so it is unlikely to be deployed in practice aside of tests or experiments.

However, considering also notes mentioned in https://github.com/anthropics/claudes-c-compiler/issues/231 there are two problems with this capability trial:

**Reproductibility**:
Ideally, anyone willing to throw away $20000 in credits should be able to perform the same. However neither this repo, neither the article contains much detail - there is no initial prompt and only very vague details about the scaffolding and code used to run it.
I'd assume at least the initial prompt should be available. Ideally the code to run this should be also available, but I understand there may be some trade secrets involved ....

**What does it prove**:
It is questionable, whether the LLM has truly produced a compiler from scratch, or whether it essentially copied existing concepts or large parts of code from compilers that are likely present in its training data (LLVM, GCC, ...), merely transliterating from C++ to Rust.
As it may be hard to train LLM and ensure none of the compiler code is in training data (let alone to proving this fact), I suggest picking up a different task for the next capability trial.

Ideally something for which a software does not exist currently.

While this may be a less useful problem solved (if that problem would be interesting enough, someone would have probably written a program to solve that already), if such a program would be written, there would be much less doubt whether it was truly created from scratch and not just plucked from memorized parts of the training data.

Guide de contribution

Aucun guide de contribution indexé pour ce dépôt

Évaluation

Cette issue n'a pas encore été évaluée.

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.