[gaz]

Code reviews - Shines

Review code - Shines

Finding bugs - Shines

Suggesting coding patterns you want and suggest to LLM - Shines

Explaining error messages - Shines

Writing code from scratch - Mediocre

Architecting a solution - Mediocre/Poor

Understanding code or solutions it has no clue about - Poor

Contextual multi-file or multi-domain code understanding - poor

[sgbeal]

There's a contradiction in that list: proper code reviews require knowledge of the problem the code professes to solve. If AI is poor at context-dependent details, it can't shine at reviewing code.

[famouswaffles]

It doesn't say the AI is poor at context dependent details. Specifically it says multi-file context or context dependent on knowledge of multiple domains. There's not necessarily any contradiction here.

[sgbeal]

> Specifically it says multi-file context

Only the most trivial of projects have only a single code file.

[famouswaffles]

It's not about one code file vs 2. It's about the amount of code that can fit into the context at any time.