Limits of Large Language Model (LLM) Code Generation

1. Analysis of large amounts of data: LLMs need a description of the data and can't analyze large amounts of data on their own. 2. Operating at the boundary of human understanding: LLMs can't push human understanding or make scientific discoveries. 3. Dealing with breaking changes in new libraries: LLMs struggle with updates in software libraries and programming languages. 4. Coping with new programming language features: LLMs may not be up-to-date with the latest features of programming languages. 5. Large proprietary function libraries: LLMs may not be familiar with specific corporate libraries. 6. Modifying large monolithic software projects: LLMs need well-defined modules to work effectively.

via Limits of LLM Code Generation (2.5)