Large language models like OpenAI's Codex and Deepmind's AlphaCode can generate code based on natural language descriptions, with tools such as GitHub Copilot commercially available. LLM-assisted programming shares characteristics with prior assistance methods like compilation and pair programming but has unique technical capabilities. Experience reports indicate that writing effective prompts can be difficult, resulting in a shift in programming activities towards code verification and debugging. Existing metaphors for AI-assisted programming fail to represent the fundamentally different process introduced by these models, necessitating a new framework for understanding their interaction.
Large language models can generate code to solve a variety of problems expressed in natural language, with tools like GitHub Copilot leading the commercial applications.
Programming with LLM-assisted programming differs from prior conceptualizations of programmer assistance in both technical capabilities and underlying methodologies.
Experience reports reveal that effective prompting for AI assistance is challenging, as programming duties shift towards verification and debugging.
Existing metaphors for AI-assisted programming, like search or compilation, are inadequate; a distinct understanding of LLM interaction is necessary.
Collection
[
|
...
]