Explain why LLMs excel at coding compared to general intelligence tasks

Ascertain the underlying reasons why large language models appear to handle software engineering and coding tasks better than tasks requiring other facets of general intelligence, thereby clarifying the mechanisms behind this performance disparity.

Background

In discussing risks and implications, the authors observe a notable empirical pattern: LLMs often perform relatively well on software engineering and code-related tasks but struggle with broader aspects of general intelligence, such as systematic reasoning or certain forms of mathematical competence.

The authors explicitly state that the reason for this disparity is not known, emphasizing a fundamental gap in understanding the capabilities and limitations of LLMs across different task domains.