LLMs reward expertise

In the 2010s, lacking technical skills meant relying on colleagues or hoping the answer was found online. Today, anyone can write basic code using large language models (LLMs). This has led many to believe there's no skill involved in working with LLMs. However, this is incorrect. The most crucial skill in prompting LLMs is expertise in the domain being queried.

A prime example is Terence Tao's discussion with ChatGPT about a recently discovered counterexample to the Jacobian Conjecture. While Tao is an exceptional mathematician, his approach highlights the importance of domain knowledge in prompting LLMs effectively. By understanding the mathematics, Tao is able to extract relevant ideas, suggest alternative approaches, and identify inconsistencies.

His ability to ask targeted questions, such as "does X work here?" or "why A?", demonstrates how domain knowledge enables better interaction with LLMs. Even though Tao's expertise is not directly transferable to programming tasks, the underlying principle remains: familiarity with a specific domain yields superior results when using LLMs.

This insight is supported by personal experience, where a deep understanding of a codebase allows for more effective guidance of LLMs. While domain knowledge may not be essential for all tasks, it is invaluable for leveraging the full potential of LLMs. The key takeaway is that human expertise will always be crucial, even as models become more advanced, as the challenging aspect lies in effectively communicating the desired solution to the model.