This article explores how domain expertise significantly enhances the utility of Large Language Models (LLMs), arguing that the true bottleneck in AI interaction is often the user's ability to communicate specific requirements. Using Terence Tao’s dialogue with ChatGPT regarding a counterexample to the Jacobian Conjecture as a primary case study, the author demonstrates that expert users elicit vastly different outputs than novices. While untrained users receive lengthy, explanatory "amateur-mode" responses, Tao’s expertise allows him to trigger a "talking-to-mathematicians" mode, resulting in more concise, high-level technical discourse. Tao achieves this by using brief, gist-focused prompts, pushing back on incorrect model logic without direct contradiction, and providing his own technical leaps rather than following the model's suggestions. The author posits that expertise enables users to steer LLMs through specific queries—such as questioning why a particular formulation fails or suggesting alternate codebase architectures—rather than relying on generic principles. Ultimately, the text concludes that as models grow more powerful, human expertise remains vital because the capacity to extract precise, high-value information from a model's vast latent knowledge depends on the user's ability to identify what "looks weird" and suggest correct alternative formulations.