Why experts get more from the same LLM

LLMs can give almost anyone a passable first draft of CSS, code, mathematics, or professional copy. But the people who get the most from them are not necessarily the best “prompters,” according to the post “LLMs reward expertise.” They are the people who understand the subject well enough to steer, challenge, and refine the model’s output.

Why domain knowledge changes the result

Before large language models, someone who could not write CSS had two options: ask a skilled colleague or search the internet for an answer that matched the problem. LLMs changed that trade-off by letting almost anyone produce “sort-of-okay” CSS through delegation. The same pattern applies to PhD-level mathematics, computer code that is often useful but can be stylistically poor, and awkward LinkedIn-style prose.

That accessibility makes it easy to assume that using an LLM requires little skill. Everyone is speaking to the same models, so a newcomer can sometimes receive an answer that looks similar to one produced by an experienced user.

The post’s argument is that the crucial difference is not a secret prompting formula. It is expertise in the domain being discussed. A subject-matter expert can recognize the important part of a long answer, identify an unusual assumption, propose a better direction, and reject a solution that is technically plausible but wrong for the situation.

Recommended reading

Flux 3 makes fake history look disturbingly real

For software engineers, that often means knowing the codebase rather than merely understanding generic software principles. Familiarity with the systems they own lets them ask concrete questions: whether X works in a particular place, why A follows from Y and Z, or whether the code already performs a proposed operation. Those questions are more valuable than broad requests for “best practices” because system-design problems are shaped by specifics.

That is consistent with our earlier reporting on why verification remains expensive after code got cheap: generating an answer and knowing whether it is the right answer are separate capabilities.

What Terence Tao’s ChatGPT exchange shows

The post points to mathematician Terence Tao’s conversation with ChatGPT about a recently discovered counterexample to the Jacobian Conjecture as a clear example. The author says Tao appears to be using a different version of ChatGPT—not because the model is fundamentally different, but because Tao’s expertise changes the interaction.

Several details stand out:

• Tao’s messages are short and precise.
• He responds to the gist of an answer rather than addressing every point line by line.
• The model responds more concisely than it does when the author asks it about mathematics.
• By signaling expertise, Tao appears to move the model from an “explaining-to-amateurs” mode into a “talking-to-mathematicians” mode.
• When an answer looks wrong, Tao pushes back indirectly, with comments such as “this looks more complex than I was hoping for.”
• He makes his own leaps and suggestions instead of simply following the model’s proposed next step.

Those behaviors cannot be reduced to a checklist. Someone unfamiliar with the mathematics could imitate Tao’s short messages or cautious corrections, but would not know which idea to extract from a multi-paragraph response, which alternative formulation to suggest, or what detail looks suspicious.

The author makes the same point from software work. A strong theory of a codebase provides an internal model of what a good solution should look like. That lets an engineer tell the LLM that a design could be simpler, that an existing component already solves the problem, or that the request should be expressed using concepts familiar to the system.

The human remains the constraint

The post does not argue that domain knowledge is required for every useful interaction. Someone without expertise can still use an LLM to get started, explore an unfamiliar topic, or produce a workable first attempt. The difference is in how far the user can push the model and how reliably they can judge what comes back.

Most people will combine both modes: expertise in some areas and dependence on the model in others. As models improve, the post argues, the bottleneck for many tasks may remain the human’s ability to communicate the desired solution precisely.

That leads to a less obvious conclusion about increasingly capable models. The relevant information may already be present in the model, but extracting the right answer still depends on a user who knows what to ask, what to ignore, and when to redirect the conversation. Stronger models do not eliminate expertise; they can make expertise more powerful by giving it a better instrument to work with.