Developers find large language models tiring because interacting with them drains social energy that could be spent on colleagues. A good tool merges with the brain; driving a car or typing on a keyboard feels like an extension of the body. Conversing with an LLM, however, remains a dialogue, demanding negotiation, persuasion, and occasional frustration.

When a keyboard or a car responds instantly, the brain treats it as part of the self. Claude and Cursor, two popular LLM interfaces, fail to achieve that seamlessness; latency and inconsistency keep users in a conversational loop. The author calls this extra effort a social tax, a cost rarely offset by the model’s output. Developers often spend minutes rephrasing prompts or correcting hallucinations, turning a simple query into a drawn‑out exchange.

The trade‑off pays off for a narrow set of tasks—generating boilerplate code, writing tests, or surfacing obscure bugs—tasks that were cumbersome a year ago. For broader problem solving, the author argues that directing the same mental bandwidth toward real teammates yields richer feedback and mentorship. The piece concludes that LLMs remain a conversation, not a true extension.

Nevertheless, the community continues to experiment, integrating LLMs into IDEs and CI pipelines to shave minutes off repetitive chores. Whether those minutes justify the ongoing social tax remains a personal calculus for each developer, and team leads must decide if the marginal gain outweighs the cognitive load.