Beyond recall and the illusion of competence

Two main perspectives exist regarding AI and programming. One argues AI is largely ineffective due to generating poor code, with debugging taking longer than coding itself. The other posits AI-generated code will eventually render programmers obsolete. However, the author contends that focusing on who writes the code is misguided, as recalling syntax and typing code from memory is not critically important.

Modern software systems typically involve various technologies, making it inevitable to rely on external resources like documentation, internet searches, Stack Overflow examples, blog posts, and colleague code. The author also notes that working on systems not originally written by the team is common, with parts contributed by multiple individuals.

Despite this, familiarity with the system develops due to understanding its functionality, behavior, dependencies, and potential issues. The author believes AI-assisted programming becomes intriguing when considering the risk of delegating understanding along with typing. Typing code is relatively easy to outsource, but debugging is far more challenging without compromising essential skills.

Debugging involves constructing a system model, identifying discrepancies between expected and actual outcomes, and working backward to understand the root cause. This process fosters an intuition for system behavior. However, AI can easily lead to a false sense of competence by enabling developers to produce functioning software without developing a mental model.

When relying solely on AI for error resolution, developers may become overly dependent on the machine, hindering the development of a mental model. This dependence can be problematic when the AI fails to provide a solution or runs out of available resources. Experienced developers may have a reservoir of knowledge acquired through years of trial and error, but for newcomers, this experience may not manifest as seamlessly.

If developers rely on AI to handle debugging, there is little motivation to invest time in understanding the underlying system. The time spent learning system intricacies is crucial for developing a mental model. Developers who focus on understanding systems, architecture, integration, distributed systems, observability, failure modes, and boundaries will stand out in the future.

AI should be used to write boilerplate code, remind users of syntax, explore unfamiliar libraries, and implement tedious parts. Developers should maintain control over decision-making, architectural choices, and system understanding while leveraging AI's capabilities.