Common Mistakes - Learn by Gautam Jha

Believing the AI “understands.” It is a statistical text predictor, not a mind. The moment you attribute intent or comprehension to it, you stop engineering guardrails and start trusting vibes. Stay mechanical in your thinking even when the output is dazzling.
Expecting determinism. Newcomers file “bug reports” because the same prompt gave two answers. That is the system working as designed. Engineer for a distribution of outcomes: evals, retries with validation, and tolerance for variation where it is harmless.
Dismissing it as “just autocomplete.” The opposite error. Yes, it predicts the next token — and that simple mechanism, at sufficient scale, produces genuinely useful behavior across an astonishing range of tasks. Underestimating it leads to missed opportunities; overestimating it leads to outages. Calibrated respect is the goal.
Using AI for things ordinary code does better. Arithmetic, exact lookups, schema validation, idempotent transforms, hard policy enforcement — write these in Go. Reaching for a model here is slower, costlier, and less reliable. Match the tool to the shape of the problem.
Ignoring cost and latency until the bill arrives. Tokens are money and time. A design that ignores them works in a demo and fails a budget review. Build cost intuition from your very first call
Forgetting the model has no memory. Each call is stateless. If you don’t send the context, the model doesn’t have it. Many “the AI forgot what I told it” complaints are really “the program didn’t resend the history.”