Senior Engineer Admits AI Still Makes Mistakes. Just Smaller Ones. The Pattern Holds.
Brandon Pedersen, a senior principal engineer at GoTo, has spent three years working heavily with AI tools including Claude Code and GitHub Copilot. He reports that errors have decreased in smaller use cases like single lines of code or simple functions, but mistakes still appear in larger tasks. He verifies AI output from the beginning and has learned to push back on AI rather than accepting its answers blindly.
This illustrates a principle I call calibrated delegation. The mental model is simple. You match your trust level to task complexity, not to your overall enthusiasm for the tool. Small context window, small risk. Large context window, large risk. The mechanism is error surface area. More output means more places for plausible-sounding nonsense to hide. Always verify proportional to scope.
Brandon Pedersen, senior principal engineer at GoTo, uses Claude Code and GitHub Copilot to build larger-scale solutions, create presentations and web pages, and help with entire applications. He verifies sources and pushes back on AI output rather than accepting it blindly.
- Open a free AI chatbot like ChatGPT or Claude and ask it to write a single Python function, say, one that reverses a string. Read every line and note any errors or odd choices. Expected outcome: you will likely find the code works but may include unnecessary steps or stylistic quirks.
- Ask the same AI to build a small application with multiple functions, like a simple calculator with a menu. Read through the entire output carefully. Expected outcome: you will notice more places where logic is slightly off or assumptions are unclear.
- Take one specific answer the AI gave and ask it to explain its reasoning. Then ask it to find a flaw in its own answer. Expected outcome: the AI will often identify at least one issue it initially overlooked, demonstrating why verification matters.