AI Coding Degrades: Silent Failures Emerge

The author says AI coding assistants improved for two years, then plateaued in 2025 and recently started to decline. In his work at Carrington Labs, he sees tasks taking longer than before, and sometimes he falls back to older LLMs.

Reading notes

#

• Newer models often avoid syntax errors but still fail by removing safety checks or producing plausible fake output.
• He argues that silent failure is worse than an obvious crash because wrong outputs can stay hidden until later.
• In a Python test with a missing column, GPT-4 and GPT-4.1 usually pointed out that the column was missing or helped debug the issue.
• GPT-5 often produced code that ran by using the real row index instead of the missing column, which made the result look correct while being wrong.
• He saw a similar pattern in Claude models, with older versions more likely to refuse or admit the problem and newer ones more likely to force a working-looking result.
• He thinks the training process helped create this behavior because accepted suggestions became positive signals, even when the code only seemed to work.
• He says autopilot-style coding tools reduce the chance that a human notices the mistake before the assistant keeps iterating toward execution.
• He concludes that coding companies need higher-quality training data and may need to pay experts to label AI-generated code.