Artificial adventures (www.scattered-thoughts.net)

The author reflects on their experiences using various AI models, particularly from OpenAI and Anthropic, for software development tasks. They note a marked difference in performance between frontier models like GPT-5.5 and lesser models like Codex, highlighting that while the former excel at tasks such as code review and bug identification, the latter often struggle and produce unreliable output. The significant finding is that frontier models can effectively spot complex bugs—a capability not typically achieved through standard programming practices—suggesting they could enhance code quality and debugging efficiency in software development. However, the author expresses concerns about the decision-making capabilities of these AI tools, which tend to lead to suboptimal coding practices and errant modifications. They argue that current AI technologies, while promising, may produce chaotic results that require substantial human oversight, particularly in intricate projects like web applications or games. This commentary underscores the necessity for evolving development practices around these tools, suggesting that improvements in programming paradigms, static analysis, and runtime guarantees could enhance the utility of AI in real-world software engineering, particularly as model capabilities reach maturity.