Developers say AI coding tools work—and that’s precisely what worries them

In recent years, software developers have closely observed the rapid evolution of AI coding tools, which have transitioned from advanced autocomplete systems to sophisticated applications capable of generating complete software projects from mere text prompts. Innovations such as Anthropic's Claude Code and OpenAI's Codex are now capable of working on software endeavors for extended periods, effectively writing code, conducting tests, and, under human oversight, resolving bugs. OpenAI has taken the step of employing Codex to enhance Codex itself and has recently shared technical insights into the inner workings of this tool. This evolution has sparked a debate: is this a genuine advancement in technology, or merely another instance of AI hype? To delve deeper, Ars reached out to various professional developers on Bluesky to gauge their sentiments about these AI tools in practice. The responses revealed a community that largely acknowledges the effectiveness of the technology, yet opinions are split on whether this is entirely positive. David Hagerty, a developer specializing in point-of-sale systems, expressed his skepticism regarding the marketing claims surrounding AI. "All of the AI companies are hyping up the capabilities so much," he remarked. "While I believe LLMs are revolutionary and will significantly impact the industry, it's unrealistic to expect them to produce the next great American novel or similar works. That’s not how they function." Conversely, Roland Dreier, a software engineer with extensive contributions to the Linux kernel, recognizes the hype but has closely monitored the advancements in AI. "It may sound like exaggerated claims, but state-of-the-art agents are remarkably proficient right now," he stated. Dreier noted a significant leap in AI capabilities over the past six months, particularly following the release of Claude Opus 4.5. His usage of AI has evolved; he now anticipates instructing an agent to "debug this failing test and fix it for me," and expects it to deliver results. He estimates a tenfold increase in efficiency for intricate tasks, such as constructing a Rust backend service with Terraform deployment configurations and a Svelte frontend.