W30 - Full-Stack Development in the AI Wave
Be cautiously optimistic about full-stack.
Full-stack has a long history; individual developers can be considered full-stack, but it has never become mainstream domestically. Large companies tend to favor depth of expertise, and many interviewers look down on jack-of-all-trades full-stack candidates. The value of full-stack is being reassessed now, and I think we should maintain a cautiously optimistic attitude toward this change.
Many of the technologies related to this wave of generative AI are near the peak of the Hype Cycle, including Vibe Coding and Software Engineering. Technologies at the peak often come with inflated expectations and impractical deployment behaviors.
Vibe Coding is different from Software Engineering. Vibe Coding—AI excels at generating one-off prototype code, quickly building runnable products to validate design and user acceptance. Because these are just prototypes, concerns about long-term maintenance and code quality are absent. In Software Engineering, coding itself is only part of routine delivery work. Research by organizations like METR shows that AI tools offer very limited productivity gains for experienced developers.
Mastering fundamentals lets you accomplish 99% of tasks; fundamentals remain important. While troubleshooting an issue last week, I noticed clear gaps in basic networking knowledge (e.g., Nginx, the HTTP protocol) and in our company middleware (the PaaS layer). When systems run normally everyone thinks it’s simple, but once something breaks people fall into awkward, blind-all investigative modes. Without solid fundamentals, no matter how advanced intelligent tools are, it’s hard to build reliable systems.
Building trust between engineers and AI takes time and leads to a correct understanding of AI’s capability boundaries. There are three boundaries to explore in practice: when you can trust AI and let it participate in tasks; when you can have AI produce initial output and engineers act only as reviewers; and when you must turn AI off and rely on humans to calmly solve the problem.
Of course these boundaries are changing rapidly and intelligence levels continue to improve, so we should be optimistic about the positive impact on business development over the next 3–5 years. Last week OpenAI announced that an unreleased LLM achieved gold-medal-level performance on the International Mathematical Olympiad. Before this, only a Google DeepMind model specifically optimized for mathematics had reached silver-medal level.
Last updated