AI is fundamentally changing career progression in engineering by removing crucial learning opportunities at various career stages, while simultaneously allowing individuals to perform at levels beyond their actual experience. Alasdair Allan highlighted this phenomenon in his talk "Engineering Progression When AI Ate the Middle" at QCon London. The industry is seeing a decrease in junior developer intake, as AI tools contribute to slower hiring rates for entry-level positions. Allan points out that relying heavily on AI for code generation might lead to a future where new engineers lack the foundational skills needed to properly supervise and validate AI-written code.
What are the conclusions from studies of how AI impacts development time and quality of generated code?
Research findings regarding AI's influence on development speed and code quality are notably contradictory. For instance, a randomized controlled trial by METR revealed that seasoned developers were 19% slower when using AI, despite their perception of being 20% faster. Similarly, Anthropic's studies found that junior engineers utilizing AI demonstrated a 17% lower mastery score without achieving faster completion times, indicating a trade-off of learning for no actual efficiency gain. Allan emphasizes that while AI is genuinely effective for discrete, well-defined tasks, its benefits diminish significantly as task complexity escalates, shifting the bottleneck to code review—a stage where senior judgment is paramount. He further noted that out of fifteen AI-generated pull requests on legitimate open-source projects, none were mergeable, even after passing automated tests. This suggests that AI excels at generating functional code but falls short in the nuanced 'craft' aspects of software engineering.
How does the work that computer programmers do change when they use AI tools?
The nature of work for computer programmers is increasingly transitioning towards a supervisory role due to the integration of AI tools. For example, engineers at Anthropic use Claude for 59% of their daily tasks, yet can only fully delegate approximately 20% of the work. This substantial gap between automated generation and human judgment is where future employment opportunities are expected to emerge, with AI handling code production and humans responsible for validating its accuracy and suitability. A significant concern raised by Allan is that 80% to 90% of engineering questions are now directed to AI instead of human colleagues. This practice circumvents the invaluable incidental learning gained through collaborative problem-solving with mentors. Allan asserts that the current career development framework, designed for code writers, is inadequate for preparing individuals for roles primarily focused on supervising AI-driven systems.
What are your suggestions for building engineers in the AI era?
Alasdair Allan acknowledges that a complete solution for training engineers in the AI era is still elusive, viewing the current stage as one of diagnosis rather than definitive answers. Nevertheless, he provides practical recommendations, advocating for structured learning paths that include intentional rotations through fundamental tasks, drawing a parallel to medical residencies. This 'scut work,' though seemingly inefficient, is crucial for cultivating sound judgment. He stresses the importance of evaluating understanding over mere velocity, encouraging a focus on how engineers think rather than solely on their output. Additionally, Allan highlights the necessity of treating 'context' as a critical piece of infrastructure, suggesting the creation of comprehensive documentation tailored for senior engineers who possess coding expertise but lack specific codebase knowledge. He concludes by emphasizing a realistic perspective: AI is a tool, not a mentor, and bypassing genuine comprehension through its use is a gamble with future skill development.
