How AI Will Drive R&D Systems in the Next 5‑10 Years

AI coding tools have progressed from generating single lines to producing multi‑file, thousand‑line code, reaching a skill level comparable to a junior developer who knows basic syntax but lacks project‑level expertise.

Such tools can join a team but still make mistakes like reinventing components, introducing security or performance bugs, or causing regressions, especially when lacking domain knowledge. The author argues that accurate and effective context is the key to elevate the AI from junior to senior developer status.

Providing precise, complete context enables the AI to perform analysis, planning, coding, testing, and deployment like an experienced engineer; insufficient context reverts it to a junior level.

Observing a polarization, some developers become “super‑individuals” who leverage AI to accomplish weeks of work in a day, while others spend time fixing AI‑generated bugs. The author stresses that future survival requires mastering new skills: understanding LLM fundamentals, prompt engineering, context construction, and building tools for AI.

Based on these observations, the author proposes the AI‑First System Development Methodology (ASDM) built on two assumptions: (1) if AI can code hundreds of times faster, solving accuracy yields massive productivity; (2) once AI handles most value creation, tooling should be built for AI, not humans.

ASDM’s core challenges are: (1) ensuring AI generates correct code; (2) constructing tools suitable for AI use. The author notes that model training alone cannot address domain‑specific knowledge due to trade‑offs between generalization and optimality, data volume limits, and compute/iteration constraints. Therefore, the solution combines the best models with context and prompt engineering to handle uncertainty.

The article introduces a self‑feedback system based on the PDCA (Plan‑Do‑Check‑Act) cycle. Instead of constraining AI, the system lets AI generate code, then automatically compile, test, and iterate based on feedback, forming inner (coding‑debugging) and outer (integration‑deployment) loops. Automating tool invocation enables AI to complete the full software production pipeline.

Finally, the author concludes that the key to accurate AI code generation is establishing this self‑feedback mechanism, and the overarching shift is from building systems for humans to building systems for AI.