2x, not 10x: coding with LLMs in 2026
LLMs provide a ~2x productivity boost rather than 10x because they cannot yet autonomously ensure long-term maintainability.
Key Points
- The ‘staircase hypothesis’: LLM adoption in 2026 is driven by models becoming reliable enough to operate within automated feedback loops; once this threshold is met, further raw performance gains yield diminishing returns on productivity.
- LLMs are highly effective for tasks with objectively verifiable acceptance criteria (e.g., creating a button that performs a specific action) via iterative refinement.
- Significant gaps remain in subjective quality areas, specifically maintainability of code structure and the creation of high-quality documentation.
- The productivity shift has changed the definition of ‘done’: a working implementation that previously signaled 80% completion now represents only about 20%, as the remaining effort is spent on heavy iteration for structure and readability.
- [AI Synthesis] The author suggests that a 10x productivity leap will not come from model scaling alone, but from industry-wide retooling of workflows and the development of sandboxed environments to route around LLM weaknesses.