Why LLMs Deliver a 2x Coding Boost, Not 10x — and Why That Won't Change Soon

A software engineer argues that the real reason LLMs took off for coding in 2026 isn’t raw intelligence but reliability: models finally became dependable enough to run inside automated feedback loops. Given an explicit, objectively verifiable goal — ‘build a button that does X, then confirm it does X’ — they iterate in meaningful steps and can predict when a human would accept the result. That capability is genuinely transformative, but the author pegs it at roughly a 2x productivity gain rather than the 10x that hype suggests.

The core of the piece is a ‘staircase hypothesis’: crossing the threshold of usefulness is like being tall enough to climb one stair at a time, and being able to take two or three steps at once matters far less. By this logic, further model improvements will yield diminishing returns because the hard remaining problems are subjective, not verifiable — questions like whether code is structured maintainably or whether documentation includes the right information. The author uses LLMs to generate rough drafts and then iterates heavily, noting that a working implementation used to mean a task was 80 percent done but now signals closer to 20 percent. He even instructs models to never write READMEs, docstrings, or comments, preferring to handle those himself.

The takeaway is that the next wave of gains will come from the industry retooling its workflows and tooling around the capabilities models already have — sandboxed environments, declarative specifications, better test practices — rather than from smarter models alone. He leaves open the possibility that vibe coding and new safety tooling could eventually route around LLMs’ weaknesses to reach 10x, but treats it as unproven, since there hasn’t yet been a long enough term to judge.