The Reinforcement Gap — or why some AI skills improve faster than others (techcrunch.com)

AI progress is becoming uneven: coding and other tasks with clear, repeatable pass/fail tests are accelerating rapidly while subjective skills like prose writing and multi-purpose chatbots lag. Recent model advances (GPT-5, Gemini 2.5, Sonnet 2.4) have unlocked new developer automation, and OpenAI’s Sora 2 shows that even video fidelity can jump when specific visual qualities are reliably measurable. The author coins this divide the “reinforcement gap”: capabilities that can be improved via large-scale reinforcement learning (RL) with automated rewards get better much faster than those that need subjective human judgment. Technically, the gap arises because RL scales best when there’s an objective metric you can run billions of times—unit, integration and security tests for code are ideal signal sources. Human grading helps, but it’s costly and slow; automated tests produce dense, repeatable feedback that drives rapid optimization. The implication is structural: products whose underlying processes are testable will be easier to automate and monetize, reshaping startup opportunities and labor markets (e.g., which healthcare or accounting tasks get automated). The gap isn’t immutable—new testing regimes can make previously subjective domains RL-friendly—but as long as RL dominates product improvement, where a task sits relative to that testability boundary will strongly determine which AI capabilities advance next.