AI coding tools are rapidly improving due to reinforcement learning (RL) and easily measurable tests. However, skills like writing emails are progressing slowly. The key factor is whether a task can be automatically graded, allowing for billions of iterations without human input.
Software development is well-suited for RL because of existing testing methodologies like unit and integration testing. These tests provide systematized and repeatable validation for AI-generated code. Conversely, subjective tasks like writing emails are harder to measure at scale.
The "reinforcement gap"—the difference between capabilities that can be automatically graded and those that can't—is a critical factor in AI development. Unexpected advancements like OpenAI's Sora 2 model demonstrate that some processes may be more testable than initially thought, likely due to robust RL systems.
As long as RL remains a primary tool, this gap will widen, impacting startups and the economy. The testability of a process will determine its potential for automation, potentially leading to job displacement in certain sectors. The future of healthcare services, for example, will depend on which aspects can be effectively trained with RL.