Artificial intelligence coding tools are improving really quickly. If you do not work with code, you may not even notice, but models such as GPT-5 and Gemini 2.5 have opened up a whole range of new tricks for developers that can be automated. Sonnet 4.5 confirmed this again last week. On the other hand, skills such as writing emails are developing more slowly. Even when a model improves, the product may not benefit to the same extent—especially if it is a chatbot that handles dozens of different tasks at once. Artificial intelligence is still moving forward, but that progress is no longer as uniform as it once was.
This difference in progress has a simple explanation. Coding applications benefit from billions of easily measurable tests that train them to produce functional code. This is reinforcement learning (RL), which has probably been the biggest driver of progress in artificial intelligence over the past six months. Reinforcement learning works best when there is a clear measure of success or failure, allowing it to be repeated millions of times without having to stop for human input.
As the industry increasingly relies on reinforcement learning to improve products, a real divide is emerging between capabilities that can be evaluated automatically and those that cannot. Skills suited to reinforcement learning, such as fixing bugs or competitive mathematics, are improving quickly, while skills such as writing are advancing only gradually. In short, there is a reinforcement gap—and it is becoming one of the most important factors determining what artificial intelligence systems can and cannot do.
Why coding is ideally suited to reinforcement learning
In some ways, software development is the perfect field for reinforcement learning. Even before the era of artificial intelligence, there was an entire subdiscipline focused on testing how software performs under pressure—mainly because developers need to ensure that their code does not fail before deployment. So even the least elegant code must pass unit tests, integration tests, security tests, and so on. Human developers routinely use these tests to verify their code and, as Google's senior director of developer tools recently said, they are equally useful for verifying code generated by artificial intelligence. Moreover, they are systematic and repeatable on an enormous scale, which is ideal for reinforcement learning.
It is not easy to evaluate a well-written email or a high-quality chatbot response; these skills are inherently subjective and difficult to measure at scale. Not every task can be placed into the categories of “easy to test” or “difficult to test.” We do not have ready-made test suites for quarterly financial reports or actuarial mathematics, but a well-funded accounting startup could create such a suite from scratch. Some test suites will work better than others, and some companies will be smarter in how they approach the problem. But the testability of the underlying process will be the decisive factor in whether it becomes a functional product rather than merely an interesting demonstration.
Where reinforcement learning exceeds expectations
Some processes are proving to be more testable than you might expect. If you had asked me last week, I would have placed AI video generation in the “difficult to test” category, but the enormous progress in OpenAI's new Sora 2 model suggests that it may not be so complicated. In Sora 2, objects no longer appear and disappear out of nowhere. Faces retain their shape and look like a specific person, not merely a collection of features. Sora 2 footage respects the laws of physics in both obvious and subtle ways. If you looked behind the curtain, you would probably find a robust reinforcement learning system for each of these properties. Together, they make the difference between photorealism and an entertaining hallucination.
This principle also appears in other fields. In autonomous vehicles, for example, training reinforcement learning agents in simulation is proving feasible, but policies often fail in the unpredictable real world because of the gap between simulation and reality and the associated safety risks. In healthcare or robotics, reinforcement learning agents could theoretically learn to perform complex operations or physical tasks, but the absence of scalable and safe reward mechanisms, together with the high risks of errors, makes improvement slow.
Reward and efficiency problems in reinforcement learning
Reinforcement learning algorithms need enormous numbers of interactions with an environment to learn effective behavior, making real-world training impractical for many tasks. Simulations help, but they often lead to policies that cannot transfer to real-world conditions—this is known as the simulation-to-reality gap. Reinforcement learning excels when success criteria are binary and easy to evaluate automatically, as in coding (the code compiles or it does not) or mathematics (the proof works or it does not). These tasks provide clear, scalable feedback, enabling reinforcement learning agents to improve quickly and autonomously.
By contrast, skills such as writing, speaking, or creative problem-solving involve subjective, context-dependent judgments, making it difficult to design precise reward functions. This factor causes slow progress in areas where feedback is unclear or expensive to obtain at scale. Even when reinforcement learning agents perform well, their understanding of action values may be superficial, masking shortcomings in skill transfer or generalization.
Implications for startups, the economy, and the future
This is not a fixed rule of artificial intelligence. It is a consequence of the central role that reinforcement learning plays in AI development, which could easily change as models continue to evolve. But as long as reinforcement learning remains the primary tool for bringing artificial intelligence products to market, the reinforcement gap will only widen—with serious consequences for startups and the economy as a whole. If a process ends up on the right side of this gap, startups will probably automate it successfully—and anyone currently doing that work may have to look for a new career. The question of which healthcare services can be trained through reinforcement learning will have enormous implications for shaping the economy over the next 20 years. And if surprises such as Sora 2 are any indication, we may not have to wait long for an answer.
The reinforcement gap affects which artificial intelligence skills achieve practical usefulness first, favoring domains with clear, automatable rewards while subjective or ethically complex skills lag behind. Bridging this gap is crucial to achieving artificial general intelligence (AGI), because it requires reinforcement learning systems to learn efficiently, generalize across domains, and derive goals from ambiguous feedback—capabilities that current systems lack. Artificial intelligence safety and ethics are deeply affected; reinforcement learning agents may behave dangerously or unethically if reward functions are poorly designed or misaligned with human values.
Source: techcrunch.com
