AI models are handling longer stretches of useful work with less human input. So competent digital work, such as a draft, a recommendation, a design change or a code revision, is getting cheap to produce. That changes where value sits inside a business, and the edge moves to taste. Matt Shumer made the point in his essay Something Big Is Happening.
The clearest evidence is how much work advanced AI can finish on its own. In March 2025, researchers at METR measured the length of tasks an AI can complete correctly about half the time without help. The most advanced models had reached around 50 minutes on the researchers’ tasks, and that length had been doubling roughly every seven months since 2019.
As execution gets easier to buy, the volume of acceptable work rises, and acceptable work alone does not create a lasting advantage. The hard part becomes recognising what is good, and what keeps its value once competent work is abundant.
That is where taste becomes a business skill. Taste is judgement about what people want, what earns trust and what creates real demand instead of busywork. It covers what deserves to be built, what is worth approving, what should never reach a customer and what is merely polished enough to look convincing.
The machines are getting better at judging too, even if that is easy to overstate. Some AI models score their own answers during training. Research on these self-rewarding models shows that, over repeated cycles, they improve both their answers and their ability to judge them. So stronger models are increasingly helping to create stronger models, which helps explain why AI’s abilities appear to compound so quickly. That does not make AI independent of people.
Even so, taste is harder to automate. A model can generate options, rank them and improve against a target, but someone has to define the target. In business, that means deciding what quality looks like in a market, a product, a brand or a customer relationship. Even the self-rewarding research depends on standards and examples that tell the model what a better answer looks like.
