Learning Cycles

A learning cycle is a round of learning consisting of four steps. First, you think about what should get better. Then you try out a change. After that, you measure whether it actually helped. Finally, you draw a conclusion from it and start the next round. The German term for this would simply be Lernzyklus, but in everyday professional life the English term has become the standard.

Why the length of a round determines the lead

Two teams can be equally smart and still progress at completely different speeds. What matters is how often they complete a round. A team that tests one idea every day gathers about thirty insights in a month. A team that spends three months building a single large version ends up with exactly one. Even if that one version is more thoroughly thought through: thirty small corrections usually beat one big leap.

With AI products, there is the added factor that nobody knows in advance what will work. Whether users find a chatbot understandable cannot be calculated at a desk. You have to measure it. That is exactly why companies like OpenAI, Google, or Anthropic talk so much about iteration speed. Iteration simply means: going through the same loop again, just with better knowledge than the time before.

A common misconception is that fast cycles are the same as sloppy work. The opposite is closer to the truth. Anyone who measures often notices mistakes after days instead of months. A mistake doesn’t become costly because it happens, but because it stays undetected for a long time.

The four steps of a round

At the start there is an assumption that can be disproven. Not “the answers should get better,” but for example: “If the model receives longer instructions, the number of wrong answers decreases.” Only an assumption formulated this way can, in the end, be clearly confirmed or rejected.

Then comes the change itself. It should be small and affect only one thing. Anyone who simultaneously switches the model, rewrites the instruction, and rebuilds the interface won’t know afterward which part had an effect. After that, measurement takes place, using a metric defined in advance. For language models, these are things like the share of correctly solved test tasks, response time, or cost per request.

The last step is the one most often skipped: recording the result. Even a failure is valuable, because it rules out one possibility. Without these notes, someone will try the same idea again half a year later. Only once an insight flows into the next round is the cycle truly closed.

Learning cycles in AI models and in school

In quarterly reports and tech news, the term usually appears as a justification for speed. When a company states that it has shortened its learning cycles from weeks to days, it means: we test more often and correct earlier. For large language models, such rounds often consist of automated test sets, so-called evaluations, plus feedback from real users.

The principle also exists outside of technology. A soccer team watches the recording after the match and specifically practices a weakness the following week. It works the same way when studying for an exam: work through a practice problem, compare the solution, redo the step that was misunderstood. Anyone who only reads and never checks doesn’t have a cycle, just a beginning.

Related, but not identical, is the term feedback loop. A feedback loop generally describes a result feeding back into the system, even unintentionally. A learning cycle is deliberately planned and has a goal. So the difference lies not in the form, but in whether someone is intentionally turning the loop.