Capability Overhang

Sometimes a tool can do more than its owners realize. That’s exactly what the English term Capability Overhang describes, roughly translated as “surplus of capabilities.” It refers to the gap between what a technology can already achieve in principle and what people have actually done with it so far. The term is mainly used for computer programs that learn from examples, that is, for artificial intelligence. Such a program is built once and then no longer changed. Nevertheless, it is often discovered months later that it can solve tasks nobody thought of during its construction.

Why the overhang throws off forecasts

Anyone trying to estimate how strongly AI will transform the economy usually looks at the latest models. That is misleading. Much of the impact often lies in systems that already exist. Even if development stopped today, new applications would keep emerging for years. The overhang therefore means: technology is ahead of society, not the other way around.

For companies, this is uncomfortable news. A competitor can suddenly gain an edge without owning a better model. They simply understood earlier what the existing model can do. This is exactly what happened after the release of ChatGPT at the end of 2022. The underlying technology was already years old and publicly documented by then. What was new was merely the packaging as a chat window that anyone could use without instructions.

The term is also central to safety questions. If nobody knows exactly what a model can do, nobody knows what dangerous capabilities might be lurking within it either. Researchers therefore systematically test published systems for unexpected abilities. Still, there remains a residue that only millions of users end up finding.

Where the hidden capabilities come from

A large language model is not programmed for individual tasks. During training, it learns from vast amounts of text which word follows which. In the process, capabilities emerge on the side that nobody directly taught it. A model that was only meant to predict text can end up writing program code or solving math problems. Such unplanned capabilities are called emergent, meaning “arising on their own.”

The second reason lies in how the model is used. How a question is phrased significantly changes the quality of the answer. If you ask a model to write out its reasoning step by step, its accuracy on calculation tasks rises noticeably. This technique was only discovered after the models had already been released. The model did not become smarter as a result. People simply learned how to access the capability that was already there.

Add to this the wiring with other programs. If a model is given access to a search engine, a calculator, or a calendar, its practical usefulness grows dramatically. Here too, the model itself remains unchanged. Think of it like a smartphone that has been sitting in a drawer since purchase. The hardware stays the same, but the apps written for it make it more useful year after year.

The term in investor meetings and the news

In business news, Capability Overhang usually appears as an argument. Analysts use it to explain why AI revenues can keep growing strongly even though no new flagship model has been released. Executives at major labs also use the expression. What they essentially mean is that the world has not yet come close to exhausting the models that already exist.

In everyday life, one experiences the overhang through one’s own usage. Most people use chatbots for simple questions, even though the same systems could analyze spreadsheets or summarize contracts. There are often years between the release of a capability and its widespread use.

An important distinction: Capability Overhang is not the same as hardware overhang. The latter means that data centers possess more computing power than current models actually use. Another common misconception is reading the term as a promise. It only states that unused potential exists. Whether it pays off economically is a separate question altogether.