The articles you just read were about a time problem: how do you build AI that keeps learning as the world changes, rather than freezing at the moment of training? This cluster is about a different problem, though it's related. It's about a space problem: how does an AI system handle situations it's never encountered before?
The naive answer is: it can't. A system trained on data from one domain shouldn't work in a different domain. A system that's never seen an example of something shouldn't be able to recognize it. And for a long time, that was roughly true. Early machine learning systems were brittle in exactly this way — highly capable within the narrow range of their training data, nearly useless outside it.
What changed is that researchers found ways to make the knowledge learned in one context transfer to another. Transfer learning is the foundational version of this: take a model that's already learned something useful — say, the general structure of language, or the visual features that distinguish objects — and adapt it to a new task, rather than training from scratch. This is why large language models can be fine-tuned for specialized applications. The general knowledge transfers; only the task-specific layer needs to be learned.
From there, the capabilities get more striking. Few-shot learning refers to systems that can learn a new task from just a handful of examples — sometimes as few as one or two — rather than the thousands typically required. Zero-shot learning goes further: a system that can handle a task it's never been explicitly trained on at all, by reasoning from related knowledge it already has.
These aren't tricks. They're evidence of something real: that sufficiently capable models develop representations of the world that are general enough to apply in contexts the training data never covered. The better the underlying model, the more it can generalize.
Transfer learning is where the mechanics of this start to become clear.
