Here's something that doesn't get mentioned enough in conversations about AI: most models, once trained, stop learning.
You collect data, you train the model, and then you deploy it. From that point on, the model is static. The world keeps changing — new products launch, language evolves, fraud patterns shift, customer behavior moves — and the model just keeps applying what it learned from data that's now a year old, or two years old, or more. For a lot of applications, this is fine. For a lot of others, it's a real problem.
The obvious fix is to keep training. Collect new data, update the model, redeploy. And that works, to a degree. But it surfaces a deeper issue that turns out to be genuinely hard: when you teach an AI system something new, it tends to forget what it already knew. Not gradually, the way humans forget things over time, but abruptly — the new learning overwrites the old. This phenomenon, called catastrophic forgetting, is one of the central challenges in building AI systems that can actually keep up with the world.
The articles in this section are organized around two related questions. The first is the forgetting problem: why does it happen, what makes it so hard to solve, and what approaches have emerged for building systems that can learn continuously without losing what they already know. The second is a different dimension of adaptability: not just learning over time, but learning across domains — how AI systems can apply knowledge from one context to another, or learn new tasks from very few examples, or handle situations they've never encountered before.
These aren't separate problems with separate solutions. They're two sides of the same question: what does it take to build AI that keeps getting better rather than staying frozen?
