ICL
Also known as: in-context learning, in context learning, learning from examples in the prompt
Before GPT-3, teaching an AI a new task meant retraining it on new data, a slow and expensive process. GPT-3 showed that a large enough model could do something surprising: learn from analogies embedded directly in the prompt. You describe a task, give a few input-output examples, and the model extrapolates the pattern to new inputs. That's in-context learning.
ICL sits underneath most practical prompting work. When you write a prompt that includes a couple of examples of your desired output format, you're using ICL. The model isn't 'learning' in the traditional sense of updating its weights (the internal parameters that define its behavior). It's using its existing training to rapidly pattern-match within the current conversation.
The limitation is that ICL lives entirely inside the context window. Once the conversation ends, the model forgets everything. It also degrades as the number of examples grows or the task gets more complex. This is one reason fine-tuning still matters: for tasks that require deep, consistent adaptation, baking examples into the model's weights beats trying to cram them all into the prompt.