Tag
Few-shot
Steering the answer with the right examples.
2 · Articles on this topic
Few-shot prompting: how a handful of examples steer an AI model
There is a simple way to make an AI model understand what you actually want, and it does not run through longer instructions: it runs through a couple of examples. Showing the model two or three cases of what you consider well done — the question-answer pair, the format, the tone — steers it far more than any description in words. This technique is called few-shot: a few examples inside the request that shape what the model writes next. It is one of the most practical levers for working well with AI, and it explains why the same model sometimes jumps from generic answers to on-target ones without changing a single line of instruction. But it has a limit no example solves on its own, and that is where the conversation leads to Arena.
AI Reasoning: What Happens When a Model Thinks Out Loud
The latest AI systems have learned to show more than the final answer: they lay out the path that leads to it, step by step. This is chain-of-thought reasoning. Exposing the intermediate steps often improves quality on hard problems and makes the path checkable, but those steps are still generated text, not a guaranteed record of what happens inside the model. Knowing the difference between reasoning as a process and reasoning as an explanation is the key to trusting outputs the right amount. And once you can compare reasonings side by side, they stop being a monologue to accept and become material to examine.
All topics