Language Models are Few-Shot Learners
Map
Notes
Library
Scale a language model large enough and it learns to adapt on the fly.
Fine-tuning requires labeled data for every new task.
In-context learning requires only examples in the prompt.
A reading of Language Models are Few-Shot Learners
by Tom B. Brown et al. (OpenAI)
Zero-shot (instruction only). One-shot (single example).
Few-shot (up to 100 examples). Fine-tuned (gradient updates).
— the prompt as a temporary program
Rather than updating model weights, in-context learning treats a handful of prompt examples as a temporary, task-specific 'program' the model executes at inference time. No gradient updates. No new weights. Just examples that shift the model's behavior in the moment.
For years, the assumption was: bigger model, better performance, done. Brown et al. showed that how you frame the task in the prompt matters as much as the model itself. The examples you include, their order, their phrasing — all shift the model's output. This is not magic. It is meta-learning in the forward pass. For builders: few-shot prompting can substitute for expensive labeled datasets. Larger models extract more signal from the same examples. But capability gains arrive bundled with bias and misuse risks that need proactive mitigation.
Choose your setting by data budget and distribution shift tolerance.
We read every message. No algorithm, no auto-reply — just Sophie and the LOOMUS team, with coffee, sometime this week.
Your note is in the inbox.
We'll write back when the kettle's on.
Saved with your account — part of a study that grows every time you distill. Nothing here expires, and nothing is lost.
Every LOOMUS reader may claim a planet — once, for life. It’s a slow choice, made when you’re ready.
Enter The Choosing →or simply look around your space →
No hurry; the sky holds your seat.