"Use few-shot examples to unlock hidden model capabilities When ESPN's Cricinfo site had a 20-dropdown query form that almost no one used, the speaker built a natural language front-end for it using GPT-3 and a custom domain-specific language the model had never seen. By feeding the model a handful of paired examples — a cricket question in English alongside its DSL translation — GPT-3 learned in real time to produce the correct output: "I had no access to internal of GPT-3, I had no access to the weights, but still, it worked." The practical takeaway is that before fine-tuning or retraining a model, try curating 5–20 high-quality input-output examples and passing them as context — you may unlock the capability you need without touching the model at all. Source: Vishal Misra, No Priors (Martin Casado), 'How LLMs Actually Work: Bayesian Inference, Causality, and the Path to AGI'"