PRE-TRAINED ERA: Within Context and Prompt Engineering in Pre-trained Models

Pre-trained models have the ability to learn within context and make good predictions by leveraging biases and knowledge from a few examples. Another interpretation is that pre-trained models model a multiverse of fictional documents, and when prompted, the model is in a superposition of multiple possible documents. Prompt engineering involves constraining the generation space to include only documents with correct answers. Few-shot prompting works by constraining the space of possible documents to those that contain more correct answers. It has been shown that performing zero-shot translation, without providing any correct examples, can be better than performing translation with a few correct examples. The key is to establish a fictional narrative where a skilled translator flawlessly translates the sentence. This approach defines prompt engineering, which involves constructing fictional scenarios that can only be completed correctly. It is about imagining what kind of document might contain the answer.

Transcript

Play full episode

Transcript

Episode notes

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.

Get the app