Recently there has been increased interest in using large language models like GPT-3 for open-ended information retrieval. However, simply prompting a large language model with a question and getting an answer is often insufficient. More sophisticated strategies are needed to fully leverage these powerful AI systems. Dave Shapiro has proposed an iterative four-step approach he calls the "Basher Loop" - Brainstorm, Search, Hypothesize, and Refine.
The first step is to brainstorm a list of potential search queries that could help answer the user's question. This takes advantage of the associative abilities of large language models to come up with relevant search terms the user may not have thought of.
Next, those queries are used to search external information sources like Wikipedia or the web. This provides a diversity of information beyond just what is encoded in the language model's training data.
The results of the search are then fed back into the language model to hypothesize an answer to the original question. With more context from the search results, the language model can formulate a more comprehensive response.
Finally, the process repeats by further refining the search queries based on what was learned. Each iteration develops a better understanding of the information landscape and hones in on the most relevant facts.
This technique combines the strengths of large neural networks with traditional information retrieval principles. It essentially turns the language model into an AI-powered research assistant.
Some key advantages of the Basher Loop approach:
- Provides more thorough and nuanced answers
- Handles open-ended questions, not just fact lookup
- Allows querying external data sources, not just the LM's training data
- Mimics the iterative process of human research
- Continuously improves queries and hypotheses
This prompt engineering strategy demonstrates the future possibilities of AI assistants. Rather than just answering simple questions, they can synthesize information and iteratively explore topics. While current systems are still limited, the Basher Loop shows the path forward.