Summary

  • This is a report on a Sunday hack project composing multiple AI tools.

  • The project aims to answer complex NBA statistics questions.

  • It uses GPT-3 as a general-purpose agent, and Statmuse, a specialized natural-language search engine for sports statistics.

  • The interaction between the two is orchestrated by LangChain, a Python library helping combine LLMs with “other sources of computation or knowledge”.

  • It combines GPT-3’s open-ended reasoning with Statmuse’s specialized database, to answer questions neither service can do alone.

  • It also fails in interesting ways.

  • The whole project was written in a few hours, and it shows what is possible now, and what remains to be done.

Original Article