Prototype then Deploy: shrinking foundation model quality into small models
Foundation Models (FMs) such as GPT-4 promise a natural language interface for data analysts to perform complicated transformation tasks and gain insights from their data. Despite the wide success of FMs, there are still high barriers to use on enterprise data such as hard to prompt, privately owned, and a high cost to run (a single run of 1M rows could cost more than $3.7K!). At Numbers Station, our goal is to bring state of the art FM technology to all data analysts. With a few natural language examples from an analyst, Numbers Station customizes a small FM on the analyst’s task and data. Our smaller FMs match the capabilities of proprietary large FMs and can be up to 2000x times cheaper. At Numbers Station, we are excited to take the first step towards making FM technology accessible to all enterprise analysts.