Keep pulling the thread on Amber Teng.
The end-to-end development time for the GPT-3 based cover letter generator, including learning the API and deployment on Streamlit, was approximately five hours.
The ramp-up time for building an application with OpenAI's GPT-3 was surprisingly quick, requiring only a few lines of code to generate a functional cover letter generator.
OpenAI provides a "playground" environment that allows users to test prompts and get results from models like GPT-3 without writing any code.
When prompting GPT-3 for a cover letter generator, a prompt that was too simple resulted in generic and factually incorrect outputs, such as claiming the user graduated from Stanford University.
When prompting GPT-3 for a cover letter generator, a prompt that was too specific caused the model to copy the input text verbatim instead of generating new content.
Among OpenAI's models available at the time, the DaVinci model was preferred for the cover letter generator because it had the highest token count, allowing for longer cover letters.
Adjusting the "temperature" parameter in GPT-3 is a key part of prompt engineering; a higher temperature (closer to 1.0) produces more creative but higher-risk results.
Using parameters like "frequency penalty" and "presence penalty" in GPT-3 can encourage the model to introduce new ideas and concepts rather than repeating words.
Common errors made by GPT-3's DaVinci model in the cover letter generator included assigning an incorrect gender to the user or fabricating user experiences, such as attending MIT.
OpenAI's Curie model, which has a smaller token limit than DaVinci, tended to generate cover letters in bullet-point format rather than the desired paragraph form.
The combination of a high temperature setting and a long requested text length is a "danger zone" for GPT-3, often leading to incoherent outputs, typos, or non-words.
At the time of the project, OpenAI's DaVinci model had training data up to June 2021, while other models like Ada, Babbage, and Curie were trained on data only up to October 2019.