Despite being impressed with the functionality, the author noted some areas for improvement. For instance, when asked to schedule a meeting without first asking about the schedule, the model sometimes tried to use the python interpreter as a function, passing in python code to execute. The author suggests that while this can be caught, turning it into a pleasant user experience might be challenging. The author concludes by expressing excitement about exploring similar functionality in popular open-source models.
Key takeaways:
- The author experimented with OpenAI's Functions API to create a module that emulates a portion of a calendar service API, using it to perform simple scheduling tasks.
- He used Elixir and Livebook for coding and demonstrated how the model could understand and execute function calls to fetch and schedule events.
- The code consists of a core module that interacts with OpenAI's API, a plugin module that exposes certain functions to the agent, and a behavior that provides the specification for callable functions.
- While the author was impressed with the functionality, he noted some issues, such as the model trying to use the python interpreter as a function, and inconsistencies in GPT-4's responses.