Agent Smith
In summer 2018, I started an amazing journey by writing a short paper, "Developing a Document Trained Automated Advisor." The idea is to use AI to develop an AI (I must admit I am a big terminator fan). I incorporated the idea, later named "agent Smith" into Jill Watson over the course of 2019 working as a research scientist at the Georgia Tech. DiLab.
The name "agent Smith" is in reference to the movie, "the Matrix." Agent Smith generates copies of Jill Watson, with each copy being trained on unique datasets. The tool is designed to be used by non-experts. Before agent Smith, it took multiple NLP experts hundreds of hours to create a Jill Watson. With agent Smith, a non-expert can create a jill Watson in less than 20 hours (as we continue to automate processes the goal is to get below 5 hours).
The above mockup of agent Smith's user interface provides guidance on how the process worked. Agent Smith is driven by a domain model (in this example we are using the syllabus domain). Using the model, the system asks the course creator a series of questions about the specified domain. The answers provided by the course creator actually generate the agent answer pool. This is by design because it means the agent never provides a response to the user that has not been vetted by the course creator. This is so important nowadays because nobody wants an agent that they created saying something awful!
The ability to train an AI agent by simply answering its questions, then having the generated agent go out and answer questions in lieu of its trainer is the gist of agent Smith. The system is an amplifier. Amplifying the reach of the trainer by answering 100's of questions a minute 24/7.
Some thoughts on utility:
- Course creators can use it to amplify their impact by reaching out to more students than a single person could possibly support.
- K-12 teachers can use it to automatically answer parent's questions about where to find material.
- User Interface designers can use it to train agents for providing guidance on how to use their web applications.
- HR could create agents to answer employee questions about policies or help new employees fill out forms.
- Drug companies could generate agents to answer questions about their drugs.
Xprize paper: https://arxiv.org/pdf/2006.01908.pdf