MilesBot 1.0
This is a multi-turn bot that is fine tuned to mimic me, Miles the page author.
This is based on the OpenAI GPT-3 project, implemented with an OpenAI.com API. The ML model is trained on 175 billion parameters. The model is fine tuned on 850 life story questions. The lookback history is about 12 turns.
It isn’t entirely accurate. It often synthesizes a response that is counter to the training data. But for only 850 life story questions to capture someone’s biography it is pretty good.
Picture Books using AI Helpers
I used OpenAI ChatGPT and OpenAI DALL·E 2 to create greeting card inscriptions and cover art, and picture books with themes that interest me. It’s an AI hobby rather than a professional activity but have a look! There are instructions in the picture books if you are interested in doing this too.
Click for the AI greeting cards and picture books!
Natural Language Understanding Bot
This is a multi-turn conversational AI bot. If you want to change the thread, type Reset.
It uses a lot of resources so I put it up on a company server. You can try it here:
Sample Dialog
You: Where do you like to travel? Bot: I like to travel to places like the Caribbean, Europe, and Asia. You: Where in Asia? Bot: I'm from Malaysia. You: What is a good book? Bot: The Book of Mormon. You: Have you read it? Bot: I have not. You: How old are you? Bot: I'm 21. You: Are you in school? Bot: I'm in college.
For techies: The NLU core is the OpenAI GPT-2 model that was trained on eight million Reddit pages in 2019. I’m using Python3 with Flask (generates Web pages on the fly) on an Apache2 Web server with the mod_wsgi module (connects Flask with Apache) on Ubuntu Unix in the AWS cloud. I tweak parameters now and then so the responses may differ.
There is a hidden front end that captures common phrases like “Who are you” and replies from a Q&A table if there is a good match using cosine similarity, otherwise it goes up to the GPT-2 model. Start a phrase with qaz like “qaz Who are you?” and you will get the table reply plus a matching score. Compare with the regular response for “Who are you?” without the keyword qaz.
Phrase Completion
This is known as Masked Language Modeling (MLM). Leave a word out by using “[MASK]” in its place, and the bot will suggest how to fill it in. It’s not meant to be particularly user friendly. It’s just a demonstration of the technology and a reminder to myself of how to do it.
Masked Language Modeling (MLM)
Ref: https://huggingface.co/distilbert-base-uncased
Story Completion
Start a story with a sentence or two and let the GPT-2 bot complete it for you. The result may not be of publishable quality, but for an author, it can break a writer’s block to see predictive text that lies ahead.
Ref: https://huggingface.co/blog/how-to-generate
AWS Lex Market Quotes
[NOTE: I disabled this on 17 Apr 2023 because the free sandbox from iexapis is gone and they want $20/mo to re-enable it for something I rarely use. I’m keeping the code here even though it no longer produces quotes, to remind myself how to do it.]
This demonstrates an integration of a bot with AWS Lex. There are a lot of chatbot builders out there, but few that integrate with AWS Lex.
This uses Javascript/HTML to connect with Lex on the back end. It spells out a ticker with spaces: “AMZN” becomes “a m z n” because this feeds a phone audiobot. This trick forces it to spell rather than try to pronounce the ticker symbol.
Training Program Management at IIUSA
I am the President and CEO of IIUSA, which has extensive management capability in military training programs, covering military communications at echelons above corps (EAC). EAC is digitized voice, data, and video carried over military LAN and WAN links. Our military training courses cover several tactical assets, COMSEC, and tactical, and COTS network nodal management.
