ChatGPT, Replit, and the AI Flywheel

Something happened to me last night that would have seemed like science fiction even a few months ago. It has me convinced that OpenAI and Replit may be the two most important startups of the 2020s.

OpenAI's ChatGPT project, which garnered 1 million users in 5 days (!), is a chat-style user interface for interacting with OpenAI's davinci large language model - a way for people to easily chat with "AI". It was trained on publicly available information to answer just about any question you may have.

What if you had a private chat assistant that could answer questions about your company or your product? OpenAI enables you to "train" their models on your own data, but that can be a painstaking, time-consuming process. I wanted to see how easy I could make it.

Enter Replit. Replit’s web-based IDE & virtual environment enables anyone to code, run, and build apps. It's like walking into a fully-stocked kitchen and cooking... well, whatever you can think up.

Last night, after the kids went to bed, I tried using ChatGPT and Replit to follow this thread:

1. Fetch historical customer support chats.
2. Use GPT-3 to parse the chats for product-specific question and answer sets.
3. Use those sets to train a GPT-3 model on your product information - perhaps alongside FAQs, help pages, and other product documentation.

You could then use that model behind a chat UI to onboard & train new employees, auto-suggest customer support responses, write blog posts for the website, do your laundry, etc.

What I did not expect was having built a working prototype in Replit in 3 hours:

1. Fetch and format historical conversation data from Intercom:

2. Feed each chat into the GPT-3 model, ask the model to structure the chat into a product-specific prompt/response output:

3. Add the output to a file for human eyes to review and grade:

After some testing and tweaking done in the Replit UI… it worked!

The kicker? The first line of Python I have ever written was for this app. Last night.

And that's possible because I didn't write the first line of of code for this app. ChatGPT did:

The ability to build and run an app like this in ~ 3 hours, combined with the recursive nature of this exercise, highlights the possibilities inherent in the flywheel. Sooner than we expect, the ability to build software will only be limited by logic and will, not time or knowledge. I was initially skeptical of Amjad Masad 's prediction that tools like these together will boost developer productivity 100x.

After last night, I have no doubt.

(Post?NOT?written by ChatGPT)