ChatWeb3

Inspiration

I was chatting with ChatGPT about crypto and it is living in the bull market of summer 2021! I wondered how it would analyze the recent crypto crash.

What it does

It is ChatGPT plus it knows about more recent news.

How we built it

I used the OpenAI fine-tuning feature. I used the "davinci" base model, their best one, and created a dataset of prompts and completions with information about recent events and characters such as FTX, LUNA, Celcius, SBF, etc.

After training, I build a proxy in Cloudflare Workers to handle the secret API key, and a simple React app hosted on Cloudflare Pages.

Challenges we ran into

Fine-tuning the model is really hard to get right, and some answers are actually worse now than using the base model!

Accomplishments that we're proud of

This LLM is totally state-of-the-art and this is a truly unique model as far as I'm aware.

What we learned

How to tune an LLM, how to pay for inference.

What's next for ChatWeb3

Train with a larger, better dataset, charge for usage to make it sustainable, explain what it is to more people who could make use of it.

Built With

• chatgpt
• cloudflare
• davinci
• llm
• openai
• react