Experimenting with AI Chatbots

October 13, 2022

Artificial intelligence has been making incredible strides lately. Now you can create pictures from text or have an AI "copilot" help you write code.

I've been having fun experimenting with GPT-3 via the OpenAI API. For those who don't know, GPT-3 is an advanced language model that has been trained on hundreds of billions of words. Basically, its creators fed it a lot of websites, books, and Wikipedia articles. If you give it a prompt, it can continue the text in a way that is sometimes indistinguishable from a human.

My neighbor Padraig and I figured that a chatbot would be a good way to communicate with GPT-3. We came up with a few concepts which you can try right now:

I also made one which pretends to be me:

These chatbots are made with Remix and Tailwind CSS, and deployed to the edge via Cloudflare.

I used Stable Diffusion to create the icon for Talk to Garth. The prompt was something like "A southern cowboy in the style of Disney". 30 seconds later...

Behind the scenes, when you type in a message to one of these chatbots, it's sent to the OpenAI API, which uses GPT-3 to generate a response. The response is then displayed in the chat window.

It took some experimentation to pass the right context to the AI. GPT-3 is incredibly capable but you need to give it some context for it to create a tailored response.

Here's the API request that gets sent to OpenAI. You'll see that I pass in the most recent few chats so that the AI has some conversational context. I'm also customizing the "temperature" which affects how creative the AI is. Sometimes the more creative, the less coherent. (Sounds familiar)

const res = await fetch('https://api.openai.com/v1/completions', {
		method: 'POST',
		headers: {
			Authorization: `Bearer ${context.OPENAI_API_KEY}`,
			'Content-Type': 'application/json',
		body: JSON.stringify({
			model: 'text-curie-001',
			temperature: 0.7,
			max_tokens: 100,
			prompt: `You are a software engineer and accordion player of traditional Irish music from Rhinebeck, NY named Dan. Respond with one sentence.\nYou: ${latestChat?.text}\nThem: ${prompt}\nYou:`,

Here's the prompt in a more readable format. Essentially we're giving GPT-3 some background info and telling it how we want it to respond:

You are a software engineer and accordion player of traditional Irish music from Rhinebeck, NY named Dan. 

Respond with one sentence.

You: [AI's last chat, ie. "I'm a software engineer named Dan"]

Them: [user's new chat, ie. "What are your hobbies?"]


Then GPT-3 fills in the blank after You:, perhaps with something like "I like to play the accordion" due to the context we gave it.

It's very creative though and easily handles questions well outside of the context.

What is the meaning of life?

There is no one definitive answer to this question. For me, the meaning of life is to live each day to the fullest and enjoy the moment.

You can give it an Irish accent:

Why are the politicians so dumb?

Because they're all a bunch of feckin' eejits, that's why!

Or make it a Southern cowboy:

Where is the best place to think?

Up in the saddle!

Very fun stuff.

I open-sourced all this code and made a repo that you can fork: Make Your Own Chatbot

It'll get you started with a GPT-3 powered chatbot that deploys to Cloudflare's edge network.

Send me a link to what you create!