Don't sleep on email as an LLM interface
Email is undervalued as an interface for LLMs.
Email is ubiquitous. We used to say at Twilio, “SMS is the only app that’s installed on every phone,” but that’s true of email as well.
Email is frictionless. No signup. No login. Just send an email to get started.
Email is natively collaborative and multiplayer via CC and FWD.
Email is ripe for virality. See: Hotmail, or that chain letter your uncle sent you.
Email is where most business conversation happens, especially outside of tech.
Email is where long-form content is created and consumed. Creating and consuming long-form content is where LLMs shine.
Email is where we all have an untapped archive of unstructured text.
Email is async and doesn’t carry the expectation of immediate response. Slow generations are more tolerable.
Email is cross-platform out of the gate.
Email based apps are lightweight and easy to iterate.
Email is where your users already live and work.
Also, despite email's long ubiquity, we've historically lacked an easy way to parse emails so that they could be used as an interface for other applications. LLMs unlock that ability. For decades, email has been the de facto way of communicating via the internet, and LLM's just came along and unlocked a brand-new way to use the medium.
So, I see parallel opportunities for Email + LLMs:
- Use email as an interface for LLMs based apps. Whenever you're tempted to build a chat-with-AI feature, ask, "would email be a better interface?"
- Use LLMs to unlock email-as-an-interface for other software applications. The application itself may have nothing to do with AI, but you can use LLMs and function calling as a middle layer to transform unstructured text into structured inputs.
To explore these opportunities, I built a gpt wrapper using FastAPI and Postmark that executes different scripts depending on the inbound email address. You can think of this as a personal platform for email-based custom GPTs – though I call each app a "HaiHai."
Here are a few HaiHais I've created. You can test any of these out by sending an email with the suggested prompt in either the subject or body:
adventures@haihai.ai - send a description of a setting and story to play a bespoke choose your own adventure style story. CC your friends if you want to them to join.
recipes@haihai.ai - ask for a recipe. Add restrictions to your heart's content.
trivio@haihai.ai - Send literally any topic to spin up a Trivia game. It's more fun if you CC your friends.
bos@haihai.ai - ask a question about SaaS and you'll get an answer based on the last fifteen years of talks from my favorite conference, the Business of Software Conference. (This is a RAG app built on fifteen years of speaker transcripts)
This platform has been a great way to quickly build and deploy a bunch of LLM based apps over the last year to learn what these new tools are capable of. This month HaiHais sent and received over 2,000 emails, and I've learned a ton watching folks from all different walks from life – from grandparents to public school principals – interact with LLM based apps via email. If there's interest, I'm happy to chat about those lessons, or dive deeper into how I built this.
When ChatGPT launched, GPT-3 had already been available via API for over a year. ChatGPT didn't set the world on fire because the underlying AI was new, but because a new UI made it more accessible.
Chat is an undeniably useful, but it's not the ultimate interface for all things AI. A lot of the opportunity in building with LLM is going to come from combining your UI with your (users') data, and meeting your users closer to where they already work. Copilot and Cursor are great examples of this.
You can reach me at: