HTTP, SMTP, IRC

Sam Altman suggested he wants to make OpenAI your core AI subscription to a platform that everybody uses to use their product to do agentic things.

I have a different idea.

Qwen with questions (QwQ) can surf the internet by making HTTP requests, having something fetch the page, and then making another HTTP request serially until it ends up traversing the page to solve the user’s query.

Here’s a YouTube video showing it in action.

Essentially, if the model is good enough to understand HTTP, and it can also understand email, and it can also understand IRC, what we should be doing is hooking these models up to these protocols directly, and then optimizing what we do on top of those protocols to tolerate errors.

So email, you should talk to a very tolerant server that just accepts the very bare minimum headers. No spam headers, none of that.

We’ll not call it email, we’ll call it virtual messaging or something.

Essentially we repurpose the format of the email address and bootstrap a new protocol that descends from the existing protocol. You must control DNS to set the protocol records for the domain. I fully believe it is possible to make a simplified version of the SMTP specification that most engineers could navigate by hand. A subset of the existing specification, but a simplified version.

I tested with SMTP and IRC too and models can navigate both with ease.

Soe we let the models have access to HTTP and SMTP in the most simple, easy-to-understand forms. And then we can easily train those models on real HTTP and SMTP sessions to get even better. This is a model that probably never thought about HTTP or SMTP, but it’s in the training data.

I put forth a simple thesis: if we’re going to reimagine the web we take what we have and dramatically simplify it. And that simplification will just work with the agents because it’s already in the training data.