The Human-Intelligence Interface

For the entire history of modern computing, software has had one job: make it easier for humans to operate machines. Every software you’ve used is an example of the human-computer interface.

That interface was a compression step. Machines were superhuman at narrow things but they couldn’t understand us. So we had to compress our intent into what a machine could parse: buttons, forms, rigid flows.

We’ve seen vertical and horizontal software compete on abstractions, where better abstractions outcompeted worse ones.

In the era of intelligence, we’re designing a new category of applications. Not the human-computer interface, but the human-intelligence interface. What does an interface look like when you no longer have to compress your intent?

There’s now a middleman in the loop: a language model that takes your intent and converts it into instructions for a machine. It sounds like one more translation layer but for the first time, the machine is adapting to us instead of us adapting to the machine.

And that flips the purpose of software: maximize how easily you can describe your intent to something far more knowledgeable than any traditional interface.

The goal now is to scale a human’s leverage: one person interacting with many machines at once.

As a consequence, a lot of applications don’t need human interfaces anymore, they need machine-readable ones. APIs and CLIs have existed for decades; what’s new is who is calling them.

Traditional software took you down a rigid pipeline to execute a verifiable task. With a language model, you can sprawl. You can go after tasks that were never verifiable enough to build a pipeline for.

Today, there is a tradeoff: you give up some certainty for a lot more surface area. But eventually, as models become more capable, surface area will not come at the cost of certainty. We wrote about this in Primitives over Pipelines.

And because intelligence can act on your behalf, it’s not hard to envision a world where the primary interface between a human and software, or the internet itself, is a language model.

So the question becomes: what’s the best interaction to build for the human-intelligence interface? Today, it’s the chatbot but it largely remains an unexplored space.

We wrote about a few interactions patterns for the human-intelligence interface in Challenging the Chatbot.

Originally published on Twitter.