Chat is a Solvent

AI has a major UX problem: Chat is a solvent.

The Sea of Text
Chat is amazing. Being able to converse with our machines in true natural language was the stuff of science fiction a few years ago. Now it’s something I do throughout the day, and the implications of the interface are barely explored.

But Chat-as-an-interface is a double-edged sword. When you’re working with AI on your desktop, just getting things done, it begins to become obvious:

Chat dissolves every object, every notion, every thing, into a single undifferentiated stream of text.

You’re faced with a wall of text ticking past in a terminal or a web app. Once the conversation has progressed sufficiently, it’s hard to remember where it began, difficult to see “where we’re up to” when building an artefact: “go back to that previous image and add the blue there” — “ok let’s go back to that todo list you had”. Everything is text, everything is a discussion, and it’s wonderful and terrible at once.

Fundamentally, you have a UX problem: you can’t build affordances when your object has no edges. And in the current paradigm there is only a sea of text. Everything is just words.

For decades, SaaS designers have established conventions for workflows, real-time collaboration, sharing and distribution, version control, prioritisation, project management - a whole host of useful noun-concepts that we take for granted in knowledge work.

And now we’ve replaced those mature UX interactions with the equivalent of amnesiac Slack DMs.

Developing the AI UX
To develop a notion of AI UX, we have to extract things from that stream of text. Skills are one that was identified early. Memory is another, currently fragmenting into many possibilities. There are very likely many, many more things waiting to be discovered.

Once we’ve identified them, we have to be able to define those things, extract them and manipulate them. Fundamental questions have to be answered that will establish the UX language of the AI era: How do I explore my memory? How do I compose my skills? And crucially — how do I share these things with others?

I say “crucially”, because one point stands out right now: how do my colleagues pick up something important that’s dissolved within my chat stream? We’re operating in silos again, decades after Web 2.0 collapsed them.

Even “Skills”, the most established reuse point in AI, feel like a proto-paradigm. We certainly need to be able to capture units of actionable knowledge and reuse, refine, and compose them. But what exactly are skills? What are their inputs and outputs? How do they glue together? A programmer would immediately reach for types and signatures — what does that look like in the world of natural language back-and-forth?

Right now, we’re missing the foundations. Imagine desktop computing without the notion of files. The web without the “post”, the “share” button, the mobile without “apps” and “pull-to-refresh”.

How do you discover these things, hiding in the sea of text? I think there are four key criteria to guide us: you can name it, you can compose it, you can share it, and — the strongest indicator that it’s an artefact in motion, on a journey — you can version it.