Podcaster Lenny Rachitsky interviewed Ian Silber, OpenAI’s head of product design.

Ian Silber discusses task-specific UIs. He shows how ChatGPT’s transcript can make room for them:

Another good example is, then once you actually ask it to do something, the chat transcript doesn’t need to be purely text, right? We’ve seen these now evolve so far from what it was with when ChatGPT just launched. It’s no longer just a pure chat bot.

It actually has many features within that, right? So a good example is what we call writing blocks. You ask it to write you a letter or an email.

And it now contains that email in a block that you can directly manipulate, and you can type with it, and you can brainstorm, and you can edit little bits and parts of it. You can copy just that. And that has made that interaction much better.

And we’re seeing the same thing with image generation and creating special tools for that kind of mode. And I think that’s where the UI is going to start to really continue to evolve. How do we make if you come at this and you’re a designer? Maybe ChatGPT should actually really look different or have different affordances. Or you’re a data scientist or you’re using this to run your home, your your personal life.

Maybe it should really kind of work or look differently.

Chat is useful at the beginning of an interaction because users can express ambiguous intent in plain language. The UI can then provide task-specific controls. These might be a writing surface for an email or image controls for a picture. A designer might need another set of affordances. Silber’s example turns the post-chat argument into a product architecture already appearing inside chat. In other words, generative UI.

Chat usually collapses two jobs: understanding what someone wants and giving them the right environment to finish it. Silber on how one entry point could handle both:

And hopefully all of this gets simpler as we do it too, right? When we think of the super app, I think eventually we really think of the idea of especially for our broad consumer base. We do think that there should be this one universal input that you can ask that knows when it needs to answer really quickly, or it should be chatty, or whether it should go off and do work for you.

And so, that’s a big excitement that we can get to. How can we continue to simplify the experience such that you shouldn’t have to think about how this works. You shouldn’t have to think about what model you’re on or what sort of speed you’re using. That goes back to the balancing the different kind of personas and user needs. Of course we want to expose that for the right people, but then we want to make sure that we can get a default experience that works great for for most people.

Quotes slightly edited for clarity.