GPT-5 got announced yesterday, and it wasn’t a particularly meaningful leap in terms of raw model capabilities. Obviously scaling laws still apply, and the added compute, better data, more reinforcement learning still result in an improved model, but nothing major was (at least obviously) unlocked yesterday.
What we did see yesterday was OpenAI flex its product muscle, by announcing one significant area of improvement that they dedicated a significant amount of work to (software development workflows), but mostly just scaffolding around the new model. We saw an improved voice mode, a ton of client work (improvements on both the web and desktop clients for ChatGPT), but frankly not much else.
Major frontier labs are AI research facilities with a product problem. They need to spend a tremendous amount of resources on training these models, but they also need to spend more and more time on the scaffolding so that users can extract value from them, and in order to distance themselves from the competition.
A large language model on its own is a spectacular piece of technology, but I would argue that a larger and larger percentage of the magic now lies around it. The magic is in when the product creates an artifact for you; in how you can interact with it while driving via voice mode; in how it integrates with the tools you use in your day to day; in how it orchestrates work that is handled by other models further down the chain.
I’ve argued several times over the past few years that this is a time for design and user experience folks to shine. The old practices of ethnography, hovering by someone’s shoulder and figuring out how they use a product, are highly relevant today. We’ve built the most amazing clay, and we now have to shape it to best serve us.
So I will make my argument again: AI labs are just regular product companies. The models are capable, but ultimately useless without the scaffolding of the right UI and the right interactions. As an industry, if we are to keep pushing the machinery of AI forward, we need designers to shape these systems, engineers to implement them, and the users to learn from. Maybe this is you, reader?