Beyond the Chat Box: OpenAI's Head of Design on Adaptive Interfaces and the Designer's Future

Host Lenny Rachitsky opens with results from his survey of the tech workforce, in which designers and user researchers were the most unhappy on nearly every dimension: overwhelmed, anxious, pessimistic and unlikely to recommend their role. Silber says he sees a mix in his own world, with many designers optimistic and excited, but he understands the uncertainty. People wonder whether expectations have changed and whether they must now ship code every day.

He points to a recurring theme from internal surveys at OpenAI: engineers have become 10x or sometimes 100x more productive, but the design team has not. Working with a coding agent is fairly binary and usually gets it right, whereas design still involves trying something, finding it bad, trying again, and collecting user and internal feedback. AI compresses idea generation, but the feedback loop and the alignment overhead of a large company remain.

That lack of clarity about what is expected of a designer is what causes unease, he suggests. Yet he describes OpenAI as an experimental company that treats itself as a research lab. Designers are hired for curiosity about the tools rather than AI backgrounds, and he says the bar and the thinking on his team are rising as they use these tools.

Rachitsky notes that about half of survey respondents are having the best time of their careers, and the strongest correlate is feeling amplified by AI. Silber says the people embracing the shift use AI at every step, for example throwing an idea straight into Codex or Claude or another agent to think it through and prototype. It does not have to be the exact right way, but trying a different way of working helps, and a community of designers experimenting together helps too.

He emphasizes adaptability because what failed yesterday may work today. He recalls trying coding agents on side projects during paternity leave late last year, hitting hiccups, then trying again a month later after Codex launched and finding that it just worked. In his view we are in the early innings, so what feels cutting edge today will look archaic in a year.

That leads to his claim that it is the best time in history to be a designer. When he was learning as a teenager, he says, you viewed page source and used tools like Dreamweaver; today you can go from an idea to a working version in minutes and anyone can be a designer. As more software is built, he argues, the human element, such as understanding users and delighting them, becomes a more important way for products to stand out.

Rachitsky brings up a previous guest's image of PMs, engineers and designers in a three-way standoff, all pointing at each other and expecting to win. Silber says the role boundaries have always been blurry, since he has seen designers who were strong product managers and others who were polished engineers. Single-silo definitions often exist only because large companies need ladders to organize people.

Still, he argues that bigger organizations need people with certain responsibilities. Expecting one person to design, rally the team, align the company and engineer the result is a tall order. A good PM can sharpen the problem space and handle resourcing, and an engineer can make sure the output works for the broader system, even if people move more fluidly between roles.

He adds that a startup is different: if he were starting a company now he would hire the best generalists he could find. Rachitsky summarizes the view as roles that stay distinct while each person can do more of the others' jobs, and Silber agrees, saying the different hats needed to build great software remain important functions.

Asked whether AI will become an extremely good product designer, Silber says it already is, and that its biggest virtue is that it is accessible to everyone. He cautions that it is not necessarily the best visual designer or best at hierarchy, typography or interaction design, and that predicting the future is impossible. OpenAI's philosophy is to build tools that let people be more creative and build more.

Where do humans stay valuable? He names truly understanding what people need by watching them use something, inventing something new, and the human point of view behind a product. Designers should hone their answer to who they are building for and why. Rachitsky relays the Codex PM's argument that AI struggles with novelty because it is trained on existing sites.

Silber agrees that great design is about the new, citing the iPhone's multi-touch, Instagram embracing ubiquitous mobile phones, and Snapchat flipping how people communicate. None had training data to draw on. Now designers face new questions such as how to talk to AI by voice and how to use an agent to carry out tasks, so human judgment about what good looks like stays vital.

Silber believes in hiring a well-rounded team: some generalists who can do everything, plus visual designers, brand experts, great prototypers, strategic product thinkers and people obsessed with the cutting edge. No single designer ticks every box. He tells candidates OpenAI does not expect an AI background, but it does expect high curiosity and aptitude for learning, including how to build tools that enable more people.

Prototyping stands out in candidates because it has become more accessible, and so does having a point of view on where the tools are going. He also stresses systems thinking. Because OpenAI is building something with many capabilities, he says, the team needs composable primitives, as Notion did early, so the product feels like one cohesive thing that the model can reason about.

Rachitsky jokes that systems thinking has come up on his last five podcasts. Silber says his own encouragement to designers has been "just do less": do not design something if an existing component can be extended, or maybe the feature is not needed. He says thinking across everything, not just your own area, matters more as speed increases.

Rachitsky cites Jenny Wen's view that the traditional process of prototype, mock, test, research and iterate has lost its time budget, and the designer's role has shifted to big-picture steering. Silber says OpenAI works both ways. Some ChatGPT features, such as the composer that changes daily, get obsessive attention where they try 100 things and ship one, while Codex and the super app involve trying things quickly and building in public.

He explains that the team A/B tests and runs user research on some things and takes big swings on others, then learns quickly. Because capabilities, intelligence or latency can shift, a needed interaction might disappear or a new one might become possible, so you should not be too precious until something has hardened. Making ChatGPT feel deeply embedded in personal life is a bigger project that gets whiteboards, Figma, prototypes and long dogfooding, while other ideas go from concept to ship in four hours.

Rachitsky summarizes it as picking your battles and focusing design effort on what will be durable. Silber agrees, with the caveat that it is hard to know what that is.

Asked about balancing a huge existing audience with the need to invent, Silber says OpenAI's users span an unusual range: someone choosing a recipe or a haircut, and someone automating a farm in Japan. He contrasts this with Instagram, where the product was simpler. OpenAI believes in a capability overhang, in which most people get only a sliver of the value available, so it builds an extremely simple experience for most users while making cutting-edge features accessible to those who care.

The team experiments early with enthusiasts through the desktop app and Codex, and ChatGPT's work features are reaching a broader audience. Ultimately, he says, things work for billions when people no longer have to think about switching modes. This is a design challenge he calls the blank-box problem: how to build an interface that can shape-shift into anything, from image generation to a PowerPoint, helping novices while getting out of the way of daily power users.

Rachitsky relays Kevin Weil's argument that conversation is actually a good interface for spanning a wide range of intelligence. Silber says the team works on helping people express intent, for instance with tappable follow-up questions to open-ended prompts, and on making the transcript richer than text. Writing blocks put an email or letter in an editable block, and he expects the UI to keep changing depending on whether you are a designer, a data scientist or running your home.

On the long-term vision, Silber expects ChatGPT and Codex to become more adaptive to the context of your life. He thinks proactivity is a big untapped part of the experience, and he points to early signs in ChatGPT work, which can offer suggestions based on a calendar or Slack. He also notes a step change in voice, and says its ability to go from answering questions to doing things for you will change a lot.

For a broad consumer base, the idea of a super app is one universal input that knows when to answer quickly, be conversational or go off and do work. People should not have to think about which model or speed they are using, though those options will remain exposed for those who want them. He also expects richer voice input, more interactive visual outputs, and durable workflows you can reuse rather than starting from scratch each time.

Silber says he was intimidated joining OpenAI after GPT-4 had just come out, and still has a lot to learn. His advice to his day-one self is to recognize that everyone is early. He also felt he was bad at leading a rapidly growing team until friends in similar roles told him they felt the same, and he recommends finding a mentor or buddy going through similar challenges.

Rachitsky mentions a post against "AI confidence theater", and Silber agrees that being humble beats pretending you have it all figured out. To a designer feeling overwhelmed, he says the feeling is rational, and the answer is to explore and have fun. He urges focusing on the outcome: the goal is great products people love, not process or maximizing tokens, and the person steering and directing the tools still has a huge role.

On personal use, he spends less time in the pixels and uses ChatGPT or Codex to prototype ideas that strike him at night, and to summarize important Slack messages before meetings. He credits ChatGPT work running fully in the cloud with letting him do this from anywhere. On failure, he cites Instagram's IGTV flop, whose lessons fed into the successful Reels, and says what matters is how you fix forward.

Looking back across Groupon, Instagram and OpenAI, he says each company has its own DNA and you succeed by leaning into it rather than trying to change it. He carries Instagram's motto of doing the simple thing first, while acknowledging that OpenAI's current stage sometimes calls for a different approach.