Ian Silber - What it's like designing at OpenAI

2026-04-08 - 45 min - source - Read full transcript

Reid (host)Ian Silber

Key insights

OpenAI designers are trained to ask whether a problem can be solved with tokens or model behavior before reaching for bespoke UI.

Silber says the team constantly asks 'can we do this without pixels, can we do this with tokens, can we do this with the model itself' before designing new screens - treating the model as the actual product and interface work as the fallback, not the default.

ai-native-design-workflow

A new direct-manipulation editing container in ChatGPT came directly from usage data, not a design hunch.

The team saw that a huge share of ChatGPT usage is writing help, and that the turn-by-turn chat loop for edits ('make this shorter') was tedious. They built a container where users can select and edit specific text spans directly while chat instructions still work for broader rewrites.

ai-native-design-workflow

The best systems thinkers at OpenAI ask how a feature extends shared primitives, not just how it solves the task in front of them.

Silber points to 'skills' as an example of an emerging primitive: designing toward the deepest reusable abstraction means one piece of work improves many other features, versus optimizing a single isolated use case.

design-leadership-in-ai-era

OpenAI is only now building a formal design systems team, roughly two-plus years after ChatGPT's launch, because speed took priority earlier.

The centerpiece is the Dynamic User Interface Library, a system designed from first principles so the model itself can interpret and assemble UI components, rather than a static component library a human always has to hand-place.

design-leadership-in-ai-era

Silber frames a widening 'capability gap' between what models can already do and what the shipped product lets users access.

He cites Codex as an example of a model spending real time and tokens to do substantial work, while ChatGPT's product surface still exposes only a limited slice of that capability - closing that gap is a central design problem, not just an engineering one.

ai-native-design-workflow

The AI coding toolchain went from manual copy-paste to live, model-backed prototypes in roughly two years.

Silber traces the path from copying code between ChatGPT and VS Code, through GitHub Copilot autocomplete, to Cursor (still somewhat inaccessible to non-technical designers), to Codex - which lets designers ship working, model-connected prototypes instead of static Figma mockups or recorded videos.

ai-native-design-workflow

Silber compares today's AI-assisted building boom to the invention of the camera: it democratizes production without eliminating the need for skilled practitioners.

Before cameras, only trained painters could produce a portrait; afterward anyone could take a photo, yet expert photographers with taste still stood apart. He argues the same holds for software - anyone can now build something, but knowing what's worth building still requires craft.

future-of-the-designer-role

Silber expects the designer's core job to shift toward editing, directing, and curating rather than hand-executing every screen.

As AI tools let more people (engineers, PMs, researchers) produce working prototypes themselves, he thinks the differentiating skill for designers becomes recognizing good software from bad and steering ideas toward coherence, not personally producing every pixel.

future-of-the-designer-role

Silber's entire founding team joined OpenAI together as a package hire after their gaming startup's founder recognized GPT-4/ChatGPT as a discontinuity.

Silber, a fellow designer, and the rest of an eight-person team building a Minecraft-meets-Roblox browser game were separately recruited by a former colleague who had become head of product at OpenAI; he didn't initially realize the two designers he was courting already worked together, and the whole team ended up joining within about a week of first being contacted.

career-path-to-openai

Working inside a research-led company changes what drives the product roadmap compared to a consumer-product company.

Silber contrasts OpenAI, which started as a research lab and launched ChatGPT as a low-key research preview, with Instagram, which started as a consumer product; at OpenAI so much of the work is figuring out what the models are newly capable of and wrapping that in something usable, rather than starting from user research.

career-path-to-openai

Silber's top hiring signal is demonstrated curiosity about the technology, not an AI background.

He says candidates don't need technical AI credentials, but he looks for people who've gone deep on a side project or genuinely spent time playing with model capabilities to understand where they're strong and where they break, since that curiosity is what drives good product instincts inside a fast-moving research environment.

hiring-and-talent

Silber uses the story of John Madden refusing to endorse an early football video game as a model for designing at the edge of current technical limits.

Citing the 30 for 30 podcast's episode on the Madden game, Madden reportedly refused to lend his name to the game because the hardware couldn't render enough players on the field; a year or two later chip advances made it possible, and he signed on. Silber uses it to argue designers must work with today's real capabilities while staying ready to push the moment the ceiling moves.

future-of-the-designer-role

Media referenced

30 for 30 - podcast - Reid brings up the ESPN 30 for 30 podcast's episode on John Madden refusing to endorse the original Madden football video game until the hardware could render enough players on the field, as an analogy for designing at the edge of current technical capability.

Companies

OpenAI - Silber is head of product design; the episode covers how design works inside a research-led lab building ChatGPT and Codex.
Instagram - Silber spent about 8 years there before leaving; cites its deep investment in design tooling (Origami) as formative.
Artifact - An AI news app Silber joined after Instagram, founded by Kevin (Systrom) and Mike (Krieger).
Meta - Instagram's parent company at the time Silber worked there; built the Origami design tool internally.
Cursor - Cited as an early step in AI coding tools that was still somewhat inaccessible to non-technical designers, before Codex.
37signals - Silber references being 'really into the 37signals crew' early in his career, when designers were closer to writing front-end code themselves.
Descript - Episode sponsor; pitched as a code-connected prototyping tool that lets teams prototype directly against their real codebase and design system.
Paper - Episode sponsor; a design tool with a Chrome extension ('snapshot') that copies live website components into editable canvas layers.
Inflight - Episode sponsor, built by host Reid; a feedback tool for sharing prototypes and video walkthroughs and collecting structured feedback.

Techniques and frameworks

Model-as-the-product design - OpenAI designers are pushed to ask whether a problem can be solved with tokens or model behavior before reaching for bespoke pixels or UI.
Direct manipulation writing container - A UI change that lets users select and edit specific text spans directly rather than routing every writing correction back through a chat turn; grew out of data showing writing is one of ChatGPT's top use cases.
Dynamic User Interface Library (DUI) - OpenAI's first-principles design system, built so the model itself can interpret and compose interface components rather than a designer hand-crafting every screen.
Capability gap - Silber's framing for the growing distance between what underlying models can actually do (e.g., what Codex can do) and what's exposed in the shipped product surface.
PD whip - An internal Slack-style channel where OpenAI designers post rough work-in-progress prototypes or videos to get fast reactions.
Doc crits - A recurring OpenAI design ritual where a group reviews written docs/ideas together to riff and refine them.

Summary

Ian Silber, OpenAI's head of product design, walks through how he ended up there and what design actually looks like inside a research-led company. He spent about eight years at Instagram, then joined Artifact (an AI news app from Kevin Systrom and Mike Krieger), then joined a small team building a Minecraft-meets-Roblox browser game. When ChatGPT/GPT-4 landed, the game startup's founder recognized it as a discontinuity worth chasing, and the whole eight-person team - recruited by a former colleague who had become head of product at OpenAI - joined as a group within about a week of first being contacted.

The heart of the conversation is how design changes when the roadmap is set by what a model newly becomes capable of, rather than by user research alone. Silber describes OpenAI's habit of asking whether a problem can be solved with tokens or model behavior before reaching for bespoke UI - illustrated by a math-rendering feature (replacing raw LaTeX output with something interactive) that started as one designer's prototype and got hardened into a shipped feature because it was "so clearly valuable" once shared. A second concrete example is a new direct-manipulation writing container: usage data showed writing help is one of ChatGPT's largest use cases, and the old edit-by-chat-turn loop was tedious, so the team built a container letting users select and edit specific text directly while chat instructions still work for broader rewrites.

Silber repeatedly returns to systems thinking as the differentiator between good and mediocre design work at this scale. The best systems thinkers, he says, ask how a given feature extends shared underlying primitives (he cites "skills" as an emerging one) rather than solving only the task in front of them - so that solving something once compounds across many other product surfaces. This is also why OpenAI, roughly two-plus years after ChatGPT's launch, is only now standing up a formal design systems team, centered on a "Dynamic User Interface Library" built so the model itself can eventually interpret and compose interface components rather than requiring a designer to hand-place every screen.

On tooling, Silber traces a fast arc: from copy-pasting between ChatGPT and VS Code, through GitHub Copilot autocomplete, to Cursor (still somewhat inaccessible to less technical designers), to Codex, which now lets designers ship live, model-connected prototypes instead of static Figma mockups or recorded walkthroughs. He draws a direct line to Origami, the prototyping tool built at Facebook/Instagram, arguing that just as that tool let designers who embraced it uplevel their craft during the web-to-mobile shift, today's AI-native tools are creating a similar opportunity - and a similar risk of building something that "feels amazing" but is completely wrong, because functional no longer means correct.

On the future of the role, Silber is not worried the job is disappearing. He compares the moment to the invention of the camera: photography let anyone capture an image, but skilled photographers with taste still stand apart from casual snapshots. He expects designers to spend more time editing, directing, and curating what gets built as more people (engineers, PMs, researchers) can produce working prototypes themselves. Asked what he looks for when hiring, he says it isn't an AI background - it's demonstrated curiosity, evidenced by candidates who've gone deep on a side project or genuinely spent time probing what models are good and bad at.

Notable Quotes

"So much of our work is figuring out what the models are good at, and then trying to wrap that in a product that people can understand and can use." - Ian Silber

"What can we do this without pixels? Can we do this with tokens? Can we do this with the model itself?" - Ian Silber

"I think most people can be doing more editing and directing and curating. I don't think the job of a designer is going away anytime soon." - Ian Silber

"It's technically not possible to put 22 players on the field. We can only put 10 total... come back to me when that's possible." - Ian Silber, recounting John Madden's response to an early Madden football game

"Things are changing underneath your feet all day long. And it's very exciting." - Ian Silber