OpenAIChatGPT Atlas and the next era of web browsing — the OpenAI Podcast Ep. 9
CHAPTERS
- 0:00 – 2:00
What ChatGPT Atlas is: a browser built around natural-language interaction
Andrew sets the stage for Atlas as a new kind of browser where ChatGPT isn’t a bolt-on, but the central interface. Ben and Darin frame it as a tool for telling the computer what you want, then letting it use the web to accomplish tasks over minutes to months.
- •Atlas is designed for natural-language browsing and task completion
- •ChatGPT is the “heart” of the browser, not a sidebar add-on
- •Supports long-running tasks and personalization over time
- •Helps users return to paused work by remembering context
- 2:00 – 3:40
Why build it now: models, computer-use, and the “slope” of capability
The team argues that recent leaps in model performance and “computer use” make integrated browsing experiences finally compelling. They compare progress from earlier agent experiences (like Operator) to Atlas’s speed and capability, and emphasize laying foundations early for the future.
- •LLM and computer-use capabilities reached a practical inflection point
- •Atlas reflects rapid year-over-year progress in agents
- •OpenAI wanted to establish a foundation for future browsing workflows
- •Goal: make ChatGPT the default starting point in the browsing journey
- 3:40 – 10:08
The browser isn’t going away: the durable, open platform for work and research
They discuss why browsers remain central despite shifts like mobile and app ecosystems. The web’s openness, lack of gatekeepers, and role as a conduit to information make it resilient—and generative AI naturally augments that experience rather than replacing it.
- •Browsers remain a primary tool for research and substantive work
- •The web’s openness enables broad publishing and consumption
- •Even ChatGPT’s rise started on the web, showing platform durability
- •AI can help users understand, navigate, and act on web content
- 10:08 – 13:30
From “semantic web” dreams to models that meet the web where it is
Andrew raises the idea that the web never fully became machine-annotated in the way early visions suggested. Darin explains that modern models can interpret human-oriented interfaces directly, reducing reliance on perfect structured markup and enabling machine help on today’s messy web.
- •Semantic web adoption was limited by incentives and effort
- •Modern models can interact with content as humans do (UI, language)
- •AI meets existing web design rather than requiring redesign
- •This capability extends naturally to browsers as “user agents”
- 13:30 – 17:35
Browsers are complex “mini operating systems,” and Atlas goes beyond a plugin
They unpack how much hidden engineering lives inside browsers and why Atlas needed deeper integration than an extension can provide. Ben explains the design aim: weave ChatGPT across the entire browsing surface to enable richer actions like writing assistance and personalized behavior.
- •Browser engineering spans rendering, GPU acceleration, process models
- •Atlas is designed end-to-end with ChatGPT integrated throughout
- •Extension-only approaches limit what’s possible
- •Integrated text-field assistance can write in your voice using your ChatGPT context
- 17:35 – 19:29
Agent workspace: separating your tabs from the agent’s tabs (and why it matters)
Ben and Darin describe a key product abstraction: agents get their own tab workspace so their browsing doesn’t clutter yours. The agent can open many pages, work in the background, and then present results and provenance without disrupting the user’s active tab strip.
- •Agents can run multiple tasks, each with their own tab collections
- •Agent tabs stay out of the user’s tab strip to avoid confusion
- •Results can be reviewed step-by-step or consumed as a final output
- •End-to-end browser ownership enables new “agent tab” abstractions
- 19:29 – 22:00
What counts as an “agent task”: acting on the web while you watch and learn
They clarify the difference between standard ChatGPT queries and agent mode: agents take actions inside web apps on your behalf. Examples include generating charts in spreadsheets or leaving comments in docs—while keeping the process visible so users can learn workflows.
- •Agent mode = ChatGPT takes action in web interfaces
- •Useful for complex tools where users don’t know the steps
- •Users can observe actions, building skill and trust
- •Examples: create charts, comment on docs, critique code and documents
- 22:00 – 25:29
Building Atlas with AI: Codex-driven productivity and faster platform expansion
The conversation shifts to how OpenAI’s own teams use coding agents to speed development. Ben and Darin share anecdotes of massive productivity gains, cross-language translation, and even using Swift for Windows to unify expertise and codebases.
- •Codex boosts PR output and enables non-traditional contributors (PMs/designers)
- •Refactors and parallel tasks can be completed dramatically faster
- •Models translate between programming languages effectively
- •Atlas expansion includes Windows work and Swift usage across platforms
- 25:29 – 28:42
Safety and user control in agent mode: sensitive tasks, stop buttons, and auth choices
Andrew asks about agent pausing when you leave a tab. Ben explains “sensitive mode” for high-stakes contexts (like email) where users must supervise and can immediately stop actions, while Darin adds options like running signed-out agent sessions for safer experimentation.
- •Sensitive contexts require user attention; agent may pause if you switch away
- •Prominent “big red stop button” provides immediate control
- •Users can start agents signed-out to avoid accessing authenticated data
- •Onboarding explains choices, controls, and how agent sessions operate
- 28:42 – 32:03
Attention, ecosystem, and serendipity: expanding the web without getting trapped
They explore how AI browsing changes discovery and incentives when agents open tabs and sites compete for attention. The team emphasizes that agents act only on user requests and that agent tabs are restricted from common spam vectors like notification permission prompts.
- •Agent runs only in response to user requests (no autonomous wandering)
- •Agent tabs are blocked from granting notification permissions
- •Sidebar Q&A can route you to better sources beyond a site’s “walled garden”
- •AI can restore productive serendipity without rabbit-hole manipulation
- 32:03 – 33:29
Browser memories and personalization: recall, context-aware search, and controls
Ben and Darin describe how Atlas uses browsing activity to build memories that help users return to prior work and reduce repetitive instructions. They highlight preference learning (e.g., defaulting to favorite travel sites) and emphasize user controls to view, limit, or disable personalization.
- •Memories help you rediscover previously viewed pages and tasks
- •Personalization reduces repeated prompting and improves forward queries
- •Examples: defaulting to preferred airline or commonly used sites
- •Users can inspect, control, or disable memory-based personalization
- 33:29 – 41:28
Search inside the browser: familiar entry points that reveal “model superpowers”
Andrew probes whether Atlas implies OpenAI’s own search direction. Darin explains they include familiar search facets (Images, News) to match user intent while blending in model responses so people learn to use ChatGPT naturally as part of everyday browsing—starting from a single input box.
- •Atlas supports familiar search intents (images/news) alongside model answers
- •Design goal: minimize retraining by keeping recognizable patterns
- •Single-box new tab experience merges navigation and conversational intent
- •Model-in-the-flow helps users discover capabilities organically
- 41:28 – 52:44
Favorite non-AI primitives: scrolling tabs, tab search, and performance architecture
They highlight pragmatic browser improvements like scrolling tabs and tab search that make large tab sets manageable, enabling workflows where users don’t constantly close tabs. Darin also explains resource management (closing background pages), fast restarts, and Owl (Chromium out-of-process) for resilience and speed.
- •Scrolling tabs improve tab hygiene and reduce clutter (optional feature)
- •Tab search enables retrieval across very large working sets
- •Atlas limits live pages by closing background tabs and reopening on demand
- •Owl runs Chromium out-of-process for fast startup and crash resilience
- 52:44 – 58:40
Why Chromium (and its lineage): compatibility, extensions, and where innovation sits
Darin explains Chromium is chosen for web compatibility and extension ecosystem realities. Ben adds historical context (WebKit, KHTML, Blink) and they clarify Atlas’s innovation focus: not reinventing the renderer, but adding an AI-native layer on top with a distinct process architecture and modern Swift app stack.
- •Chromium maximizes compatibility with real-world websites
- •Chromium extension ecosystem works out of the box
- •Rendering engine lineage: KHTML → WebKit → Chromium/Blink; plus Mozilla code influence
- •Atlas innovates above the engine: AI integration + out-of-process embedding + Swift/SwiftUI
- 58:40 – 1:14:21
Long-term bet and what’s next: platforms, mobile UX, and the agentic web in 5–10 years
They address whether Atlas is a serious product (not just an experiment) and outline a long roadmap with frequent updates and broader platform support. Looking forward, they envision users expressing intent while agents handle “toil,” with increasing agent-driven web traffic—while the open web remains the publishing substrate.
- •Atlas is positioned as a long-term investment with iterative improvements
- •Roadmap includes Windows and mobile, with UX adapting to form factors
- •Mobile may blend “bring the model to the web” and cross-device continuity
- •Future vision: intent-first computing; agents do the grunt work and return choices
