OpenAIChatGPT Atlas and the next era of web browsing — the OpenAI Podcast Ep. 9
CHAPTERS
What ChatGPT Atlas is: a browser built around natural-language interaction
Andrew Mayne introduces Atlas with the team behind it, framing it as a shift from browsing and searching to simply stating intent. Ben Goodger describes Atlas as a “new kind of browser” where ChatGPT is the core, helping users understand content, take actions, and carry long-running tasks across days or weeks.
Why build it now: models, computer-use capabilities, and the “slope” of progress
The team argues Atlas is possible now because model capability, speed, and reliability have improved rapidly, especially in agentic computer-use. Darin points to how quickly agent performance has advanced since earlier tools like Operator, motivating a foundation that can evolve over years.
The browser isn’t going away: why the web remains a durable, open platform
They discuss why browsers remain central despite new AI interfaces—similar to how desktops persisted through mobile. The open web’s lack of gatekeepers, plus the fact that ChatGPT itself launched on the web, reinforces the browser as a durable “conduit” to information and apps.
From semantic web dreams to AI that “meets the world where it is”
Andrew raises the unrealized promise of a machine-readable semantic web. Darin argues modern AI models can interpret human-designed interfaces directly, reducing reliance on perfect annotations/accessibility implementations, and extending this human-like interaction to browsing.
Browsers are secretly complex: rendering, GPU acceleration, and modern architecture realities
The conversation shifts to why browsers are so complicated under the hood, from rendering models to GPU-accelerated pipelines. Darin notes the depth of engineering involved and how Atlas’s internals reflect lessons from decades of browser evolution.
Why Atlas isn’t “just a sidebar”: deep integration, writing everywhere, and browsing memory
Ben explains the design goal: ChatGPT woven throughout the entire browsing surface, beyond what extensions can do. They highlight personalized writing assistance in any text field and “memory” that helps users return to forgotten pages or resurfacing prior research.
Agent mode explained: taking actions on the web with a separate agent workspace
They define agent tasks as ChatGPT operating the web UI on your behalf—clicking, navigating, editing documents—while showing its work. A core product idea is that agents have their own “tab workspace,” preventing the user’s tab strip from being flooded during automation.
Safety and user control: sensitive mode, stop buttons, and signed-out agents
Atlas introduces guardrails for tasks involving sensitive contexts like email. The agent may require the user to keep the tab in view, offers a prominent “stop” control, and supports running agents signed-out to experiment without access to authenticated sessions.
Expanding discovery without trapping users: breaking out of site rabbit holes
They discuss how AI can restore positive serendipity by helping users compare sources and escape walled-garden experiences. Sidechat can answer questions using broader web context, suggest alternatives while shopping, and help find content missed by platform search tools.
Personalization and “browser memories”: better defaults with user controls
Atlas extends ChatGPT personalization into browsing behavior, creating “browser memories” that inform future queries and agent actions. They note this can reduce repetitive instructions (e.g., preferring United Airlines) while providing controls to inspect or disable memory use.
Search inside Atlas: familiar modes (Images/News) plus model responses
Andrew observes Atlas feels like both a browser and a step toward OpenAI search. Darin explains the design keeps familiar search affordances (chips, verticals like Images/News) while steadily teaching users that a model response is part of the default browsing journey.
One-box UX and the shift from “modes” to intent: telling the computer what you want
They describe the conceptual move toward computing where users express intent without juggling separate boxes or modes. The team draws a parallel to Chrome’s “one omnibox” revolution and aims for Atlas to route half-formed thoughts to the right outcome with minimal friction.
Favorite features and performance primitives: scrolling tabs, tab search, and fast startup
Beyond AI, they highlight productivity features like scrolling tabs to manage clutter and support huge tab counts, plus tab search as a retrieval layer. Darin details performance strategies: limiting live pages, smart caching, and a fast restart model where pages reload on demand.
Under the hood: Owl + Chromium, compatibility, resilience, and a Swift-based app shell
Darin explains choosing Chromium for web compatibility and extension support, then describes a different embedding approach: Owl (Chromium) runs out-of-process from Atlas. This separation improves responsiveness, crash resilience (Atlas can restart Owl), and developer productivity via a thinner Swift/SwiftUI codebase.
Roadmap and the agentic web future: platform expansion, trust-building, and publishers’ perspective
Ben frames Atlas as a long-term investment—“Netscape 1.0” for an agentic era—with weekly improvements and broader platform plans (Windows, mobile). They predict more agent-driven internet traffic over time and discuss helping users build trust by watching agents work, while still connecting users to publishers and navigational intents.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome