Aravind Srinivas on Comet, AI Browsers and the Future of Work

Matthew Berman meets Perplexity CEO Aravind Srinivas shortly after the launch of Comet. Berman says he has switched to the browser and asks why a company known for AI search decided to build one.

Srinivas begins with distribution: many searches start in a browser’s combined address and search bar. He recounts the importance of Google’s toolbar before the modern omnibox and argues that control of this entry point mattered to Google’s growth.

He recalls accounts of Microsoft changing Internet Explorer and Google seeking to preserve its toolbar distribution. Srinivas says this history is based on accounts by former employees, rather than a public confirmation from Google.

Perplexity had its own experience with platform dependence, he says: after a Chrome update, an extension that set Perplexity as the default was removed on his machine, and another extension could not do everything the team wanted. Those experiences made control of a client a defensive motivation.

A browser was also a major distraction for a small company that still had to improve its search product. Srinivas says the decisive question was whether there was an offensive opportunity as well: a browser could support AI tasks that would be harder to do elsewhere.

He points to deep research and Perplexity Labs as early examples of systems that spend time researching or creating outputs. His vision is to apply that capability to daily browsing, using context across tabs and services to help a person complete work.

Examples include auditing a calendar, reviewing email, comparing prices for a small business and tailoring marketing using public feedback. These are proposed workflows, not a promise that every one can be completed reliably today.

Srinivas says the team began work near the end of the previous year and shipped in about eight months. He acknowledges that Comet builds on Chromium rather than a browser engine made from scratch and credits Google’s open-source work.

In the Google antitrust context, he says Perplexity argued against separating Chrome from Google, partly because it considered Google a good steward of Chromium. That is Perplexity’s position in the dispute; he says operating a browser at the scale of a billion users would require earning that responsibility.

Even with Chromium, Srinivas says, browser development requires substantial work on testing, bugs, imports, logins and basic reliability. He is pleased with the launch but says the product still needs to work consistently for many more daily users.

Berman says he could import bookmarks and existing logins and hand an in-progress task to an agent. He contrasts that with a cloud-hosted agent session that begins in a separate environment.

Srinivas argues that maintaining logged-in sessions on a third-party server creates security risks. He describes Comet as keeping logins for services such as email and shopping on the user’s client instead of maintaining those authenticated sessions on Perplexity’s servers.

For a particular prompt, Srinivas says the assistant extracts relevant information from open tabs and sends the needed context to powerful cloud models. He describes controls for deleting queries and using incognito mode. These are his descriptions of the product and data handling, not an independent security audit.

Srinivas cautions that AI cannot yet perform every task autonomously with complete reliability. He would not trust Comet to conduct Perplexity’s financial accounting, for example, so people need a shared workspace where they can perform some tasks themselves and delegate others.

Berman asks whether Perplexity might eventually build an operating system or device. Srinivas says the company is not interested in hardware at that point, citing limited hardware expertise and the need to stay focused. He considers an operating system more achievable after the browser experience, but does not announce one.

Srinivas says Perplexity earned the opportunity to build a browser by developing its search product and research workflows. He claims its work influenced other products, while acknowledging that its usage was below ChatGPT’s; these are his competitive assessments.

He believes OpenAI has already won the chat-interface race and says he does not want Perplexity merely to compete for second place. His bet is that an end-to-end agent workflow in the browser can be a more useful layer above chat.

The browser is a sticky product, Srinivas argues, but serving agents at scale is costly. At the time, Comet was available to Max subscribers and invited waitlist users. He expects scaling, security and ongoing maintenance to require years of work.

Berman asks how Comet can compete as Google and others add AI to browsers. Srinivas argues that if an agent does the research and purchasing for a person, advertisers may get less value from paying for search clicks. This is a hypothesis about future behavior, not proof that advertising has already lost its value.

He imagines a user asking an agent to ignore sponsored links, review sources, select a product and seek approval before buying. In that scenario, he argues, businesses might shift spending from performance advertising toward brand advertising.

Srinivas suggests that Google must balance agent deployment against its existing business, inference costs and security standards. He also argues that a large organization must coordinate separate browser, advertising and AI teams. His account of Google’s incentives is a competitor’s interpretation.

Berman asks whether an agent might sit between people and an increasingly noisy web. Srinivas hopes the web does not become dominated by AI-generated spam, but says filtering low-quality material and summarizing useful information could be a valuable agent task.

They discuss filtering social posts, connection requests or event applicants against user-specified criteria. Srinivas sees this as a way to reduce repetitive work, though the interview does not establish how accurately an agent will make every such judgment.

Berman asks how Perplexity chooses models and post-trains them, mentioning Srinivas’s interest in Kimi. Srinivas describes an internal evaluation set called PPLX Bench, which grows as users report problems through product feedback and community channels.

Rather than changing a prompt for each isolated complaint, the team groups issues, considers prompt or post-training changes, and evaluates them against the larger set. Because the benchmark is private and reflects real product use, he says it provides a useful complement to public academic scores.

For post-training, the company combines prompts on which its product struggles with general capabilities it wants to preserve. Srinivas describes supervised fine-tuning and reinforcement learning using GRPO, and says the company was working with DeepSeek and smaller Qwen-based models at the time.

Different models also perform classification tasks: deciding whether a query needs personal context, a finance interface, a chart, a sports card or structured shopping results. Srinivas says many models may participate in a single query, even when the user selects one model for the main answer.

He argues that a strong model is necessary but not sufficient. The application must collect relevant context, select and orchestrate tools, and package the result in a useful workflow. He expects capabilities that are frontier-level today to become more widely available over time, a forecast rather than a guarantee.

Berman asks whether Perplexity needs to build a frontier model itself. Srinivas says the challenge is not producing one good model once but sustaining the enormous effort: hiring and retaining researchers, planning compute, acquiring energy and continuing to train new generations.

Instead, he says, Perplexity focuses on models that matter for its products. Its Sonar models target search-oriented summaries, references and formatting. He says the company wants to develop specialized models for controlling browser tabs and actions as well.

At the time of the interview, Srinivas says Perplexity did not yet have its own model making the browser-action decisions. Its internal models handled tasks such as assembling context and summarization, while the action model remained an area for future work.

A smaller, faster browser-control model could reduce a task taking several minutes to around a minute, he hopes. Running such a model locally would also have privacy benefits, but he acknowledges that sufficiently capable, power-efficient local models were not available to meet his desired standard then.

Srinivas expects distillation and other techniques to help smaller and open models catch up, though perhaps with a delay of six months or a year. He discusses companies’ incentives to release models and argues that challengers may open models to attract developers and attention. These are his predictions about a competitive market.

Berman asks about a previous interview in which Srinivas was quoted about tracking and highly personalized ads. Srinivas says he was answering a hypothetical question and insists he would prefer a future that does not rely on advertising.

He argues that Google’s established advertising network would be hard to displace on its own terms. Instead, he wants Perplexity to deliver value through agents and charge subscriptions, usage fees or potentially fees tied to outcomes.

Srinivas sketches other possible revenue streams, including APIs, browsing infrastructure and transaction commissions. He offers illustrative user and revenue figures to explain why a smaller business than Google could still be attractive. These are scenarios, not reported results or firm product commitments.

Berman asks about people who cannot afford a paid service. Srinivas points to people who already pay for tools they find valuable and argues that saved time may justify a subscription. His examples do not resolve how people unable to pay would obtain access.

Berman returns to the tracking question. Srinivas says the company would let people avoid retaining data on its servers and says logins, passwords and card details stay on the client. To summarize unread Slack messages, however, relevant message content must be sent to the cloud model for that request.

He describes deletion and incognito or zero-retention options, then says his longer-term goal is to move more intelligence to the device. His view is that the quality of local models was not yet sufficient for the desired experience; the interview does not independently verify the stated data practices.

Berman raises concern that AI may automate jobs that form an important part of people’s identities. Srinivas expects people skilled at using AI to become more employable, but says the pace of change makes adaptation unusually difficult.

Telling people simply to learn AI is not enough, he says. Tools can change within months, educational materials can become outdated, and some people may struggle to keep up. He expects this transition to lead to job losses for some workers.

Srinivas calls for more entrepreneurs and new companies to create work. Some displaced workers may start businesses; others may learn to use AI in new roles. He expects a difficult transition rather than asserting that everyone will immediately benefit.

Berman cites Box CEO Aaron Levie’s argument that highly productive teams could attract more investment and hiring. Srinivas accepts this as a plausible long-term argument but questions whether enough people can adapt to rapidly changing tools quickly enough.

Srinivas closes by encouraging people to spend time learning and using AI rather than ignoring the shift. He expects some tasks, such as routine form filling, to require less human labor; the scale and timing of that displacement remain forecasts.