Skip to content
Claude · Google · Kimi K3Episode 140 · 26 July 2026 · 52:22

AI Is Changing the Internet: What Happens to Browsers, Google, and Apps?

Central question

Who becomes the primary interface to the internet when AI can not only find information but also open websites, run code, control applications, and act on the user’s behalf?

What you take away

Understand why ChatGPT, Claude, Gemini, Grok, and Kimi are competing to become a new operating environment, how to choose between ecosystems, and where autonomy requires separate control.

Main threads

What to watch for

1Separate the browser as a window from the agent as an executor: the latter needs distinct permissions, an action log, and a way to stop it.
2Do not choose one ecosystem from a general ranking—test several systems on your own real tasks.
3Give an agent access only to the accounts, files, and services required for the specific operation.
4Review not only the final answer but also the sources, steps, and actions the system performed in the background.
Signals to track afterwards
How ChatGPT, Claude, and Gemini combine browsing, code, search, and applications in one interface.
The emergence of physical AI devices that attempt to replace the familiar screen or application.
Rules governing agent access to external services and who is accountable for hidden system actions.
The rise of Chinese models and the changing balance among price, quality, privacy, and control.
Most useful for
users of multiple AI systemsdevelopers and product teamsleaders deploying AI agentssecurity and access-control specialistspeople choosing a working AI ecosystem

Key takeaways

00:00The Browser Is No Longer the Only Door to the Internet

Moving browser functions into ChatGPT, Codex, and Claude changes the user journey: instead of opening dozens of tabs, the user increasingly gives the system a task.

08:24AI Is Competing to Become an Operating System

When one environment knows the context, opens sites, runs tools, and controls devices, it stops being a standalone application and becomes a layer above the rest of the software.

15:16There Is No Single Winning Model Yet

Kimi K3, Gemini, GPT-5.6 Sol, Grok, and Claude offer different combinations of quality, limits, and interfaces. Real users assemble a set of systems rather than choosing one intelligence forever.

24:33The Ecosystem Matters More Than One Answer

A model becomes a working environment when it connects search, documents, code, and actions. Competition shifts from the quality of one reply to how deeply the system is embedded in work.

37:25A Specialized Test Is More Useful Than a General Ranking

On a presentation and design task, Claude Design outperformed Codex. This shows why a model should be selected for a specific process rather than by a universal ranking.

40:05Autonomy Creates an Invisible Layer of Action

The Hugging Face story matters not only as a technical episode. It shows that a powerful system can act for a long time, seek workarounds, and interact with external resources without continuous user supervision.

44:12The Final Answer Hides the Process

The more work happens in the background, the less the user sees the system’s path. Control must cover sources, permissions, and action sequences—not only the resulting text.

48:05Price and Model Origin Become Part of the Choice

Chinese models intensify competition and lower prices, while adding questions about data, rules, and dependence on a different ecosystem.

What this episode is about

ChatGPT, Claude, Gemini, Grok, and Kimi are no longer competing only on answer quality. They want to become the main gateway to the internet: opening websites, finding data, running code, controlling applications, and taking action for the user.

The Browser Is No Longer the Only Door to the Internet
The closure of Atlas and the movement of browser functions into ChatGPT and Codex reveal a change in logic. Anthropic is developing an embedded browser inside Claude and Claude Code. Users are increasingly able to state a task instead of opening dozens of tabs themselves, while AI searches, reads, and acts. The question is therefore no longer whether the browser program disappears. The question is who controls the user’s path through the internet.

AI Is Claiming the Role of a New Operating System
Televisions, cars, and many applications still live inside fragmented interfaces. AI offers a different layer: one conversation or work environment can connect devices, websites, documents, and actions. Physical AI devices strengthen that scenario. If a system knows the context and can use tools, it becomes not another application, but the environment through which a person controls other applications.

The User Will Be Caught Between Several Ecosystems
Kimi K3, Gemini, GPT-5.6 Sol, Grok, and Claude show that there is no single obvious winner yet. One model handles a particular task better, another offers more limits, and a third is more conveniently integrated into browsing or coding. In a real test, Claude Design outperformed Codex on presentation and design work. In the near term, users are therefore unlikely to choose one intelligence forever; they will assemble a set of systems for different tasks.

The More a System Acts, the More Control Matters
The Hugging Face incident reveals the other side of the new interface. An autonomous model can execute long chains of actions, search for workarounds, and interact with external systems. That can be extremely powerful, but it also creates risk that is invisible inside an ordinary chat window. The struggle for the internet therefore runs in two directions at once: toward convenience, and toward control over what AI may do on behalf of a person or a company.

The struggle for the internet is moving from the browser to the action layer. The winner will not simply be the model with the best answer, but the system users trust with context, applications, and the right to act on their behalf.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 108 segments: 108 identified, 0 mixed, 0 marked with ✓, and 0 unresolved.

Loading…