Google CEO: Agents, Open Source, Race to AGI, Cybersecurity, Chips, China

Agents as the Web’s New Front Door

Matt Berman opens by asking whether agents become the main way people use the internet, and Sundar Pichai mostly says yes — especially after you’ve felt the “superpower” of an agentic workflow. He points to developers as the first group already living in this future, orchestrating agents and building with them, but stresses that agents should remove drudgery, not flatten the whole web into a single interface.

Trust, Control, and the Waymo Analogy

When Berman pushes on whether people can trust agents to shape their information diet, Pichai reaches for two familiar examples: Gmail’s spam filter and riding in the back seat of a Waymo. His point is that trust is earned through safety data, careful deployment, and user control — which is why Google is introducing Gemini’s stronger capabilities first in first-party tools like Gmail and Calendar before opening up third-party actions, browser use, and full computer control.

Don’t Worry, Exploration Isn’t Dead

Berman worries that agents add yet another buffer between humans and the “raw internet,” and Pichai doesn’t really deny the abstraction layer — he just argues that people still want discovery. He uses YouTube creators, shopping, and trusted news sources to make the case that humans won’t outsource everything, because exploration and connection are part of the point.

Cybersecurity Is Becoming an Agent-vs-Agent Battlefield

The conversation then turns sharply practical: are AI-enhanced cyberattacks ramping up? Pichai says Google has been serious about security for years, but over the last two years model progress has clearly improved its ability to detect vulnerabilities and patch them, and he calls Anthropic’s “Mythos” a genuine inflection point for cyber-focused AI.

Code Mender, Wiz, and Google’s Case for Responsible Cyber Release

One of Pichai’s more concrete product moments is Code Mender, a system Google uses internally and is preparing to share externally; he says it can identify vulnerabilities, generate patches, test them, verify them, and deploy them continuously. On whether dangerous cyber models should be released, he lands in the middle: if a model doesn’t dramatically move the frontier, release is fine, but if it makes a big leap, the security world’s norms — responsible disclosure, vendor coordination, and government engagement — should apply.

Open Source, Gemma, and the China Question

Berman presses on why Google won’t release a large frontier open model, and Pichai answers with a familiar Google line: the company loves open source, has contributed massively across Chromium, Android, and Kubernetes, and sees Gemma as part of that tradition. But he’s equally blunt that frontier training consumes enormous capex and R&D, so Google is trying to support open ecosystems without giving away everything at the very top.

Why Chinese Open Models Don’t Automatically Win

Asked why US enterprises wouldn’t simply adopt cheap, near-frontier Chinese open-source models like DeepSeek, Pichai shifts the lens from geopolitics to buyer behavior: companies care about predictability, reliability, safety, and support, not just benchmark scores. His most interesting line here is that he worries less about whether open models come from China and more about whether the US is doing enough to stay at the frontier.

Flash Models, AGI Race Talk, and Google’s Compute Reality

In the final stretch, Pichai gives one of the clearest defenses of Google’s Flash strategy: most businesses are not solving Math Olympiad problems, they need cheap, fast, capable models that can run repeatedly in agentic workflows without blowing the budget. He pushes back on the idea that AI is just a winner-take-all sprint to self-improving systems, saying any true recursive self-improvement moment would require a broader societal response — then closes by admitting the obvious: even Google has more demand for compute than it can serve, and the bottlenecks range from data center construction and power to memory and chips.