Sundar Pichai on Gemini 4 and the AI compute crunch
Sundar Pichai explains how Google gates and safety-tests Gemini 4 Argon, why it keeps spending aggressively on compute, and where agents, Waymo, space data centers and quantum are heading.
Watch on YouTube ↗Charlie's summary
- ▶ 0:30
Argon launched compute-constrained, so Google gates it and relies on distilled versions to widen access
Sundar says Argon drew almost 25 trillion tokens in its first 72 hours, about 3x what 3.0 saw. It is large and expensive to serve, so it started with Ultra and Pro tiers and the API. Google can distill big models into versions that are “nearly as capable but much more efficient”, and those 4.0-series versions are in progress for broader access.
- ▶ 4:03
Safety for agentic models means real-time monitoring, sandboxing and telemetry, which cost compute too
Sundar says Argon's coding and agentic strength made Google slow down. It ran its widest internal testing, gave early access to cyber researchers, and submitted it to the US government. The key piece is monitoring models while they work, which is compute intensive. Google built sandboxing and telemetry so customers know where an agent runs and what access it has, and so Google can intervene early.
- ▶ 7:37
Google fixed its release cadence by hill-climbing agentic coding on Flash, which is why 3.5 Pro never shipped
Sundar says Google found gaps in agentic coding and decided the fastest fix was the smaller Flash model, which allows more iterative loops. That produced 3.5, 3.6, 3.7 and 3.8 Flash in about three months. By then Argon pre-training was done, so Google skipped a short-lived 3.5 Pro and applied the learnings to Argon's post-training. He expects faster cadence with the 4.0 series.
- ▶ 9:10
Argon brings a step change in internal productivity, and Sundar expects recursive self-improvement to accelerate over the next 12 months
Sundar says Google points Argon at problems spanning millions of lines of code, such as a Chromium feature to drag a tab into an AI experience without reloading, a change that cuts through every layer. He calls 2027 a golden age of productivity. Researchers already co-work with models that suggest recipes and evaluate checkpoints. Some pipeline parts are closer to recursive self-improvement, and humans remain in the loop for safety.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.