The Pomp Podcast: Google's Billion-Dollar AI Push, Gemini 4, and the Race to Frontier

**Transcripted.ai** (0:01)
When AI strategy, capital, and competition collide, the stakes get enormous fast. Anthony Pompliano recently sat down with Logan Kilpatrick from Google DeepMind for a fascinating look at Google's position in the AI race. And honestly, Kilpatrick didn't shy away from the tough questions about whether Google is falling behind. Right, and his response was pretty nuanced. He admitted the criticism is fair in one sense because expectations are so high for Google. But he pushed back hard, saying they're laser-focused right now at the frontier. He even teased that Gemini 4 will be their largest, most ambitious pre-training run so far. That's interesting, especially when you look at their progress across Gemini 3.5 through 3.8. Kilpatrick suggested that steady improvement might mean recursive self-improvement is already taking shape. But what really struck me was his point about how deeply frontier work runs through Google. It's not just about flashy models. Exactly. According to Kilpatrick, it's about powering billions of users across Search, Workspace, Cloud, even drug discovery work. In his view, frontier models are the accelerant that makes everything else inside Google stronger. That's a very different strategy than a startup approach. Which brings up resource allocation. Kilpatrick says Google doesn't just build the smartest model and hope to monetize later. They make hard trade-offs across coding, research, science, and product work. Though he did admit, "We should have probably put more resources into coding sooner."
At least they're adjusting. And they're both building and buying talent strategically. Like launching the Gemini CLI and bringing in the Windsurf team through an acqui-hire.
Kilpatrick made the point that products that make sense today would have looked completely different twelve months ago. That speed is reshaping the competitive landscape too.
What did you make of his comments about Chinese labs? He was diplomatic but respectful. He pays attention to Chinese open-weight and open-source labs, not out of fear, but because they're contributing real research innovation. He sees competition as healthy.
But a recurring theme was focus—filtering through the noise and avoiding what he called the "TMZ of AI" drama to find "actual, actionable signal." That discipline shows up in their product strategy too. Kilpatrick says they have to decide whether the future belongs to general-purpose models, specialized workflows, or both.
He believes vertical products still matter because users like specific tools for specific jobs. And form factors matter. Chat is resurging because it's familiar, voice helps in demos, but typing creates a different kind of thinking space. The interface is becoming part of the intelligence workflow itself. Behind all this, DeepMind operates as a two-part engine: foundational models like Gemini, and a science group working on AlphaFold, weather, mathematics, and genomics.
Those lessons feed back into the main model, but that flywheel takes enormous effort to maintain.
The conversation closed on measurement. Kilpatrick emphasized that "one cannot improve what one cannot measure." He argued benchmarks miss real-world nuance, and the next big bottleneck is data itself—"the dark art of transforming raw data into usable formats."