Google DeepMind Leaders on Gemini Strategy: Model Consolidation, Agents, and Self-Learning

On May 30, Google released a two-hour conversation featuring four of its most influential AI figures: Jeff Dean (founder of Google Brain), Noam Shazeer (co-author of the Transformer paper), Oriol Vinyals (who leads Gemini), and Koray Kavukcuoglu (DeepMind CTO), hosted by Logan Kilpatrick. Framed around Gemini 3.5 Flash, the discussion was less a product announcement than an internal strategic retrospective on where Google's AI is heading over the next five years.

1. Gemini Is an Organizational Restructuring, Not Just a Model

Jeff Dean was blunt: before Gemini, Google's internal teams—Pathways, PaLM, PaLM 2—each ran their own compute and models. "This was stupid," he said, in his own words. Building a sufficiently strong model required consolidating people, compute, infrastructure, and data. Koray added that early AI research resembled academia with parallel exploration, but as models scaled, it had to become one centralized engineering effort. The name Gemini essentially marks the merger of Google Brain and DeepMind's scattered capabilities.

2. Flash Is About Coding and Agents, Not Just Speed

Oriol said the headline improvement in 3.5 Flash is coding ability. Koray added: "Coding capability and agent experience are what define the AI experience right now." Competition has moved past chat quality—winning means getting models into real workflows: writing code, calling tools, executing tasks end to end. Noam noted that Google's playbook is to make Gemini an internal R&D tool first, then an external product.

3. Benchmarks vs. Real Users

Noam's sharpest line: "The real test is whether users are actually using it. If you optimize in a box against benchmarks, you end up optimizing the benchmark—and possibly contaminate your data." This explains why Google pushes Gemini into Search, Workspace, Android, and glasses. Koray called product reach part of the "frontier"—not just research capability, but what products let users do.

4. World Models: From Reading the Internet to Understanding the Physical World

Koray described Omni as a model meant to understand all modalities of the physical world—vision, dynamics, physical laws—and to simulate future states for decision-making. Oriol noted that after large-scale joint training, temporal consistency, 3D world understanding, and sound now emerge naturally in video generation, which previously required manual specification. Jeff extended multimodality beyond text, image, audio, and video to genomic sequences, chemical structures, robot grasping data, and LiDAR: the goal is understanding the world, not just internet content.

5. Squeezing the Lemon: Distilling Pro into Flash

Oriol said his biggest surprise is that Pro-level capability can be compressed into Flash generation after generation—sometimes a new Flash surpasses the previous Pro. Koray's metaphor: "Like squeezing a lemon. The juice goes into a glass—the glass is the small model." Jeff explained that distillation's core method is unchanged from the original papers: a teacher model and a student model; what was once a 50-model ensemble teacher is now a single top-tier teacher. Google's ambition is to make strong capability cheap enough for billions of users—the compute bill eventually makes every bold claim very practical.

6. What's Still Unsolved

• Continual learning (Jeff): Current MoE experts are structurally similar; he wants more "organic," fluid structures. "Not achieved—but the current approach works well enough."
• Noam, half-joking: "Models can't yet invent cancer therapies."
• Evaluation (Oriol): Evaluation is underrated. With models in real products, questions multiply: how to decompose capabilities, detect data leakage, validate user agreement with scores, and predict next major capabilities.
• Sample efficiency (Jeff/Noam/Oriol): Humans hear roughly a billion words in a lifetime; models need trillions of tokens. Noam noted humans run on only a few GB of "source code" (genes). Oriol pushed back: humans are pretrained too—they aren't the first human. This remains debated.

7. Self-Learning: Gemini Improving Gemini by ~2027

Koray predicts self-learning will become a talking point within about a year: as models grow more agentic and better at coding, Google already uses them internally for research, and a point will come where parts of Gemini are improved by models themselves. Jeff agreed: "You may be able to point to an important improvement in Gemini and say the models and agents did that." Noam was more direct: instead of telling a team member "try this experiment and report next week," you hand the task to a model. Gemini is becoming part of Google's own R&D system, not just a product.

8. The 30-Day Agent: A Tooling Problem, Not a Model Problem

Logan set a goal: by I/O 2027, a model capable of 30 days of continuous autonomous operation. The immediate obstacles: memory systems, continual learning, better hardware, and astronomical token consumption. Jeff noted better hardware compresses latency—if a nominally 30-day task finishes in a day, that's far more valuable. Noam added that maybe 29.5 of those 30 days would be spent waiting: today's tools are designed for human latency. The agent era requires rebuilding the entire software toolchain for machine invocation—many tools are just "electronic furniture" designed for slow human clicking.

9. How Many Products Will Google Have in Five Years?

Asked whether Google would have 3 products or 10,000:

• Koray: Just one. The product *is* the model.
• Jeff: Many product outlets, but fewer, more concentrated foundations underneath—Search and glasses are both products transformed by the same model.
• Oriol (cautious): Humans actively choose what to do with digital devices; this division of labor may be a human factor, not a technical limit. He wouldn't bet on a single product.
• Noam: Information products will have many consumption modes—text, visual, glasses, brain-computer interfaces—and further out, models may enter physical products, "moving atoms," not just moving bits.

Conclusion

Gemini is not a single-point model competition. It is increasingly the core operating layer of Google's future AI system: models handle understanding and generation, agents handle long-horizon execution, products collect real feedback, hardware cuts cost and latency, and toolchains let machines actually do work. Whoever plugs models into their own R&D, product, and infrastructure loop—and lets them keep improving—takes the lead. Google is playing this game at massive scale.

*Source: Official Google conversation video (May 30), hosted by Logan Kilpatrick; compiled from multiple sources.*