Koray Kavukcuoglu: Building AGI Through Products

Redefining AGI Progress

• 💡 The development of Artificial General Intelligence (AGI) is framed as a joint effort with the world, not a theoretical research project confined to a lab.
• 🎯 Real-world utility and user feedback are considered the ultimate measures of progress, surpassing traditional benchmarks.
• 📈 While initial benchmarks provide validation, they are fleeting; models like Gemini 3 quickly make them irrelevant by achieving significant advancements (e.g., HLE and RKGI2 from 1-2% to over 40%).
• 🧠 The GPQA Diamond benchmark remains a challenge, requiring expert-level scientific reasoning and multi-step thinking, indicating that hard problems in reasoning are still unsolved.

Key Development Focus Areas

• ✅ Instruction following is paramount, ensuring models precisely understand and execute nuanced requests rather than just inferring intent.
• 🌍 Internationalization is a huge strategic focus, expanding model utility and feedback loops to diverse languages like Hindi, Portuguese, and Swahili.
• 🚀 Agentic actions and code are seen as a major intelligence multiplier, enabling models to perform tasks through function and tool calls.
• 🛠️ The concept of "vibe coding" allows non-programmers to generate working applications from high-level ideas, democratizing creation.

Engineering Mindset for AGI

• ⚙️ Building AGI requires an engineering mindset, integrating AI technology into every Google product to force robustness and gather real-world user signals.
• 🤝 Platforms like Anti-gravity facilitate the deployment of AI agents, providing crucial data on model performance, weaknesses, and areas for architectural improvement.
• 🔒 Safety and security are first principles, built into the entire development process from pre-training through post-training, rather than being added as an afterthought.
• 🌐 This effort involves massive global coordination across Google teams, co-designing software and hardware (chips, data centers, networking) for global scale robustness.

DeepMind's Legacy and Future

• 🔬 DeepMind's history with large, specialized teams on projects like DQN, AlphaGo, and AlphaFold forms the cultural backbone for current AGI development.
• ✨ Multimodality (understanding text, images, audio) is naturally emerging as underlying architectures converge, allowing world knowledge to transfer between domains.
• 🖼️ The Nano Banana Pro image generation model, built on Gemini 3 Pro, exemplifies this by leveraging the text model's deep world understanding for complex tasks like creating infographics from dense documents.

The Biggest Risk to Innovation

• ⚠️ A humbling insight reveals that the biggest risk to AGI development is running out of innovation, not a lack of money or compute resources.
• 💡 DeepMind's aggressive innovation was partly driven by feeling like an underdog two and a half years ago, pushing them to join the "leadership group" in LLM development.