Koray Kavukcuoglu: “This Is How We Are Going to Build AGI”

Full Title

Koray Kavukcuoglu: “This Is How We Are Going to Build AGI”

Summary

The episode discusses the advancements in Google's Gemini models, particularly Gemini 3, and how these models are being developed and integrated into products.

It emphasizes the collaborative, engineering-driven approach to building Artificial General Intelligence (AGI) and the importance of user feedback in this process.

Key Points

• Gemini 3 has received a positive reception, demonstrating advanced capabilities and pushing the frontier of AI development, validating the team's efforts.
• The rapid progress in AI is attributed to continuous innovation and new ideas stemming from real-world application and user feedback, which broadens the surface area for learning.
• Benchmarks are crucial for guiding AI development but can become saturated as technology advances, necessitating the creation of new benchmarks to define future frontiers.
• Key areas of focus for improving Gemini models include instruction following, internationalization (especially for underrepresented languages), function/tool calls, agentic actions, and coding capabilities.
• Product integration, such as through tools like Antigravity and AI Studio, is vital for gathering user signals and directly informing model improvements, highlighting a co-building approach with customers.
• The dual role of DeepMind CTO and Google's Chief AI Architect signifies a commitment to making advanced AI technology accessible and impactful across Google products.
• The development of AI is framed as a joint effort between Google and its users, not just an isolated research endeavor.
• An engineering mindset, prioritizing safety and security from the ground up, is essential for robust and responsible AI development, with these considerations integrated throughout the model development lifecycle.
• The development of Gemini models is a massive, collaborative, global effort across all of Google's teams, reflecting a shift from pure research to product-focused engineering.
• Multimodality is a natural progression, with architectures and ideas converging across domains like text, image, and audio, leading to more capable and nuanced AI.
• The naming of models like "Nano Banana" is presented as an organic outcome of the development process rather than a forced marketing strategy.
• DeepMind's journey, from being the first deep learning researcher to leading AI development, highlights a shift from pure academic research to an engineering mindset focused on productization and user impact.
• The success of DeepMind's scientific endeavors like AlphaGo and AlphaFold has informed their approach to organizing large-scale AI projects, now merging with an engineering mindset for continuous model development.
• The balance between scientific exploration and scaling existing models is critical, with innovation remaining the core driver for achieving the ambitious goal of building AGI.
• The culture at DeepMind, characterized by scientific rigor, kindness, and trust, fosters collaboration and enables tackling complex problems.
• Google's "underdog" journey in LLMs involved significant innovation to catch up and find unique solutions, ultimately positioning them at the leadership level.

Conclusion

The rapid advancement and integration of AI, particularly with models like Gemini 3, signify a transformative era where technology and user collaboration are key to building AGI.

Google's approach emphasizes a grounded, engineering-driven strategy that balances cutting-edge research with practical application, safety, and user needs.

Continuous innovation, a collaborative spirit, and adapting to evolving user demands are crucial for navigating the future of AI development.

Discussion Topics

• How can the development of AI benchmarks keep pace with the rapid advancements in AI technology?
• What are the ethical implications of co-building AGI with users, and how can user feedback be best incorporated?
• How can companies foster an engineering mindset that balances ambitious scientific exploration with the practical demands of product development and user experience in AI?

Key Terms

AGI

Artificial General Intelligence. AI that possesses the ability to understand, learn, and apply knowledge across a wide range of tasks at a human-like level.

LLM

Large Language Model. A type of AI model trained on massive amounts of text data, capable of understanding, generating, and manipulating human language.

Multimodality

The ability of an AI system to process and understand information from multiple types of data, such as text, images, audio, and video.

Benchmarks

Standardized tests or datasets used to evaluate and compare the performance of AI models on specific tasks.

Agentic actions

Capabilities that allow AI models to act more autonomously, make decisions, and perform tasks in an environment.

Function calls/Tool calls

The ability of an AI model to interact with external tools or functions to retrieve information or perform actions beyond its core capabilities.

Timeline