Interview

2026 Frontier & Pioneer Symposium on Aug 7 - Jeff Dean and Dawn Song

In his first public talk since leaving Google after 27 years, Jeff Dean takes the stage at Stanford to discuss the future of AI, scientific discovery, and his new venture, Discovery Loop. In this fireside chat at the 2026 Frontier & Pioneer Symposium, Dean reflects on his work at Google — from foundational computing infrastructure and TensorFlow to Gemini — and shares why he believes AI’s next frontier is moving beyond answering questions toward making discoveries. Topics include mixture-of-experts models, lessons from TensorFlow, the evolution of Gemini, AI safety, scientific discovery, and why Dean chose to build Discovery Loop outside Google. Jeff Dean Co-founder & CEO, Discovery Loop Former Chief Scientist, Google Moderated by Dawn Song, Professor at UC Berkeley.

Source attribution: Original source

Interview details

Publication date: Aug 19, 2026
Duration: 49m

Speakers

Dawn SongProfessor, UC Berkley Jeff DeanCEO, Discovery Loop

Viewpoints in this interview

Sparse experts can combine large capacity with computational efficiency

Dean says mixture-of-experts architectures allow models to maintain very large capacity while activating only the components most useful for a particular token or request, improving efficiency over dense models.

Jeff Dean Mixture of Experts Source at 7:00

Large efficiency gains signaled mixture-of-experts would matter

Dean says the researchers recognized mixture-of-experts as an important idea because experiments showed roughly a tenfold improvement in training-compute-to-quality ratios compared with dense models.

Jeff Dean Mixture of Experts Source at 7:53

Scaling neural networks produced major gains across several fields

Dean says Google observed early that substantially larger neural networks produced major improvements in areas such as image classification and speech recognition, indicating that scaling would broadly matter.

Jeff Dean Deep Learning Scaling Source at 9:28

Deep learning unified previously separate technical disciplines

Dean argues that deep learning provided a common learning paradigm for fields that previously relied on separate specialized techniques, enabling researchers to approach varied problems through a shared framework.

Jeff Dean Deep Learning Scaling Source at 10:24

Machine learning frameworks should abstract away computing infrastructure

Dean says a central TensorFlow design principle was allowing researchers to express machine-learning computations independently from the specific CPUs, GPUs, or distributed resources used to execute them.

Jeff Dean TensorFlow Source at 12:47

TensorFlow should have adopted eager execution earlier

Dean identifies the absence of eager execution in early TensorFlow as a design weakness and says the later addition improved the framework's abstraction.

Jeff Dean TensorFlow Source at 14:45

TensorFlow contrib created unnecessary fragmentation for users

Dean says allowing many alternative helper libraries inside TensorFlow's contrib directory confused users and that those extensions would have been better maintained separately on top of a smaller core.

Jeff Dean TensorFlow Source at 15:05

Gemini benefited from being multimodal from its beginning

Dean says bringing previously separate Google research efforts together around a model designed as multimodal from the start was an important and successful architectural decision.

Jeff Dean Gemini Source at 17:38

Improving coding can strengthen broader reasoning capabilities

Dean says Gemini initially lagged in its emphasis on coding and argues that improving coding ability also improves reasoning skills used to decompose and solve complicated non-coding problems.

Jeff Dean Gemini Source at 19:35

Broad exposure helps researchers connect previously separate ideas

Dean advises researchers to sample many papers and abstracts so they develop a broad map of emerging possibilities that can help them connect ideas not previously combined.

Jeff Dean Research and Engineering Strategy Source at 22:04

Ambitious research should combine tractability with unresolved challenges

Dean says good long-term research problems contain many components that appear increasingly solvable alongside a smaller number of genuinely hard unknowns, avoiding both impossible projects and merely incremental engineering.

Jeff Dean Research and Engineering Strategy Source at 22:42

First-principles estimates are a valuable engineering skill

Dean emphasizes back-of-the-envelope calculations for evaluating whether proposed technical approaches are feasible and for comparing their likely performance before implementation.

Jeff Dean Research and Engineering Strategy Source at 24:22

Trying uncertain ideas is part of successful research

Dean says his successes coexist with projects that failed and recommends attempting many ideas that might not work because some eventually will.

Jeff Dean Research and Engineering Strategy Source at 25:29

Agentic AI brings benefits alongside growing security risks

Song says society should capture the benefits of increasingly capable agentic AI while addressing growing risks, highlighting rapidly improving frontier-model capabilities in cybersecurity.

Dawn Song AI Safety Source at 26:56

Autonomous cyber behavior could create substantially greater harm

Song describes an agent performing a prolonged chain of vulnerability exploitation and warns that similar autonomous behavior could cause substantially greater harm in other circumstances.

Dawn Song AI Safety Source at 28:03

AI cybersecurity capabilities strengthen both attackers and defenders

Dean characterizes AI-enabled cybersecurity as a dual-use capability because models may help attackers exploit vulnerabilities while also allowing defenders to discover vulnerabilities humans might miss.

Jeff Dean AI Safety Source at 30:24

Some AI risks may require legal and regulatory responses

Dean says technical safeguards are not the only possible response to harmful AI uses and suggests that laws and regulations may also be appropriate for discouraging unwanted behavior.

Jeff Dean AI Safety Source at 31:40

Machine learning improving machine learning is not a new concept

Dean says recursive improvement has precedents such as neural architecture search, where machine-learning systems generate and evaluate architectures and improve their design choices through iterative feedback.

Jeff Dean Recursive self-improvement Source at 34:41

Automated loops could improve many components of model development

Dean describes recursive self-improvement as automating improvements across data selection, evaluation, architecture, and other ingredients of model development through interconnected feedback loops.

Jeff Dean Recursive self-improvement Source at 36:31

Scientific discovery can be organized as iterative feedback loops

Dean says many scientific and engineering problems share a structure of decomposition, hypothesis generation, implementation, evaluation, and feedback, making them suitable for increasingly automated iterative systems.

Jeff Dean Scientific Discovery Source at 37:24