Interview, Fireside Chat
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
- Global GDP is predicted to increase by orders of magnitude due to the deployment of artificial engineers, supported by an explosion of capabilities where personal compute usage per individual becomes astronomical.
- By 2030, the world is expected to solve unlimited energy and carbon issues while deploying millions to billions of robots to construct data centers, driven by a strategy where a fraction of global GDP is allocated to AI and inference compute demands.
- A feedback loop between automated researchers and AI systems could precipitate an "intelligence explosion" within a two-year window, potentially marking the period between Gemini 4 and Gemini 5 as the most significant in human history by surpassing human intelligence levels.
- Hardware development timelines are projected to shrink from 12–18 months to a few months through automated search processes, reducing team sizes from 150 people to a few and shortening the hardware design outlook from 2.5 years to 6–9 months.
- Model capabilities are expected to improve substantially over the next two to three generations, shifting from handling 5–10 step sub-problems at 80% accuracy to breaking down 100–1,000 step tasks at 90% accuracy, with compute investment yielding incremental 5–10 IQ point gains for every 2x spend increase.
- Strategic plans include utilizing a million automated researchers, exploring modular model architectures with frozen versions and specific modules for languages or private data, and distilling large models for efficient phone-based serving via systems like Pathways.
- Future inference strategies will focus on "thinking harder" at inference time, allowing for 100 to 1,000 steps of sub-problem solving, while training objectives may shift to include active learning, self-play, and multi-token prediction rather than passive observation.
- Publication and deployment protocols will become more selective, with new techniques potentially released in products first followed by research papers, and a mix of open sharing and restricted access for customized base model variants.
- Safety and mitigation efforts will rely on engineering safeguards, self-checking mechanisms, and continuous monitoring of model capabilities, aiming to maximize benefits in education and healthcare while maintaining flexibility in the deployment space.