Conference Presentation, Keynote
The Physics of Thought & Architecture of Intelligence | Amin Vahdat, Google | RAISE Summit 2026
- The speaker characterizes the current era as the early stages of a societal transformation into the "intelligence era," paralleling but accelerating beyond the Industrial Revolution's "muscle multiplier" role.
- Historical analysis of the Industrial Revolution highlights that James Watt's 1765 steam engine improvements increased efficiency from 1% to 5%, eventually leading to modern efficiencies exceeding 60% over centuries.
- Jevons Paradox is cited as a key historical trend: making resources 5x more efficient did not reduce coal consumption; instead, demand for coal exploded due to new, expanded use cases that rewrote civilization.
- The speaker posits a "fallacy of finite work," arguing that human ambition is unbounded and that efficiency gains in cognition will trigger exponential increases in demand for decision-making and problem-solving capabilities.
- AlphaFold3 is identified as a specific breakthrough where protein ligand structure prediction time was reduced from years to hours, demonstrating the potential for accelerated scientific discovery.
- Google demonstrated the construction of a fully functional operating system using 93 subagents in 12 hours, generating 15,000 model requests at a cost under $1,000.
- A partnership with France's Institut Curie is underway to utilize AI for identifying new biomarkers to enable targeted treatments and specialized medicine.
- Rich Sutton's "The Bitter Lesson" is referenced as a core tenet: 70 years of AI research proves that scaling compute and data consistently outperforms human-encoded expert systems or brittle rules.
- Demand for machine learning compute is growing at a rate of 10x year-over-year, a trajectory significantly faster than the internet era's 2x growth every 18–24 months.
- The 2017 introduction of the Transformer architecture shifted performance efficiency curves left by 4x–5x, effectively breaking previous barriers and triggering an explosion in AI use cases.
- The AI infrastructure stack is defined as a sequential dependency chain: energy → data center land/enclosures → AI hardware → AI software → models → services.
- Specialized hardware can deliver at least 100x greater system efficiency compared to general-purpose CPUs when optimized for specific workloads like AI.
- Google's Tensor Processing Unit (TPU) roadmap has reached its eighth generation (TPU v8), split into two distinct lines: TPU v8i for inference and TPU v8t for training.
- The TPU v8 split addresses the specific lead times and optimization needs of training versus inference, targeting a 2x improvement in performance per watt through specialized on-chip memory and latency optimization.
- Data centers are shifting reliability models from 99.999% (five nines) to 99.9% (three nines) to effectively double usable capacity without redundant idle infrastructure.
- Google has established gigawatt-scale demand response agreements, allowing power down periods during peak utility grid stress in exchange for reduced baseline provisioning costs.
- The future inflection point is "agentic computing," where swarms of autonomous agents coordinate computer-to-computer tasks, vastly increasing model invocation rates beyond human interaction limits.
- The speaker emphasizes a "messy middle" transition period requiring responsible management of economic turbulence and societal disruption to ensure AI augments human capability ethically.