Jeff Dean

Jeff Dean appears in E228｜谷歌TPU能撼动英伟达吗？前TPU工程师首次揭秘 as one of the high-level Google figures [[HenryTPUEngineer|Henry]] associates with deciding future TPU direction. The episode frames Dean and DeepMind as closer to the workload and model-strategy brain, while the hardware organization implements those bets through chips, systems, and software support.

The source uses him to connect TPU roadmap choices to internal model and product demand. That matters because ASIC Workload Prediction Risk depends on what senior technical leadership believes future [[TransformerArchitecture|Transformer]], [[MixtureOfExperts|MoE]], reinforcement learning, and inference workloads will need several years later.

「模型能力已经够了，要卷就卷 infra」｜对谈戴冠兰：Runta 创始人 adds Jeff Dean as an angel investor and as the person whose question frames [[DaiGuanlan|Dai Guanlan]]’s agent-infra thesis: what should infrastructure look like if the lowest execution unit is probabilistic? The source therefore links Dean not only to model and hardware roadmaps, but also to Probabilistic Software and [[AgentRuntimeExecutionLayer|agent runtime infrastructure]].

Connections

• Google, DeepMind, Google DeepMind, and Gemini — organizational and model context.
• TPU, AI Chip Specialization, and ASIC Workload Prediction Risk — hardware-roadmap context.
• Training Compute Allocation and MaaS Infrastructure — strategic compute-allocation frame.
• Runta, 戴冠兰 / Dai Guanlan, Probabilistic Software, and Agent Runtime Execution Layer — agent runtime question added by the Runta episode.