Google's Jeff Dean on the Coming Era of Virtual Engineers (sequoiacap.com)

At AI Ascent 2025, Google Chief Scientist Jeff Dean predicted that within about a year AI systems will reach the capability of junior software engineers and sketched how the field will get there: continuing scale of large models plus smarter algorithms, targeted distillation to create lightweight specialists, and an increasing role for multimodal models and agent-based systems trained with reinforcement learning and simulated experience. He reiterated that while a handful of well-resourced players will build the most capable base models, techniques like distillation let startups and product teams turn those giants into fast, focused models for real use cases. Dean stressed that infrastructure and specialized hardware remain critical differentiators—accelerators for reduced-precision linear algebra, high-speed interconnects, and careful attention to compute efficiency, memory bandwidth and data movement drive both training and inference cost and performance. For practitioners and founders the near-term opportunities are concrete: build targeted agents and education/productivity tools, optimize workload-specific hardware/software stacks, and prepare for more “organic” systems that mix levels of compute intensity, specialized components, and continuous learning. The implication: technical strategy (model shaping, hardware-aware design, RL-driven agents) will determine who converts AI capability into real products and scale.