Demis Hassabis Reveals Google's 'Secret' Behind Benchmark-Topping Gemini 3 (officechai.com)

Demis Hassabis credited Gemini 3’s benchmark dominance not to a single trick but to Google’s long-built ability to marry “world-class research AND world-class engineering AND world-class infra.” Responding to Oriol Vinyals, Hassabis and Vinyals framed Gemini 3’s leap as a combination of two technical gains: meaningful pre‑training scale returns (Vinyals says the jump from 2.5→3.0 is unusually large) and a vast “greenfield” of post‑training algorithmic improvements. That mix — better ways to extract value from scale plus new post‑training methods — underpins the model’s step‑change capabilities. What makes this significant for the AI community is Google’s vertical integration: decade‑long research depth (including the transformer lineage), unified Google DeepMind/Brain organization, custom TPUs (Gemini 3 was trained on Google’s 6th‑gen TPUs rather than NVIDIA GPUs), huge datacenter and networking assets, and optimized software stacks (TensorFlow/JAX). Those elements let Google iterate on hardware, kernels, training pipelines and data in lockstep — a coordination advantage few competitors currently match. The implication is that leadership now often hinges on systems-level alignment (research → engineering → infra), not just novel algorithms; competitors can close gaps, but doing so requires massive investment and orchestration.