Behind the scenes with Google's Gemini team - 3 insights that surprised me the most (www.zdnet.com)

Inside Google’s San Francisco office the Gemini team walked a small group of testers through the rapid-fire launches of Gemini 3 (just released) and Nano Banana Pro (about to debut), revealing why Gemini 3 took longer than expected and how the team balances speed with quality. Rather than rushing model updates, engineers extended post-training work—tuning reasoning, multimodality, tool use and model persona—after learning that frequent experimental releases caused heavy churn for developers. Coordinating a simultaneous ship across the Gemini app, Search and AI Studio added infrastructure complexity for serving “hundreds of millions” of users. The team also uses Gemini to cluster and triage massive feedback volumes and to accelerate coding and UI tests—helping build features for Gemini 4—while intentionally keeping humans close to raw user comments to preserve empathy. On the multimodal front, Nano Banana Pro shows a notable leap: AI-generated text in images is much more accurate, dramatically boosting the “cherry-pick rate” for usable infographics in single-shot generations. But limits remain—multi-turn editing still degrades text fidelity (models invent plausible-looking fake words or drift into fragments of other languages), and longer iteration chains require conversation resets. The takeaway: Google’s quality-first, feedback-driven approach yields stronger multimodal and reasoning gains and tighter internal tooling (including using AI to build AI), but persistent multi-turn and edge-case failures mean work continues.