Gemini 3: Launch day reactions

Introduction to Gemini 3 Pro's headline features including state-of-the-art reasoning, enhanced multimodal understanding (especially video/image), superior coding abilities, and agentic use cases. The model achieves 1501 Elo on LM Arena while maintaining well-rounded usability across different product surfaces.

• •Model achieves 1501 Elo on LM Arena, demonstrating both technical excellence and real-world usability
• •Enhanced multimodal understanding enables deep context integration between text, images, and video
• •Best-in-class for agentic use cases, tool use, and code generation (vibe coding)
• •Launching simultaneously across most Google surfaces ever on day one: Gemini app, AI Studio, Vertex, and new Google Anti Gravity coding product
• •External partners shipping the model on launch day based on early feedback collaboration

Deep dive into how DeepMind collaborates with product teams to iterate on model capabilities. The team discusses managing trade-offs between different product requirements (developer needs vs. consumer app vs. AI mode) and using live experiments with real users to drive model improvements.

• •Strongest partnership yet between DeepMind and product teams, with continuous feedback loops from real user base
• •Different products have different requirements: developers prioritize different features than Gemini app or AI mode users
• •Team runs live experiments and silent testing (like in Canvas) to gather feedback on specific model characteristics
• •Example: One checkpoint had great personality but regressed on tool calls, requiring careful balancing
• •Product feedback now seamlessly flows back into model development, enabling rapid quality improvements

Discussion of when and why to ship models, balancing speed with quality. The team explains their approach to setting increasingly complex quality goals while maintaining relentless shipping cadence, including extensive testing with customers and internal products before launch.

• •Philosophy: Ship as fast as possible while meeting quality bars for real-world usability
• •Extensive testing period with customers and in products to calibrate expectations vs. reality
• •Quality goals span benchmarks, product experiences, and specific use case performance
• •Each model release sets higher bars than previous versions, making meeting standards harder
• •Turnaround from model completion to product deployment measured in days/hours, aiming for minutes

Team shares specific 'wow moments' with Gemini 3, focusing on vibe coding (creating interactive web apps from descriptions), multimodal content transformation, and multilingual capabilities. Examples include creating games from single prompts, transforming handwritten Korean recipes into interactive apps, and analyzing pickleball videos.

• •Vibe coding: Describe something in your head, instantly get interactive website/game - major compression of time and skill
• •Multimodal transformation: Convert video lectures, handwritten notes into any format (e.g., Korean recipe → English interactive app)
• •Strong multilingual performance in less common languages like Gujarati, enabling personal use cases
• •Video understanding for detailed analysis (e.g., breaking down pickleball moves with critical feedback)
• •Model 'vibes' and response style feel noticeably smarter and more refined than previous versions

Introduction to generative interfaces where the model acts as a design agent, creating custom UI layouts and interactive experiences on-the-fly rather than using pre-coded templates. This represents a fundamental shift from static, engineer-designed UIs to dynamic, model-generated interfaces.

• •Model now makes design and implementation choices autonomously, not just following pre-coded templates
• •Example: Trip planning query generates custom carousel, table, or magazine-style layout based on context
• •Model acts as design agent, choosing optimal presentation format for each specific query
• •Experimental feature in Gemini app: Describe anything, get interactive website streamed as pixels
• •Demo: 'History of Van Gogh by period with artwork' generates full interactive timeline automatically

How external user feedback (like SVG art reactions on Twitter) influences model checkpoint selection and development priorities. The team discusses tracking subjective qualities like persona, style, emoji usage, and conciseness based on real-world usage patterns.

• •Team actively monitors social feedback (e.g., SVG art quality) and runs internal tests across checkpoints
• •SVG art quality varies by checkpoint and type - not universal across all artistic styles
• •Tracking persona metrics: emoji usage, conciseness (addressing 'long-winded Gemini' feedback), house style
• •Persona and style are highly subjective, making user feedback critical for calibration
• •User creativity with models inspires new demos and encourages broader exploration of capabilities

Overview of new agent capabilities in Gemini app, including multi-step task orchestration, Gmail integration for inbox management, and research assistance. The model can execute complex workflows while asking for user confirmation on critical actions.

• •Experimental agent feature in Gemini app for Ultra plan members on launch day
• •Use case: 'Create to-do list from inbox' - analyzes email, generates tasks, adds to Google Calendar
• •Research agent: Performs multi-step research across disparate sources, similar to Deep Research
• •Safety mechanism: Model pauses before critical actions (sending messages, purchases) to ask user confirmation
• •Built on Project Mariner foundation from nearly a year ago, now production-ready

Behind-the-scenes look at managing compute resources across products during launch. The team discusses prioritizing experiences, creative solutions like TPU conversions, and balancing demand across consumer apps, developer tools, and enterprise customers.

• •Daily requests in Gemini app tripled in last quarter, creating massive compute demand
• •Launch planning involves complex prioritization: P0 (must-ship) vs P1/P2 experiences across different products
• •Creative solutions: Converting chips to different TPU types to unlock additional capacity
• •Challenge: Comparing 'apples to oranges' - Notebook LM, Gemini app, and AI Studio showcase different model strengths
• •Subscription plans (Google AI Pro) get highest rate limits; free tier for university students ($20/month value)

Discussion of why Gemini 3 Pro launches first, with Flash and other variants coming later. The sequential release strategy allows the team to learn from real-world Pro usage to inform Flash development, ensuring the workhorse model addresses actual user needs.

• •Gemini 3 Pro is first of the Gemini 3 family; Flash and other models in active development
• •Sequential release enables learning: Pro usage patterns inform Flash optimization decisions
• •Flash designed as 'workhorse model' - needs to account for cost, speed, and specific use case feedback
• •Shipping in sequence supports 'relentless shipping' philosophy - get models to users faster
• •Real-world feedback reveals where Pro resonates vs. where users need different trade-offs (cost/speed)