Gemini app: Canvas, Deep Research and personalization

Gemini Updates: Canvas, Deep Research, and Personalization

Key Concepts:

• Gemini Canvas: An interactive UI within the Gemini app for collaborative content creation (docs, code, web apps).
• Deep Research: A long-running, agentic feature that automates in-depth research and generates synthesized reports.
• Gemini 2.0: The latest generation of Gemini models, including "thinking models" with improved reasoning capabilities.
• Personalization: Tailoring Gemini app responses based on user data, starting with Google Search history.
• Thinking Models: AI models with enhanced reasoning capabilities, allowing them to plan, search, synthesize, and generate more comprehensive and relevant outputs.
• Agentic Features: AI functionalities that can autonomously perform tasks over extended periods, such as Deep Research.
• Audio Overviews: AI-generated summaries of content that can be listened to on the go.

Gemini Canvas: Collaborative Content Creation

• Problem Addressed: The iterative process of refining content in traditional chatbots can be cumbersome, requiring users to specify which parts to modify in each prompt.
• Solution: Gemini Canvas provides a familiar, interactive UI (similar to Google Docs) for direct editing and collaboration with the model.
• Functionality:
  • Doc Editor: Allows users to edit text, change formatting (font size, styling), and collaborate with Gemini in real-time to refine the document.
  • Code Editor: Enables users to create and modify code, subselect specific pieces of code for iteration, and preview web apps inline.
  • Web App Creation: Users can create interactive web apps with minimal or no coding experience, preview them, and share them with others.
• User Journey: Users can start in either the Gemini app or Google Docs. Starting in Gemini is beneficial for brainstorming and initial content creation, after which the content can be exported to Google Docs for further refinement.
• Example: Creating a chemistry study guide for a daughter, then transforming it into an interactive periodic table web app.
• Limitations: Initially focused on sandboxed-type examples, but the platform will evolve to support more APIs and sophisticated web apps. Single URL, but can build comprehensive navigation structures inside of the page and dynamically update.
• Real-World Applications: Solar system visualizer, particle simulator.

Deep Research: Automated In-Depth Research

• Problem Addressed: Manually conducting in-depth research can be time-consuming.
• Solution: Deep Research automates the research process, leveraging Gemini's long context windows to analyze numerous sources and generate synthesized reports.
• Key Improvement with Gemini 2.0: The upgrade to Gemini 2.0 thinking models has dramatically improved the output quality and depth of research.
• Functionality:
  • Agentic Task: Deep Research is a long-running, agentic feature that autonomously performs research based on a user's query.
  • Pre-Planning Step: The model outlines the research plan, allowing users to review and modify it before the research begins.
  • Real-Time Insights: Users can see the model's thought process, including the websites it's visiting and the questions it's asking itself.
  • Inline Citations: The generated reports include inline citations at the paragraph and sentence level, linking to the original sources.
• User Experience: Users can initiate Deep Research with a prompt, review the research plan, and receive a notification when the report is complete.
• Example: Researching integrating AI agents.
• Availability: Deep Research is now free for everyone to try.
• Mobile Experience: Users can start a Deep Research task on desktop and continue it on mobile, receiving notifications upon completion.
• Audio Overviews: Deep Research reports can be converted into audio overviews for on-the-go consumption.
• Developer API: While a dedicated Deep Research API is not currently available, the trend is towards making it easier for developers to build similar agentic experiences as the base models improve.

Personalization: Tailoring Gemini App Responses

• Problem Addressed: Traditional chatbots lack personal context, leading to transactional and less helpful interactions.
• Solution: The personalization feature allows Gemini app to access user data (starting with Google Search history) to provide more relevant and personalized responses.
• Vision: To transform Gemini app from a transactional chatbot into a truly personalized AI assistant.
• Functionality:
  • Google Search History Integration: With user permission, Gemini app can access Google Search history to understand user interests and preferences.
  • Thinking Model Reasoning: The thinking model analyzes the search history to identify patterns and tailor responses accordingly.
  • Transparency: Users can see how the model is using their search history to generate responses.
  • User Control: Users have full control over the personalization feature, including the ability to connect or disconnect it at any time.
• Example: Recommending vacation destinations based on past searches for South Korea, Japan, and Hawaii.
• Evaluation: Evaluating the effectiveness of personalization is tricky because only the end user knows whether the answer is actually good.
• User Experience Considerations: Ensuring users feel in control of their data and that the personalized responses are helpful and not annoying.
• Future Directions: Integrating more data sources from across the Google ecosystem to further enhance personalization.

Synthesis/Conclusion

The Gemini updates focus on enhancing user experience through collaborative content creation (Canvas), automated in-depth research (Deep Research), and personalized interactions. The integration of Gemini 2.0 thinking models has significantly improved the capabilities of these features, enabling more comprehensive, relevant, and user-friendly AI experiences. The emphasis on user control and transparency ensures that these features are both powerful and responsible.