A year of Gemini progress + what comes next — Logan Kilpatrick, Google DeepMind

Gemini Updates and Future Directions: A DeepMind Developer Perspective

Key Concepts:

• Gemini 2.5 Pro: Latest iteration of the Gemini Pro model, focusing on improved performance and addressing user feedback.
• Multimodality: Gemini's design as a single model capable of handling audio, image, and video inputs.
• Proactivity: AI systems that anticipate user needs and take action without explicit prompting.
• Agentic Models: Models that are increasingly self-sufficient and capable of reasoning and decision-making.
• Infinite Context: The challenge of scaling models to handle vast amounts of input data.
• AI Studio: Google's developer platform for building AI applications.
• VO: Undisclosed feature or product experiencing high demand.
• Embeddings: Numerical representations of data used for tasks like semantic search and recommendation.
• RAG: Retrieval-Augmented Generation, a technique for improving the accuracy and relevance of generated text by grounding it in external knowledge.

1. Gemini 2.5 Pro: The Latest Model

• Announcement: A new Gemini model, Gemini 2.5 Pro, has been launched. This is intended to be the final update to the 2.5 Pro line.
• Performance: Gemini 2.5 Pro boasts significant improvements across various benchmarks, including state-of-the-art (SOTA) performance on ADER and HLE. It addresses previous feedback and aims for overall enhanced performance.
• Significance: 2.5 Pro is seen internally as a turning point for Gemini, setting the stage for future developments.
• Availability: The model is accessible through ai.dev and the Gemini app.
• Feedback: Users are encouraged to provide feedback on the model's performance.

2. A Year of Gemini Progress: Rapid Innovation and Adoption

• Pace of Development: The past year has seen a remarkable amount of progress in Gemini development, with the speaker likening it to "10 years of Gemini stuff packed into the last 12 months."
• DeepMind's Advantage: DeepMind's diverse research efforts in areas like science, robotics, and AlphaGo contribute to the mainline Gemini models, enhancing their capabilities in specific domains. Examples include AlphaFold and AlphaGeometry.
• Adoption Rate: There has been a 50x increase in AI inference processed through Google servers compared to the previous year, indicating a surge in demand for Gemini models from both internal and external developers.
• Organizational Restructuring: Google consolidated various AI research teams into DeepMind in late 2022/early 2023. Later, product teams were integrated into DeepMind, creating a unified structure responsible for research, model development, product creation (Gemini app), and developer platform (Gemini API).

3. The Gemini App: A Universal Assistant

• Vision: The Gemini app aims to be a universal assistant, unifying various Google products.
• Unifying Thread: Gemini is envisioned as the thread that connects all of Google's products, moving beyond the traditional Google account-based integration.
• Proactivity: A key focus is on developing proactive AI capabilities, where the system anticipates user needs and takes action without explicit prompts.

4. Model Development: Multimodality, Reasoning, and Context

• Omnimodal Model: The goal is to create a single multimodal model that natively handles audio, image, and video.
• Audio Capabilities: Native audio capabilities have been introduced, including text-to-speech (TTS) and audio input, powering experiences like Astro and Gemini Live.
• Video Integration: Efforts are underway to integrate video capabilities into the mainline Gemini model, potentially leveraging technologies like VO.
• Agentic Models and Reasoning: Models are becoming more self-sufficient, with increased reasoning capabilities. The reasoning step is seen as a critical area for future development.
• Model Scaling: There will be more small and large models.
• Infinite Context Challenge: The current model paradigm struggles with infinite context. New innovations are needed to scale the amount of context models can handle.

5. Developer Platform: Tools and APIs

• Embeddings: A state-of-the-art Gemini embeddings model will be rolled out to developers, enabling applications using RAG.
• Deep Research API: A bespoke API is being developed to bring together research tasks and consumer product features.
• V3 and Imagine 4: These are coming soon to the API.
• AI Studio Repositioning: AI Studio is being repositioned as a dedicated developer platform, moving away from a consumer-oriented feel. This includes integrating agents and developer coding agents like Jewels.

6. Conclusion

DeepMind is rapidly advancing the Gemini ecosystem, focusing on model performance, multimodality, proactivity, and developer tools. The organizational restructuring has enabled faster innovation and product delivery. Key areas of focus include scaling context, enhancing reasoning capabilities, and providing developers with the tools they need to build AI-powered applications. The speaker encourages continued feedback from the community to further improve Gemini.