AI Dev 26 X SF | Paige Bailey: Research to Reality

Why It Matters

By making powerful multimodal models accessible on‑device and offering substantial cloud credits, Google DeepMind accelerates AI adoption for startups and enterprises while enhancing privacy and productivity across robotics, AR, and software development.

Key Takeaways

• Gemini 3 is natively multimodal, handling video, audio, text, code.
• Gemma 4 open‑source models run locally on phones and laptops.
• Gemini powers robotics tasks like making salads and cleaning spills.
• Augmented‑reality glasses use Gemini for live directions and object insights.
• Google anti‑gravity IDE lets agents generate 75% of internal code.

Summary

Paige Bailey, engineering lead for Developer Relations at Google DeepMind, introduced the latest Gemini and Gemma model families during AI Dev 26. She highlighted Gemini 3’s native multimodal capabilities—processing video, images, audio, text, and code simultaneously—and outlined the tiered lineup from high‑performance Pro to lightweight Flash and Nano variants. The open‑source Gemma 4 series, available in 2‑31 billion‑parameter sizes under an Apache 2.0 license, can run on mobile devices and laptops, offering strong performance for vision, audio, and code tasks. Key insights included the integration of Gemini models into real‑world applications: robotics agents that can prepare meals or clean spills, augmented‑reality glasses delivering live navigation and contextual insights, and the Genie 3 world‑building system that generates playable video scenes from natural‑language prompts. Bailey also showcased the anti‑gravity IDE, an agent‑first development environment where AI now writes roughly three‑quarters of Google’s code, and announced a startup support program offering up to $350 k in cloud and Gemini credits. Notable examples featured a Raspberry Pi‑powered 3D‑printable robot using Gemini for vision and speech, AR demos with Google Maps integration, and real‑time speech translation via Gemini APIs. The Genie 3 demo illustrated dynamic, physics‑free video generation, while anti‑gravity demonstrated seamless code‑generation workflows across multiple model providers. The announcements signal a shift toward on‑device, privacy‑preserving AI and lower barriers for developers and startups to embed advanced multimodal models into products, accelerating innovation across robotics, AR, and software engineering.

Original Description

At AI Dev 26 x San Francisco, Paige Bailey from Google DeepMind delivered a presentation on the latest advancements in AI models and their real-world applications.