Build & deploy AI-powered apps — Paige Bailey, Google DeepMind

The Expanding Google DeepMind Ecosystem

Google DeepMind has rapidly released a suite of powerful models, creating a comprehensive ecosystem for developers. This includes:

• Gemini 3.1 Series: Featuring the large-scale Pro model, the efficient workhorse Flash, and the exceptionally fast and low-cost Flash-Lite.
• Gemma 4: The latest family of powerful open models.
• Generative Media Models: Including NanoBanana 2 for images, Lyria 3 for music, Veo 3.1 Lite for video, and Genie 3 for interactive world generation.
• Embeddings Model: A unified model supporting video, images, audio, text, and code in the same embedding space.

A key differentiator for the Gemini family is its native multimodality for both inputs (video, images, audio, text, code) and outputs (text, code, audio, images), unlike many models that are limited to text/code outputs and static image inputs.

Prototyping in AI Studio

AI Studio (aistudio.google.com) serves as a free, web-based playground for rapidly prototyping with these models. It provides a suite of powerful, configurable tools that can be enabled with a toggle.

Multimodal Analysis with Tools

A core strength of AI Studio is combining Gemini’s multimodal understanding with powerful tools like grounding and code execution.

• Video Analysis & Grounding: By providing a YouTube URL and enabling “Grounding with Google Search,” you can ask complex questions about the video’s content. A demonstration showed Gemini 3.1 Flashlight analyzing a dinosaur video, creating a timestamped table of all dinosaur appearances, and using Google Search to pull in a fun fact for each one, citing its sources directly.

• Code Execution: This feature provides Gemini with a sandboxed Python environment pre-loaded with data science libraries. In a demo, Gemini 3.1 Flashlight was prompted to “draw bounding boxes around all of the green Lego bricks using Python” from an uploaded image. The model wrote and executed the necessary Python code to identify the bricks and display the resulting image with bounding boxes, all for a fraction of a penny. This illustrates how tools can augment smaller, faster models to perform complex tasks.

• URL Context: To ground the model with up-to-the-minute information, you can provide a list of URLs. A demo compared and contrasted the recently released Genie 3 and Gemma 4 models by feeding their respective blog post URLs into the context window, allowing the model to generate an accurate comparison and cite its sources.

Gemini Live: Real-Time Interaction

Gemini Live enables dynamic, conversational interaction with a model using your screen, camera, and microphone as inputs.

• Screen Sharing: The model can see and answer questions about your screen content.
• Multilingual & Style Adaptation: It can switch languages (e.g., from English to Italian) and adopt different accents or styles (e.g., a Texan poet) on the fly within a single conversation.
• Real-World Applications: This opens up possibilities for interactive kiosks, augmented reality guides that describe the environment, or real-time translation and transcription tools.

AI Studio Build: From Prompt to Full-Stack App

The “Build” feature in AI Studio allows you to generate and deploy entire web applications from a single prompt. It now supports integrated databases (Firestore) and authentication (Firebase with Google Login).

A full application was generated live from the prompt: > create an app that allows me to upload a picture of a bookshelf… The app should use Google search grounding to add more information. So what we should get is like the author name, the title name, a description of the book… the app should ask the user to log in… it should save all of that information for the user to a database… and we should be able to have that persist.

Within minutes, AI Studio generated the full-stack application. It successfully handled user login, image upload, identified books from a photo (even with obscured titles), used grounding to fetch details for each book, and persisted the user’s library in a database. The final app was shareable via a public URL.

Frontiers of Generative AI

• Project Genie 3: DeepMind’s world model can generate new, navigable, interactive worlds from a text prompt. Each frame is generated dynamically by the model, pixel by pixel, without a traditional physics or game engine. A demo created a playable world described as “Big Bend National Park in Texas… but all of the rock formations are made out of Lego bricks” with a “pink ostrich with a rocket blaster” as the character. While not yet a public API, it showcases the future of interactive content creation.

• Image & Video Generation:

  • NanoBanana 2: Can be used for complex image editing and creation, leveraging image search for grounding. A demo combined a photo of a dog and a photo of a park with the text “with a can of Celsius” to generate a new, coherent image.
  • Veo 3.1 Lite: A cost-effective model for generating high-quality stock video footage. A prompt for a “vegan basketball-themed food truck” (itself generated by Gemini) was used to create a short, thematic video clip complete with custom food names like “Chef Curry’s Tofu Scramble.”

• Music Generation with Lyria 3: Using an app built in AI Studio, a theme song was generated for the “vegan basketball food truck” concept, with a prompt specifying an “electronic, dancable” style with “lyrics in Spanish.” The model produced a fitting, high-quality audio clip.

Gemma 4: Powerful and Open

The newly released Gemma 4 family of open models (Apache 2.0 license) punches far above its weight class in terms of performance. These models are available for free in AI Studio for experimentation and can be downloaded for local use, fine-tuning, and self-hosting. Their efficiency allows them to run on devices ranging from mobile phones to browsers.

• Tags:
• Prototyping
• Generative AI
• Gemini
• Feature
• Genie
• Multimodality
• Gemma
• AI Studio
• Lyria
• Veo