AI Studio can turn a prompt into an application with Google login, Firestore storage, model calls, and deployable code.
2
Google's generative media models can turn a public-domain book into character images, chapter illustrations, video scenes, music, and narrated dialogue.
3
Gemma 4 makes multimodal agents, parallel subagents, and coding workflows possible on phones, laptops, and single-GPU systems without cloud APIs.
Summary
Paige Bailey gives a tour of Google's current developer tools, starting with Gemini in AI Studio. She demonstrates video analysis with Gemini 3.1 Flash-Lite, code execution for image bounding boxes, and Build creating a bookshelf catalog app with Google login and Firestore. Gemini Live handles screen and camera input, while Genie creates a playable world from a text description. Guillaume Vernade then builds a media pipeline around The Wind in the Willows. Gemini extracts characters and chapter prompts, Nano Banana creates consistent reference images, Veo animates scenes, Lyria generates music, and text-to-speech produces a multi-character reading. He explains how structured outputs, chat history, reference images, retries, and service tiers make the workflow more usable. Ian Valentine closes with Gemma 4 running locally. He shows on-device agent skills, parallel SVG generation, local coding with Open Code, and a model fixing a game from its own error message.
AI Studio gives developers a direct path from a working prompt to application code
Paige describes AI Studio as both a model playground and a Build environment similar to v0.dev or Lovable. In Build, she prompts an app that accepts a bookshelf image, identifies titles and authors, uses Google Search grounding to fill in missing details, saves books to Firestore, and attaches records to a Google account. The generated workspace includes an IDE, file uploads, secrets, version history, OAuth integrations, and GitHub syncing. When Firestore permissions fail, the model inspects the relevant rules and edits them. Paige's practical claim is simple: once an interaction works in AI Studio, the Get Code button provides TypeScript, Python, or another language for reproducing it in an application.
Small Gemini models can inspect video and use code execution at low cost
Paige uses Gemini 3.1 Flash-Lite to inspect a YouTube dinosaur video frame by frame. She limits the input to the first 300 seconds, with sampling at about one frame per second, then asks for timestamps, dinosaur names, and a fact about each one. AI Studio exposes the uploaded video, prompt, token count, and grounding controls, and Get Code generates the API setup. In another example, Gemini receives an image of Lego pieces and uses a sandboxed Python environment to draw bounding boxes around green bricks. Paige says the same setup can count objects, locate them across video frames, or calculate their orientations. A larger model checks its first result, while the lightweight model completes the basic task for pennies.
Genie creates interactive worlds from pixels rather than traditional game assets
Paige explains that Genie is a composition of models including Nano Banana, Veo, and Gemini. It turns a description into a frame-by-frame playable experience instead of producing a Unity or Unreal Engine project. Her live prompt describes Regent's Canal on a sunny day, dolphins in the water, pirate flags on the boats, and a pink sparkly squirrel with purple feet and a pirate hat. The resulting world responds to movement and jumping for a 60-second experiment. The squirrel can hop across the canal and boats, while the dolphins and people appear in the generated scene. Paige contrasts this with World Labs, which builds actual Unity or Unreal environments. Genie stores neither 3D assets nor a conventional physics simulation, and generates raw pixels dynamically as the player moves.
Reference images and structured outputs improve character consistency in book illustrations
Guillaume's workshop uses The Wind in the Willows from Project Gutenberg as input for a generative media pipeline. Gemini reads the book and produces descriptions for the main characters, then Nano Banana creates portraits in a selected style. A first attempt asks for chapter illustrations, but Guillaume notices that the toad does not match the earlier reference image. His improved design makes each chapter result include its name, image prompt, and a list of characters who should appear. The generation step then passes only those characters' reference images into the prompt. He says this is more sustainable for a book with many characters, and suggests creating several views of each character, such as front, back, and side images, before generating scenes.
Gemini can coordinate separate media models through prompts and shared context
Guillaume uses the chapter descriptions to drive several models. Veo receives a chapter image and a prompt to animate it, although he finds that prompts written for still images do not always explain what should happen next. A second approach asks Gemini to write a motion prompt from the chapter and image first. Lyria then generates a song for each chapter from Gemini-written music prompts, including peaceful folk music, an adventurous road song, and a suspenseful piece with pizzicato strings. Text-to-speech turns book dialogue into a play transcript. Guillaume assigns a narrator voice and asks the other characters to share a voice with different speaking styles, accents, and pacing. He notes that Gemini is good at writing prompts for these models because much of the media-model training data was made with Gemini.
Gemma 4 makes local multimodal agents practical across several hardware levels
Ian Valentine introduces four Gemma 4 models. The E2B and E4B effective models target phones, Raspberry Pis, and small devices. The 26B mixture-of-experts model activates four billion parameters, while the 31B dense model targets laptops, desktops, and single-GPU cloud instances. The models include built-in thinking and support multimodal input, with audio support on the two smaller models. Ian shows the Google AI Edge Gallery running skills on a phone, where the model can interpret a request and call Android intents, JavaScript skills, or a research-tracker function. He also shows a small calculator generated on-device. Gemma 4 works with tools such as Ollama, LM Studio, vLLM, and SGLang, and can expose an OpenAI- or Anthropic-compatible local endpoint.
Parallel local subagents can divide visual and coding work without a cloud API
Ian runs a 26B Gemma model through LM Studio on an M4 Mac and points Open Code at its local endpoint. One orchestrator creates ten subagents, each assigned a different SVG to draw. The agents work in parallel, then the orchestrator combines their outputs into a page. Ian says the same pattern can divide file sorting, coding, research, or data analysis. He then gives a local model a specification for a game called Nebula Drift. Open Code lets the model read and write files, use shell commands, select editing tools, and call subagents. The first version has a syntax error, so Ian feeds the error back manually. The model edits the files and eventually produces a playable starfield with a movable ship. A separate 31B run also creates a small game in which the player builds another game.
"Usually if you see everybody sprinting to do the same thing, that's a great indication that it's the wrong thing."04:47
Who should watch
You are building a multimodal application and want concrete examples of Gemini APIs, structured outputs, file uploads, grounding, and sandboxed code execution.
You want to turn documents or books into coordinated image, video, music, and speech outputs, with better consistency between generated assets.
You need models that can run locally on phones, laptops, or a single GPU, and want to see how local agents can use tools and repair generated code.
