This video features a deep dive into Google's latest AI technologies, with Paige Bailey from Google DeepMind joining Amit Maraj on The Agent Factory podcast. Through introductions and demos of Gemini 3's evolution, the new AI-native IDE Antigravity, and the latest image generation model Nano Banana Pro, the episode shows in detail how developers can leverage these tools to build innovative AI applications.

1. Podcast Introduction and Guest Welcome

Podcast host Amit Maraj welcomes viewers to "The Agent Factory," emphasizing that it's a podcast that dives deep into the technical details of the AI world. This episode features Paige Bailey from the DeepMind team as a guest. Paige introduces herself as the UTL (Unit Team Lead) of the DeepMind Developer Relations team, overseeing all Google AI products that developers can use in their projects, including the Gemini API, Bio, Nano Banana, and Gemma.

2. Google's New AI Releases

Paige highlights that Google has been releasing new models and features at an incredible pace, especially over the past year leading up to May 2025. "We've been shipping a new model or feature every single day, sometimes more than one a day, for the past year," she says, adding that this rapid innovation will continue.

Among the most recently launched products, she mentions the highly anticipated Gemini 3 Pro and the next version of the Nano Banana series, Nano Banana Pro. She also introduces Antigravity, a new developer IDE built by the DeepMind team, noting positive developments across all parts of the development stack. Amit notes that as of November 24, 2025, these models had launched and made a huge impact in the AI industry, describing it as being "like Christmas."

3. What Is Gemini 3? (The Evolution from 1.0 to 3.0)

Paige explains that Gemini 3 is the next step in the Gemini model series.

• Gemini 1: Focused on understanding various types of content. It could simultaneously understand video, images, audio, text, code, and more.
• Gemini 2: Advanced with a focus on reasoning capabilities. It introduced step-by-step planning, detailed thought tracking, and concepts like thinking tokens.
• Gemini 3: Achieved a massive leap in code capabilities and acting ability. Paige says "it's gotten much, much better at code and acting," making it highly useful for agent-style workloads.

Gemini 3 significantly improved tool use and function calling, enabling "more composite architectures" where various models work together or interact with systems to complete tasks. Amit notes that it's intuitive to understand why Gemini 3 excels at coding, since coding is fundamentally about taking action, writing code, and changing functionality.

Paige explains the model-building process in two phases: pre-training and post-training.

• Pre-training: The stage where Gemini is exposed to as many tokens (data pieces) as possible, utilizing the entire internet and vast amounts of synthetic datasets including video game footage and synthetically generated code.
• Post-training: The stage of providing much more specific examples for particular use cases that users expect. This includes carefully curated examples of people performing tasks using multiple tools, multi-turn conversations including website editing, and techniques like reinforcement learning. Paige emphasizes that "tens of thousands of people at Google have worked really hard to make these models that much better." Amit draws an analogy comparing it to Gemini learning basics at school (pre-training) and then learning workflows through on-the-job experience (post-training).

4. "Vending Bench": AI Running a Business

Paige introduces Vending Bench as her favorite performance benchmark for Gemini 3. Vending Bench measures a model's ability to run a passive business -- specifically, a vending machine operation.

"Vending Bench measures the model's ability to run a passive business."

The model performs the following roles:

• Deciding what items to stock
• Reordering when inventory runs low
• Dispatching people to refill machines when orders come in
• Managing all components of the business and making long-term plans to maintain maximum uptime

Performance is evaluated by how much money the model earns over a year. Currently, Gemini 3 Pro generates approximately $5,462 in revenue per vending machine. Paige jokes that "we're operating on an exponential trajectory, so someday we can all retire and run passive businesses," expressing optimism about the future. She adds that it's really exciting to see how models have improved in use cases requiring more strategy and long-term decision-making.

Amit lightheartedly suggests a meeting about the vending machine business after the podcast. Paige says she wishes someone would create a "laundromat bench" evaluation tool, emphasizing that these passive businesses have enormous potential. She mentions that recent AI startups are generating significant monthly revenue through nearly autonomous businesses. Amit agrees that vending machines and laundromats can be safe and good starting points for entrepreneurship.

Paige adds that Gemini 3 is the first model to break 1500 on the Elmarina benchmark (currently at 151 points) and shows outstanding performance in Webdev Arena. Replit's design features are powered by Gemini 3, and the quality of apps generated in AI Studio has improved significantly. It also leads in reasoning, multimodality, tool use, and planning capabilities, and produces much better results than previous models in voxel art experiences like Minecraft. She notes that it's "really great at building games with interesting game mechanics and beautiful UIs."

5. Demo: Building "Nordic Shield" in AI Studio

Paige now demonstrates using Gemini 3 in AI Studio's build feature.

She wants to create an app that talks to users via webcam and microphone. This "Nordic-themed" insurance app lets users show various items while the app:

• Catalogs items: recording name, description, and wear condition
• Uses Google Search grounding to estimate item values and display them in a table
• Should be well-designed overall and display statistics about items
• Uses Gemini Live for video conversation

When Paige enters all these requirements in natural language, AI Studio immediately transitions to an IDE-like environment. Expanding the "Thinking" section on the left reveals Gemini's reasoning process -- determining function calls, model names, data structures, and other actions needed to build the app. Gemini creates the app using React Native, making it accessible on mobile. Paige says it's incredibly useful that Gemini writes much better prompts than she does and generates apps leveraging the latest models and API features, eliminating the need to look up documentation or figure out which model to use.

One of the most impressive moments is when Amit points out that the app came up with the name "Nordic Shield" entirely on its own, without Paige requesting it.

6. Debugging and "Auto-Fix" Feature Demo

Paige runs the "Nordic Shield" app. Initially, the microphone doesn't work properly due to an audio session conflict. At this point, AI Studio's Auto-Fix feature activates, showing Gemini's reasoning process for resolving the issue -- checking "whether the microphone is on, whether it's active, how to understand media streams," and so on.

Paige says that Gemini's ability to think and debug on its own is personally very helpful. "I can't claim to be a JavaScript or TypeScript expert. So getting this help in figuring out what's wrong, debugging, and finding out which libraries to include in the latest stack is really useful for me personally." Amit emphasizes how important it is that AI can recognize errors, analyze them, and transparently show its thought process.

7. Using the Annotate Feature for UI Design

Paige shows that the bottom of the app displays warnings, build process insights, and recommendations for website ergonomics and accessibility (such as display issues on mobile devices).

She then demonstrates the Annotate App feature, which lets users add feedback directly onto the app UI. For example, you can leave a comment like "Make this catalog item section more visually appealing." Gemini interprets these annotations and applies changes to the website, much like working with a designer to refine the UI. She notes that being able to fix bugs and annotate design feedback simultaneously is very efficient.

Shortly after, the app is successfully updated, revealing an "insurance catalog assistant" named Sven. Paige starts a conversation while showing her smartphone:

Sven: "Hello. I'm your insurance catalog assistant, Sven. Show me your items one at a time and I'll help catalog them." Paige: "How about this?" Sven: "I see a dark smartphone with a ring grip on the case. It looks to be in good condition. Do you agree?" Paige: "Sure. It's a Pixel 7, and if I turn it around, there are scratches and a screen crack." Sven: "Ah, thanks for describing that. Now that you're turning it, I can see scratches and a screen crack. Based on that, I'd classify the condition as 'fair.' Do you agree?" Paige: "Great." Sven: "I've cataloged the Pixel 7 smartphone. Anything else?"

Paige then adds a Diet Dr. Pepper can and a computer monitor in the background to the catalog. Sven records each item's name, description, and condition but notes it cannot provide direct valuations. However, the app provides "estimated values" for cataloged items along with citation links from Google Search.

Paige explains this worked as a two-step process. The Live API assistant (Sven) catalogs items, and once items are added to the list, a second agent uses Google Search to find estimated values. Amit says the new capabilities of Gemini 3 and the API that allow grounding in search results are extremely useful, adding that it would be helpful for organizing belongings or selling things on Facebook Marketplace.

8. Nano Banana Pro: Cutting-Edge Image Generation

Paige showcases various use cases for Nano Banana Pro:

• Synthesizing Pinterest-style item collages into single images
• Generating orthographic projection blueprints of real locations (like castles) from various viewpoints with high fidelity
• Creating detailed physics explainers, all backed by tools including Google Search grounding
• Image generation at 1K, 2K, and 4K resolution in various aspect ratios

Paige says she's "really excited to see what people will create" with Nano Banana Pro, mentioning she's seen creations ranging from reimagined game experiences to NeurIPS conference presentations. Amit emphasizes that combining reasoning capabilities with multimodal output opens up endless possibilities.

9. Demo: Website Redesign Using the Antigravity IDE

Amit shows his own project built with Gemini 3, demonstrating the process of redesigning an existing website. He had been struggling to share resources and articles through short-form video content on TikTok, Instagram, and YouTube Shorts. He input his new website design ideas (doodly, vibrant pastel neo-brutalism, notebook-like theme) into AI Studio, which created a beautiful new website for him.

The problem was that his existing website looked nothing like this new design. Amit uses the Antigravity IDE and Gemini 3 Pro's multimodal capabilities to solve this. He feeds two screenshots of the new website into Antigravity and instructs Gemini 3 Pro to "rebuild this site following the new design philosophy seen in the included images."

Antigravity interprets the images, identifies the most important design characteristics, and applies them to code. The AI generates a task list that includes:

• Exploring the existing codebase
• Analyzing the new design
• Setting up the new design system

Amit says the implementation plan generated during this process is useful because it serves as documentation, helping the AI reference past work as the codebase grows.

While Antigravity works, Paige shares examples of using Nano Banana Pro to generate images of physicists or computer scientists, and cases of using it for game asset generation at hackathons.

After the task completes, Amit runs the app to show the results. His new website follows the brand guidelines well, featuring "pill" elements that wobble on hover, along with information about resources, prompt libraries, and agent blueprints. Amit says even a single prompt produced pretty good results and asks for Paige's opinion. Paige exclaims, "That looks amazing! Thank you, Gemini 3 and Antigravity," adding "I love the attention to detail and the creativity in the UI design. It seems better than ever at creating these kinds of experiences."

10. The Full Google AI Stack

Finally, Paige thanks Amit for having her on the podcast and shares that it was great to talk about Google's full AI stack. She emphasizes that every part of Google's AI stack -- from hardware to compilers, machine learning frameworks, and final app deployment tools like Antigravity and AI Studio -- "helps build the best models." Amit wraps up by mentioning that apps can be deployed to Google Cloud for actual production use.

Conclusion

This podcast clearly demonstrated that through Gemini 3's remarkable evolution, AI has moved beyond simply understanding and reasoning to now possess the 'agency' to generate code and take real action. Tools like the Antigravity IDE and AI Studio enable developers to build complex AI apps much more easily and quickly, while Nano Banana Pro is pushing the boundaries of creativity in image generation. Together, these technologies represent Google's full AI stack offering limitless possibilities to developers, heralding a future where AI becomes even more deeply integrated into our daily lives and businesses.