How Gemini 4 Argon & Carbon were built (ft. DeepMind’s Head of Gemini Product)
From Google Cloud Tech · · 41 min

The player loads from YouTube only when you click. It uses YouTube's privacy-enhanced mode.
If the video does not play, the owner may have turned off embedding. Watch it on YouTube instead.
Description
Try Gemini 4 Argon & Carbon in Google AI Studio → https://g.dev/ai/studio Build with AI Agents in Google Antigravity → https://g.dev/ai/use-antigravity
How do you build a frontier model that can work autonomously on complex codebases for days while bringing indirect prompt injection and cheating rates down to 0%? In this episode of The Google Cloud Show, we talk to Tulsee Doshi, Senior Director and Head of Product for the Gemini model family at Google DeepMind to unpack Google's back-to-back model releases of Gemini 4 Argon and Gemini 4 Carbon through an exclusive Google DeepMind interview and a hands-on developer demo showing how persistent, long-horizon models are powering the next generation of AI agents and software engineering workflows.
In our main segment, Smitha sits down with Tulsee to take you inside the lab to share how Google dogfoods early checkpoints through internal experiments. We will also discover why Gemini 4 Argon (built for deep reasoning and multi-step workflows with a 1M token output capacity) and Gemini 4 Carbon (#1 on the Artificial Analysis Intelligence Index) rolled out first to trusted cyber defenders via the Fairwind program. You'll also hear how Carbon improves web development aesthetics, reduces fabrication, and achieves state-of-the-art defensive cybersecurity performance, plus why agentic code submissions inside Google have jumped 30% since Argon launched, how the Chrome team saw Carbon perform over 10x better at generating high-quality code patches, and how Tulsee uses audio models in Antigravity alongside Argon to turn voice rambles into structured documents and vibe-code a custom nanny app.
Stick around after the interview for a hands-on builder segments that put Gemini 4 to work in Google AI Studio by using voice-to-text prompts to remix a 3D chessboard with coordinate labels and bright neon pieces, and adding Firebase Authentication and a live Cloud Firestore leaderboard to a 3D racing game in minutes instead of weeks.
Finally, Rody Davis from the Antigravity DevRel team demonstrates how to build an automated "Software Factory" inside Google Antigravity powered by Gemini 4 Argon by using Markdown-defined custom subagents with isolated context windows and read-only permissions to audit Web (Node.js), Android (Jetpack Compose), and iOS (SwiftUI) code in parallel, strip out legacy polyfills, and steer subagents directly using /learn, /boost, and Teamwork Preview.
Chapters: 0:00 - Intro 0:46 - Gemini 4 Argon (1M token output) & carbon (#1 intelligence index) overview 2:24 - Interview: Tulsee Doshi, Head of Product for Gemini at Google DeepMind 4:07 - Relentless shipping: How Gemini 4 iteration compares to Gemini 2 & 3 6:23 - Gemini 4 vs. Gemini 3.1 Pro: Model persistence on long-horizon tasks 7:46 - Argon vs. Carbon: Less fabrication, web dev aesthetics & SOTA Cyber 13:36 - Why Argon & Carbon rolled out to defensive cyber teams first 14:43 - DeepMind PM’s workflow: Voice rambling to structured docs in Antigravity 18:01 - How a DeepMind PM tests every new model checkpoint 21:21 - Why you should treat Gemini 4 as a thought partner, not the thinker 22:24 - Why Gemini 4 models are named after periodic table elements 25:27 - Remix 3D apps in Google AI Studio 28:28 - Adding Firebase Auth & a Live Firestore leaderboard to a 3D racing game 30:26 - How to build a multi-agent software factory in Antigravity 31:09 - Defining custom subagents in markdown & platform version policies 34:20 - Using the /learn command & parallel subagent context windows 35:31 - Enforcing read-only subagent permissions & direct subagent messaging 38:00 - Scheduling monthly codebase audits & using /boost and teamwork preview 40:07 - Wrap up & where to start building with Gemini 4 & Antigravity
🔗 More resources: * Read the Gemini 4 Argon announcement → https://g.dev/ai/intro-gemini-4-argon * Read the Gemini 4 Carbon announcement → https://g.dev/ai/intro-gemini-4-carbon * Explore Google Antigravity custom subagents…