The Future of AI Development ft. Gemini, Robotics & Space [Chain of Thought]
Google DeepMind's Paige Bailey on Gemini 3, vibe coding, and why the next frontier is "bits to atoms"
Something New for Test Lab
You’re used to getting occasional essays from me on AI strategy, dev tools, market trends, and mental models. Starting now, I’m also sharing episode breakdowns from Chain of Thought: my podcast where I interview the people actually building the AI infrastructure we’re all navigating.
What to expect going forward:
1-2 podcast emails per month (key takeaways, timestamps, full transcript)
Continued occasional essays when I have something worth saying
No fluff, no daily spam
I think you’ll find these conversations valuable. The guests are people like Paige Bailey (Google DeepMind), Chip Huyen, and the teams building the tools we’re all learning to use.
First up: episode 48 with Paige on where Google’s AI platform is heading - and a bit on what data centers in space may look like 👀
Also available on Spotify, Apple Podcasts, or wherever else you prefer to listen.
Why This Episode
Paige Bailey leads AI Developer Relations at Google DeepMind, where she sits at the intersection of frontier research and what developers actually ship.
In the past few months alone, her team has launched Gemini 3, shipped Anti-Gravity (Google’s new agentic IDE), and released what they’re calling the largest open-source interpretability toolkit ever. She also found time to build a gamified NASA curriculum using AI Studio, just to prove what’s possible.
This conversation covers the practical (when to use Flash vs Pro, how to think about model selection) and the ambitious (AI-powered robotics on Raspberry Pi, data centers in orbit). It’s an excellent window into how Google is thinking about the next phase of AI development.
Note: Paige demos several tools during this episode. If you’re listening audio-only, I’d recommend checking out the video version on YouTube to see the screen shares.
Key Takeaways
1. “Vibe coding” is real, and Google is betting big on it
Paige demoed Space Math Academy, the gamified NASA curriculum she built using AI Studio’s Build feature.
The workflow: describe what you want in natural language, point at UI elements to request changes (”make this button blue”), and deploy to Cloud Run.
Developers are being unblocked to aggressively ship, non-technical people are being unlocked to build, and Google is seeing massive interest in their new Build features.
2. Project Suncatcher: Google is exploring data centers in space
Google is actively researching orbital data centers that would run on continuous solar power, eliminating the energy sustainability concerns around AI infrastructure. It sounds like science fiction, but Paige positioned it as a serious long-term investment, not a moonshot. The narrative around AI’s environmental impact (which Andy Masley debunked on the podcast in December) may look very different in 5-10 years.
3. The next frontier is “bits to atoms” - and it’s already happening
Google’s robotics work is now accessible to hobbyists. Paige showed us Stanford Puppers—quadruped robots running Gemini on Raspberry Pi that developers can build and program themselves. Google is opening up a robotics trusted tester program.
Her framing: we’ve spent years on AI that understands and reasons. Gemini 3 is about AI that acts, and acting in the physical world is the logical next step.
4. Gemini 3 Flash outperforms Pro for most agentic use cases at a fraction of the cost
The Flash vs Pro decision isn’t just about capability tiers anymore. Paige showed that Flash actually beats Pro on several agentic benchmarks, particularly browser operation and tool use.
Practical guidance: start with Flash ($0.10/1M input tokens) and only move to Pro when you hit specific reasoning limitations. Most developers are over-provisioning.
5. If you’re feeling overwhelmed by the pace of releases, you’re not alone
Paige’s closing advice resonated: even people inside Google feel overwhelmed by the release velocity. Her recommendation isn’t to try to track everything. It’s to build intuition by actually testing the tools yourself. Pick one thing and go deep rather than trying to stay current on everything.
It’s going to keep changing anyway.
Full Transcript
Timestamps
00:00 — Introduction
01:24 — Paige’s background & connection to Modular
02:29 — Gemini integration across Google products
03:11 — Jules, Gemini CLI & Anti-Gravity IDE overview
03:56 — Gemini 3 Flash vs Pro: live demo & pricing
05:51 — Google’s hardware advantage: TPUs & JAX
06:19 — Choosing the right Gemini model
09:51 — TensorFlow history & evolution to JAX
11:54 — NeurIPS 2025 & Google’s research culture
14:14 — Google Brain to DeepMind: the merger story
15:34 — PaLM II to Gemini: scaling from 40 people
17:55 — Anti-Gravity IDE deep dive
20:15 — Gemma open source models
20:56 — How to use Anti-Gravity for development
24:04 — MCP Protocol & Chrome DevTools integration
27:03 — Gemini CLI in Google Colab
27:46 — Image generation & AI Studio traffic spikes
28:15 — Space Math Academy: gamified NASA curriculum
31:39 — Vibe coding: building with AI Studio & Anti-Gravity
36:18 — AI from bits to atoms: the robotics frontier
37:14 — Stanford Puppers: Gemini on Raspberry Pi robots
38:16 — Google’s robotics trusted tester program
40:26 — AI in scientific research & automation
42:41 — Project Suncatcher: data centers in space
45:17 — Sustainable AI infrastructure
47:31 — Non-dystopian sci-fi futures
48:05 — Closing thoughts & resources
Links & Resources
Connect with Paige:
Website: webpaige.dev
LinkedIn: linkedin.com/in/dynamicwebpaige
X/Twitter: @DynamicWebPaige
Tools & Projects Mentioned:
AI Studio — Try Gemini models and build apps
AI Studio Build — Deploy apps from natural language
Space Math Academy — Paige’s gamified NASA curriculum project
NASA Space Math — The original curriculum
Stanford Pupper Documentation — Build your own AI-powered robot
Related Episode:
Related Article:
🤫Sneak peek
For those of you who read to the end, here’s an early look at some upcoming episodes:
Angie Jones - VP Engineering, AI Tools & Enablement - Block (with a live demo of Goose!)
Intercom’s Chief AI Strategist, Fergal Reid, takes us inside their wildly successful AI agent Fin
+ many more exciting guests!
If you know someone we should have on the show, reply and let me know.
Subscribe to Test Lab for episode breakdowns, occasional essays, and strategic thinking on AI, developer tools, and the business of technology.
Paige's point about Flash outperforming Pro on agentic benchmarks while costing way less is a game-changer for budgeting. I'd been defaulting to Pro for everything agent-related, but those browser operation stats clarfiy that most teams are overpaying. The Stanford Puppers robotics demo on Raspberry Pi shows how accessible hardware-software integration has become tho.