Gemini 3 & Robot Dogs: Inside Google DeepMind's AI Experiments | Paige Bailey

Show Notes

Google DeepMind is reshaping the AI landscape with an unprecedented wave of releases—from Gemini 3 to robotics and even data centers in space.

Paige Bailey, AI Developer Relations Lead at Google DeepMind, joins us to break down the full Google AI ecosystem. From her unique journey as a geophysicist-turned-AI-leader who helped ship GitHub Copilot, to now running developer experience for DeepMind's entire platform, Paige offers an insider's view of how Google is thinking about the future of AI.

The conversation covers the practical differences between Gemini 3 Pro and Flash, when to use the open-source Gemma models, and how tools like Anti-Gravity IDE, Jules, and Gemini CLI fit into developer workflows. Paige also demonstrates Space Math Academy—a gamified NASA curriculum she built using AI Studio, Colab, and Anti-Gravity—showing how modern AI tools enable rapid prototyping.

The discussion then ventures into AI's physical frontier: robotics powered by Gemini on Raspberry Pi, Google's robotics trusted tester program, and the ambitious Project Suncatcher exploring data centers in space.

00:00 Introduction

01:30 Paige's Background & Connection to Modular

02:29 Gemini Integration Across Google Products

03:04 Jules, Gemini CLI & Anti-Gravity IDE Overview

03:48 Gemini 3 Flash vs Pro: Live Demo & Pricing

06:10 Choosing the Right Gemini Model

09:42 Google's Hardware Advantage: TPUs & JAX

10:16 TensorFlow History & Evolution to JAX

11:45 NeurIPS 2025 & Google's Research Culture

14:40 Google Brain to DeepMind: The Merger Story

15:24 Palm II to Gemini: Scaling from 40 People

18:42 Gemma Open Source Models

20:46 Anti-Gravity IDE Deep Dive

23:53 MCP Protocol & Chrome DevTools Integration

26:57 Gemini CLI in Google Colab

28:00 Image Generation & AI Studio Traffic Spikes

28:46 Space Math Academy: Gamified NASA Curriculum

31:31 Vibe Coding: Building with AI Studio & Anti-Gravity

36:02 AI From Bits to Atoms: The Robotics Frontier

36:40 Stanford Puppers: Gemini on Raspberry Pi Robots

38:35 Google's Robotics Trusted Tester Program

40:59 AI in Scientific Research & Automation

42:25 Project Suncatcher: Data Centers in Space

45:00 Sustainable AI Infrastructure

47:14 Non-Dystopian Sci-Fi Futures

47:48 Closing Thoughts & Resources

- Connect with Paige on LinkedIn: https://www.linkedin.com/in/dynamicwebpaige/

- Follow Paige on X: https://x.com/DynamicWebPaige

- Paige's Website: https://webpaige.dev/

- Google DeepMind: https://deepmind.google/

- AI Studio: https://ai.google.dev

Connect with our host Conor Bronsdon:

- Substack – https://conorbronsdon.substack.com/

- LinkedIn https://www.linkedin.com/in/conorbronsdon/

Presented By: Galileo.ai

Download Galileo's Mastering Multi-Agent Systems for free here!: https://galileo.ai/mastering-multi-agent-systems

Topics Covered:

- Gemini 3 Pro vs Flash comparison (pricing, speed, capabilities)

- When to use Gemma open-source models

- Anti-Gravity IDE, Jules, and Gemini CLI workflows

- Google's TPU hardware advantage

- History of TensorFlow, JAX, and Google Brain

- Space Math Academy demo (gamified education)

- AI-powered robotics (Stanford Puppers on Raspberry Pi)

- Project Suncatcher (orbital data centers)

What is Chain of Thought | AI Agents, Infrastructure & Engineering?

AI is reshaping infrastructure, strategy, and entire industries. Host Conor Bronsdon talks to the engineers, founders, and researchers building breakthrough AI systems about what it actually takes to ship AI in production, where the opportunities lie, and how leaders should think about the strategic bets ahead.

Chain of Thought translates technical depth into actionable insights for builders and decision-makers. New episodes weekly.

Conor Bronsdon is an angel investor in AI and dev tools, Technical Ecosystem Lead at Modular, and previously led growth at AI startups Galileo and LinearB.

Disclaimer: All views, opinions and statements expressed on this account are solely my own and are made in my personal capacity. They do not reflect, and should not be construed as reflecting, the views, positions, or policies of Modular. This account is not affiliated with, authorized by, or endorsed by Modular in any way.

[0:06] Conor Bronsdon:
Back to Chain of Thought, everyone. I am your host, Conor Bronson, head of technical ecosystem at Modular. And I'm delighted to be joined today by Paige Bailey. Paige is the AI developer relations lead at Google DeepMind, which maybe a few of you have heard of, and has had one of the most interesting careers in AI. She started as a geophysicist at Chevron, become a developer advocate at Microsoft and Google,

[0:28] Conor Bronsdon:
helped ship GitHub Copilot as a principal PM, led product for Gemini, is a huge fan of open source, which I'd love to talk to her about, developer tools, and much more. Paige now runs developer experience for DeepMind's entire AI platform. And in the past couple of months alone, her team has launched, what, Gemini three, the anti gravity, Google's new IDE. They've released what they've called the largest open source interoperability

[0:53] Conor Bronsdon:
tool kit ever and has also found time to build a gamified NASA curriculum using AI Studio, which I'm hoping she'll show off.

[1:01] Speaker:
So, Paige, welcome, Chain of Thought. Where are joining us from? I am, joining from Texas. I'm here for the holiday season visiting some family and then also taking in probably too much barbecue, if there is such a thing. But it's, it's beautiful weather out here. I are you based on the West Coast?

[1:18] Conor Bronsdon:
I am in Seattle, and it is a little less beautiful here. We've had an atmospheric river coming through. Thank god. Our house is not flooding like some folks. Very thankful for that. But Yeah. Definitely a a little less, sunny than I expected it. Yep.

[1:32] Speaker:
Awesome. Well, thank you so much for having me. I'm also a huge fan of the team over at Modular. I think y'all got all of the or most of the machine learning compilers contingent. So, Chris and Tim are are good friends and worked closely with them on the,

[1:47] Conor Bronsdon:
TensorFlow and kind of the early days of many of the machine learning frameworks and compilers at Google. Yeah. I'm excited to see what the future holds because I think the the team here at Modular is is really excited to keep building alongside the incredible innovation that the whole team at Google is doing. And I will I will note here, this is, as we're recording, this is the start of week four for me. So I'm still very new, figuring things out, but, having a lot of fun with it for sure. It's it's pretty incredible seeing what,

[2:14] Conor Bronsdon:
you and Chris and Tim and everyone else involved have built over the years. And now seeing the results of that come to fruition with, you know, Gemini really taking center stage among the the grouping of frontier models, the variety of product releases Google's done. I mean, even just some of the interesting stuff that's happening with agents within Google products. Like I talked to the Google Shopping team a little while ago,

[2:40] Conor Bronsdon:
and some of the new agents they have that can track prices for you and help you identify when you want to purchase an item, or even call a business to get, answer questions answered like, hey, do you have the right type of guitars I'm looking for? What are your hours? It's really cool to see how Google is, collectively throughout the entire product strategy, beginning to integrate agents, to integrate Gemini within the product suite.

[3:02] Speaker:
And you've been a big part of that. Yeah. Well, it's it's a huge, huge team effort all over Google. I've been I've been really excited to see, you know, tools like Joules for asynchronous agent work, Gemini CLI, which is is more about kind of synchronous operations. So you're working with the agent in real time in order to to accomplish various tasks within the terminal, and then also, of course, anti gravity.

[3:28] Speaker:
And it's been really especially cool. Let me pull this up just

[3:33] Conor Bronsdon:
just a second. I will note briefly for anyone who is listening to this, Paige is gonna be sharing a variety of things from her screen. So I recommend either trying to watch on Spotify video or YouTube if you can just to get a a little more discussion since we're gonna be showcasing a few things. Absolutely. And so so this the first thing that I pulled up is is basically our our overview of Gemini three Flash. So this is one of our latest kind of big releases.

[3:59] Speaker:
Flash is the companion for Gemini three Pro and is even better at Gemini three Pro at a variety of use cases, specifically things like being able to operate a browser, many of the capabilities that you might expect out of a Gentex systems. But one of the things that I'm most excited about is this kind of Pareto curve that we see here on the screen. For folks who haven't seen this before or who might be listening along,

[4:27] Speaker:
it's basically a chart showing that the the capabilities to cost performance for the Gemini three Flash and Gemini three Pro series are pretty pretty much blazing a path for for frontier models in general. And that's one of the things that I love most about Google is that we're we're very good at making things go very fast and very cheaply so they can scale out to billions and billions of users. I was listening to,

[4:56] Conor Bronsdon:
some of the Acquired episodes recently that are covering Google, and there's, I mean, listeners who who followed me for while know that Acquired is one my favorite podcasts. They do incredible storytelling. And one of the the early conversations there is about the differences between how Google built its first data centers, which were, they were not going for shiny machinery. They were going performance. Like, how much performance could we fit into a square foot? We really need to optimize for this. And I think that ethos has carried through with Google,

[5:25] Conor Bronsdon:
I mean, for for decades now. And part of that is some of the early engineers who were helping build that. People like Jeff Dean have been here a long time and are, having major impacts on the culture, part of it simply, like, the the talent being recruited. But but I am really continually impressed by how Google has kind of stepped up within the AI race and obviously has been a research leader for a long time.

[5:48] Conor Bronsdon:
What do you see as the way developers should be thinking about the different models that Google's offering? So, you know, you've got three Pro, you've got three Flash. You've gotta get obviously, you still have the 2.5 models, and then there's also open source work that's being done. What are the considerations that should be in the heads of developers as they think through, like, hey, I'm building something.

[6:08] Speaker:
Which model should I be using? What are the pros and cons? Oh, that's such a juicy question. So, and we can show a couple of demos for this in just a second. Gemini three Pro, obviously, our flagship model. It has thinking baked in, which means that the answers you're going to get are probably going to be much more fully expressed, but they'll probably take a a lot longer to get you the answer that you're hoping for

[6:33] Speaker:
due to to kind of all of the the thinking tokens as well as it's just a much larger model. And it also comes at a much higher price point than something like Gemini three Flash, which does give you the ability to turn on thinking, but comes at a significant discount compared to Gemini three Pro. And I think I saw yep. So so this is the input price comparison

[6:57] Speaker:
between Gemini three Flash and Gemini three Pro. For folks who can't see, the input is about 25% of what you would expect to pay for Gemini three Pro, even a little bit lower if you have significant number of tokens in your context window for inputs. And then the output is also significantly less expensive, just coming in at about 25%. But if we kind of zoom out,

[7:26] Speaker:
I'm gonna head over to AI Studio real quick. So let me go ahead and share a different window so we can see the AI Studio interface. This is for folks who are unfamiliar or uninitiated, I suppose, this is the place where all of the latest DeepMind models get released. You can test them out. You can try them out. We even have something called compare mode up here on the right.

[7:53] Speaker:
It looks like two little arrows pointing at each other. And so if I click compare mode, I select Gemini three Pro preview. I select the Gemini three flash preview, and maybe I turn off thinking. So I'm going to make thinking level minimal. Thinking level is high by default using the the Gemini three Pro model. And then we can ask something ask something like,

[8:20] Speaker:
please write a Python function to give me the lat long coordinates of a given city. And then hit run. You can see that the two different models kind of take a little bit of time to little bit of time to think. But Gemini three Pro is still thinking. Gemini three Flash got its answer very, very quickly. And when you look at the Python that's returned, it's more or less the same thing.

[9:02] Speaker:
So there isn't an example usage, but but you can see there are sources cited. They're both incorporating GeoPy. They both basically have the same Python code. This just got you the answer in, like, one tenth of the time. Well, and, crucially, there's a difference in tokens used as well. Absolutely. Because you aren't using all of the intermediate thinking tokens.

[9:25] Speaker:
And I love that when you hover over the tokens in this compare mode or in AI Studio in general, you can see kind of the input cost, the output cost, and the total cost. And it's coming in at less than a penny in order to do in order to do this work for Gemini three Flash. And I think this is part of where Google's

[9:45] Conor Bronsdon:
hardware advantage, frankly, really comes into play too. The fact that I mean, we talked very briefly about TensorFlow earlier, but, I mean, the fact that TPUs are what powers Google architecture, and it it's a vertical stack from hardware on up compared to most competitors that are largely building off of NVIDIA GPUs. Mhmm. There are limits to the optimizations you can do when you are using someone else's stack compared to something you've you've built from the ground up yourself for performance in mind.

[10:17] Speaker:
Yeah. Absolutely. And I and I think it's actually funny to remember because when TensorFlow was first released or at least first open sourced, which was way back in kind of November, December time frame of 2015, the when it was first released, it only had CPU support. So it was it was really intended to be doing machine learning at scale, but across clusters of CPUs.

[10:43] Speaker:
Not GPUs, not TPUs. And then all of the the the kind of different back end support was was effectively retrofitted. So you had different code paths to use CUDA implementations, to use XLA as the target compiler, which would interface with TPUs, ultimately. And then really, the the thing that was really built around XLA is JAKs, which is what we use in production and in in all of our research work and have for a a very, very long time.

[11:19] Speaker:
And I think part of the reason, to your point, why people love JAX so much other than JAX being just like a friendly NumPy like interface to do work is because it can use TPUs so optimally. And you can scale out not just to a single TPU, but also many, many t use all at once, even across multiple data centers, which is what we need for when we train the Gemini models.

[11:46] Conor Bronsdon:
And you mentioned research, and I think this is a interesting area because Google increasingly has had a a larger share of what's happening at major conferences like NeurIPS, which I know you were recently at the 2025 edition. Yep. I'm curious what your takeaways have been about how Gemini and Google's broader AI ecosystem are penetrating the research sphere, not only from internal folks at Google, but also

[12:10] Conor Bronsdon:
the perspectives you're seeing externally?

[12:13] Speaker:
Yeah. So so I I I feel so fortunate. Like, we have we've had Jeff Dean for so long, and he's been, like, the the biggest champion you can imagine of research and of kind of helping grow the academic community as well as the applied machine learning community and also open source. So he he and Jeff Hinton actually just had a talk. It was sponsored by Radical Ventures.

[12:42] Speaker:
Let me see if I can pull this up really quick. But they they gave they gave a talk at, and I'm going to pull up that pull up that really, really quickly so we can see it. The collaboration that built modern AI, Jeff Hinton and Jeff Dean in conversation with Jordan Jacobs. I strongly, strongly recommend this if you're in need of if you're in need of a podcast or or something similar

[13:13] Conor Bronsdon:
to listen to this weekend or Finishes this episode first, to be clear. Yes. Yes. Yes. Yes. Follow on. Follow on. Absolutely.

[13:20] Speaker:
And and but it it talks through everything from like, I remember Brain in the early days was it felt so close knit. Like, you could walk around the hall and see pretty much everybody. And, like, you know, Gnome and the Transformers team and everybody was just kind of, you know, all working very, very closely together. But this walks through so many of the anecdotes of Brain in the early days of what it felt like to work

[13:47] Speaker:
on TensorFlow and on some of the machine learning frameworks when they were just released. And then it also touches on kind of the the launch of the Gemini program. And Jeff was really the one that was, you know, pushing in the direction of DeepMind and Google needing to collaborate and to join forces.

[14:06] Conor Bronsdon:
So it's it's just such a good thing. And collectively, you've been along for, what, six years of that ride now at this point?

[14:13] Speaker:
Longer longer than that. So I the I joined Google Brain in 2017, 2018. Oh, wow. Okay. And then and then spent a year away at GitHub helping with machine learning features in Versus Code and some of the early, like, user experience work for for Copilot. But but it's been a wild and very exciting ride. Like, really, really enjoy really, really enjoy both DeepMind and then, of course, love, love, love the time at Brain. I would love to

[14:43] Conor Bronsdon:
get your thoughts on the the obvious change there where I mean, it used to be that Google had these differentiated teams that were working on different AI projects before it came together. And I think it's really clear externally that there's been massive gains from combining that brainpower, Google and and Unincent. Enabling focus. Yeah. That was intended. I and

[15:07] Conor Bronsdon:
I'm just curious from your perspective, having kind of been on both sides of that, where you've seen the team improvements. Because, I mean, externally looking in, that is a a huge improvement that Google has made just structurally to enable a lot of the innovation happening today, like antigravities and other things. Oh, absolutely. Like, I I think that it makes perfect sense that we were able to join forces, and then obviously, the organization has just grown

[15:33] Speaker:
so much. Like, I'm I'm not sure I'm not sure if this has been mentioned externally, but, like, the Palm two model, as an example, which was a precursor to Gemini, I think the the team that was building it was really, like, 40 people, 42 ish people. So super, super small. It was like, myself is, like, a 20% assignment, and then it had a 20% PM or TPM, Erica Marrera,

[16:03] Speaker:
but super, super tiny. And now when you think about the size of the Gemini effort, you know, you need multiple pages to just include the names of all of the team. And and so I I think it's just a kind of a a difference in scaling from zero to one and one to 10 and then 10 to to many, many more. And I I think it's absolutely been for the best in terms of, you know, making it possible for us to build these great models and making it possible for us to accelerate them into production.

[16:34] Speaker:
But it's just very much a different a different style of vibe. Maybe a little bit like moving from an early stage startup to a to a later stage startup. That's a great comparison.

[16:45] Conor Bronsdon:
And it's also enabled Google to, as we've mentioned, put Gemini and AI powered, enablement into so many Google products that are already either seeing massive growth. I mean, we can look at Chrome where I've got this little Gemini app at the corner for me to leverage. You can look at the Google Workspace product, obviously, and the new abilities to agents within it.

[17:07] Conor Bronsdon:
And then you can also look at new products like anti gravity, the IDE, the agentic IDE that Google has developed, post Windsurf acquisition. And I'd love to get your thoughts on how you see developers working with anti gravity in the coming months and and hopefully years. And then also, an interesting point of the strategy around anti gravity, which is that it's not only Google models you can use within anti gravity. My understanding is you can also use open AI and anthropic models as well. And so would love your thoughts on on that piece,

[17:40] Conor Bronsdon:
especially as it relates to I mean, are all closed source models, but this this idea of, like, openness within AI.

[17:46] Speaker:
Yep. Absolutely. So I and that's also a a very nice segue to the to the last thing that I'll that I'll point out in in terms of pricing for our models. But I I feel like this is a much a much underutilized aspect of our APIs. We don't offer the Gemma models within AI Studio, but it is available in our APIs. So if you don't want to download the the Gemma three or three n models and use them

[18:17] Speaker:
use them locally, you can use them free of charge via the Gemini APIs in AI Studio. Like, a i.google.dev

[18:26] Conor Bronsdon:
is the URL that you can go to to get access to. And to use them, all you have to do is toggle out the model name. I feel like a lot of folks don't know that much about Gemma, frankly, either. I mean, and it sounds so similar, Gemini Gemma, but Gemma being this collection of lightweight open source models that Google has released. Yep. Absolutely. So Gemma three is

[18:45] Speaker:
an Gemma three is our latest open model family. Gemma 3N is 4,000,000,000 parameters in size, though there are some other variants, so 1,000,000,000, 2,000,000,000 parameters, small enough to fit within a browser. But they're super, super fast, very speedy, comparatively, comparatively strong performance. Gemma three is 27,000,000,000 parameters in size. And if you need to do things like summarization

[19:16] Speaker:
or, basic translation between one language to another or, a kind of categorization, things like toxicity detection, the Gemma three family is very, very strong as a choice. And, again, like, since they're free tier, relatively fast, available via the API, I would strongly suggest think about thinking about using some of those for the for the lighter weight tasks that you might be you might be experiencing

[19:49] Speaker:
with the the Gemini the Gemini models. And they are built on many of the same research ideas. Though the expectation is that the Gemma model family will always be kind of smaller in size and not kind of the the same frontier capabilities that you would get with the Gemini models.

[20:06] Conor Bronsdon:
So particularly when it relates to coding tasks, and this agent enabled IDE that is anti gravity, which I'll note, you know, has some really cool capabilities. You know, folks on this show know that I'm a a Cloud Code fan, but I I really enjoyed some of the approaches anti gravity has taken around how you can actually work through examples and the engagement with Chrome and how that's been integrated.

[20:34] Conor Bronsdon:
I'd love to just understand I mean, I I know you have a great example of, like, what you've built with it lately, but how do you approach using anti gravity? And, like, what are the recommendations you have to developers who are looking to explore?

[20:47] Speaker:
Yeah. So so anti gravity is our latest agent first IDE that's been released by by some of the team who came over from from Windsurf. This is I I've got kind of one of my favorite extensions pulled up here on the right where you can see the different utilization that you have for the the different models. As you mentioned, there's Gemini Flash, Gemini Pro, and then also

[21:13] Speaker:
some of the Anthropic model family that you can use as part of your work. You can also select the different models here off to the right to incorporate into into your projects. And for this one, I was just doing kind of a a basic walkthrough where you can ask anti gravity to to build kind of a step by step workflow for for any app that you wanna stress test. So you can say, like, accomplish this task within the website and then periodically take a screenshot and and do the

[21:46] Speaker:
do the walk through for me. It will also give you the ability to to kind of comment on on any of the the plans that it makes. So just like you would comment in a Google Doc, you can say, okay. This plan shouldn't be using LiveKit. It should be using something else, or it should be using kind of not just video and audio, but also maybe some other things. And then

[22:13] Speaker:
each one of these are capable of of being kind of modified and stress tested. And like you said, you can even invoke a Chrome browser in order to accomplish many of these tasks. So if you did want to to say something like, go through and on rei.com, like, try out 10 of the of the top user journeys that you think would be most important, and then give me feedback on each of them.

[22:43] Speaker:
It would be able to do kind of like poor man's QA testing in a in a really seamless way that's directly baked into your IDE. So lots and lots of good work happening in the IDE space for antigravity. You can also incorporate the Gemini CLI, just invoke it within a terminal. And and Joules as well can be invoked via Joules So the the Jules agent, for folks who haven't seen it, I'll go ahead and pull it up on my screen.

[23:20] Speaker:
The Jules is intended to be an asynchronous agent. So if you work a lot in if you work a lot in GitHub, this is something that allows you to pull up your GitHub repo, kind of describe describe some changes that you would like to see, maybe some documentation that you would like to add. And then Jules will do the task of of baking that in for you. So you can select different repos and then just ask for for any of the any of the changes that you might make. And I'll I'll say these aren't the only dev tools Google has put out, obviously. Like, I've really enjoyed the,

[24:00] Conor Bronsdon:
the Chrome dev tools, MCP that, Google has released as well. And it's it's interesting to see how these different protocols and approaches are, you know, collaborating. Obvious obviously, like, MCP by originally by Anthropic has really blown up, I think been integrated into a lot of approaches. It's cool to see CLI just just being leveraged a lot more here and and the growth of the Gemini CLI.

[24:23] Conor Bronsdon:
And then Google is always taking this, like, very agent first approach. What are your thoughts about how you see development happening within the Google ecosystem around building with agents and and building with these new tools that are accessible to developers.

[24:39] Speaker:
Yeah. The the hardest thing is, I think, trying to navigate which ones to use and when. The the and and that's or at least that that feels just just coming at it from a user's perspective that, you know, given that there are so many different lovely things to choose from, how do you select which one for for your specific tasks? I think that if you're more of a fan of an IDE, like if you're a cursor user, if you're someone who really, really likes having that that kind of immersive support, anti gravity is absolutely

[25:11] Speaker:
the place that you you might want to go. I also really, really love RootCode and Klein, which are two open source Versus Code extensions that you can use with Gemini, and they also incorporate agents. So if you want if you want something that's a little bit more open that you can kind of hack on and tweak and change, those are also two great options. If you if you enjoy working in a terminal,

[25:38] Speaker:
so so maybe you're more of like a a Claude code enthusiast, then Gemini CLI is probably the place to go. And if you just want something that you can invoke kind of similar to maybe a GitHub action or, you know, something that you can just tag into a GitHub issue, then I I think Jules is a is a solid option. Joules is really intended to work in the background while you do other things,

[26:05] Speaker:
whereas Gemini CLI is intended to be something that you work with in real time, and so is so is the anti gravity IDE. Like, really launching, you know, launching browser windows, like, doing all of these options within the browser, and then also intended to be more of kind of like a a work partner than than maybe some of the other tools. I also quite like if you go to collab, and I'm going to

[26:38] Speaker:
to just pull up Colab really quickly. For folks who haven't seen it before, it's a Python environment where you can write code in a notebook or pull up a terminal. We've also just recently integrated Gemini CLI within Colab. You can also see the the handsome little crabs that are dressed up like Oh, I gotta love the crabs. Yes. But if you type in Gemini Gemini within the terminal,

[27:04] Speaker:
it should pull up the Gemini CLI. So you have that agent embedded within your collab instance. And then there's also the collab data science agent, which is something that you can invoke as well, either based on Gemini two dot five Flash or Gemini three Flash. And then, yep. There we go. So Gemini CLI, as soon as the, as the machine spun up, it was able to to invoke it.

[27:31] Conor Bronsdon:
And this isn't even getting into things like the Nano Banana models and image generation, which drove, you know, the these massive traffic spikes to AI Studio on their release. Absolutely. And I know you've been building with all these models and experimenting with them. I my understanding is you've done some interesting stuff recently as far as leveraging, the Google tool suite and models to

[27:57] Conor Bronsdon:
build out a a gamified NASA curriculum.

[28:00] Speaker:
Oh, heck yeah. So so I am the biggest I am the biggest NASA fan girl. Like, my my early research was all focused in space sciences. So my background was more like geophysics applied math. And so all of the all of the work that I did from a research perspective was very, like, planetary sciences related. But there's a curriculum that NASA released quite a while ago called Space Math,

[28:32] Speaker:
which is as cool as it sounds. It's basically, like, pulling in all of the different math problems that are related to the the kinds of problems that NASA scientists tackle every day. Everything from, like, coronal mass injections to or coronal or CMEs, so coronal mass ejections. Things like understanding lunar ultraviolet, understanding things like interplanetary

[29:01] Speaker:
dust, like hundreds and hundreds and hundreds and hundreds of PDF problems that students can kind of go and see. But, you know, from from a a present day perspective, you know, this is this is very dry. If you're if you're a kid who's, you know, maybe eight years old or 10 years old, like, my my first expectation is that they would download this PDF document or they would take a picture of it with their phone, and they would send it to something like Gemini app or ChatGPT or Claude or whatever. Notebook LM and have it build me a little podcast to listen to and flashcards. Yeah. Exactly. And so so I I feel like the the curriculum

[29:44] Speaker:
is much in need of a revamp. You know, all of these things are really, really interesting and are the are the kinds of things that we need folks to be folks to be prioritizing. But but I built something that that took many of these PDFs. So all of the ones kind of associated with the first missions. And I built something called the mission control training program.

[30:11] Speaker:
So it welcomes the cadets. They enter into mission control. Cadet, welcome to mission control. I'm Doctor. Thorne. We're tracking anomalous readings from the deep field. I need fresh eyes on the data. Process your missions. Accuracy is paramount. Do not trust the automated flags. Trust the math. And so, which I also aspire to do, like trust the math as always. But each one of these

[30:42] Speaker:
each one of the PDFs has gotten changed into more of an interactive assignment. The first one is kind of questions and answers, but the but the next ones are all much more interactive. So you can, like, measure things like tree rings, measure things like the number of stars per quadrants. You can sort of decrypt mission telemetry from from different planets.

[31:07] Speaker:
And as you go along and solve each one of these problems, you're collecting more and more kind of audio snippets for a story that that kind of bakes it together into into something that feels much more exciting and immersive,

[31:21] Conor Bronsdon:
as opposed to just playing with PDF documents. So talk me through the process that you used to build this. Is this something you vibe coded? Did you do some active coding as well? What what were you using?

[31:32] Speaker:
Yeah. So so the I started within AI well, I actually started with Colab. So each one of those PDF documents, I went through and I web scraped. So it got all of the the different links, downloaded the PDF files, kind of put them into a directory format. Then I went through, and in a loop, I used, like, a very hyper specific prompt to convert each one of those PDF assignments into a TypeScript file

[31:59] Speaker:
that would accompany the that would kind of accompany the app. The the app itself was built in AI Studio for the front end. So the front end was all AI Studio. The the audio snippets were all using Gemini text to speech, though the scripts were all made with with Gemini and AI Studio. And I saved each one of the the kind of text to speech files so that they would immediately load as opposed to being generated on the fly,

[32:32] Speaker:
which helps assist with the gameplay. I actually ran into so AI Studio has a couple of limitations. So if you're building apps within AI Studio, you might or might not run into these as you as you start creating. One limitation is that it's actually a drive limitation. The size of the app has to be less than 10 megabytes in size. And this is significantly higher than 10 megabytes.

[32:57] Speaker:
So so it would not have been feasible to to create within AI Studio. Again, the front end is is is okay, but the but the the kind of all of the audio files, the the the PDFs, like many of the other component parts, you you need to be using a more fully featured IDE in order to do development. So I built the first part in AI Studio, did a whole bunch of the the kind of web scraping and

[33:29] Speaker:
speech snippet generation within Colab, and then also stitched it all together using antigravity. And I open sourced all of it as well, because why not? And the the open sourced version also has a link to the website itself, which is Space Math Academy. But if there's if people are significantly interested, I would happily, like, I would happily expand it. Right now, it's just one mission,

[34:01] Speaker:
but there's definitely content for, like, up to 20 of them. So so I I think this would be very, very fun to see adopted in in k through 12 curriculum as opposed to the the things that that students have to learn from today, which is, again, you know, like PDFs,

[34:20] Conor Bronsdon:
not the most exciting thing. Yeah. I mean, I think this brings up a couple of interesting trends. So one being the opportunity to I guess, almost like gamify education where I I'm not saying that the entire educational experience needs to be a game, but it's extremely useful for teaching at times. Like, if I go back to, you know so my education where it's like, oh, typing games, for example, have been a standard for ages. Like, how do I get students to learn these to type? It's kinda boring, like, but it's, you know, very important. Great. We can we can play a game. And

[34:51] Conor Bronsdon:
it's really clear that we are enabling so much more with whether it's, you know, fully Vibe coded applications for this, some of the auto gen work being done, or for developers like yourself to to build these interesting opportunities. So education is just, like, such an interesting space here.

[35:07] Speaker:
A 100%. Like, the I I feel like I I also all of the made me made me go and find it. But my very first one of the first educational games that I played was Grade Builder Algebra one. Very and it was my favorite for the longest time. Very similar in vibes to this space math situation, in that you had Hallie, who was an AI, and then a whole bunch of, like, teachers

[35:35] Speaker:
that were teaching you math, and you had to play games along the way to learn things like the coordinate planes and how to launch rockets and have the appropriate trajectories. Also things like gizmos and gadgets and the incredible machine and all of the Sims games. So strong, strong agree. Like, this is the reason why I love math. This the learning company is absolutely

[35:57] Speaker:
like, they're doing god's work lately. Absolutely

[36:00] Conor Bronsdon:
most formative thing I could imagine. I I love to hear that. And, know, speaking of space, I think another trend that I've heard you talk about is this idea that we are gonna see AI move from, bits to atoms in the future. Yeah. And that, you know, we're gonna start seeing that already with robotics. I think there's some really exciting stuff to to kind of highlight there. And then I suspect both of us feel that there are massive opportunities of space as well. So we just love your thoughts on this move to,

[36:28] Conor Bronsdon:
I guess, physical AI.

[36:30] Speaker:
Yeah. So Gemini Gemini already runs on robots.

[36:36] Conor Bronsdon:
There's I'm sorry. It's just so cool. Well,

[36:38] Speaker:
it's We live in the future. It's wild. Heck yeah. And so so this is from from an event that we had just about a week and a half ago, but we are using something called Gemini Live to operate this robot. It's completely open source, by the way. Like, the Stanford team is so badass. They they have these things called Puppers, which are you can either build it yourself, so you can three d print all of the components.

[37:06] Speaker:
It's running a Raspberry Pi. You can also order all of the hardware components just like off the shelf from various locations. It ends up being like $2,000 if you order and kind of three d print everything yourself, or you can buy one that's already been preassembled for like $2,500. But it's running it's running Gemini for all of the vision. You can also make it use Gemini Live for all of the text to speech and robotic actions.

[37:32] Speaker:
But the pupper can do all sorts of things. So you can say like, hey, pupper, shake my hand. Hey, Pupper, follow me. Hey, Pupper, do the spider or, like, go swimming. Or, hey, Pupper, tell me a joke. Like, all of these things are are possible today just with the Gemini APIs by themselves running and invoking all of the different robotic components on device on a Raspberry Pi. Like, how cool is that thing? It it's it's really rad just to see

[38:04] Conor Bronsdon:
how much things have improved the last few years around robotics. And, I mean, this is just not even getting into some of the incredible stuff happening in, like, Shenzhen and China and and and some of the insane things happening around, like, personal robots for home use Mhmm. Like with Sunday recently announcing The US. And and it's so interesting to see how this tech and the kind of combination of AI and

[38:27] Conor Bronsdon:
hardware is coming to fruition. What do you see in the next couple years as this continues to build out since you're clearly optimistic about it? Oh, absolutely. I I think the

[38:38] Speaker:
we're working really closely with the we have a robotics trust to tester program within DeepMind for the Gemini models and for some of our open robotics models that are available on Hugging Face if folks wanna check them out. But everybody it's kind of like a who's who of who you would expect. So Enchanted Tools, Boston Dynamics, the Figure Team, like, they're all kind of testing out or using

[39:04] Speaker:
Gemini models in order to operate robotics. I think in the future, this will become increasingly important. We'll start to see many more kinds of robots built with commodity hardware. So things like these Raspberry Pi, but also, like, other options that are that are perhaps a little bit easier to get than than, you know, sometimes you might have to wait a while for Raspberry Pi Zero, whereas, like, there there are options, like you mentioned, from Shenzhen that come a little bit faster

[39:33] Speaker:
and that are a little bit cheaper. And then I I I'm also really excited about companies like Periodic Labs, which are using AI to to control robotics, but also to to have those robots do really interesting material science work. So they can design experiments, run the experiments, test out, you know, the likelihood of different component parts being successful for the creation of semiconductors.

[40:05] Speaker:
And I and I think we're going to see a lot more of that automation in the science world, going forward.

[40:12] Conor Bronsdon:
Yeah. I know, you also do some angel investing. And I'll say, like, one of my recent very, very small indolent expense was in a company called Trilo Bio, which, they're building basically automated biology labs, in a box where they can, you know, work together modularly. And I think we're gonna keep seeing this and and other opportunities in the sciences where it's like, oh, look. This is very

[40:36] Conor Bronsdon:
fine tuned work that needs to get done. It needs to be repeated quite a bit so that we can actually experiment across the board. How can we speed this up and enable it to happen async just the way developers are now able to have agents running in the background? And and these seem like clear areas of opportunity. It's it's kind of crazy seeing how rapidly this is all changing.

[40:57] Conor Bronsdon:
Yeah. Honestly, to me, it's it's pretty wild.

[40:59] Speaker:
Yeah. And I as a person who you know, their earth sciences labs are are pretty or at least at my university, they were pretty close to the biology labs. There was a lot of work happening in terms of, like, biogeochemistry and and also ecology and biology, you know, as it relates to things like carbonate geology. But it was always bonkers to me. Like, so much

[41:25] Speaker:
so much of a grad student's life was just very tedious, you know, either counting things in a petri dish or, like, counting things in a satellite image or, like, grinding by hand different rock samples before putting them into, you know, some sort of instrument in order to analyze them. And all of those things could be much more easily accomplished and delightfully accomplished by robotics.

[41:52] Speaker:
Whereas grad students, you know, obviously it's boring work, which is not the best, but it's also very, very easy for humans to make mistakes. So I think the more that AI is incorporated into the scientific process, both for designing experiments, but also executing on them, I think the better off we're all gonna be. So speaking of,

[42:15] Conor Bronsdon:
space age applications of AI and excitement there, I do have to ask about this idea of Project Suncatcher.

[42:25] Speaker:
Oh, the the

[42:27] Conor Bronsdon:
data centers in space. Yes. Yes. Yeah. Yeah. I I don't know if I'd I'd love any thoughts you have on that. Actually, I'm gonna ask you at the end of this episode, who can I talk to on the Google team to answer them, because I I would love to do deep dive? Find you I can find you people. Like and I I am personally, I I think it's a really, really interesting idea.

[42:47] Speaker:
One of the interview questions at Google, I remember, was how would you upgrade how would you upgrade Linux in a data center that was orbiting the moon or a data center that was on the moon. So so I think it it ends up leading to a lot of really, you know, just kind of strange strange and interesting and delightfully crunchy technical questions, like how do you maintain infrastructure

[43:18] Speaker:
that's so far away from the earth? How do you deal with things like, you know, we were just talking about CMEs, like, how do you make your infrastructure resilient to those kinds of challenges? But I I do think that not just Google, but many, many other companies are investing in that space, and it'll be really interesting to see what outcomes are driven by having data centers in the space.

[43:41] Conor Bronsdon:
Yeah. I mean, the build out around AI has been incredible to see the infrastructure that's been put together. And the opportunity to essentially adjust the inputs that are being put in the data centers so that you have cleaner energy fueling them, that you're hopefully making them much more efficient. It's it's really exciting. There's obviously huge challenges to make it work.

[44:06] Conor Bronsdon:
I will not pretend that I know the physics of how you deal with gamma waves in space and how that's going to impact, the data centers, the radiation side of it. I'm excited to see how it gets figured out. But folks who I have briefly talked to about it seem very optimistic. And it's also cool to see some of the simple efficiencies happening for physical data centers

[44:28] Conor Bronsdon:
in, earthbound ones, I should say. Maybe it's the the phrasing I'll start using here. But just seeing the efficiency gains in recent ones, whether it's, you know, the Microsoft, Fairview work, or I'm miss I'm misremembering the name of that one, and or the work being done with some of the recent Google set data centers to increase, like, water efficiency. I think we're we're seeing the industry just continually improve

[44:52] Conor Bronsdon:
both on the model side and then, on the technology that's actually fueling those models as well. And it gives me a lot of reason for, optimism.

[45:01] Speaker:
Yep. 100%. And speaking as a very proud Texan, like, we're we're also quickly accruing many, many data centers in Texas. So as a person who also knows just how scarce water is in the state, and then also we have a very, very interesting and bespoke electrical grid. Making sure that all of that is done responsibly is top of mind. And like, I would I would much prefer to have,

[45:32] Conor Bronsdon:
you know, data centers in space as opposed to terrestrial data centers if we can if we can finagle it. Totally. And I I will say, you know, folks who have listened in this show a bit, we know we did a recent episode with Andy Masley, breaking down some of the myths around water usage and energy usage with data centers because some of these concerns are very overblown compared to things like corn usage for biofuels, for example, which is extremely inefficient. So not to say that the industry shouldn't get more efficient. We absolutely should. I'm very optimistic about what we're doing with space as the next frontier for it. But

[46:07] Conor Bronsdon:
it is important to remind us all that, like, look, like, we can also push back on some of these narratives because they are not as accurate as they are portrayed to be, I think, in media sometimes. Not to say that we're not using some resources, and we shouldn't be paying attention to that. But, you know, there there's a lot of work to be done here. And I think some of the fears around

[46:27] Conor Bronsdon:
privacy and IP usage and folks' fears around their own jobs tends to seep into the discussion in other areas. And they think, oh, well, I feel morally opposed or I feel scared, and therefore, it must be bad in these other areas. And so it just becomes a very challenging, squishy discussion at times. And if we can start to offload some of those costs in a way where,

[46:52] Conor Bronsdon:
you know, the trade offs are are much easier to make, where it's like, oh, you know, this is we this is in space. You don't have to worry about this impacting the ground here. Like, okay. We can still talk about, like, what does it take to do a launch? That's fine. But those are much more efficient than they used to be. I think that's just better for everyone to to start doing that. And honestly, like, I I love a good non dystopian sci fi novel Yep. Future for us. So I think that's very exciting.

[47:15] Speaker:
100%. It feels like it it feels honestly like it should be part of, like, a Kim Stanley Robinson book or something similar.

[47:23] Conor Bronsdon:
Well, and speaking of your your robotics conversation that you brought up earlier too, like, this also, I think, directly aligns there. It's like, oh, like, this is also where we're gonna have to figure out how do we get robots to repair these things. Like, how do we approach this in space? So it just speaks to the excitement of the future. And, Paige, I can't thank you enough for joining me to chat through this today. It's super exciting to kind of see your vision and Google's vision. Do you have any closing thoughts as we wrap up here?

[47:49] Speaker:
Nothing other than definitely if folks are feeling overwhelmed with the number of models, the number of releases, like, you are not alone. I I think everybody, even people who are operating in the space, are are feeling similarly overwhelmed, though excited and energized and enthusiastic. And so so I think the the best path that you can that you can have for understanding the models and the capabilities and how you can build your own intuition around any

[48:19] Speaker:
performance curves is by testing out the things yourself. So test out all of the the things that we were talking about today, and then also make sure to try the models at ai.dev, which is AI Studio, and to build apps and deploy them at ai.dev/build.

[48:36] Conor Bronsdon:
I will also highly recommend anyone listening who has enjoyed, Paige's conversation here, which I think most of you probably will have, to check out Paige's website at webpage.dev. Find Paige's LinkedIn, which we'll probably linked link in the episode show notes. She's at dynamic web page on x/ Twitter and elsewhere around the Internet sharing her incredible content, talking all about the innovations Google's doing here. It's been a ton of fun having you on the show, Paige. Can't thank you enough. Thank you, and have a happy holiday season. Yes. Yes. Happy holidays, everyone. I know this is gonna be our first episode coming out of the New Year.

[49:12] Conor Bronsdon:
We're recording here the week of Christmas, so hopefully everyone has a lovely and restful time to close things out. A couple of brief housekeeping notes from my end as we wrap up here. One, a big thank you to our presenting sponsors, Galileo, for making this episode happen. Super excited about the future of season three. Two, if you are listening to this episode and you haven't watched the video of it, just you're missing a few things. High highly recommend checking our YouTube, which is linked to the show notes.

[49:40] Conor Bronsdon:
Highly recommend checking out this out on Spotify, whatever your best usage is here, but at Chain of Thought AI on YouTube. And just hope everyone who's had listened to this had a wonderful holiday season, having a great start to the year. And, Paige, thank you so much for making this happen. Yeah. Thank you. Thanks to Galileo for sponsoring this episode. Their new 165

[50:04] Conor Bronsdon:
page comprehensive guide to mastering multi agent systems is freely available on their website at galileo.ai and provides you the lens you need to understand when multi agent systems add value versus single agent approaches, how to design them efficiently, and how to build reliable systems that work in production. Download it for free at the link in the show description to discover how to continuously improve your AI agents,

[50:34] Conor Bronsdon:
identify and avoid common coordination pitfalls, master context engineering for agent collaboration, measure performance with multi agent metrics, and much more.