Gemini 3 Flash (Smartest, Cheapest AI) with Google DeepMind's Logan Kilpatrick

27 Feb 2026 · 1 h 59 min · 51 chapters

Ask about this episode

Ask anything about it. ChatGPT or Claude reads this page and answers with the times it was said.

Connect VO and ask about every podcast you hear, including the moments you saved. Add to ChatGPT · Add to Claude

In short

Podcast Notes: The Neuron - AI Explained

Episode Title

Gemini 3 Flash (Smartest, Cheapest AI) with Google DeepMind's Logan Kilpatrick

---

Episode Summary In this episode of The Neuron, hosts Grant Harvey and Corey Noles sit down with Logan Kilpatrick from Google DeepMind to discuss the launch of Gemini 3 Flash, a new AI model that outperforms its predecessor Gemini 2.5 Pro while being faster and cheaper. The discussion covers its implications for developers, knowledge workers, and the broader workforce, focusing on its capabilities through live demos, use cases, and the changing landscape of AI in business.

---

Key Topics Discussed

  1. Introduction to Gemini 3 Flash
  2. Performance Metrics:
  3. Outperforms Gemini 2.5 Pro across benchmarks.
  4. Approximately 3x faster and costs less than 1/4 of Gemini 2.5 Pro.
  5. Context window of 1M tokens.
  6. Advanced visual and spatial reasoning capabilities.
  1. Live Demos of Gemini 3 Flash
  2. Multimodal Understanding:
  3. Demonstration of coding capabilities and understanding of multimodal data.
  4. Use Cases:
  5. Opportunities for tasks previously considered impractical due to cost/complexity.
  6. Increased efficiency in task execution.
  1. Economic Implications of AI
  2. Impact on Hiring:
  3. Analysis of how the affordability and capabilities of AI tools could shift hiring dynamics.
  4. Current trends show AI is increasing wages for affected roles, but there are concerns about future job viability as AI becomes more capable and cheaper.
  5. Labor Market Dynamics:
  6. Discussion on the concept of "intelligence as a commodity" and its potential to change the value and demand for human labor.
  1. Feedback and Iteration in AI Development
  2. Importance of User Feedback:
  3. The significance of user interaction in refining AI tools.
  4. Discussion on the iterative nature of developing new AI applications and models.
  1. Predictions and Future Outlook
  2. Concerns about Job Displacement:
  3. Exploration of the potential negative impacts of AI on entry-level positions.
  4. Debate on whether the market can absorb those displaced by automation into new roles.

---

Key Takeaways

  • Gemini 3 Flash represents a significant advancement in AI technology, allowing for faster and more cost-effective solutions in various industries.
  • The economic landscape is poised for changes as AI tools become more accessible and impactful, leading to potential shifts in employment patterns.
  • Continuous learning and adaptability will be essential for professionals to thrive in an AI-augmented work environment.
  • User feedback and iterative development are crucial in shaping AI solutions that meet real-world demands and challenges.

---

Additional Insights

  • Curiosity and Lifelong Learning: Emphasis on the importance of cultivating a mindset of curiosity and adaptability in order to leverage new tools and technologies effectively.
  • Community Engagement: Acknowledgment of the audience's role in discussions and the value of diverse perspectives in shaping the understanding of AI's impact.

---

Conclusion The episode provides a deep dive into the advancements brought by Gemini 3 Flash and the implications for developers and businesses alike. As AI continues to evolve, understanding its capabilities and preparing for its economic impact will be essential for navigating the future workforce landscape.

---

For more information about the podcast, visit [The Neuron Daily](https://www.theneurondaily.com/subscribe).

Written by AI. May contain mistakes. Listen to the episode to check what was said.

Chapters

Tap a time to open that second in VO

Recap of Previous Episodes

0:46 to 1:54

Hosts recap their previous live episode and share excitement for today's guest.

“And here in just a few minutes, we'll be joined by our good friend, Logan Kilpatrick from Google DeepMind, who's done a live with us before, done an interview before.”

Introduction of Logan Kilpatrick

1:55 to 4:35

The hosts introduce Logan Kilpatrick from Google DeepMind and discuss the upcoming topics.

“But today we're going to talk about Gemini 3 Flash.”

Discussion on Gemini 3 Flash

4:36 to 9:08

Logan discusses the capabilities of Gemini 3 Flash and its advantages over previous models.

“We joked it would be nice if all the companies would just sign a 14-day truce for the next two weeks and everybody breathe.”

Implications of Cost and Efficiency

9:09 to 11:59

The hosts talk about how Gemini 3 Flash's pricing impacts its use cases and market potential.

“And it's been really interesting to watch efficiency start to play a role in those kind of builds where like, you know, not every step.”

Emerging Use Cases for Flash Models

12:00 to 14:01

The conversation explores new applications enabled by Gemini 3 Flash and the shift in user preferences.

“of pictures, most of which they never look at, like many millions of pictures and video and audio clips and all this other stuff around the internet and in your sort of like work corpus and all these other things.”

The Shift in AI Model Selection

14:01 to 15:00

Explore how user preferences are changing towards faster AI models.

“Actually, one of the interesting trends that I think we're going to see, and I saw a bunch of early examples of people talking about this online.”

Evolving Interaction with AI Models

15:01 to 18:09

Understanding how to effectively collaborate with AI models has transformed.

“engaged given that the models inherently have this latency, not problem, but characteristic about them.”

Ambitious Prompting Strategies

18:10 to 19:10

Learn about the shift to demanding more from AI models to unlock their potential.

“That's a strategy we lean on really heavily when working on shows and newsletters and websites and all of the things.”

Building More with Faster Models

19:11 to 23:20

Discover how faster AI models inspire more creativity and productivity.

“And that when a new model like a Gemini 3 Flash comes out, is that the first instinct of like the AI Twittersphere, for example, is like, okay, how do we break it?”

AI's Economic Impact on Software Development

23:21 to 26:50

Discussing how AI changes the landscape of software development and job roles.

“bit of time and we'll probably see this, but it is, it's interesting to see that trend play out.”
Show all 51 chapters

The Future of Affordable AI Solutions

26:51 to 28:00

Examine how cost-effective AI models can revolutionize production processes.

Exploring AI Cost Efficiency

28:00 to 28:50

Learn about the cost efficiency of AI models and their impact on production.

“fun thing about us having this stuff and putting it in an API is that like we get this feedback from developers all the time.”

Scriptwriting with AI

28:50 to 30:40

Discover how AI can assist in generating screenplay elements.

“And I've also tested that on AI Studio as well to try it out.”

Advancements in AI Capabilities

30:40 to 33:15

Understand the significant improvements in AI models over the past year.

“I'm actually curious to see how well, how close it gets.”

Real-Time AI Applications

33:15 to 35:00

Examine the implementation of real-time features in AI applications.

“So we'll see how quickly this one makes it happen, but it'll sort of calm us down at the end.”

Text-to-Speech Innovations

35:00 to 37:30

Learn about the latest upgrades in text-to-speech technology and its functionalities.

“these days and I'm like it knows what I mean.”

The Future of Speech Interactions

37:30 to 41:20

Discuss the future of speech-to-speech interactions and their potential applications.

“Your speech to speech seems to have taken a decent leap as well in recent months.”

Key Advice for Using Gemini 3 Flash

41:20 to 42:04

Hear key insights on how to effectively utilize Gemini 3 Flash models.

“Logan, before we let you go, I know you've got other appointments and a busy day ahead of you and hopefully some time off in the near future.”

Maximizing Potential with Gemini 3 Flash

42:04 to 42:52

Learn how to leverage the Gemini 3 Flash model for enhanced user engagement.

Meditation App Discussion

42:52 to 44:36

Hear about the potential of a meditation app and holiday plans from Logan.

“I'm going to ship a meditation app over the holidays.”

Education Access to AI Products

44:36 to 46:04

Discussion on the availability of Gemini 3 Flash for educational institutions.

“Hope you and yours have a good holiday season.”

Introduction to Cat Doom Benchmark

46:04 to 46:51

Exploring the Cat Doom benchmark as a testing ground for AI capabilities.

“And if you have any questions or any requests, please share them in the chat.”

Testing Gemini 3 Flash with Cat Doom

46:51 to 48:20

Watch the demonstration of Cat Doom using the Gemini 3 Flash model.

“alright so here I have a very basic version of my prompt which is our benchmark here at the Neuron let's make a version of the game Doom starring cats I am using Gemini 3 Flash.”

Feedback on Cat Doom Gameplay

48:20 to 49:59

Discussion on user experience and feedback for the Cat Doom game iteration.

“there's kind of some scary cats in here.”

Enhancing Gameplay with AI Insights

49:59 to 51:51

Explore the improvements in gameplay and user instructions within the game.

“I'm showing Cat Doom, our benchmark here for testing games in AI Studio.”

Creative Gameplay Concepts

51:51 to 56:00

Discussion on creating engaging gameplay themes and character dynamics.

“I love that you have lives, a score, health.”

Exploring Omnidirectional Control Features

56:00 to 57:50

Learn about control tips for an AI-driven game design experience.

“And if you look in any reflective surfaces, you should see yourself as a mouse.”

Iterative Feedback in Game Development

57:50 to 1:01:40

Understand how iterative feedback improves game mechanics and user experience.

“And you're doing this all with Flash, right?”

Conceptualizing an AI Screenwriting Editor

1:01:40 to 1:05:20

Discover the potential of an AI-powered screenwriting tool for real-time storyboarding.

“Yeah, if I wanted to be very precise, which obviously if I was creating a game that I wanted to iterate on and build on, I would definitely be more prescriptive with how I wanted to do these things.”

Building a Visual Storytelling Application

1:05:20 to 1:10:01

Learn about creating an application combining screenplay writing with visual planning.

“So for people who either missed when we were talking to Logan, or didn't quite understand what I was asking for.”

Creating Engaging Video Narratives

1:10:01 to 1:11:19

Learn about the process of generating videos and the importance of planning.

“But that way you're not frivolously generating video clips as you're writing.”

Storyboard Techniques for Visual Scenes

1:11:20 to 1:13:01

Discover how to visualize and storyboard scenes effectively.

“silence is heavy, broken only by the rhythmic thump, thump, thump of pause.”

Tiny Wins: A Habit Tracking App

1:13:02 to 1:15:00

Explore the features of a new app designed to help track and build habits.

“I wonder if I hit visualize if it'll run.”

Enhancing User Experience in Habit Tracking

1:15:01 to 1:18:19

Learn about user feedback and suggestions for improving habit tracking apps.

“And for that, we'll use this little runner dude here.”

Discussing App Functionality and Future Potential

1:18:20 to 1:23:48

Engage in a discussion about app capabilities and potential mobile conversion.

“For example, if a task requires 30 minutes, they should be able to mark that they did do 20 or for eight glasses of water to show that they drank six.”

Exploring AI Project Ideas

1:24:00 to 1:25:00

Discussion on personal projects and challenges in AI integration.

“If anyone has any questions, I mean, I think at this point, we're kind of just playing around.”

Visualizing Storytelling with AI

1:25:00 to 1:26:40

Review of different AI tools for visualizing narrative scripts.

“And I'm trying to come up with a, I say trying to come up with, I have a list of questions to ask it in terms of investigating like my code base and determining, you know, do I have any visible hard-coded API keys?”

User Experience in AI Tools

1:26:40 to 1:28:20

Comparison of user interfaces in different AI applications.

“Please like, and subscribe if you haven't yet.”

Gemini's AI Capabilities

1:28:20 to 1:31:00

Discussion on the advancements and features of Gemini AI.

“I've never done anything with an app and I'd like to know how.”

AI's Impact on Job Markets

1:31:00 to 1:36:40

Analyzing the effects of AI on job roles and wage dynamics.

“conversation that i started with when we covered gemini 3 flash earlier this week which is the idea that you know one we're we're at the pareto frontier which i'm i don't think we really talked about that with Logan.”

The Lump of Labor Fallacy Explained

1:36:40 to 1:38:00

Exploration of labor demand shifts due to AI innovations.

“So there was a lot of optimism in there.”

Understanding the Lump of Labor Fallacy

1:38:00 to 1:39:14

Learn about the misconception that AI reduces job opportunities by examining the lump of labor fallacy.

“as many entry-level engineers as there are today?”

Reframing Job Roles in an AI World

1:39:14 to 1:41:40

Explore the need to redefine job roles beyond specific tasks to adapt to the changing job landscape.

“I don't know what the right framing is, but It's something that's more like, I am a marketer.”

The Individual Impact of Job Market Shifts

1:41:40 to 1:45:45

Discuss the personal implications of job market changes due to automation and AI, highlighting individual struggles.

“If your identity is tied to that task that now can be automated, then I think it's harder for you to switch to a new skill set.”

Preparing for the Future of Work

1:45:45 to 1:47:29

Learn strategies to prepare mentally and emotionally for changes in the job market brought on by AI.

“prepare ourselves for this with whatever quote unquote tasks that we do now?”

The Importance of Curiosity and Adaptability

1:47:29 to 1:50:15

Understand the significance of curiosity and adaptability in facing the rapid changes in industries due to AI.

“Some people read, some people need to be hands-on, some people like to listen to a lecturer.”

Becoming an AI Power User

1:50:15 to 1:52:00

Discover the qualities and skills needed to thrive in an AI-enhanced work environment.

“I think a lot of people feel a little hopeless right now that AI is going to replace everyone in two years.”

The Importance of Learning with AI Tools

1:52:00 to 1:53:25

Discover how to effectively learn and use AI tools to enhance your career.

“And I have a better understanding of why it matters.”

Sharing Unique Learning Prompts

1:53:25 to 1:54:17

Explore an interesting Richard Feynman prompt that aids in learning.

“I think it's over 3 % wage increase if you are someone in an AI-exposed field and you know how to use the AI tools that companies want their employees to use.”

Predictions and Pessimism for the Future

1:54:17 to 1:55:04

Learn about the contrasting views on the future of AI and societal impacts.

“And I was like, this way, when I want to learn something, I can go in and have it taught to me like Richard Feynman.”

Navigating Concerns and Optimism

1:55:04 to 1:56:51

Understand the importance of balancing concerns with optimism in AI discussions.

“And I want to address that and talk about that realistically.”
Hear the part that matters, and keep it.Open this episode in VO. Double tap your headphones to save a moment as you listen.
Get VO free

Transcript

Automatic transcript. May contain errors.

0:06I sent it to him. Okay, cool. I don't see why he won't be on. How many people do we have waiting? Let's see. 34. Yeah, let me send it. All right. Well, hello, everyone. Looks like we're getting up and rolling, just kind of checking some quick logistics, making sure everything's set up good to go. uh welcome to the neuron live we're so great to have you here um these are always a ton of fun and uh we had one yesterday do what i don't think we're live are we oh it didn't click the youtube button yeah hey what's up everybody cory here from the neuron with grant this morning good to see you all we thought we were live about a minute ago and just sat here talking away like we knew what we were doing but hey you live you learn um today was kind of a surprise thing we were able to put together at the last minute and uh thanks to everyone who joined us yesterday for the great live with versell that was an absolute ton of fun and uh learned a lot and hope you all enjoyed it as well and today we're going to shift gears a little bit.

1:25And here in just a few minutes, we'll be joined by our good friend, Logan Kilpatrick from Google DeepMind, who's done a live with us before, done an interview before. And every time he comes, people seem to really enjoy it. So we're excited to get him back today to talk about, excuse me, to talk live. Sometimes you burp. I didn't even hear it. You could have gone away with that. I could have just pretended it didn't happen. But today we're going to talk about Gemini 3 Flash. What do you think, Grant? Do you have any thoughts? Yeah, it's really awesome so far from my testing. I actually have a project I'm working on right now spinning up.

2:10Cool, cool, cool. So I'm excited for that. I've been tinkering with it a little bit too in between. It's really funny this week. We were just joking about what an interesting time in business it is that right now we're essentially watching Google and OpenAI drop competing products in back-to-back days all week. There's like something new and cool from everywhere every day right now. I hope it lets up a little bit next week with the holiday. Or we're not going to have holidays. We have evergreen content ready to go because there's so much stuff that happens in the AI space. But everyone should take a break, man.

2:53No kidding. Y 'all go have some Christmas. Let's all breathe. We promise we'll come back. Yeah. We do have some videos that will be dropping as well. If you're alive for me, I can't remember who. I'll have to look at a calendar. Let's see who's coming out on Tuesday while we're waiting for a few here. from albert invent nick talking from albert invent that was a good conversation uh and then on the 30th we have karina hong of axiom math coming out who's really interesting they're essentially building a mathematical super intelligence uh is the goal and uh and they've done some things made some headway and it's it's really interesting she is um she's absolutely brilliant she's like 24 years old, runs Axiom Math, founded the company, extremely intelligent.

3:47It's a good conversation. A lot of times, you know, when you talk to folks like that, it's a little intimidating because, you know, you're like, wow, I am so dumb. But, you know, it's a good one. And I think on that note, it looks like our guest is here. I'm going to bring on Mr. Kilpatrick. Logan, how's it going, man? Good to see you. Hello. Happy Friday. Apologies for being a few minutes late. I was just jamming with the team about a bunch of cool new stuff coming in February. So the targets are already set for next year, which is exciting. There's always something new coming, isn't there? I'm relieved because for a second I thought you were going to say next week.

4:29Not next week. Next week. Folks are rightfully taking time off. So it'll be nice for people to relax and have a moment of calm before the storm reemerges. We joked it would be nice if all the companies would just sign a 14-day truce for the next two weeks and everybody breathe. I think you'll roughly get that. That's my guess. I'm sure someone will try to make noise while other people are not. But I think mostly people are – everyone's off, which is the nice part about the end of December, one of the few nice parts. It is. It's like the guaranteed mellow at the end of the year usually and good chance to reset and start it all over.

5:07Yeah, for sure. Well, Logan, you guys had some more big news this week with Gemini 3 Flash. It's fantastic. Thank you. Yeah, the model's incredible. I feel like it's a testament to the work that's been happening on research. It's a testament to what happens when you have great frontier capable models with Pro as well, sort of like being the model that then distills down into Flash. um i think a lot of the i've seen a bunch of comments online of folks being like this this model sort of like defies the paradigm of what a flash model actually has historically meant which i think is actually a good reasonable comment just because flash used to mean like small and fast and now it's like actually the flash model is um in some cases like more capable than the pro model which is really really cool so that's wild that's right that was actually the surprising thing for me especially uh i believe in sweet bench and agentic coding tasks right that that flash is actually better than pro yeah and i mean part of this there's like a very a very simple answer to this which is just given um like one of the ingredients of making great models is time and the pro model was done you know probably like i don't know like two months ago now at this point from like a training perspective so within that last like month and a half two months whatever it was month there's been like post-training innovation and a bunch of that like new recipe and stuff like that made it into the flash model that didn't make it into the pro model so a bunch of that stuff will go into the next version of the pro model but it just like from a time sequencing standpoint like didn't end up making it in so there's no like uh yeah i mean obviously in the distillation worked really well from pro to flash as well but um you'll you'll see those gains also trickle back to pro which is great you know working with some of these small models even small open source models that are out it continues to impress me how much better they're all getting and i i wonder i've often wondered like like what's what's the deal are we looking at like you know more curated data that is is cleaner unless we've scooped up the internet or something yeah good question i I mean, I think some of this is, you see this on the whole stack, like when you distill, there's like a bunch of innovation and distillation when you like take large models and sort of extract a bunch of the like core intelligence and put it into small models.

7:36And like historically that was like a more lossy process, which is interesting. And I feel like it's gotten like much better at like actually making sure a bunch of those capabilities transfer were over. So I think that's part of the story is just like innovation. I think there's definitely like a data curation aspect. I don't think that's actually unique to the flash models for what it's worth. I think that's happening like across everywhere, which is like, how do we do more with less data? How do we find the data that's actually like really useful and amplify it? How do you, you know, all that, that whole story I think is playing out.

8:13I also think there's, and uh i talked to aureole and jeff and noam yesterday and i think they made the comment which is which is 100 true which is also you can just do more iteration with the flash models just because they're smaller so like you get more shot versus like you can't do as much research experimentation with pro because it's big and actually if you look back to when we did our first reasoning model um to flash to flash was our first reasoning model back in i think december 2024 and it was explicitly a flash model at that time with reasoning capabilities because the team wanted to be able to get a bunch of shots on goal very, very quickly and sort of scale up from that perspective.

8:53So they figured out a bunch of stuff that worked really well, brought it over to Pro and then sort of have used that now as the sort of baseline for this stuff. So it is nice to have small models. I agree. And there's so much, so many applications for them, especially as we're, We're becoming more agentic in business and in life, frankly, in some ways. And it's been really interesting to watch efficiency start to play a role in those kind of builds where like, you know, not every step. I don't need a frontier model to classify a thing, maybe. I need the frontier model to analyze the thing so we can tick through and really dial in where you're spending your money and your time.

9:36A hundred percent. It is interesting. And I think as you, something that's fascinating, you actually see this happening. And we, I had this conversation with our customers all the time, as you sort of go up the capability curve, um, there's like a floor at every step as you go up the capability curve. And there's a whole, uh, there's a whole suite of applications that like actually just work at that floor level. Um, they don't need the next level of intelligence. They actually just need like that current level of intelligence to work well at that cost price point. And I think that's actually what's happened with Flash, like continually, like actually 2.0 Flash still has like a massive amount of users and a massive amount of like token consumption and all that stuff.

10:192.5 Flash even more so. So I think you'll see that with 3Flash as well, where like, you know, it just it's it works. It's the right model at the right price point with the right level of intelligence. And there'll be a lot of customers who are just like happy to be there and keep their keep their product experience going. um so it'll be interesting to see how that like you sort of you see this in the mic maybe not micro right now but like play that out over like three years and i'm curious how that trend continues of like whose products require frontier versus not uh yeah it's interesting yeah actually a lot of my questions had to do with this idea of it being you know fundamentally you know cheaper faster, outperforming 2.5 Pro, one fourth the price.

11:08The big question for me is what use cases does this enable that perhaps we weren't able to do before or weren't affordable to do before, but now it's a lot cheaper to do things that are a lot more complicated? What kind of use cases do you think that this enables? Yeah, I think agents is definitely one of them. I think actually running agents in production and folks have like tried to do this with, with like frontier intelligence models is actually really expensive. You have to be like careful not to, not to sort of shoot yourself in the proverbial foot. Um, so I'm excited for that just cause like the models are great at tool calling.

11:45They sort of have good general intelligence can do a bunch of this stuff. Multimodal I think continues to be just like a theme and a trend that works really well for Gemini. Um, it's also something that inherently has lots of scale by default. If you, for example, like everyone has million, you know, maybe not millions of pictures, everyone has many thousands of pictures, most of which they never look at, like many millions of pictures and video and audio clips and all this other stuff around the internet and in your sort of like work corpus and all these other things. So I think those use cases, like historically, it's actually where we saw a lot of initial flash traction for previous iterations of flash was like, how do I understand all this multimodal data that my company has or that I personally have, and then actually have a system that I can take action or make use of that data where applicable.

12:34So I think this is a continuation of that story. And I think coding, both of you mentioned this already, but I think coding is one of these use cases where it really is, you can sort of feel how good this model is. It's better on Sweebench than, which the caveat of that story is Sweebench measures a very, very specific type of coding. If you've ever looked under the hood of what Sweebench is doing, it maybe doesn't translate towards what my personal coding workflows are or what other people's might be. But you see this Revive coding, which I'm really excited about. I tweeted the other day, I think 3Flash is the default model for Revive coding because historically the cost was actually a real limiter.

13:18If you look at a lot of products, it was like, hey, you can use it once and then you have to pay because it's expensive to have like great frontier coding models. And I feel like Flash is challenging that idea, which is exciting because it means more people are going to be able to build stuff. Yeah. Yeah. And stuff that's good. Yeah. Stuff that works. Stuff that they're excited about, which I think is like a special moment. It is. You might have touched on this just a little bit, but as far as like the knowledge worker side of using Flash? You mentioned multimodal, but is there any other use cases that come to mind or things that you've seen in the first 24 hours since it's launched that's, wow, that's actually a great use case of this?

14:01Yeah, that's a good question. Actually, one of the interesting trends that I think we're going to see, and I saw a bunch of early examples of people talking about this online. One of them was I saw somebody make the comment that they have historically always opted for the frontier intelligence model as their sort of daily driver model. So whatever the most intelligent model was, was the one that they would go with. And they were sort of willing to pay the sort of latency hit on that because they wanted frontier intelligence. And they said that Flash was the first model that has made them change that perspective.

14:35Because really being able to sort of have a model that's smart enough that you can sort of quickly, that like iteration speed actually matters a lot. And like, it keeps you in flow state and makes it so that you're not, you know, having to like pull in and out or get distracted and go check your phone while the model is doing something or do some other random task and lose track of it. So I think we're going to see that, which is really interesting. And I think about this a lot for AI products, but like, how do you keep people engaged given that the models inherently have this latency, not problem, but characteristic about them.

15:15So I'm excited. I'm curious to see how that trend goes. And I think the fun thing is Flash can be your daily driver model and it's free and it's available everywhere and all that stuff, which is really cool. So you don't have to be paying$2 ,000 a month or$200 a month for one of these like pro subscription products you can just like use the default experience across many products in in the industry and ecosystem and and sort of have that frontier intelligence which is really really cool you know as as models have gotten better like across the industry and uh the bottom end of what isn't impressive now is way higher than it used to be like like like the base level it just keeps rising and and we're reaching this point where i think that is where most normal tasks are is in that range that's it's really dialed in now too and uh it's funny because like you know before we come on we're always sitting here and we're throwing prompts at whatever model we're working on and our questions are feeling a little more lame than they used to and a little more dated.

16:27It's almost more about looking at how it answers the question and its behavior for me with some of the simple questions more than it is about can it do it. And it's trickier to find questions that stretch it and that we understand as people. Yeah, this is a good question. I made this comment when we launched 3 Pro, which is, I think these models are now sort of turning the corner of how you have to think about working with AI systems. And I think historically, you would sort of, you know, how can I make the most perfectly crafted thing, this whole notion of prompt engineering, I need the model to do this very specific thing, I have to sort of artfully explain what I need, all this stuff.

17:16And like, if you sort of mess up a couple of words here and there. Maybe it won't do the thing that you needed to do. Um, and you would also like principally give it like the smallest bite-sized chunk possible. Like do this one very specific thing. Don't do anything else, please. Um, because if it doesn't, it goes on lists and things and very, yeah. And now, and I think that's changed. Like now I'm like, um, and I had this conversation with Sundar and we were talking about like an example of a product that he, he wanted. He was like, okay, well, I'll just give you the five things. I really have these 30 things that I want, but I'll tell you five for the sake of this example.

17:55And I was like, no, you got to tell me the 30 because actually that's how capable these models are. And it's a corner to turn because I think we've all gotten so used to like, don't ask for too much because the model's not going to be able to do it. And I feel like now I've gotten myself back into the habit of being ambitious with the model and getting it to do more and actually abdicating some of this like state space search directly to the model i'm like hey just like come up with 10 new things uh you know do those ideas yourself like be inspired and come up with some of this stuff and then like of course you know come show me what you did later eight out of ten are are mediocre um at best but like maybe the one in ten is like actually pretty solid and i've gone through that continually now and it continues to work.

18:41That's a strategy we lean on really heavily when working on shows and newsletters and websites and all of the things. It's like, give me a bunch of options. I don't ask for one of anything anymore. It's very like, if I'm trying to come up with a headline and I'm just stuck, I'll be like, give me 25. Or 50. Or 100. Or 100. It really doesn't matter. It will give you however many you want. That's the thing. And it's fascinating how that works. And that when a new model like a Gemini 3 Flash comes out, is that the first instinct of like the AI Twittersphere, for example, is like, okay, how do we break it?

19:26And it's very much like throwing the hardest questions they can come at and seeing like, all right, where's it going to snap? Where's it going to snap? And it's very different from how I think people interact with larger models. And I like that it's just there. It's the easy button. It's the default. If I want more, I'll ask for more. I can click the button, go down, hit pro, and I'm good. I think that's a good strategy. Yeah. So do we want to pull up and check it out? Would you be game to do that? Yeah, let's do it. Let's do it. Let me pull up a AI studio window and we can build something really quick and we can sort of see and I'll, maybe we can actually do a side by side of pro and flash.

20:16I actually did this earlier. I was deciding whether or not I should show you my version. I would love to see. Okay, we'll do that at the end. I have an idea for a project, but if you want to, why don't we ask the audience I'll ask the chat real quick to see if anyone has a better idea. You good for me to bring that on up, Logan? You need me to wait a minute? No, no, I think I'm good. Well, I'll leave it on this pitch. I always like to make sure I'm not like blasting company secrets up on a screen or something before I ask, before I click the button. No secrets. We're all good. okay so i'm gonna go ahead and uh share this prompt that i sent yesterday which is basically just a a question i was asking like on on twitter basically and i was like why doesn't this exist and then i sent this to um gemini and maybe we can translate it but i'll go ahead and send it in the private chat here to you if you want to throw this in the window.

21:20But basically what I wanted to create is where is the video edit model or video generator that works like a script editor? So for example, you essentially have like a, you know, like a final draft UI where you type into it and each sentence is basically a shot. And every time you write a sentence and you hit like period uh it will then like basically spin off a uh an image that is like the frame of that and then you can be creating a bunch of frames and like shot listing storyboarding your script at essentially like the same time you're writing it

22:02i like cool i'm gonna kick this off i just took your raw prompt yeah so we can see how good it is and i'm actually going to kick off a version with uh with pro as well nice you know what i think it got cut off halfway so i will send you the other half right now the other half oh yeah i see this i see this uh more on this first time you introduce a character sets the laura character consistency based on your description yeah yeah send me the full the second half and i'll we'll do that as the second turn um cool we'll see how good flash is at it cool cool let's see we've but we're at 30 seconds.

22:38It seems like the model is making great progress. It's already building the app.tsx file, which would be great. Anything unique or surprising you've stumbled across in AI Studio using it yet compared to 3 Pro? Yeah, I think the biggest piece is what happens when you give people a faster model. We were running a bunch of these tests. What happens when you give people a similarly capable model that's actually a lot faster? you get people build more stuff, which has been really interesting. So I think that's like, people build more stuff. They're like more sticky. They want to build more because the sort of like experience is real time and iterative.

23:20And I think we're like, obviously flash still takes a little bit of time and we'll probably see this, but it is, it's interesting to see that trend play out. And I think that's actually like a story that's like similar to Google as well. Like, you know, the reason the google search page is so fast is because like they've seen this data before like people are inclined to look up information if it's for if it's frictionless just like people are inclined to build stuff if it's frictionless which is which is really cool and exciting so easier it is the more they'll do yeah uh i i actually was reading a reddit thread last night that i wanted to get your opinion on because i think that this is a an interesting idea that flash kind of enables somebody was doing the classic you know oh software engineers we're all going to lose our jobs because of ai and somebody had a reaction to that where they basically said what ai is really doing is changing how much we can build in a given time period so any company with creativity and vision has a massive backlog of features bugs and automations but if you think about prioritization through rice which is reach impact confidence effort AI significantly reduces the effort, making things that weren't worth doing before all of a sudden feasible.

24:31And I was wondering, one, what your thought is on that. And number two, if you think Gemini 3 Flash could possibly have that same impact on Google internally. Yeah, it's a good question. I mean, my sort of broad take on this is, I think as you keep seeing these capabilities increase, it really does increase the floor for everyone. And that's what continues to get me excited about this AI moment is, you know, for all the talk of potential downside and negative impacts of AI on society. And I think there's lots of interesting and good conversations to be had about that. You look at the, as just one example of this, you look at the creation of software, which is one of the most economically transformative actions that you could possibly take in this world right now.

25:21and to think about the fact that hundreds of millions of more people have access to be able to do that and to bring their ideas to life and to bring like the business that they want or the product that they want to life um without having to like spend you know what what was the barrier to entry five years ago or three years ago or even a year ago uh was like still still so significantly different yeah and at the same time you know it's expensive to build software yeah i think it was like$250 ,000 minimum for like a new product that you're going to make something like that. Yeah. 100%. It's crazy.

25:55And I think like at the same time that more and more people at the same time that the floor is going up, I actually also still think that like the value of being someone who deeply understands software is important because like your expertise is in, there's like a, the market is actually much larger. Like the total volume of software being created is like is on an exponential. The number of people who like deeply understand what's happening in software is not on that same exponential. Yes, you have access to AI tools, but like, I think the total addressable market for like really talented, deeply, deeply engaged engineers who understand how to use AI tools and how to build software, I think is going to be like a incredibly good market.

26:39So I'm excited for people who like build software every day and are using these tools and want to spend their time doing that because I think the coding era is going to, is going to increase the opportunity space as yeah, as there's more, there's more cool things to build. Absolutely. And then do you find that, you know, with flash three being so cheap that you're having new conversations inside Google about things that you could build or put in production that maybe weren't feasible to do before, or is it too early for that? No, no, a hundred percent. I think this is the, um, I think we'll, and again, just to belabor the vibe coding point, I think it's a great example because like again like actually like it takes a lot of tokens to build software it's like on the order of like i think our average generation is like 50 to 75 000 tokens or something like that which is like that's a reasonable amount of tokens could be on the order of like a few dollars per generation every time you hit enter you know every time yeah every time you do it and like again we're in a studio providing this like vibe coding experience to people for free which is really cool but like at a certain point like you know you can't always provide something for free or maybe we'll be able to with flash.

27:44But like that is literally the conversation is like maybe with these flash models and maybe as models get better at coding, like that code generation piece actually will just be free because it's like so it's so economically possible. And I think you see this in a lot of developer use cases, actually, where like I think the fun thing about us having this stuff and putting it in an API is that like we get this feedback from developers all the time. was like my use case wasn't possible before. I needed a model that was like$5 per million input tokens and$20 out. And now I have that capability with three flash, 50 cents in,$3 out.

28:22And like my use case that was 10X too expensive to put in production is not possible. So you see with like how interesting the like cost intelligence latency trade-off is for Flash, which is really exciting. um are you uh are you staying till uh 9 30 or do or that's a good question actually go ahead i can remember if it was 9 30 or 10 i can stay for i can stay for a few more minutes but i do i have to run to another meeting so let's let's finish vibe coding this yeah let's test this out yeah i only put in so the prompt i put in originally was um like the first half of my message i set the other two halves of a year model where you can prompt like an actual screenplay inter-ascendence into a unique shot you establish scene heading at the top character speaking indicate shot uh you should be able to write a whole page as a single prompt and get a whole page with all the shots um so i wonder what is needed to actually make this happen parse script interesting oh we have things yeah that's generate all shots i don't know how easy this is actually going to worry this might actually take a minute to um generate everything and i go in to do something like this and then i and then after it generates i look at it and i'm like oh this is a bigger project than i thought i was getting into yeah yeah that'd be a copy in your other prompt every time a character talks the model knows what the character looks like etc etc i by the way i I wrote a 40 ,000 token version of this last night to get all of my ideas into it.

30:00And I've also tested that on AI Studio as well to try it out. I love that. What is it? What did it end up looking like? It looks good. The problem is I was doing it on my work account and I didn't have my API fired up. But I have my personal account here and I did two different versions so I can show you at the end. Oh, we're getting somewhere. the eiffel tower sparkles amongst a deep navy sky let's see there you go how cool yeah i don't know how this this kind of uh script is pretty lame right now so we'll see uh there's not a lot of like i don't know what the story is here the continuity i guess it's the artist and artist paints on a canvas near the seine river artist it's never quite right He dips his brush into the crimson paint close up.

30:54First touch of the canvas. Interesting. I'm actually curious to see how well, how close it gets. I'm intrigued. I feel like it's. It's going through them one at one after another, isn't it? Yeah, yeah, it is. It's doing it sequentially, which is interesting. And I don't know. Is it doing it with VO 3.1? Yeah, I think it is. Yeah. If we look at, I can look at the code really quick just to see. Cause we could even tell it to do. Yeah. maybe the via flash. It is. It's already using fast. Okay. Fast generate. Nice, nice. Oh yeah, there it is. And then I forgot, you know what happens? And this is a horrible bug that I'll get fixed.

31:34But when you click code and you come back to the preview, it automatically refluxes, which is terrible. So we will, I'll get it fixed. But this is basically what I was going for is, you know, every time you write a sentence, you can then see what that sentence looks like. And that's not possible without everything that you've built here. Yeah. So to very easily put it together, to have it easily, you know, without me having to set up all the API keys to VO and Imogen and all this stuff, or Nano Banana now, it just does it for you, which is the coolest thing about AI Studio, I think. It is. Yeah.

32:10And this was the version with Pro. So like actually a really similar sort of like UI experience. There's like a text editor over here that I can like change. And there's like some UI visualization. I think the pro one like looks maybe like slightly better. There's like a way that you can like put in images and actually bring them into the videos themselves, which is pretty cool. So like generally the same vibe from these two generations. And it looks like the pro one took 164 seconds and the flash one took 124 seconds. But it's because it ran into an air the first time that it was generating. So still that I'll be honest that a flash model can do it is, is impressive in itself.

32:57You know, especially at the level that it works in like build where it's very much like, we're going to build out all the code. You're going to see your preview right now. This is going to happen. We're going to ask questions back and forth. And that it a year ago, that wasn't a thing a small model was going to do. like it's crazy how far we've come over the last 12 months you know we tend to look at it very much through this microscope and always think god it feels like it's been forever but we were on artificial analysis visiting with them a couple of weeks ago and and one of the things we looked at was like well look look how far back like the first models of this year were you know it's it's actually a pretty big jump it is crazy yeah yeah and the artificial analysis folks are crushing into i just saw them last week as well uh in san francisco and um it is cool to see there's just like so much depth in understanding the model capabilities it's cool to see i i like their i like all the like depth that they have in in their website to understand models it's it's super cool it yeah they've just grown it and and it's just gotten better and better every time i go that website there's more stuff to help you understand yeah um close us out with a guided meditation app to generate sessions with unique visuals and voiceover.

Read the full transcript

34:12So we'll see how quickly this one makes it happen, but it'll sort of calm us down at the end. Well, this is a great example, like 64 seconds, fully working app. Let's see. That's awesome. Select your API key. Yeah, it's a banana, which is, what do I want? how can you relax right now forest chilling with squirrels there we go can't spell the word squirrels but it's okay it doesn't matter anymore spelling will soon be a lost art unfortunately we're always joking about the fact that you know what I used to put a lot of time and effort into my props and now I just prompts and I just throw slop in a window for normal stuff these days and I'm like it knows what I mean.

35:03Yeah, it is true. I think there's something interesting about the refinement piece. You trade off how much you have to iterate versus how much you have to do the work up front to come up with some of these ideas. So I feel like that's where I am at these days where I'm like, if I'm going to be annoyed by spending the time doing 10 iterations to make sure I will put in the work up front. But there's a lot of cases where I'm like, maybe I won't do the iteration and it's worthwhile to just take a shot at it.

35:33Logan Kilpatrick:So yeah. Yeah. I guess it's pretty much a task by task thing. Do you find that you're iterating mostly with Flash now? Yeah. I mean, it works really well for iteration. I think our code agent right now basically like keeps everything in context, which makes it like, it's like, it's a kind of greedy from that perspective. The results are pretty reasonably good, But given like pro is definitely a little bit better when you keep everything in context. So it just depends on the app when I'm, whether I'm using flash or pro, like how complex is it? But this is a great example of like, I feel like this is almost exactly what pro would have come up with as well.

36:11Yeah. And it works, it works pretty good. So this is good. Let's try it. Oh, it's a party. Find a comfortable position. that okay to be honest whispering like that just like nails on a chalkboard to me but you can give it the feedback you can say like please don't make it whispering make it a calm peaceful voice um you know what though it was gorgeous it was a very very chill vibe uh right right up and right up until she started talking yeah we can't change the problem it said speak in a calm whisper uh which actually is a great example of like how good the tts model is our text-to-speak the fact that it was able to like whisper like that um i just personally don't want the the whisper so is that a relatively recent tts model or or this is the new one so we actually just upgraded this model i think like a week or two ago or something like that so it is um it's not yet on the three model the three model family but it's like i'm pretty sure it's actually state of the art the TTS model is right now, which is really cool.

37:17So you can try it in ASU. You can build experiences with it. And it has a lot of the upgrade was around like control and steerability and the audio fidelity, which is super cool to see. Awesome. Your speech to speech seems to have taken a decent leap as well in recent months. More to come on speech to speech too. We just rolled out last week or the week before the real-time translation capability as well, which is based on the same speech-to-speech stuff. So we'll have that for developers early next year, which is really exciting. So you'll be able to build real-time speech-to-speech apps that translate between different languages.

37:57So I think that would be super cool to see people come up with new ideas there. I think that's an area a lot of people sleep on, is that that's such a natural way to interact while you're working, like you're sitting there with someone helping you. And I think speech-to-speech has a lot of room for implementation, especially with people who really haven't spent time with AI yet. Agreed. Yeah, yeah. I think it's a natural, we're betting a whole lot on audio as this sort of like next form factor i think like audio and code together going into 2026 like there's something interesting and magical there um so i'm i'm really excited to see what what ends up happening um there are cases where like i just want to type and like i don't want to speak because i have been speaking too much or i'm in an environment where talking out loud would be weird i feel that yeah you just want to have the option to do whichever one is most convenient in that moment Yeah.

39:00I think you'll, you'll sort of see this ubiquitously across all the products. Like we actually have this in AI studio sort of to a certain extent already. We have this like yap to app thing that I think I've shown you all before. And I love it. It's nice because like sometimes, especially when I'm, uh, according to your point, like not being deeply thoughtful and I'm sort of stream of consciousness coming up with my idea. Um, I just oftentimes like verbally say what I want and just sort of enumerate all the things and then let, let the model do the dirty work of like coming up with um a well-formulated prompt based on my make that a plan yeah yeah list is my favorite i think that's a great way to do it because sometimes the friction is the typing and and trying to put it in some rigid structure so if you can just talk through what you're trying to get across then you can iterate and especially with flash if it just can spin up something super quickly 64 seconds i mean that's great 64 now the now the barrier is like doing the tts and generating the all the other stuff so this is uh we you you should do these generations offline from a uh from a tts perspective so that your users don't have to wait a bunch of time in order to do to do that's a good call um are is there anything that you you could foresee or predict on the horizon that could lead to that time latency gap shrinking that perhaps you know new chips enable or some new technologies TPUs and TPUs yeah TPUs will help more TPUs is always is always part of the solution um one of the interesting things is like we're and we're working on this for the the audio stuff right now and actually this is the difference between the live audio model and the TTS model the TTS model at least in this example is not like streaming the stuff back.

40:50It's like doing a full generation and then it's like giving me the fully generated stuff. So I think you can actually like kind of bridge this like hybrid real time text to speech, not text to speech, um, or like asynchronous synchronous use case by just like streaming back the audio tokens. Um, so I think we actually are going to do that, which will make like this example that I'm trying to build here, like a lot faster because it'll just work out of the box. Um, So I'm excited for that, to be honest. I think it'll be awesome to not have to wait to generate stuff. Logan, before we let you go, I know you've got other appointments and a busy day ahead of you and hopefully some time off in the near future.

41:32If you had one piece of advice to share with viewers about how to think of interacting with Flash 3 Pro, with, excuse me, Gemini 3 Flash, what would that be? And I have one addition to this. If there was one use case where you would rip out 3 Pro right now and instantly replace it with Flash, what would that be? I think the use case is like 3 Flash should be the default model. Like I think for most things actually, like it depends, unless your use case is like truly the most frontier intelligence thing, you should by default be using 3 Flash. like 3pro is the default like from and we actually see this across google search and ai mode and the gemini app and ai studio vibe coding like your retention from a user perspective like engagement is all much better when you give people flash model um so you should definitely do that i think the advice i have for people is like go and be ambitious with this model like i think i think it really does at the at the price latency cost trade-off uh intelligence trade-off like it unlocks a bunch of new stuff that wasn't possible before so like make use of it the only way you benefit is like actually going and building something on top of it um so make use of it start building something and we can close out with a calm city morning meditation i love it

42:57Logan Kilpatrick:find a comfortable position sitting tall but relaxed gently close your eyes or soften your gaze Take a deep, slow breath in, feeling your chest expand. This is actually great. I'm going to ship a meditation app over the holidays. Here you go. I'll use it every day to sort of calm myself with all the chaos of launches and shipping things. So thank you. Jared, I could use it. Yeah. Well, you two can be my first users and we'll make it happen. So this is what's going on. One two second question. We just got in. Do you know, is Flash 3 Pro or Gemini 3 Flash, God, I'm sorry I keep stumbling today, will it be available to schools as part of their free product experience that they already have in the works?

43:52That's a good question. I would have to double check. My guess is like 3 Flash will be the model that sort of powers all of Google products. I don't know what the like explicit timeline is or like the packaging from a school perspective. It is available in Workspace, which is sort of part of the product suite that's available to EDU customers. But I think it probably will become more available early next year as well. I think the timeline from launching this to doing the rollouts across Google is very difficult end of year because there's lots of code freezes to make sure we don't break people's businesses and stuff.

44:26There's a lot of Google. Yeah, there's a lot of Google. So if it's not yet available in whatever the product is, my guess is it'll be available hopefully early in January. Cool. Cool. Logan, thank you so much, Logan. I appreciate you taking the time. Hope you and yours have a good holiday season. Likewise. Same to you both. See you, man. Yeah. All right. So I have a couple things I can share on my side. Corey, if you have something you want to share on your side, we can do that. Let's go with yours first. Okay, great. One of the first things I want to do is do our typical benchmark, which is cat doom.

45:07So, and I'm going to give it a very basic problem. how much I love that Logan comes on here. He's such a neat guy who's got his hand in all of, frankly, Google's coolest products, in my opinion. And I love that we're able to come on, talk about it, ask some questions. And I'm not kidding. I really would use the meditation app. Grant and I work at a pace that is, I don't know what the word is. Pretty rapid. frantic yes that's the perfect word actually i was like there's a good word for this somewhere i just got to think about what it is but uh yeah big thanks to uh to logan kilpatrick and google for sharing him with us for the morning here uh and for all of you who came and asked questions and now we're going to dive in and go break some stuff as we tend to do love to do it Love it.

46:05And if you have any questions or any requests, please share them in the chat. We will test any prompt ideas that you have. We are your humble servants. So let us know what you want to test and we will test it. 100%. So I'm going to go ahead and present. So Corey, I have two screens. I have my personal screen and I have my current computer screen that is sharing as of right now. I'm watching. I will add it to the stage and while you're showing this I'm going to do something that's probably not the best but I'm going to run and grab a bottle of water go ahead you can drop off I'll bring you up and I'll be right back yeah alright so here I have a very basic version of my prompt which is our benchmark here at the Neuron let's make a version of the game Doom starring cats I am using Gemini 3 Flash.

47:03I don't have any system instructions. I'm just going to test how well it does it. In a separate tab, I am also going to use the prompt we used last time to test Cat Doom. And I'm going to share this tab. So this is the prompt here where we put our objective neatly formatted. Then we put the game design requirements and the mechanics. So we gave it a lot more specifics. you know even down to the visual style the audio the rendering probably don't need to give it that level of information and probably be able to do it without this the controls the performance and I think that's level structure and success criteria so I'm going to spin up both of these at the same time.

47:54And just for a reminder of how this last one performed, this was our last version of Cat Doom. And this was with Gemini 3 Pro. And I'm going to go ahead and play it here. I remember I have, it's very dark. I remember this about it. And it's hard to see. but it does work. So for example, there's kind of some scary cats in here. Oh yeah, this, I need this. Let's see. Yeah, there's some, oh gosh, there it is. Got it. And I'm shooting it with a yarn ball is what I'm shooting it with. Oh my gosh, it's like so close. Ah! Okay. So that's, that's, I'm going to go ahead and pause this. So that was the first version.

48:53Now, this is the one where we went to Gemini 3 Flash. And I said, let's make a version of the game Doom starring cats. So this is, that's the only instruction I gave it. Let's see how it performs. Oh, my gosh. This looks a lot more like Doom, I'll tell you what. I also like that it used emojis. But I also have no idea how to fire this at all. So I'm a little bit running around scared at the moment.

49:30Logan Kilpatrick:How do I attack? Oh, gosh. so okay so one thing I why don't I give it the feedback like this is great I'm gonna say okay pause it no I'm gonna dash die this is great but I need this this really looks like doom that's cool this is great but I need to know the controls on the screen

49:58because I didn't know actually why don't I do this I'll use the speech to text i didn't know how to fire my weapons and i was being chased by a bunch of dogs i also i didn't see any weapons anywhere to pick up i love it oh no it didn't it didn't um it didn't send it okay i'll just stop this and i'll send this one instead okay so now it has all the context um Hi, Corey. Welcome back. I'm showing Cat Doom, our benchmark here for testing games in AI Studio. Love it. So if you remember correctly, this was the one that we did last time, which is our version of Cat Doom with Gemini 3 Pro. It was nicely formatted.

50:46I think we gave it some feedback so that... The evil cats have been defeated. So this version over here is one where all I said to Gemini 3 Flash was, let's make a version of the game Doom starring cats. And it's pretty good. And it has dogs, which I like better. Your cats three. So this has emoji dogs that are trying to kill you pretty much everywhere you go, which is very terrifying. but I told it that I didn't have anywhere where I could find weapons to pick up or any sort of feedback of how to find weapons. Oh, it looks like maybe I was able to, oh, it looks like I shot it actually, but there's no onscreen indication that I'm shooting at it.

51:39So that was maybe perhaps some feedback I need to give it. then does it create content that could be played in meta quest three goggles uh i bet you we could ask it to do that i don't see why we couldn't now just for sake of argument i gave the exact same prompt that i gave to uh gemini 3 pro to flash as well so i i walked through all of the you know it's got success criteria game design requirements objectives here are 100 features go yeah kind of going for the for the logan uh suggestion and here's what this one created i feel like i'm very claustrophobic and something is attacking me oh my gosh oh my god you're absolutely getting attacked somewhere possibly whoa through i'm getting killed i'm getting killed and then it just spawns right on top of me which is horrifying can i shoot oh okay i can shoot oh hold on you should be doing that probably yeah yeah the problem is i don't have uh my mouse right now so it's kind of hard to control this oh that's funny yeah this one's hard okay why don't i tell it i'm gonna give it this feedback I don't know the controls, so it would be nice to put those on screen somewhere.

53:11I love the displays, though. I love that you have lives, a score, health. Yeah, this actually, I think, looks a lot better than this one, which was the Gemini 3 Pro one. And to Logan's point about, they had a lot of improvements in post-training on Flash. Oh, my gosh. the only thing i like about this one is that i can very easily move my mouse to steer which makes it a lot easier to control and look around it does but going back to oops i'm actually okay to be fair a yarn ball is not an effective weapon against a cat well it's perhaps you could argue that it's distracting the cat maybe it's distracting the cat.

54:02Yeah, maybe the goal is not to kill the cat, but for the cat to curl up in a ball and play with the yarn.

54:09This one's really weird because all I can do is kind of just move forwards and backwards. Yeah. So I'll give it the feedback. I don't know the controls, so it'd be nice to put those on screen somewhere. And I'm playing on my MacBook, so it would be nice if I could easily move the camera with my trackpad. And I'll send that to them.

54:41Let's see if this one... So this is the one where I gave very little feedback. And I just gave it some feedback about I didn't know how to fire my weapons. I was being chased by a bunch of dogs. I also had to pick up my dog. I meant to say, also didn't know where to pick up a gun. what are you by the way like I don't know you know what you should be you should be like a Rambo style mouse oh I like that cats are coming after you I feel like there's an opportunity to lean into that okay let's take this one because I think it's quite good

55:23let's do this one more time with Flash I'll open another tab and I'll say I'll give all this and then by the way as far as themes go I want the main character to be oh let me do the let me demo the speech attacks oh we got a good tip from Lumen Labs here that I'm going to share to you a Rambo style mouse so the theme should be all about you as a mouse trying to escape a bunch of cats and your little icon in the corner should be a mouse face. And if you look in any reflective surfaces, you should see yourself as a mouse. Cool. I'm dropping some control tips in here from Lumen Labs for you into our studio chat.

56:19Okay, great. Omnidirectional control. That are really good ideas. Yep, this is great. I'll throw this in here. Why don't I do the exact same prompt, and I'll do it again. But this time, I'll add those suggestions. So let's go ahead.

56:47So we're going to share this tab.

56:57All right.

57:02So while that's spinning up, let's see what else we have. So this is the one where I gave it some feedback on, I don't know the controls, it would be nice to put those on the screen somewhere. So let me see how that did. Okay. Yeah, this is way better. Although I should give it the feedback that it would be nice if an enemy doesn't spawn right on top of me. Yeah. the enemy enemy could stand to be to have more fluid motion yeah let's say that it's it's it's it's sort of it's also too fast like flashes of images yeah Yeah.

57:54And you're doing this all with Flash, right? This is all with Flash, yep. No lives left. And you can see how quickly it's creating all of this. It is. Oh, wow. I didn't even read what it requested. It was gone too quick. Yeah. Okay, this is interesting. Okay, but now the problem is I can't even... Okay, this one's very buggy. Allow it to access your camera? It wants your camera. Why? That's a good question. This one is weird. I don't like that one. Let's do this one. Oops, I think I took down my screen. Sorry, I'll do another one. got another good tip there too yeah let's let's hear it

58:56let's see so here's the one where we use the omni-directional controls right so although this one doesn't feel as fluid as the last one oh i'll be honest the cat head's almost scarier it reminds me of the uh the demons in dune that that it's just like a flaming head that flies after you. Yeah, totally. Oh, there's ninjas. Oh, oh, here's a fire button. I missed it. Sorry. There's no visual feedback that it's taking. Yeah. You need. Yeah. What is this? Oh, is this let me. Oh, okay. Oh yeah. Like a joystick right on.

59:43Typically.

59:50I also don't like that when it I need to give it the feedback that I don't like I'll just type this out or speak I don't like when the enemy gets close to you that it disappears in your view it should stay in front of you and keep hitting you so that you can keep looking at it because it's a bad user experience for it to just disappear behind you into the void. Also, I like this, but I would really like it to be easier to control with my trackpad. And also there's no feedback when I'm firing, I'm not getting any visual cue that it's firing, I'm not getting any audio cue that it's firing. Or that it's hitting.

1:00:36And I'm getting no indication whatsoever that it's hitting the enemy. so let's add all of that feedback to the game

1:00:50all right so i'll send that um anything else anyone wants to see we're just having fun over here but uh but happy to test other prompts we could do something a little more practical work related let's see i had an idea right as we came on this morning yeah and it's gone um it's very cool meow doom is a great game great name yeah i really like this one this is the best layout by far it's the best functionality which one is that so this one is the one where we uh gave it the iterative feedback so we prototyped it we found it kind of difficult to use the the characters were coming at you too quickly they spawned on top of you and yeah now is that a fish that was a fish uh yeah yeah i think so oh there it is oh okay okay it worked we we do have an indication the balls uh are a little flashy it would be great if if you know you you could see the traject see the ball fly through the air as opposed to just a purple flash let's say that the yarn ball projectile is a bit flashy it would be nice if we could see it complete its trajectory through the air instead of just sort of flash at the enemy also i think i want the enemy to have a reaction when it's hit and it should probably have four to five yarn balls before it gets to full damage perhaps also it flashes red when it's hit something like that

1:02:32i see lumens adding a lot of um good good feedback to this how how to do this technically what's really awesome about ai studio and and gemini 3 flash is that I can kind of just describe what I want and it can handle a lot of that. And it generally knows, yeah. Yeah, if I wanted to be very precise, which obviously if I was creating a game that I wanted to iterate on and build on, I would definitely be more prescriptive with how I wanted to do these things. But just because I'm able to just give it feedback and have it change it, it's great. Also, I think those cats are too small. That's better?

1:03:19I think this is an updated version. It's updating it iteratively, I think. Yeah, because you saw the ball fly that time, mostly. Yeah. Might change it again. Oh, okay. Okay. Let's see what it did different. Okay, running around, running around. This is really fluid. I really like... We need the opportunity to, like, find more yarn balls or different weapons. Yeah. you're right oh there we go oh i like the pushback that's good yeah yeah that's great i'm gonna say the cats so the cats aren't very intimidating they're a little small could you make their character icons bigger and also maybe make them look a little meaner like really really imagine that these cats are trying to kill this commando mouse in a doom-like scenario also what was the other thing we were going to change oh also we need more items that we can pick up like new weapons strewn about throughout cool let's give it that I like what Lumen saying here you can mitigate drift hallucinations or revisions with more explicit clues and terminology while still being in a creative mode totally yeah I completely agree with that I think that's spot on I regularly a thing I do when I feel like I'm asking something dumb is like I'm doing a lot right now in V0 and I pretty regularly keep either chat GPT or Claude or Gemini for that matter open in a side window.

1:04:56And if and we'll ask, what do I call this? What's the what's the right way to ask for this? and sometimes I'll get better results because I guess because I'm asking the we're asking it the right question and it's not having to interpret what is this idiot talking about he's trying to say something with a lot of dumb words it definitely helps don't don't get me wrong

1:05:24Logan Kilpatrick:Okay, let's see. It's still working on that. I wanted to show this other tab here. So for people who either missed when we were talking to Logan, or didn't quite understand what I was asking for. So I had shared an idea for an application I think should exist, but doesn't yet, which is a screenwriting editor, like, you know, imagine final draft, but as you're typing each sentence, you're shot listing and storyboarding. your project on the side, which is like a perfect use case of VO3 and something you could definitely should be able to build in AI Studio. And we tested a version where I just took like my raw thoughts on it and just dumped it into AI Studio.

1:06:08And this version, I actually worked with Cloud 4.5 Opus, don't tell Logan, to create a plan for this whole thing. And it created a 43 page document. It also outlined exactly how we, you know, essentially like making ASCII drawings for the architecture and how it should lay out. It doesn't look quite as good in this Google Doc, but I pasted it in here because I wanted to test it and share it with other people. So I spun up actually four different versions. Let's see. Let's see which one of these is. i'm going to just show perhaps three different versions i uh i have an idea i want to try to spin up here in a few too when we're in between not a rush by any stretch of the imagination though yeah so here is one that i uh where i did the exact same thing that we had logan do where i dumped just the raw tweets of me saying hey this should exist why doesn't why hasn't anyone made this and then it basically put it together here where it's got the sentences and then it's got on this side the real-time scene visualization now it's not going to work right now because i don't have the api key laid out but this is just to show you the ui from from just doing this raw dump and i believe how it would do if we went and generated like you know just just a two-minute script and say um here's the script flesh all this out maybe let's do it let's do it so i'll go to ai studio not build go to playground i'll do gemini 3 pro preview because it's pretty good writer and then i'll say let's gen or help me generate a two-minute script in screenplay format for a test run of an application and building.

1:08:17Any thoughts on the topic?

1:08:22What should the topic be? Anyone, Corey?

1:08:29The topic should be about commando mouse trying to escape. dangerous warehouse full of evil cats. Kind of like a cat doom video game movie. We'll do that. Well that's generating. I'm going to show you the alternate version. So this version I gave it the real long all of the I gave it the 40 page version of the prompt. And you can see here essentially what's happening is it has this editor and it's already spinning up everything in it. And it looks like it's going to show it over on this side. But then when you go to visualize mode, what happens is it essentially puts a timeline together of all of the lines.

1:09:31Oh, look. oh this is perfect this videos this one's working okay hold on let me show you this one so this is pretty cool so this is one where i followed a lot of like prompting best practices i work with claude 4.5 opus to make a plan and then i gave it all the features i could think of and basically based on this script it created this timeline and what i did is i had it generate all of the shots as still images and then if i wired this up with my api it could then you could then generate the video based on this. But that way you're not frivolously generating video clips as you're writing. It's more like...

1:10:09It may be terrible. Yeah, it's more like... Narrow down your first frame first. Yeah, exactly. That you want the video to follow. I like it. You know, something I do when trying these little projects that's really helpful for me is I'll come in and I'll share a plan and talk with the terminal a little bit there first like does this plan work you know any hangups or generate your own plan of action to make this happen and then kind of have it take it in small nuggets

1:10:49but this is impressive that you've gotten to hear with the full thing that's it's pretty sick that's a lot of different pictures yeah i need to add an option so that you can make it uh big because right now i can't really see what's going on but it's took it took this script that we wrote about our mouse um rando mouse yeah commando mouse a dimly lit industrial labyrinth towering shelves block out the moonlight filtering through the grim grimy skylights the silence is heavy, broken only by the rhythmic thump, thump, thump of pause. Close on, whisker, three inches tall, clad in a Kevlar vest made of duct tape, a toothpick strapped to his back.

1:11:32He crouches behind a rusted paint can, breathing hard. He taps his ear, a light blue, a tiny blue light blinks. And so you see, it's got, you know, dimly lit industrial labyrinth, towering shelves block out the moonlight filtering. The silence is heavy, broken only by the rhythmic thump, thump thump of paws close on whisker three inches tall clad in kevlar made out of duct tape and so it it storyboards the whole scene as you go uh looks like it got cut off at a certain point maybe you can only do 12 at a time or something um but but yeah this is exact pretty much exactly what i was picturing um so that's great what what um which one was this let's see

1:12:20your apps

1:12:27okay does it say made with flash yeah so that one was I think might have been with Gemini pre uh three pro this one is with flash let's see what it created cool

1:12:46Yeah, so this is the one that Flash created. I'm just going to click out of this. And so it's a little bit more basic of the UI. It's got the same scripts. I could paste in the script that we wanted. But it looks like it. I wonder if I hit visualize if it'll run. Yeah, this is not quite how I imagined it, where I imagine if you switch to Visualize Mode, it would just automatically create it for you. Okay, get out of here.

1:13:23But then it kind of shows it's the same process, right? Where basically, and then you can see it on a timeline, that's broken, so that didn't load all the way in. Yeah. Wait, I think you're showing Are you showing the wrong screen? We're just looking at Oh, I'm sorry Yeah The front screen of Google Build Isn't that fun? Here, I'll show you what I've been working on real quick Yeah, I'll show this Because this one has some bugs I can fix while that's happening Actually So I just This is just one shot It's very simple, it's nothing super impressive but a little app called tiny wins. The idea being that it's helping you form habits throughout the day or throughout your life.

1:14:12So it'll do up to three habits that you can track each day where you'll go in and punch like, yep, I did the thing today. Yep. I did the thing or no, I didn't. And it will run streaks. Ideally. It's got good. It's got street lot streak logic built in. And it took 52 seconds and I haven't scrolled through or clicked anything yet. I was waiting until I brought this up, but cool. I also had it so you see it's got a daily motivational quote here but I also had it include a salty mode so it says like congratulations on being basic

1:14:49I was like let's have a rude mode too that's fun I like that because some people don't love the like super toxic positivity uplifting I don't either it annoys me Okay, let's pick an emoji. I'm going to do 30 minute walk.

1:15:09And for that, we'll use this little runner dude here. New habit, I'm going to do 30 minutes of guitar because I'm not playing my guitar enough right now. It's been a minute and I need to dedicate some time to that literally. anyhow let's make that yeah we're gonna give it an apple because that's the best option i see third habit is going to be drink eight glasses of water or something that may be too much but who cares um so we'll start the habit we now have three different things in here and today is what is today? The 26th? No. It's the 19th. Okay. What I don't know here is how do we say yep, I did it today.

1:16:06Maybe click on a date. Can you click on a date? Oh, hang on. Oh, that's easy. I like that. Click the icon. It clicks the date. Click the icon. It clicks the date. And alright. This is cool. I like it. It needs to send the user daily push notification reminders so they don't forget the text of those reminders should be super casual like oh yes Lumen just talked about goblin tools yeah that is an awesome resource it's basically just a bunch of fun tools like this that anyone can access and it's built for neurodivergent people who maybe find other tools more difficult to use or prefer these types of tools and user experiences yes that is awesome i i have tools yeah i have heard of goblin tools another piece of feedback i would give you on this gory is let's say for example uh you walk 20 minutes or you drink three glasses of water perhaps there's a way to say to like incrementally edit it um do you know what i mean where yeah you yeah like this one here to be able to say i did four though yeah like maybe maybe it automatically knows you have to hit it as many times as the number objective so although that would be kind of annoying for a 30 minute walk uh maybe ask him and i how to help with that too offer the ability to include to mark incremental progress throughout a day when creating a task.

1:18:25For example, if a task requires 30 minutes,

1:18:38they should be able to mark that they did do 20 or for eight glasses of water to show that they drank six. also this needs to be gamified

1:19:09gamified with achievements and badges

1:19:20okay now let's see where we're at try not to fail tomorrow I love salty mode I think salty mode was a good idea it's my favorite thing about the app and here's my notification bell enable reminders of course it doesn't know where to send them we should have it ask that so that's not actually integrated I don't think but yeah it says it's crunching it's noodling it is grounding

1:19:55deciphering would you would you say that this is making it your iteration faster like because it's quicker to change things that you're more likely to engage with that you're more likely to keep using it noticeably so i am what i will say from from my limited amount of experience at doing this with other tools is that I care less about the time and more about that it does what I want. If it can do what I want and do it in less time, that is awesome. I would love that. I don't know if it would make me use it more. Like I would say I'm usually a couple of minutes in what I'm doing with V0, but I'm also, doing small tasks at a time.

1:20:42I'm doing bite-sized elements, not not a little bit.

1:20:54So maybe do like fun. And I like it. You cut in and out a little bit there. I don't know if it's just on my side or what. I think it freaked out as it was reloading this. My monitors both blinked. oh this is cool okay we've got a few badge opportunities log any progress for the first time complete your first goal seven day streak log progress before 8 a.m for an early bird so we got some options what's our history we have no history because we're brand new does this do anything yet other than just say it no we'll have to fix that but you know okay so let's see this okay play more guitar goal five times maybe we need a per what element like maybe the answer instead of 30 minutes a day is maybe it's five times a week maybe a user would like that as a functionality yeah you can set the frequency.

1:22:05Yeah. Yeah. So let's just for simplicity and looking at this, we're going to use broccoli because why not use broccoli? Okay. So there it is. Play more guitar. And we've got a plus button here. Look. Boom. Two out of five. Boom. Three out of five. I like that. That's funny. Five out of five. And we got badges. What does it say? First step. Log any progress for the first time. Goal Crusher, complete your first daily goal. This is awesome. I forget that this is set up as daily goals. So maybe per whatever isn't necessary here. Unless maybe we don't want to be daily goals. Maybe we just want it to be a goal tracker.

1:22:51This could be a cool free app. This is kind of a cool little app, isn't it? How did it design this? Could this be an Android app? That's a good question. You know, one thing I don't know about this, and it's in the thing we talked about with Tom from Vercel yesterday, is that at this point, there's not a functionality to make them mobile apps. and I don't I should have asked Logan today can you do that in build um and from what I've seen most of them can't if any of you watching know the answer to this are there tools where you can uh tell it you know I need a mobile app version of this you know and it'll buzz it out and react or something and give you something you could spit out and send to an app store.

1:23:48Okay. I'm going to remove my screen. Do you still have yours up, Grant? We're actually down to not too many people left. Yeah, we can wrap things up here. If anyone has any questions, I mean, I think at this point, we're kind of just playing around. Yeah, we're just being goofy. feel free to ask your questions in the chat I did have a conversation topic if we wanted to chat about that I'm a fan the last thing I'll do before we get to that is I just want to circle back on what I was talking about earlier and I will share my screen to do that the blind dragon 13 says replet maybe Lumen Lab says everything can be ported to a PWA and any PWA can be ported or transliterated to a mobile app.

1:24:41Swift slash cop. There you go. You heard it here first. It's good to know. I've got a project I'm working on that I'll share in here one day soon. Cause it's, it's up, it's working. My integrations are working. I just, I've got some people I need to spend a little time with it and come back with some suggestions and make sure it doesn't, you know, break the world. And I'm trying to come up with a, I say trying to come up with, I have a list of questions to ask it in terms of investigating like my code base and determining, you know, do I have any visible hard-coded API keys? Do I have any, you know, the kind of things that can be really problematic?

1:25:25You want me to bring you back up, Grant? Yeah, go ahead. So this was the version that Flash created. And it works slightly differently than the one we saw before. I'll go full screen. So basically, I've pasted in our script, which is a dimly lit industrial labyrinth, towering shelves block out the moonlight filtering through grimy skylights, blah, blah, blah, whisker three inches tall clad in a Kevlar vest made of duct tape, a toothpick strapped to his back. He crouches behind a rusted paint can breathing hard. So that's kind of the idea, right, of this. Let's go ahead and visualize. So what you do here on the side is you visualize the shots by shot.

1:26:13So I like the other UI experience better. It's going to bug me to connect my API key for a minute. Cheapo. I'm just kidding. Me? I can't connect it on this computer. I can show you a lot. This month has gotten unreasonable. Oh no. It's like really strict with me. Okay. Well, anyway, the, the difference. Thank you for the kind words, by the way, glad you were here. Please like, and subscribe if you haven't yet. Same to everybody. It really, it, it helps us make the case for why we should be doing these lives all the time. And the truth is they've been really good for us and you all have been too.

1:26:54So if you take a minute to like and sub, Maybe go check out the newsletter at the neuron.ai. We'd love to have you over there too. And thanks for coming. And back to Grant. I'm going to shut up now. Yeah, this one's toast. But basically, it just had a different experience. I think the one I liked better for this specific project was this. I think it was this one. I think it was it this one? No, it was not. The script looks more like a script. I'll give it points there. Yeah, I like. Well, you should be able to switch between dark mode and light mode. but yeah this definitely looks more final draft-esque um but the one that i liked the best was and i'm pulling it up we have a lot of cat doom i think it's cine script is that the name of it interesting yeah i think it was this one yeah okay so are you sharing the screen you think you are no yeah so this one where this one that i've outlined here which is basically you go to visualize mode and then it automatically visualizes i think this one's the best the closest so i would probably work with this one as my starting point where it generates automatically the images on the side i think that's the best experience i like that yeah it's really cool i'm gonna play with this over the weekend that's cool i'm gonna play with my my tiny winds app too i uh yeah it's it's super simple i may leave it super simple maybe there's i don't know it'd be cool if there's a way to I don't know.

1:28:27I'm going to investigate. I've never done anything with an app and I'd like to know how. So I may play around with a couple of these tools that can kind of port it over. Yeah. I have one more. Suggested and you did too. Oh, you got another one. I'm sorry. No, I pulled my screen down. I should have to switch to this tab. So this was the final version of MouseQuest where we were giving it lots of feedback. Oh, okay. What the heck? Bite? Yeah, I don't know about this. This is not the good one. Where's the good one? This is the good one. I'm going to turn off. Yeah, Mow Doom. That was the one. Mow Doom is the best.

1:29:06Yeah, this one is so fluid, so smooth. Yeah, this one's the best.

1:29:15That is cool. Yeah. Can you raise and lower your crosshairs and stuff? Do you have functionality of that? Okay. Yeah, it's very smooth. All I'm doing is just going to be my mouse. Yep. So there's a fish. What does the fish do? Do you even pick it up? Good question. I didn't notice if your health went up, you got lives back or anything. Oh, we made the cat definitely scarier. I love it. I love when it peeks around the corner. Okay, you killed the cat. They should drop loot because, dang it, we're gamers, we like loot. you're right you're right oh here's something attack mode engaged uh click click to add your feline senses click what what i don't you're also using reinforced yarn now oh catnip burst okay okay where's the enemy oh attack mode engaged wait no come on i think catnip burst should be a trap like a rogue.

1:30:23It would be really cool if I don't really know what that does. If you could set a trap and the cat comes to it and then you could attack it. Yeah, that would be cool. Caddoom. Alright, yeah. I feel like this is a good place. Definitely impressed with Gemini's abilities here. We will be back to Meowdoom. It feels a lot smoother than previous iterations of cat it does i agree i agree it is it is making significant progress um so the last thing i wanted to talk about here was kind of a a continuation of the conversation that i started with when we covered gemini 3 flash earlier this week which is the idea that you know one we're we're at the pareto frontier which i'm i don't think we really talked about that with Logan.

1:31:21So maybe I'll share my screen just briefly.

1:31:29To show this. Actually, I'll bring up this page. Okay.

1:31:43So obviously they have all their benchmarks here your mileage may vary on whether benchmarks are useful for you this is interesting though um and unfortunately well let me okay i can share this tab yeah so what you can see here is that gemini is kind of dominating what's called the pareto frontier which is essentially the best cost and speed and intelligence that you can get um it's sort of like like the best most efficient combination of those three yeah exactly and so it's it's the highest intelligence lowest latency you know fastest um cheapest model you can get and for a while tokens assuming five to one input output tokens ratio right yeah so it's dominating this this section here um and gemini 3 pro is obviously the highest intelligence but then gemini 3 flash this is so it's the lm arena score is this is 3 pro still the highest intelligence and then this is price per million tokens um the still depending on your benchmark of choice yeah let's actually go to artificial analysis we can look at that a little bit more in depth so gemini 3 pro is still the highest on their index here.

1:33:08Well, it's tied. Oh, good call. It is tied. You're right. So they are neck and neck. But if we're looking at that and compared to cost per API call or cost per token, yeah, I was just curious. I wasn't sure. You know, when those benchmarks first launch and seeing, I don't feel like 5.2 is lower performing than 5.1. Now, that may be the case in some specific areas, but in what most of what I do, which is writing and creative endeavors and video stuff. I get madly impressive results compared to what I did from 5.1. But I have seen that. I have seen like, you know, there's always disagreement within the community.

1:34:00And I love following those discussions because, you know. And maybe that just comes down to what's right for me, what's right for you, what's right for Tom and Tammy and whoever else, I guess. So often I think, and we did it again today and talked a lot about software engineering without talking about other things people use AI for. There's this tendency to lean really heavily into, look how great it codes. And I think we tend to forget that, you know, how well it does research, how well it works in a spreadsheet, how well it does this can really matter. And as it gets better at one thing, it seems that quite often the adverse reaction is it gets less good at something else.

1:34:54At least that's been my experience. And I talked with, well, I don't want to put him on blast, but I talked with a researcher a couple weeks ago while I was at Amazon. And he talked about that, that there's this give and take there that, you know, if the more you're optimizing for any one task, the less optimized you are for many, many others. and and i'm guessing we all look at every one of these through the eyes mostly of what we do yeah yeah i guess the the point that i wanted to make here specifically and talk about is that it's just it's really it's really fast it's really cheap and it's really smart and at At a certain point, that changes the calculation that we're seeing right now, where right now, jobs that are, quote unquote, exposed to AI are seeing their wages increase, according to data that Vanguard just put out basically over the last two years.

1:36:03jobs that supposedly were likely to be automated are actually the people in those roles are commanding higher salaries because they're learning how to work with the tools and they're able to do more. So you're able to do more, you can argue for being paid more. And I wonder at what point that the AI model itself gets so intelligent and so capable and so cheap that that calculation switches and then there's less people who work at less companies or there was a report that I saw recently about a bunch of copyright freelance copywriters and how their work has the market for their skill set has basically imploded and the reddit thread that I was talking about with Logan earlier was a lot of programmers worrying about the same thing happening to them now there was a lot of programmers in the thread saying the amount of vibe coding that has been unleashed has led to a lot of freelance work for them fixing slop.

1:37:04So there was a lot of optimism in there. Yeah. You know, one catch though, I think is something we talked about and we've talked about it with, we talked about it with Logan. We've talked about it with a few other people. It's this idea that, you know, for people who are really good, there's going to be a lot of opportunity around those things. But I think a thing that is a legitimate concern is what does that mean for entry-level people? Like, it is true that there's going to be more experienced cleanup needed that, you know, your real ideators that are able to do the big tasks are going to be probably just fine for a good long time.

1:37:44But what happens when we have a smaller entry-level funnel of people, you know, what's the long-term effect of that? And do you think, this is a fair question, do you think there will be a need for just as many entry-level engineers as there are today? Like, or as there were three years ago? The conversation, the comment that stood out the most to me was this comment, which I'm going to lead. it's like, I'm going to read, it's called the lump of labor fallacy, which is there's a core mistake that we make in treating labor demand as fixed. So the implicit model is there's X amount of work. So AI does more of it so that humans do less of it.

1:38:33But this is the lump of labor fallacy because what actually happens is efficiency gains, lowers costs, increases demand that creates new applications and there's actually net labor expansion deployed differently. So the example that they give is agriculture went from 40 % of workers to 2%. ATMs increased bank teller jobs because operating branches got cheaper to run. Maybe less people work in agriculture, but then they switched to other roles. But more people are working over here. But more people are working in banks, for example. When building software gets 5 to 10x cheaper, more software gets built and so therefore demand is elastic not fixed i think the the i think that while that is true there's also the the problem of just because you can do a task doesn't mean that task will be like your wages will not be guaranteed at a certain rate forever, especially as more people can do that task, if your wages are dependent on that task.

1:39:37And so I actually believe that we need to reframe ourselves away from roles that are task specific and start to position all of society and, you know, at least in our companies or how we position ourselves to roles that are not task specific. I don't know what the right framing is, but It's something that's more like, I am a marketer. I am a brand storyteller. If you're a writer, you're no longer a writer because the task of writing can technically be done by an AI. You are a storyteller. You are more than the tasks that you do. So if we think of a software engineer as just being someone who codes, I think that that is absolutely going to change because if everyone can code, like the demand for someone who codes goes down.

1:40:27Yeah, exactly. the truth of the matter you know the one thing i would say is that that i think that's that's a really interesting macro look the the my my one not criticism but but area where i where i kind of fall off is that you know like yes these people here we're going to do other cool things and uh and there will these jobs will be replaced by these jobs but that doesn't mean tom the software engineer is going to become tom the banker like tom is still screwed you know well other people will shift into these new areas i mean like as a grant on the grand scale there there there is this shift but how do you how do you and and then software engineering was a bad example as was probably banking but that was just spitting out ideas but like yeah whose job let's say it this The person A whose job is deleted over here for one reason or another, not even AI, but like that person specifically doesn't go over to these new jobs probably.

1:41:39I think if your mentality and your identity is tied to the specific tasks that you do, if you went to school for something for multiple years, specifically for a task. It was a big industry. If your identity is tied to that task that now can be automated, then I think it's harder for you to switch to a new skill set. Yeah. And is there a new skill set directly adjacent to the one you have that is an obvious transition? Like you're not going to necessarily go back to zero and be like, all right, where are the jobs now? And then go learn that entirely new path. You know, and I always the reason I mentioned TV repairman is every town had them multiple appliance repairman.

1:42:30Every town had them multiple of them. The milk guy, you know, those are those are professions that absolutely died as industries and consumer demand changed. Like with regard to like TVs or throwaways. Now, TVs aren't a thing you buy and you get repaired every few years when something happens. Like it's not a one-time purchase. It's a, it's a, it's a four year at best investment that you're going to throw figure, have to figure out how to get rid of later because nobody fixes them anymore because they're cheap. And I don't know, like, I I'm thinking of very much at a more micro level, like those individuals, like there are, like there wasn't a natural transition for that.

1:43:17Like I I've interviewed some We did a special series years ago when I was in the newspaper business on, I don't remember what we called it, but basically it was dead jobs. Like careers that were booming, good careers where people had a nice life and pretty quick it vanished. And absolutely over time there are more jobs than there ever were. There are definitely more opportunities. But at the individual level, that's a catastrophic shakeup. Yeah. When it's when it's your job, you know? Yeah. Yeah. And that's why I think to the extent to which that it is possible for us humans to prepare ourselves mentally, emotionally, I think starting to reframe your thinking around, okay, what am I if I'm not X, the task, like the task, the X task doer?

1:44:16and what could you be like how how do you open your mindset to change and being open to that if that's where all this goes because i do think that there is going to be a at least a short-term reckoning with the job market and maybe we'll get more into this in our prediction episode but but this type of release really puts this in perspective for me it's like at a certain point, there will be an economic incentive to, you know, or let's put it this way, there will be labor market ramifications where you now all of a sudden don't have as much work as you had before if you're a freelancer, or it'll be, you know, you can charge less because more people can do the task.

1:44:57So how do you prepare yourself for that transition? I think Lumen said something, what happened to the transition between those who were to follow in their parents' footsteps for working in a trade craft when machines replace that craft what did we do then um there will be new roles in demand humanity will always find new things keep busy i agree with that i think that's true but but in order for you to accept that and not being be like thrown out with the bath water mentally like say like i'm a tv repairman and that's all i do and then when tv repairman jobs go away you all of a sudden are left adrift with no no other path you have to fall into a depression hole for six months first.

1:45:36Yeah. To a degree that's unavoidable, but I'm just thinking about this, like how do we prepare ourselves for this with whatever quote unquote tasks that we do now? And how do we position ourselves and think from a personal perspective and also a society perspective on how to get ready for this shift? I keep telling people, go study philosophy, Go learn to learn Like the God's honest truth is That was a degree that I acquired Because I kind of had a passion for it But did not have Career Aspirations for You know like I never Studied philosophy with this idea that I'm going to be a philosopher I'm going to teach philosophy in college It really wasn't about that Like my minor was my job My minor was in journalism but what's interesting is that it has paid dividends over the years in in changing how you look at problems and i think i think skills like like problem solving the ability to see the big picture and look at problem from every perspective i really really think those are going to be skills that will accompany an ai power user in a valuable way uh and uh and the idea of being like i'm gonna i'm gonna go get a degree in marketing and learn to write press releases and ad copy and how to build ads i mean like that's it's not gonna look like that i mean you'll need skills in that way but it'll be skills you need in the way that you know you still take geography as a gen ed course like you want to understand the why of why it's done but you don't necessarily need to be able to create a map you know um i wonder if ai compresses the 10 000 hours that you need to master something i mean certainly it compresses the 10 000 hours you need to do something but obviously you're not just because you can do something with ai just because you can code with ai doesn't mean you understand the code doesn't mean you know what you're doing might depend on the learner too you know the truth is everybody learns a little differently.

1:47:49Some people read, some people need to be hands-on, some people like to listen to a lecturer. It's one of the things that's been a struggle with education for many years is that not everybody excels in the same learning environment. And I had a really good discussion about that with someone recently. Yeah, self-awareness, self-love, grace, gratitude, forgiveness, perpetual curiosity, for God's sakes. Curiosity is, is like, I think like I hear it discussed in we're opening a role. What are the qualities we want? Curiosity. We want people who aren't afraid to try something, to make a mistake and are capable of learning from it when they have.

1:48:35And you know, in an era where, and you know, just look at Google and open AI this week, those two companies have have batted back and forth like a game of ping pong with releases every day like oh well i've got something cool well i got something cooler no i've got something cooler and you know and they're all great they've all been great and frankly i enjoy uh basking in all the new toys that that make covering the space so fun but you need to be ready to like accept that that quick change is just a passing a fact of life now like industries are going to shift there are going to be processes that shift rapidly there are uh if you're a person who really struggles with that i i that's going to be really really rough uh you know maybe try to find a way to learn to adapt a little more easily and And it's not necessarily a bad thing.

1:49:33I mean, some of it would be a good thing. I think we do need people who don't want to change and don't want to adapt to push back. To push back, yes. So that the companies don't just get to do everything that they want. Yes. For God's sake, always have a dissenting voice on your team. The smartest thing. Even if you want to wring their neck once in a while. A dissenting voice is an important part of any decision making. And that's just a realistic aspect of systems, right? is that there's going to be stocks and flows and there's going to be pendulum swings and backlash and backlash to the backlash.

1:50:07So it's not going to be this accelerated takeoff where everyone just lets it happen. I think a lot of people feel a little hopeless right now that AI is going to replace everyone in two years. There's no way that's going to happen. But how do we therefore or protect ourselves in that process where we're learning so that we're making ourselves employable, regardless of what AI capabilities can do. So always uploading that perpetual curiosity is so important. And then also protecting the things that are really, really great about the technology, which is it can let us do things that we couldn't do personally, not just for our own skills, like, oh, now I can code when I couldn't before, or I can write when I couldn't before, I can draw when I couldn't before, but also, you know, all the things that it's good things that it's doing in science and how it can benefit everyone there.

1:51:05How do we protect those things while also protecting people from the downside of this labor market crashing? You know, I do think that the labor market will have this demand effect where like the cheaper - And$12 a pound ground beef. I can't write enough about the cost of ground beef right now. Yeah, that's messed up, man. Yeah, no. It is. And, uh, and it's, it's literally, it's$12 a pound. Like that's, that's, that's it. Like our Walmart, you know, that's not like at a Whole Foods. That's, uh, and yeah, you know, you prepare for it. You learn, learn something new every day. That's, that's the best gift you can give yourself.

1:51:45You know, this morning I watched a video on what exactly an NPU is and how it works. I won't say I know the answer. Now, please, for the love of God, don't ask me to explain it. But it's neat. And I have a better understanding of why it matters. And I think it's fun. What is that? Just default to going, go ask an LLM, go find a video, go whatever you want to do. I use voice mode for that. I love to use voice mode when I have something I want to learn because it's a two-way conversation. I can ask questions. It can ask me questions to help me understand better. And yeah, that's my best advice. Go learn something new every day.

1:52:30Try something new with an AI tool. The truth is there's a very, very, very limited number of people who are really good with these tools as users. We always talk about the people building them and about people using them for coding. But the fact is every single field right now in hiring is looking for people who are HR rock stars and are also really good with AI and are power users. You know, be that person in your field, for God's sakes. You know, just decide I'm going to be that person. take a couple classes on Coursera or something and spend some time with the tool and figure out how it impacts what you know about your industry and your specialties that you know well already.

1:53:15I think that's how you level up and become one of these people getting those cool salaries that we keep hearing about. In case you missed it earlier, it's 3 % at least. I think it's over 3 % wage increase if you are someone in an AI-exposed field and you know how to use the AI tools that companies want their employees to use. And if you look at job listings, it's a lot higher than that in a lot of places. There are big premiums. Yeah. Give that environment a week. We just – hey, Lumen, did you see? Yeah, Corey, you found it. Tell them about that. Yes. Was it in today's Neuron? No, two days ago.

1:53:59Two days ago. I found a Richard Feynman prompt that will teach you like Richard Feynman. And it was in the newsletter in a prompt tip of the day yesterday or the day before. And it's really neat. Basically, I took it through it in a project. And I was like, this way, when I want to learn something, I can go in and have it taught to me like Richard Feynman.

1:54:24it's nothing major, but it was just a cool little thing to share out that we thought was fun. Well, I think that's it for Gemini three flash and us today. I think, thanks for everyone for staying on. Yeah, it's been, it's been, it's been a really fun one. And we've run the gamut now from, from tiny models to employment problems. It's just what I'm thinking about. I think a lot of people are thinking about it. one thing I want to address in our predictions episode is I was reading a futurology thread last night where everyone was predicting what the year what the world will look like in 2065 40 years from now and everyone was so pessimistic and I understand where that comes from but it needs a it needs a balancing force of what could happen if we get some of this stuff right and so I'm excited to talk about that we're not going to predict 2065 but but I just see I know that a lot of people feel really pessimistic about a lot of this stuff right now.

1:55:25And I want to address that and talk about that realistically. Yeah. Yeah. And I think we should, too. And that, you know, there's and navigating through the things that there are to be excited about. But always having, you know, an ear to what we should be concerned about as well. And, you know, and trying to learn that full picture, I think, is really important and seeing it as a whole because it's become rapidly clear in recent months that the political approach is going to be, let's just divide it dead down the center by yes and no. And I feel like that is such a wrong approach, as is traditionally the way of politicians.

1:56:16see the world as you and make your own My favorite movie is Princess Mononoke for anyone who's seen it and basically the whole thing is like trying to see the world unblinded by hate and see it truthfully and that is what I try to do That's good, that's a good guiding principle Everyone who's here and stayed here we actually got a few more people now I'll be dang Well, to everyone who's here, thank you so much. We really appreciate that you joined us for these. Lumen, thanks for being so active today. And to everyone else who left a note, we really appreciate you. Please, please, if you have, take a moment to like, to subscribe, to do all of the things.

1:56:58It really helps us out and helps us grow the channel. We hit some massive milestones this year that we did not have on our radar. And we look forward to making those look like chump change in 2026. So lots of cool interviews, things going on, you know, low key. Some point we're going to have a robot on here. Not going to say who's yet or when, because honestly, I don't know when, but it's definitely coming. And yeah, we're excited to just be here learning with you all because we're learning. You know, I mean, the fact is when a model came out yesterday, we're going live with not much better of an idea than anyone else has.

1:57:37So we're, you know, we have played with things, but not everything we are talking about are we experts at. So thank you. I mean, as is everyone. Like, unless you work in the company where they were testing this for two weeks before, you're learning it as soon as it comes out like everyone else. Just like the rest of them. Yeah. You know, and I really appreciate that you all come in here and share what you've learned. And the fact is, you know, so often we do these and your questions, prompts, and ideas are so much better than ours. And we're really grateful for that, that you're here and doing it with us.

1:58:09So with that said, I hope you all go have some cool downtime. Enjoy the holidays with your friends and family. Get some vacation. At least get some time off work. Watch some movies on the couch. Be lazy. That's legit my plan. So, yeah. And we'll be live on the 30th. but we still have another episode coming out next Tuesday. So be on the lookout for that. It's a good one. We'll be publishing content all through the holidays as well. So we'll try to keep it relatively light. Just stuff you might have missed, cool tools you can check out in the new year, stuff like that. Yeah, the newsletter will still be coming out every day, I think, except maybe New Year's Day.

1:58:53Is it New Year's Day? I think we're going to just publish every day and see if people read it. Because so far, every day we've published. So far, if we publish, people read it. You know, and it's so many people that it's pretty easy to justify. It's really humbling. It really is. It's amazing. We couldn't do any of this without you guys who are listening and engaging and reading every day. So thank you. It's been really fun. Cool. Well, everybody, have a good week. Have a good holiday. We'll see you here next time. Farewell for now. Farewell for now, humans.

From the publisher

Google just dropped Gemini 3 Flash—a model that outperforms Gemini 2.5 Pro (their last top model) while running 3x faster at less than 1/4 the cost. It's frontier-level reasoning at Flash-level speed, and it's rolling out globally right now.


We're sitting down with Logan Kilpatrick from Google DeepMind to explore what this actually means for developers, knowledge workers, and anyone trying to figure out how AI fits into their workflow.


What we'll cover:
🔥 Live demos – Logan will show us Gemini 3 Flash in action, from coding to multimodal understanding


⚡ What's now possible – Use cases that weren't practical with previous models (or weren't possible at all)


🛠️ Building together – We might wire up a tool live if Logan's game (we've got ideas)


💰 Intelligence too cheap to meter – We'll dig into the economics: when AI gets this powerful and this affordable, does it change the hiring calculus?


On that last point: right now, data shows AI is raising wages for AI-impacted roles because workers who use AI effectively can command higher salaries. But what happens when frontier intelligence costs $0.50 per million tokens? When does “intelligence as a commodity” flip from “AI makes workers more valuable” to “why hire a human?” We’ll see if we can get Logan’s take on this topic!


Key specs on Gemini 3 Flash:
Outperforms Gemini 2.5 Pro across most benchmarks


3x faster than 2.5 Pro


Less than 1/4 the cost of Gemini 3 Pro


1M token context window


Advanced visual and spatial reasoning with code execution
78% on SWE-bench Verified (agentic coding)


Rolling out globally in Gemini app, AI Mode in Search, and developer platforms


Logan has been at the center of Google's push to make frontier AI accessible to millions of developers. If you're shipping products, building with AI, or just trying to wrap your head around where this is all going, this conversation will give you clarity.

More from The Neuron: AI Explained

All 106 episodes
Gemini 3 Flash (Smartest, Cheapest AI) with Google DeepMind's Logan KilpatrickThe Neuron: AI Explained · 1 h 59 min
Listen in VO