Episode Categories:
Resources:
Join the discussion: Ask Jordan and Logan questions on Google AI
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Try Our Free AI Prompting Course: Register for our free Prime, Prompt, and Polish AI Course!
Google Gemini 2.0 - Why It Matters
Gemini 2.0 is more than just a simple update—it’s an expanded family of models designed to integrate seamlessly into numerous AI use cases. Earlier releases like Gemini 1.5 garnered attention for their high performance, but 2.0 surpasses its predecessor by introducing new variants:
- Flash
- Flash-lite
- Pro (Experimental)
Each model serves different performance and cost requirements. For instance, Flash-lite is more cost-efficient for high-volume tasks like summarizing web pages or handling basic emails, while Pro excels at coding and intricate problem-solving. By offering these diverse models, Gemini 2.0 aims to address the wide-ranging needs of both enterprise and startup users.
“The cost per million tokens has gone down by an order of magnitude, allowing more developers to quickly ship AI products,” notes Logan Kilpatrick.
Cost-Effectiveness Drives Adoption
Over the past two years, using top-tier large language models (LLMs) could be prohibitively expensive, especially for indie developers or lean startups. Today, Gemini 2.0 changes that equation:
- Lower per-token costs
- Comparable or better performance to older, pricier models
- Improved logic and reasoning capabilities for real-world applications
For businesses worried about rapidly depleting budgets, lower overhead costs open up more room for innovation and experimentation. Meanwhile, entrepreneurs can turn big ideas into MVPs within days (or hours) rather than months.
Progress in AI and the Rapidly Evolving Landscape
With headlines often declaring that AI has “peaked” or is “hitting a wall,” Gemini 2.0 stands as a compelling counter-narrative. According to Kilpatrick, the innovation pipeline continues to surge forward:
- Larger context windows
- Better “chain-of-thought” or reasoning processes
- Tighter integration with agentic workflows
In short, the leaps we’ve witnessed over the past six months show no sign of slowing. Where older models might struggle with complex tasks—like debugging multifaceted code or parsing intricate, multi-step reasoning—Gemini 2.0 handles them more fluidly, thanks largely to new architecture and optimization.
The Role of Reasoning in AI
A critical selling point of the latest Gemini models is improved reasoning capacity. Previously, AI tools often relied on best-guess word prediction, which could result in hallucinations or nonsensical replies. By contrast, Gemini 2.0 introduces a more robust chain-of-thought feature. This ensures the model follows logical routes, referencing external data and re-checking its own conclusions, thereby lowering error rates.
“What we need are agentic LLMs that not only gather facts but also make sense of them autonomously,” explains Jordan Wilson.
Multimodality: Beyond Text-Only Interactions
One of the most anticipated features of Gemini 2.0 is its fully multimodal capability. While earlier LLMs focused predominantly on text, Gemini 2.0 supports:
- Text
- Images
- Video
- Audio
Why Multimodal is a Game-Changer
Modern workplaces operate in an information ecosystem that includes audio notes, images, videos, and text documents. Historically, AI could process these formats but struggled to connect them seamlessly. Gemini 2.0’s multimodal approach bridges these gaps, empowering developers to build apps that interpret, analyze, and generate different media types in a single workflow.
For instance, a healthcare startup might feed MRI images alongside textual patient records into Gemini 2.0 for more accurate diagnoses. Meanwhile, content creators can drop raw video clips and text outlines into an AI platform that processes both to craft a cohesive storyline—complete with subtitles and audio descriptions.
“The model’s ability to handle complex audio-video tasks straight ‘out of the box’ unlocks new applications in medicine, entertainment, and beyond,” Kilpatrick observes.
Paving the Way for AI Agents
While multimodality and improved reasoning garner most of the headlines, Gemini 2.0 also sets the stage for AI agents. Tools like Project Mariner (by Google) or Operator (by OpenAI) demonstrate how LLMs can handle tasks without continuous human supervision—from scheduling meetings to managing entire codebases.
Common Roadblocks
Despite major progress, fully autonomous AI agents still face hurdles:
- Trust and Validation
Users remain cautious about handing over tasks that directly impact business operations or finances. - Proactivity
Many AI tools still wait for user commands rather than proactively surfacing valuable insights or alerts. - Complex Productization
Building robust, user-friendly interfaces for agentic AI requires significant time and specialized design skills.
In Logan Kilpatrick’s words:
“We’re crossing off the final technical barriers, but building a great product experience still takes time. Agents are coming soon, but they’ll need more polished user flows and robust guardrails.”
How Developers Can Leverage Gemini 2.0
From independent creators to enterprise software teams, Gemini 2.0 lowers the barrier to creating powerful AI-driven applications. Here’s what developers can do right now:
- Experiment with Model Variants
Start small with Flash-lite for prototyping or large-scale text summarization, then migrate to Pro for deeper logic tasks like advanced code generation. - Tap Into Multimodal Inputs
Break free from text-only pipelines. Incorporate images, voice notes, or short video clips to see how Gemini 2.0 can unify insights across mediums. - Prototype Agentic Features
Build AI workflows that proactively scan data, highlight anomalies, or draft full solutions before you even ask. Gemini 2.0’s improved chain-of-thought can handle more complex instructions. - Iterate Rapidly
The big advantage of cost-effective LLMs is how quickly you can test, gather feedback, and refine a product. Don’t over-engineer your first iteration—get an MVP out to users fast.
The Future of Work: AI as Your Copilot
Both Jordan Wilson and Logan Kilpatrick envision a near-future where AI quietly, yet significantly, shapes our day-to-day tasks:
- Personal Email Filtering
Your inbox might be auto-managed each morning, surfacing only the high-value threads. - Proactive Summaries
A specialized agent could review project documents and deliver daily summaries without extra prompting. - Deep Organizational Insights
AI could integrate across multiple departments—HR, Finance, R&D—finding patterns and offering actionable data in real-time.
“We’re not fully there yet,” Kilpatrick notes, “but the path is clearer now that we have better reasoning models. By the end of the year, I expect to see more fully-fledged agentic workflows.”
Advice for Entrepreneurs and Tech Enthusiasts
Build for Niche Solutions
While Google and OpenAI compete in expansive domains like chat-based search, there’s ample opportunity for verticalized solutions. Smaller startups can tackle sector-specific challenges—be it agriculture, cybersecurity, or real estate—using Gemini 2.0’s advanced features.
Evolving Dev Paradigms
Developers no longer need years of coding experience to launch software products. AI-driven coding assistants drastically reduce the learning curve, enabling technical founders—and even some non-technical ones—to swiftly test new concepts.
“It’s not about being overshadowed by Big Tech; it’s about finding your lane and solving real problems efficiently,” explains Kilpatrick.
Leverage Community & Support
With over four million developers reportedly working with Gemini, robust forums and documentation exist to fast-track your learning. Whether you’re an individual builder or a corporate team, taking part in open communities can yield tips, free resources, and best practices.
Final Thoughts: Seizing the Gemini 2.0 Moment
Google’s Gemini 2.0 marks a critical inflection point in AI’s evolution. By blending lower costs, stronger logic, and truly multimodal capabilities, it paves the way for widespread adoption—unlocking new applications in healthcare, finance, education, and beyond.
For businesses, the message is clear: Innovate now or risk being left behind. The days of burning through massive budgets just to build a small AI prototype are over. Gemini 2.0’s flexible model variants make it feasible to scale incrementally, only paying for the performance you need.
For individuals, there has never been a better time to learn, experiment, and innovate. Whether you’re coding a Chrome extension in an afternoon or exploring entire agentic workflows for enterprise clients, the resources and community support have never been richer.
Topics Covered in This Episode
1. Google’s Gemini 2.0 Update
2. Rapid Progress in AI and LLM Capabilities
3. Multimodal Features of Gemini 2.0
4. Google's Agentic AI Projects
5. Future of Work & Personal Productivity
Podcast Transcript
Jordan Wilson [00:00:17]:
If you follow AI developments at all, it has been an extremely exciting couple of months. And I think that excitement even peaked even more in the past couple of days as Google just released and expanded its Gemini two point o family of models. And I think that there's a lot of aspects of Gemini two point o that a lot of people are overlooking, and I think it can fundamentally change how we work. So we're gonna be talking about that and a lot more today on everyday AI, and I'm extremely excited for today's guest because let me just say this. If you're not following AI every single day, you can learn a lot from our guest. I think he is probably one of the smartest people in AI and has really shifted how the world works, whether you all realize it or not just with his work background. So, thank you for tuning in. Maybe it's your first time.
Jordan Wilson [00:01:12]:
What's going on? My name is Jordan Wilson, and welcome to Everyday AI. This is for you. It is your daily livestream podcast and free daily newsletter, helping us all not just learn AI, but how we can leverage all of these, new updates, new model upgrades, how we can use this to actually grow our companies and our careers. It doesn't make sense to just know what's going on. You have to be able to put hands at keyboard or voice, right, voice in microphone and be able to actually grow something. So that's what this is all about. And your next best friend, if this is your first time here, is our website, youreverydayai.com. There, you can sign up for our free daily newsletter.
Jordan Wilson [00:01:50]:
We're gonna be recapping today's conversation and a whole lot more. But while you're there, we've talked to hundreds of AI experts around the globe. You can go listen to it all for free there. So make sure you go check that out. And if you are looking for that normal AI news, that's gonna be in the newsletter. Normally, we go over in the beginning, but I wanna make sure to squeeze every single minute we can with our guests. So, enough chitchat from me. I'm excited, to have on.
Jordan Wilson [00:02:15]:
Another Chicago guy. Love to see it. Alright. Please, Live stream audience. Help me welcome Logan Kilpatrick, senior product manager at Google DeepMind. Logan, thank you so much for joining the Everyday AI Show.
Logan Kilpatrick [00:02:27]:
Jordan, thank you for having me and for the, for the overly generous, intro and for getting me out of bed at, 07:30 in the morning. So I'm excited for this conversation.
Jordan Wilson [00:02:35]:
Yeah. It's it's it's always funny. Right? When people are like, yeah. Yeah. I'll do the podcast. Then I'm like, oh, it's this time. Yeah. So you definitely coffee in hand.
Jordan Wilson [00:02:43]:
But you know what? I'll say this, Logan. I don't think it was overly generous. So people that maybe don't know your background. Right? You started, you you know, before this, you were at OpenAI, and now you're leading a lot of, you know, AI projects at Google. So before we jump into the new two point o updates, what's new, what's shiny, what's helpful? Can you just quickly tell everyone about your background for those that don't know?
Logan Kilpatrick [00:03:02]:
Yeah. Yeah. Yeah. So currently lead product for Google AI Studio, one of the colleagues for the Gemini API. We're we're really focused on how do we enable developers to be successful building with Gemini. Before this, led developer relations at OpenAI, was there, joined when it was a small start up, third people, and got to see that scale out to the the crazy the crazy company that it is today. And then before that, was actually doing another start up and, machine learning and deep learning for digital pathology. Before that was a was a machine learning engineer and and a software engineer.
Logan Kilpatrick [00:03:34]:
So I've I've sort of done the whole spectrum of work from training models to writing software to building products, and, I love this stuff. It feels like this is the, in hindsight, all of that experience has been, like, the perfect thing to help me be successful in this moment. But, yeah, it's been a ton of fun so far.
Jordan Wilson [00:03:52]:
Alright. Before we get into all the new updates, I just wanna get your take on just the pace of AI right now. I think especially over the last two months. Right? I've been doing this every day for more than two years. The last two months for me have been so hard to keep up with even though I do it every day. So between, you know, we've seen a lot of new stuff from your former employer at OpenAI, Google. I mean, between what you announced in December, end of January this week, it's been nuts. Where are we at right now with the state of AI and and large language models and capabilities? And, you know, is it hard for you to keep up too, or is it just me?
Logan Kilpatrick [00:04:28]:
Yeah. That's a great question. I feel like you you and I are both in the same boat that we're we're there's so much going on. Like, it depends on, like, what you're trying to keep up with. I I feel like I have in the sort of developer world. There's a substrate of stuff that I care a lot about, and there's a lot of stuff that's happening in in perhaps, other domains that is just, like, less applicable to me on a daily basis. So I can filter some of the noise out. But it is it is interesting you talk about this pace of innovation, and it's funny how the narrative like, the broader narrative shifts from, like you know, two months ago, everyone was talking about AI's hit a wall, and, like, that was the narrative in the media.
Logan Kilpatrick [00:05:02]:
And, like, there was, you know, there's no more progress to be had on the model side. And then, like, you look at what has turned out to be, like, a bunch of actually incredibly substantive model progress in the last two months, and it's like it it is yeah. It it's interesting to see that, and I think the takeaway for me is, like, the model progress is not going to stop anytime soon. And, actually, I think there's going to be a lot of, product level progress. And if you look at, like, some of the new agent products that people have been putting out recently in, like, Mariner, OpenAI's operator, etcetera, etcetera, it's this, product experience that's, like, actually powering the sort of, like, in combination with a frontier model, this new state of the art experience that wasn't possible before. And I think there's we're we're gonna just see more and more of that as people as the model capability starts to really unlock a bunch of new use cases that weren't possible. And, like, that's literally happening today. Like, like, the Gemini two point o model, like, might have been the thing that unlocked a bunch of the use cases that you as a consumer developer wanted to have it do and it couldn't do before, and, like, now it just works, which is crazy.
Jordan Wilson [00:06:08]:
Yeah. It is it is crazy in some of the capabilities. So let's jump in and, hey, live stream audience. If you have a question for Logan, let's get it in now. But, Logan, like, walk us through what's new in Gemini two point o. And and maybe, you know, for for our audience, if you don't follow it every day, don't worry if it's confusing because there's always so many developments because, you know, we had some Gemini two point o, you know, back in December, but now we have a whole new wave of updates. Logan, what's new and what does it mean for how we work?
Logan Kilpatrick [00:06:37]:
Yeah. So the this moment, for this week was really an expansion of the Gemini model family. So we released the initial experimental version of two point o Flash back in December. That was sort of the initial moment. The reception was super positive. Developers have been getting pinged every every day being like, hey. We love this model. We wanna go and build with it.
Logan Kilpatrick [00:06:55]:
This is the thing that's gonna enable our business, which has been super exciting. So we've been pushing really hard on how do we make the two point o flash model ready for production. The model is actually, it's a upgraded variant of that model, so it's a further improved version of that model. But we didn't stop there. So there's two other additional new models. The Flash-lite model, which is a smaller, lower cost variant of the main two point o Flash model. And this is really for, like, the high scale, you know, production workloads where you're you're the use case is very cost sensitive and doesn't need to have, like, super, super high accuracy. And you can imagine a bunch of, like, basic examples for for this model, like, you know, email summarization or, you know, you're trying to summarize a bunch of different web pages or something like that for, putting together a report, things like that that, you know, the model doesn't need to be super capable.
Logan Kilpatrick [00:07:46]:
And then the other end of the spectrum, we released Gemini two point o pro experimental. So this is the first sort of officially branded two point o pro model. It is the, the the predecessor of the Gemini 12 o six model that folks were familiar with that, which was our previous strongest model. And it's, again, an updated variant of that model. It's a further improved version of that model, which really excels. Like, the main use case that two point o Pro excels at is, like, I think coding is, like, the thing that is, like, far and away better at. And I'll I'll make one other quick comment, which is, you know, for folks who have been following some of the discourse around some of the new models, I've I've seen a lot of reactions from folks being like, you know, the two point o pro model, like, doesn't actually and you can see some of the metrics if you're watching, this conversation on somewhere where there's video. But the metrics are, you know, 3%, four %, five % better, between Flash and Pro.
Logan Kilpatrick [00:08:42]:
And I think there was questions from the the community being like, hey. Why is it why is that the gap right now? And I think there's two things that are true. The first thing that's true is, to me, this is like a success story of small models. Like, there's been all this work and innovation that's happened of, like, how do we take the frontier capabilities, bring them into smaller models so that, like, at scale and at a low cost, developers can put this stuff into production. So I think we've we've been successful doing that, and to me, that's why there's sort of this continually shrinking gap between, like, the total frontier and and some of the smaller sized models. But separately, like, I don't think you know, if if you're not someone who thinks about, like, metrics or, like, looks at evals and, like, really understands what's happening, there is this nonlinear amount of additional work, but also capabilities that come as you, like, continue to move up the frontier. So, like, moving from, you know, moving from, you know, 80% to 85% could be, like, could be this, like, not only is it this massive amount of work, but also, like, the order of magnitude of use cases that, like, that 5% increase unlocks is actually pretty crazy relative to the previous 80%. So it's like it's this, like, exponentially difficult curve that you have to go up.
Logan Kilpatrick [00:10:01]:
So lots of interesting stuff around that, and happy to chat more about it.
Jordan Wilson [00:10:05]:
Yeah. And and, yeah, the benchmarks are obviously very impressive. Gemini, you know, it has been kind of this back and forth battle over the last couple of months, but I think right now, it's it's it is the highest, you know, bench model in the world in terms of, you know, Elo scores, which I think are important. You know, it's it's which do, you know, real world humans prefer. But, you know, I wanna get into some of the, maybe overlooked aspects of Gemini two point o because Logan, like, I know AI, you know, runs so fast, but it is the only true multimodal large language model out there right now, right, where you can input video. Right? You don't have to just input a photo or text. Can you talk a little bit about some of the improved multimodality capabilities of this two point o model?
Logan Kilpatrick [00:10:50]:
Yeah. No. A %. So I think when we initially put out Gemini back at the end of twenty twenty three now at this point, The sort of the sort of title was, you know, building a model for the multimodal era. And at that point, it was I think, like, we we had the research direction to enable this to happen, but the models at that point were not, like, fully multimodal input output. We're now actually at that point. So, like, Gemini two point o is sort of the delivering on the mission of, like, actually making the models fully multimodal. So being able to take in, text, audio, video, images.
Logan Kilpatrick [00:11:23]:
And now actually and and this is still an early access, but rolling out to more folks, broadly, hopefully soon. The model can actually output, audio and images as well. And I think this next step is super important. And the example that I was responding to some to some tweet threads yesterday because we released our imagine three model, which is a sort of separate domain specific image generation model. And folks were asking, hey. You know? They're we're not they were saying, we're not super interested in imagine three. We really want this native multimodal capability. And, really, there's this interesting, there's this interesting trade off between these different types of models.
Logan Kilpatrick [00:11:59]:
Like, the imagine domain specific image generation models are really good at generating these, like, beautiful, like, sort of artistic picture perfect images. Mhmm. And if you look at what the the Gemini models are able to do, it's like, really, they benefit from all this world knowledge. And if you think about, like, like, lots of image recognition, task where there's, like, a lot of complexity and you're maybe looking at a I'm looking at my kitchen table right now and there's, like, lots of objects sitting on it. You have to understand the relationship between these objects and understand physics and understand, like, bounding boxes and all this stuff, like, really complex. And you can't do those tasks unless you have the baked in world knowledge of of a large language model. And, again, those, like, domain specific image models don't do that. So there's all I'm really excited about this because there's all of these interesting use cases which are now sort of just going to work out of the box because of how good these models are, and also for Flash specifically because of how, like, cost effective it is for for developers.
Jordan Wilson [00:12:59]:
It's yeah. It's getting you know, if if if you were to show me these benchmarks, like, two years ago, I'd be like, no. Not possible. Right? It's not possible to be that fast, that cheap, and that powerful yet here we are. But, you know, Logan, I wanna talk a little bit. You know, we kinda just talked about multimodality, which I think was, you know, all the buzzword, you know, maybe two years ago. But now the conversation has really centered around, agentic capabilities. And, obviously, you know, Google, has had a lot of different, you know, kind of announcements in the space.
Jordan Wilson [00:13:30]:
And, you know, I think, OpenAI's operator has has caught a lot of people's attention so far. I'm personally excited to see Mariner. Can you tell us what that is and maybe even your thoughts on agentic work in the future, and and how does that change the human's role in all of this? Are you still running in circles trying to figure out how to actually grow your business with AI? Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on GenAI. Hey. This is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft, and NVIDIA have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use GenAI. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you can partner with us too, just like some of the biggest companies in the world do.
Jordan Wilson [00:14:36]:
Go to your everydayai.com/partner to get in contact with our team, or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on GenAI.
Logan Kilpatrick [00:14:55]:
Yeah. No. That's a great question. So for folks who who didn't see, when we did the two point o Flash release back in December, we also released a bunch of sort of experimental research around the direction we were going with agents. So we released project jewels, which is a coding agent, and you can sort of have it living in your GitHub and and helping you solve issues and writing code and doing pull requests for you, which was really cool. We released project Mariner, which allows you to, basically have the model, you know, coexist with you in your Chrome browser and, like, actually sort of you know, you can talk to it and have it take actions on your behalf, and you could sit there with the model and and sort of it does work for you in some cases. And then we also released a a data science agent, inside of Colab, which is a a sort of, Jupyter notebook if folks are are familiar with that, and you can do sort of data science tasks. And I think all three of these things were sort of validation of of the sort of fundamental research that was being done, which is how do we enable the next step in this in this journey, which is being able to build agents.
Logan Kilpatrick [00:15:58]:
And I think there's a lot of different ways that people have tried to build agents and will build agents in the future. The thing that's most interesting to me is how intertwined all these different pieces are. So I think, like, if you think about, you know, why do we not already have large scale at deployments of, of agents? Like, why why do I not have a hundred agents? Like, I'm I'm I feel like I'm living on the cutting edge of AI, and, like, I don't have a hundred agents doing my bidding every day, you know, showing up to podcast to talk for me, sending all my emails. Like, I don't have that today. And if you think about, like, why don't you have that today? I think it's just been historically that a lot of the fundamental capabilities weren't there. Like, for multimodal, for example, like, the way I interact with the world is in a multimodal capacity. Like, I I see things. I hear things.
Logan Kilpatrick [00:16:43]:
You know, there's some processing video all the time, and that just historically hasn't been good enough. Like, what else hasn't been good enough? The models can't actually reason. Like, they kinda just have to you know, it'd be like if you were forced to not think anymore and you had to just, like, give your gut take at every single question or every single interaction that you've had. And I think we're sort of getting that we're solving that problem now with our with our reasoning model. So I think we're the sort of the dominoes are falling in order to make this agentic workflow actually happen, which I'm excited about because I feel like if you look two years ago at at sort of what was being promised to the world with this technology, like, it really is all those agent workflows. It really is like, hey. I have this thing, and the model is going to go and do work for me, and it sort of removes that burden from my life. I don't feel like we're fully there yet, but I am I am excited because I think the the direction, at least on my capability side from a model perspective, is, like, very clear.
Logan Kilpatrick [00:17:38]:
Like, we're going to get there. I think people need to build the products around it to make it to work well, but, yeah, it's this is gonna be a great year for for agents.
Jordan Wilson [00:17:47]:
Yeah. And and, you you know, you kind of mentioned reasoning and and thinking models. You know? I was playing right after the announcement. I was playing with, Flash thinking. And although it didn't get every single thing right, I was actually less worried about that, and I was more taken aback and more impressed just by its level of reasoning and in kind of looking at that chain of thought. But, you know, you kind of mentioned Logan like, hey. Like, I'm I'm not currently, you know, orchestrating hundreds of agents myself, and you live on the cutting edge. So I'm wondering maybe why or what are the hurdles that we have to clear because we have power even just at one company, Google.
Jordan Wilson [00:18:23]:
Right? We have extremely powerful multimodal models. We have agentic capabilities, and we have very capable reasoning models. So it's like all the individual pieces are there. So maybe what are those hurdles, and and kind of, what has to be done to get over those until we do have that more, you know, quote, unquote, future version of work?
Logan Kilpatrick [00:18:43]:
Yeah. No. This is a great question. And I actually had a conversation with someone yesterday trying to dig into what what is the answer to this to this question. I think maybe some of this is personal preference, but part of what, part of what I think has been missing is the models really being proactive in in a bunch of use cases. And I think there's there's two there's two layers to this. The first is, like, what what I want is not to have to instruct the model in a bunch of cases. Like, hey.
Logan Kilpatrick [00:19:13]:
Go do this work for me. I think that maybe happens, like, one time as, like, a startup cost. But then what I really want is the model proactively, you know, for the email use case, going through my email every day, looking for, hey. Here are some maybe high value things that I missed or that I didn't have time to to look at and, like, compiling that into a report and then actually pushing that information to me. Like, I I, if I have to be in the driver's seat of that of that agent experience, like, at least on my personal life, like, it's not going to happen. Like, I don't have enough time. There's too much other stuff going on. Like, what I need is the model to sort of take the burden of, like, following up with me and and making this happen, sort of what you would expect, like, a really good assistant to be able to do for you.
Logan Kilpatrick [00:19:54]:
And if you just look at, like, what are the products that people have built so far using this technology, it isn't that experience. And I think this is, like, one of the big, the the big gaps that still exist. And I I think there's, like, parallels to what has happened in the last two years with AI going into production and, like, you know, in a lot of demo use cases and a lot of, like, really simple chat use cases, and then it's, like, kind of people are sitting there thinking, you know, was this really all worth it, all this effort and stuff like that? I think a lot of this is, like, the product experience needs to be better. And I think the models and the technology and the scaffolding is finally good enough to enable this to happen, but, like, it just takes time to build great products. I think it's the simple answer to this. And, like, start up start ups and, like, entrepreneurs and, like, you know, even the large companies, like, just need time to make a really great product experience. And I think we're gonna see those in in in in 2025 and, like, Mariners, hopefully, sort of the early work to make that happen. Operators, hopefully, the early work to make that happen for OpenAI and and for others.
Logan Kilpatrick [00:20:53]:
So I think we're gonna see more of it, which I'm excited about.
Jordan Wilson [00:20:57]:
So one, one thing that I love asking people who are building the AI that we all use is, you know, I think there's a lot to be said on, you know, learning how they eat their own dog food. Right? Like, right now, I can't do anything without using Google deep research, without using NotebookLM. Right? Even this very podcast, every podcast I do. Right? I use those two tools every single time. What are some of even whether it's your own internal Google tools, but, you know, what are some of those kind of AI tools or processes that you can now not live without that maybe three years ago weren't even on your radar?
Logan Kilpatrick [00:21:32]:
Yeah. I think for the thing that I use AI for probably most successfully, I do the deep research stuff for my podcast, and it's incredible for that. But it's still coding for me, honestly. Like, I think I think that's just the use case that provides the most value, and I think it's also the one that's, like, fundamentally changed the way I operate. Like, I think I was a software engineer, so, like, I lived the life of, like, I am writing this code from scratch and solving this problem from zero by myself, which is just crazy to think about. And I think now and I was having this debate, with with folks internally this week actually about, like, what what is our interview process actually like for some of these, like, future jobs? And, like, should we be you know, how much do we wanna actually check that a software engineer should be, like, doing this stuff from scratch versus, like, using an AI tool? And it's, like, a really interesting, like, philosophical debate as we're, like, in the middle of this fundamental shift for how software is created of, like, should we be sort of going for people who are really good at using AI tools and, like, know how to, like, get the most out of them but are still, like, good at software engineering? Should we go for people who can just, like, do it all from scratch but maybe, like, don't use the AI tool? So it's this really weird situation, but I think I'm I'm in AI studio all day. I'm using Code Assist, which is our our sort of coding extension. I'm in Cursor all day using the latest Gemini models.
Logan Kilpatrick [00:22:57]:
Those are probably the things that are are most, are most pertinent to me. But, yeah, I'm I'm also curious for you. Like, what are the what is the, like, top three tools other than Deep Research and NotebookLM that you're that you're working with?
Jordan Wilson [00:23:10]:
Yeah. Those those two I can't live without. Right? Obviously, I've I've been loving OpenAI's version of Deep Research that just came out as well. I I like using those. I like using tools in tandem. Right? And and and I look forward to the day when there's just one system that, you know, you can use all these different tools and features. But, you know, I I I love just anything deep research, anything, agentic. Right? You know, it's my workflow.
Jordan Wilson [00:23:37]:
I use way too many AI tools every single day. Right? So I'm actually trying to use less, and and and kind of, build more. But even another example of that. Right? Like, I was using, Google's, in AI studio, the stream real time. I was running into an issue and I'm like, you know, I'm using deep research to to try to troubleshoot this issue. And I'm like, no. You know what? I'm gonna I'm gonna do. I'm gonna go into stream real time.
Jordan Wilson [00:23:59]:
I'm gonna share my screen inside Google AI studio, and I'm just gonna quickly build, in a a Chrome extension. Right? You you mentioned, like, you know, coding and development. So I'm curious. You know? That's something I did. I never thought I would be building Chrome extensions, but it takes me, like, minutes now. How do you see the future as as someone with a background in development? How do you see that that future? Is it going to be where, you know, people like that are just gonna be spinning up their own apps on the fly?
Logan Kilpatrick [00:24:24]:
Yeah. I I do I fundamentally believe this. I think if you look at how like, the number of people who we would consider, like, software developers today, I think that's going to transition to, like, AI builders or software creators where the artifact that is actually being created in that process is code. How much of that code is being written by the person who's actually creating it, I think is going to continue to go down over time. I do think, like, there'll be some you know, people who know how the software works behind the scenes are going to have, like, this, you know, competitive advantage, essentially. Like, if you can go in and troubleshoot and like, I've used a lot of the AI tools that, like, help you go from, like, text idea to, like, working app and, like, sometimes things break and, like, being able to know and, like, look at code and understand what's happening is incredibly useful. But I do think, like, that is the new frontier to me. The new frontier is, like, letting and enabling every single person in the world to be able to to make their idea come to fruition.
Logan Kilpatrick [00:25:23]:
And I I had a bunch of with with friends in Chicago, actually, like, people who have really interesting ideas that they wanna see come into the world, and these folks, like, aren't software engineers. Like, it's actually really hard if you're not using one of these AI tools to bring any of your ideas to life. So I'm excited. And, like, this is actually something we think a lot about for AI Studio as well. It's like, how do we remove this barrier to people who wanna build? So I'm excited. Hopefully, we'll we'll have more to share around that soon and, and keep building, experiences that remove the barrier for people to create software, and and actually, like, make it accessible to the rest of the world.
Jordan Wilson [00:25:59]:
You know, speaking of building software, you know, I I I know a lot of, you know, entrepreneurs, start up people are are probably listening in. And, you you know, I think over the last two years, it's become increasingly both easier, to to build a company. Right? I think they're I I read, like, 4,000,000 developers or something like that are are using, Gemini. Right? Something like that. Crazy. So what advice do you have for for those people in a time where it's technically easier than ever to validate an idea or bring something to market yet at the same time, moat is so hard because a company like a Google or a Microsoft or an OpenAI can, you know, have one new feature update and kind of, you know, kill that moat or or sync a startup. How should, you know, entrepreneurs be building and, you know, maybe how does Gemini help in that?
Logan Kilpatrick [00:26:47]:
Yeah. No. This is a great question. I I think there's like and and we'd probably need, like, a thirty minute conversation just about this to to do the deep dive, but I think the really quick lens that I look at this through is, like, what is the scale of company you wanna build? I think there's, like, really interesting problems to be be solved, that would create a lot of value for yourself and perhaps, like, a small business. And, like, you could make a bunch of money doing that, and, like, you don't need to be venture scale. You don't need to raise money. I think the the, like, really interesting thread of the AI moment is, there's sort of these a bunch of intersection intersecting curves. And, like, one of them is the cost of building with AI and using AI has gone down, like, 99.9% in the last two years.
Logan Kilpatrick [00:27:29]:
So if you look at, like, the cost per million tokens of GPT four at, like, $30 per million tokens to today, the cost per million tokens of Gemini two point o Flash, 10¢ per million tokens, like a massive cost reduction, while the capabilities have actually improved. Like, the models are better and can do more for you. And at the same time as that curve is going down, you look at what is the consumer willingness to pay for AI technology, and the awareness of AI. And it's like, actually, consumers are willing to pay a lot more if you can actually provide them value, and more and more people are realizing, hey. These things can actually do stuff for me. So this is like this for people building stuff, it's this beautiful situation where, like, your costs are going down. You're you have more users showing up, and your users are willing to pay more money. And to me, like, that that gets me so like, the the reason I love my job is because I I think we get to build stuff for people who who wanna who wanna build.
Logan Kilpatrick [00:28:25]:
We're building for builders, and and I think that I I don't think that this is going to change, this, like, willingness to to invest in AI, while the cost continues to go down for people building. And and that just means, like, the value creation is going to builders. I think to hit on the point of, like, you know, are you going to be disrupted by Google or OpenAI? Like, not not in this context, really. Like, I think if you're going after like, if you're trying to compete with OpenAI at being a chat app or you're trying to compete with Google at being searched, like, I think that's a different situation. I'm like, I have this very specific problem I'm trying to solve, and, like, it's not one of those two, like, very, very horizontal problems. I think there's just, like, this massive amount of value to be created, and, like, it's never been a better time in the history of the world. And the, like, last really quick curve is, like, the time like, how quickly some of these new AI startups are, like, scaling to and, like, getting profitable and monetization is crazy to me. Like, I just saw a chart yesterday that Cursor is, like, I think, the fastest company one of the fastest companies now to reach, like, a hundred million ARR, MMR, whatever the number is for them, which is just crazy to think about.
Logan Kilpatrick [00:29:36]:
And, like, it's it's this, like, massive breakout success that's enabling builders to go and build stuff, which is so cool.
Jordan Wilson [00:29:43]:
Alright. So, we we we have to land this plane here in a second, but hoping we can get rapid fire, questions here. So real quick, a question from LinkedIn. LinkedIn, Someone asking what, outside of Google models, what are some of your favorite LLMs? Also, what are some of your go to AI tools that you use regularly?
Logan Kilpatrick [00:30:03]:
Yeah. Yeah. Yeah. That's a great question. I mean, I I've spent a bunch of time using the OpenAI models. I think the anthropic models are great for coding. I think those are probably the three. I use Gemini the most, and then I've dabbled with a bunch of the anthropic models and then built a bunch of stuff with OpenAI models and go to AI tools.
Logan Kilpatrick [00:30:24]:
I'm in AI studio all day, obviously, because it's it's part of my job. But the Gemini app, I think deep research, I it's rolling out in iOS, which I'm super excited about next week. And I'm I'm waiting for that to happen because I I don't I'm always on the go and have my ideas, and it's hard to be strung to a desktop. So I'm excited about that.
Jordan Wilson [00:30:45]:
Oh, can't wait for that one. Alright. Last one. Pedro from LinkedIn asking, Logan, what will agents be doing for you in quarter four? So by the end of the year. Right? You said we're not there yet, but what are agents gonna be doing for you by the end of the year?
Logan Kilpatrick [00:30:57]:
Yeah. This proactive stuff. Like, the the product experience I want is I wake up in the morning, I have an app that shows me the list of all these potential tasks that the agent wants to go and execute for me. I sort of review and see, yep. I want this to be done. Nope. I don't think this is worthwhile for you to do or change the plan or do whatever. Like, that's the product experience I want, and it's, like, material, like, actually high value work, not just, like, you know, AI that's auto drafting a response to a spam email that I got.
Logan Kilpatrick [00:31:26]:
Like, I don't need that product experience. I needed to, like, know that, oh, no. This is actually, like, a conversation with Jordan about AI or, you know, whatever it is. Like, something that's actually high value. And I think we're gonna get there, which is exciting.
Jordan Wilson [00:31:38]:
Alright. We covered a ton in today's episode, Logan. Can't thank you enough for coming on, to share with our everyday audience. So, thank you so much. Last last takeaway. What's the most important thing people need to know about Gemini two point o?
Logan Kilpatrick [00:31:53]:
It's available today. I think you can go start building with it, the best price per performance of any model on Earth, which I'm happy about.
Jordan Wilson [00:32:02]:
Love to see it. Alright. Logan, thank you so much for joining the Everyday AI Show. We really appreciate it. We covered a lot, everyone. We're gonna be recapping it in today's newsletter. Logan talked about a lot. All those links are gonna be in there, so please go to youreverydayAI.com.
Jordan Wilson [00:32:15]:
Sign up for that free daily newsletter. Thanks for listening. Thanks for tuning in. We'll see you back for more everyday AI. Thanks, y'all.