← Marketing Against The Grain
Loading summary
Transcript65 lines
- [00:02]
Kip
Hey everyone. I am excited. We have a very, very special guest. Logan Kilpatrick from Google DeepMind is here and he's going to break down the aspects of Google Gemini AI that people aren't using that they should be going to talk about where the real value is, going to talk about their brand.
- [00:18]
Phil
New image generation API, their new reasoning.
- [00:21]
Kip
Models, deep research, so much. If you are trying to learn about Google AI, this is the show for you. Let's get to today's episode.
- [00:31]
Phil
We'll be right back to today's show. But first, here's a quick word from HubSpot Marketing in 2025 is wild. Savvy customers spot fake messaging instantly, privacy changes make ad targeting a nightmare and everyone needs more content than ever. That's why you have to have HubSpot's new marketing trends report. It doesn't just show you what's changing, it shows you exactly how to deal with it. Everything's backed by research but focused on marketing plays you can use tomorrow. Ready to turn your marketing hurdles into results. Go to HubSpot.commarketing to download it for free.
- [01:23]
Logan Kilpatrick
Logan, really excited to have you on the podcast. Big follower of yours on Twitter for a long time or X for a long time. Excited about your move to Google. Have been following that as well. But let's maybe start not at the start for you, but maybe at the start for some of us who started following along your journey with being an early employee over at OpenAI, like one of the most transformative companies there is. One of the things we want to start with is it must have been a pretty wild time to be part of that company. What was one of the most exciting launches that you were there for? One of the ones where you were like, wow, this is just going to be much more impactful than anyone really gives a credit for.
- [01:56]
Greg Brockman
Yeah, that's a good question. I think the one that always stands out for me and obviously in hindsight, we all know how impactful this launch was. It was really simple and sort of had all the fanfare as far as like actually being a part of it on the ground was honestly the GPT4 launch. When we launched GPT4, if folks remember the like live stream that Greg Brockman did where he sort of talked about the models and he sort of did the infamous example of like drawing a little picture of a website on a napkin and showing the model and then it writing the code. Like if you think now what that model was able to do, you know, two years ago versus today, it's like we're in an entirely different world than we were two years ago. But I think that example, I worked with Greg a bunch on that demo and was actually, I have a great picture, like sitting right in front of him as part of that demo. And we were going through it. And I think the reflection for me is just like, on how far we've gotten from that moment, both from an AI actual capability standpoint, but also just, I think a lot of the innovation has been the infrastructure to bring AI to actually be useful. And I think even today, I was having a conversation last night with someone who was talking about the raw capabilities of the model versus this sort of AI building harness that has been created in the last two years and how much that actually makes a difference for the raw capabilities that you can get from these models. And I think it still feels like we're so early in that sense that you can get those GPT4 level. And we were talking off camera before about native image generation with Gemini and having this GPT4 level moment of people seeing, wow, this is an incredible experience just out of the box. So it's still so fun to see. You think all the juice has been squeezed out and actually you're right around the corner from something that's going to change how people think about the world.
- [03:42]
Logan Kilpatrick
Yeah, there was a couple of quick pointers about that specific video. Kip and I covered that video. I feel like that demo was the nearest we've ever had to the iPhone moment. It was very similar to that iPhone moment, but for AI, where people were like, holy smokes, I get it right? The collective world who watched that demo, that one succinct use case, really kind of sparked creativity across such a wide range of folks internally. Were you always all kind of constantly surprised by how big those launches were, just how transformative they were, or did you kind of know this is going to rapidly change the way that people think about the world? Like, when we launched GPT4, everything is different.
- [04:24]
Greg Brockman
Yeah, I think GPT4 actually was one of those moments, and there's sort of somewhat differing perspectives about ChatGPT and how much of an unknown success story that was going to be. I think with GPT4, folks knew, like, the real reason that ChatGPT came out was so that OpenAI could try to experiment with different experiences of bringing capable models to the world knowing that GPT4 was coming. So, like, ChatGPT was sort of intended to be the early, you know, get feedback from the world. Does this chatbot thing actually end up being useful for models because at that point GPT4 had already finished training and OpenAI was trying to figure out how do we productize this really, really cool technology. But I think folks knew GPT4 was going to be that useful. They'd been sitting on it for a very long time. I think the model finished training in summer of 2022 and sort of made its way out to the world in March of 2023. So there was a long while where folks had almost fully wrapped their head around this technology and what it was capable of. And I also reflect back on that sort of experience of getting to sit on the technology for that long and actually having it be differentiated with what was available externally in the world. And thinking about now, I think about, for us, it's like, you know, model comes out of the oven from being trained and it's like it has all the safety stuff baked in and like, let's get it out to the world in 24 hours. And I think about like, actually there's a real trade off and I think, I'm not sure how relevant this is for people who aren't releasing models to the world. But like, you do just not get as much time to like, figure out the story and like explore and build it. And like, we're actually figuring out a lot of that stuff, like with the external world as we're all using this model together publicly. And maybe that's actually the best thing for the world because you don't want to be like sitting on the alien technology for, you know, six months before it actually makes its way out. But I think you only get that level of like really cool demo that Greg was able to do by being able to sit on the technology and like really internalize. And I think Greg, to his credit as part of that demo, like he drove that whole thing and like he was able to put it together because he had fully internalized like what the model was actually capable of. And like, I don't see a lot of that happening with like today's era of launches, which is really interesting.
- [06:36]
Kip
Yeah. I think the point you bring up is really important is that like 2025 is like a year in which a decade is going to happen. Right. Like, the pace is very aggressive and if you're watching the show and you may be a distant observer of AI, what it generally is doing for everybody, any company, is that the expectations of pace and speed have just gone up. It's not just the AI models, it's literally every company. And so you have to Know what? Your stories are kind of your core principles of what you're building so that like you can kind of continue to build the story and the product in parallel because it's rare you're going to have the like, hey, I know everything. I've got this smooth six months launch period. Gone are those days right now. Right. Especially if you are out there building stuff because that six months, somebody might build something way better and the work you have is just completely obsolete. And that is, I think, just what, 20, 25, I will probably remember the most. Is that just kind of core change in speed and trajectory?
- [07:34]
Greg Brockman
Yeah. I wish for my own sanity that that was not the case because it feels like this is like the three year sprint that never stops. So it's also at a very human level, feels more and more important than ever. And this is what gets me so excited about, you know, the human experience. This world of AI and all the innovation and the pace that's happening, like at the end of the day, it, it just like further exacerbates how important it is for like to do all the things that it means to be human and to like have those experiences. And, and I, I so fundamentally believe that, yeah, it gets me happy on both ends of the spectrum because I love to see all the cool new stuff coming out in AI, but like, I'm also just so bullish on like the human experiences that only humans can have and, and create and all that stuff, even in a world where AI is intelligent.
- [08:21]
Logan Kilpatrick
So you changed, you went to Google. Google's pace is also, I think, expediated. Like they have some really killer launches. I think maybe they're a little more understated than OpenAI or some of these other companies because they have such a whole host of other challenges to navigate when they release AI. But what are maybe a couple of subtle differences that you've seen between how OpenAI are approaching AI in general and what Google's approach is like? What are some of the subtle differences you see in how those two companies are trying to expedite AI to the world?
- [08:50]
Greg Brockman
Yeah, I'll maybe take two ends of the coin here. First, from the core technology standpoint, but second, how we show up in the world and talk about our AI stuff, which I think will be helpful for folks who are thinking about this. On the core technology side, one of the things that Google really benefits from is the breadth of all the work that's happening. So our team's now a part of Google DeepMind. On one end of the spectrum, we have AlphaFold and protein folding and Nobel Prize science happening on the other end of the spectrum. There's image generation models and there's weather models and there's Gemini and there's this huge breadth of different stuff that's happening from an AI perspective. And DeepMind is really the only place in the world where that depth is actually happening. You can look at all the other labs and you can see the models that they're creating and the models that are being created at other labs is one pillar of what DeepMind is doing. And I think the really exciting thing to me is if you buy into the idea that this multimodal, multi capability enabled model is going to be the thing that enables humans to do all the things that we want to do, like there's only one place in which it's possible that that's going to be created and it's inside of DeepMind because of this breadth of all the work that we're doing. And we see this actually happening in practice with you know, the cross pollination of research from AlphaFold again to weather models to alpha proof, which is our math model and like how all that actually trickles back into the mainline Gemini model that is available to consumers and is available to developers. So I'm super excited about that. I'm excited for us to like lean more into that story and lean more into that advantage that DeepMind has on the sort of other end of the coin after the fundamental research and sort of product creation happens. I think there's just a huge difference in how we approach going and telling the world about the product and the models that we're building. And I think a lot of this is just like grounded in the positions that these different companies are in. Like, you know, there's a lot of factors at play. I think like people perceive Google as a company very differently than people perceive OpenAI as a company. OpenAI also like has a very different product offering. Like Google has like many, many, many products across like many, many different domains and like there's just like a lot of downstream impacts. Beautiful thing for OpenAI is they have a clean slate in many ways. So people like either haven't formed a prior about, you know, some specific angle of what they're doing or even something very tactically like they have an open namespace, like they can call their products anything, they can use whatever URL they want, you know, et cetera, et cetera because like there's nothing conflicting with that. They don't need to worry about the sort of crossover between these different products. And I think about this a lot because we get a lot of feedback from the external world that, like, hey, we wish this thing was simpler. We wish this sort of naming schema was a little bit easier to follow. And it's like, a lot of this is just the artifact of the complexity that Google has because of how large of a company it is. And then in turn, like, we have to find ways to try to more authentically lean into the things that matter to us. And I think this is one of the biggest challenges for Google. Like, I'm sure there's folks in your audience who have thought about this or experienced this at different companies. Like, it's just hard to tell a really authentic story as Google, not because there's not interesting authentic stories happening inside of Google, just because it's the artifact of the size of a company that you're a part of. And I think as the size of a company increases, the ability to tell an authentic story to the world decreases in a lot of ways. And, like, you have to, like, actively fight against this. And I'm happy that, like, I'm in a bunch of threads with a bunch of incredible people on our marketing team who I talk with and work with to, like, think about how we can tell this authentic story, because I feel like we miss telling the magic of why this technology is so important when you sort of don't go the authentic route. And I, like, feel these so deeply inside of me. These, like, really authentic, interesting AI stories that again, that only Google can tell and, like, but the only way to land that message is in a really authentic way. And there's just, like, so many different angles of this tension to reconcile. And I just don't think that OpenAI is an example. Like, they don't have to deal with this because they're just not the size of a company where, like, they have all those different dimensions of tension right now.
- [13:04]
Kip
The one thing I would probably add, Logan, as a consumer of both products, is that one of the advantages Google has is that the breadth of all of their tools and being able to just seamlessly integrate Gemini and even if the store isn't there. When I opened up Google Maps and I looked at a place and Gemini was now on the place listings, and I could just ask Gemini anything about that place. I was like, holy cow, this is incredible. I could just ask it, oh, I'm going to this place in New York. If I get there at 10am, how long am I going to have to wait in line? Right? Basic things like that that would have been impossible or I would have read like 30 Reddit threads to find out. The answer is now like a 10 second answer which I think is the nature of having a big platform with lots of different products is that you can just integrate that into those wow moments. It's almost the story is getting told by the user just discovering those features too.
- [13:53]
Greg Brockman
Yeah, I think that's a great point. And I think this speaks to again, one of those advantages from the breadth perspective is like when DeepMind builds Gemini, we're not building it for a chat app. Like if you think about like where Gemini is actually being integrated, it's like across some of the largest product suites that touch the most users in the entire world and like that has a very different set of constraints that would potentially be built and like Gemini is powering search. Gemini's in YouTube and like these are like billion, billion, billion user products that like all of these very nuanced characteristics matter a lot. And actually we've seen lots of really, really great examples of the requirements for a model to be really good for search actually leading to something that's really great from a developer perspective. We've seen this with a few of the last 1.5 series of models where the search team needed something and it ended up being a really great trade off of capabilities that developers also ended up wanting those things and it's cool to have those levels of internal engagement from these teams.
- [14:54]
Logan Kilpatrick
One of the things you mentioned, Logan, was there's just like a plethora of different AI tools available. Kip mentioned that we're going to see a decade of progress in a single year. And so one of the things we wanted to do in this show is to try to distill it down into like, what do you think are great AI use cases for our audience to take away from the show and start to implement. And maybe we can specifically think about Gemini and Google. But what are some of the AI use cases today? Like that you can actually just go and start using Google Gemini for today that you think are like wildly underestimated or underused by the average consumer of AI, the person who maybe isn't in the details and in the weeds day in, day out.
- [15:32]
Greg Brockman
This is a good question. I think one of the challenging things about this question is that it is a ever evolving answer because like literally the capability flywheel is spinning as we speak, which is awesome. I think today some of the things that are getting me most excited is in December we launched Deep Research. So we launched sort of the World's first iteration of deep research, which if folks haven't used it, is essentially a research assistant. You can put in whatever your query is and the model will go off and search. And in the context of our deep research visit, you know, thousands, potentially of different websites to answer the question. I think that simple product artifact of like showing you the number of websites that the model browse through in order to get you the answer is the thing that just makes that product experience work for me. I'm like in 0% of any of the research I've ever done in my life have I looked at more than 10 websites. So the fact that the model went out and looked at a thousand websites just gives me a lot of confidence in the model. I'm happy to bubble up and we can talk more about this, but I think this is one of the biggest challenge for AI products, which is in many cases AI products are asking the user to basically do all this upfront work in order to provide them value. And I think Deep research is this great example example of the model and the technology just sort of doing the heavy lifting for you. And you as a user get to just ask your silly question or your serious question and then the model goes and finds it. So I love Deep Research. I think lots of really interesting use cases, very underutilized today and it's now powered by our line of reasoning models, which is awesome.
- [17:05]
Kip
I agree with you Logan. That far and away the number one thing, if nobody's really using AI for much, then just some random questions, they should use Deep Research. Deep research is incredibly powerful and the amazing thing about Google Gemini, and you correct me if I'm wrong, is that deep research is a free feature. I don't think there's any hard limits on deep research yet. And really the biggest thing there is you then rolled out these reasoning models which makes it think through the ability of what sources and what follow up questions to ask around those sources far, far better than it was just a couple of months ago ago.
- [17:39]
Logan Kilpatrick
So you used to go down to this dropdown and you would pick whatever model you want to use. I don't think people fully always grok when do I use this deep research? Like what's the perfect way to delineate between Flash, which is also incredibly great, an AI model. Now I need the deep research. Plus you mentioned something really important which is now you've integrated reasoning into that maybe just explain what we mean to users that you've integrated reasoning into this Deep research.
- [18:03]
Greg Brockman
Yeah, I think that basically the land Is like your everyday questions, you know, if it's a simple question, just use 2.0 flash. It's going to be very quick. It'll get you an answer like almost instantly. If you really do need something that is not surface level. Like, if you're looking for like, you know, who won the Cubs game yesterday, you know, you don't need deep research for that. If you're trying to understand like, why the Cubs, and I'm based in Chicago and I'm not even a Cubs fan either, so I don't know why I'm.
- [18:28]
Logan Kilpatrick
Using the Cubs example.
- [18:30]
Greg Brockman
If you're trying to understand why the Cubs, you know, build whatever the ivory wall is around the back of the field and the technique that they used to build that and who the people were who worked on it and like, what type of permitting they needed in order to put that together, like that level of depth in the question that you have. There's no product that can do that besides deep research. And I think the reasoning models is the sort of key enabler of this. And we initially launched Deep Research back in December with Gemini 1.5 Pro and like, it was really powerful. But a lot of the techniques being used by 1.5 Pro were actually like, trying to get it to do what a reasoning model actually does, which is be able to sort of have essentially this like, inner monologue of, you know, thinking through different pieces of a question, like actually reflecting back on the initial answer that it's given and like trying different versions of this. And you can sort of think about how we as humans think through this process. By default, AI models just kind of spit out an answer as quickly as they can. Is basically the way that models are trained today. And the thinking models are trained to like, don't actually spit out the answer as quickly as you can iteratively go through this process. Try a bunch of different things, make sure that you're sort of covering the breadth and depth of what a user might be asking for and it actually leads to some like, pretty substantively different outcomes. So if folks have tried stuff before with AI and you're like, eh, the models are just dumb and don't really have the ability to do these things, try the new sort of generation of reasoning models. I think there's a lot of use cases that just were not possible before that like all of a sudden just work.
- [20:04]
Kip
Today I am very convinced that most humans do not realize what like deep research especially is capable of. For example, like I was wanting to get rid of like a tree in My yard. And I was like, I don't know what the permitting process is. I don't know what you would need to do. It did all of that, told me exactly what I could and couldn't do, what the rules are. And I literally just said I wanted to remove a tree and gave it my address. And it did everything else. Right. And it's like, people just wouldn't think that it could do things like that. Like, I had to estimate an entire construction project. It's like, well, all right, somebody's giving me this estimate. What does deep research say? And it's like, it's incredible how detailed it is, and its ability to go through complex documents and frame things in simple ways is really, really good now.
- [20:51]
Greg Brockman
Yeah. And I think it's wild to just reflect on that. This is like the V0 of the product. Like, truly, like, this is the bare bones version of what deep research can possibly be. And, like, today, it's essentially just using search and there's more. And I actually think at the time of this going out, because it's rolling out today, the ability for deep research to be combined with audio overviews from NotebookLM is also rolling out. So now you can sort of take that deep research experience that you had and then actually just click a button, turn the entire thing into an audio overview, and then you have a podcast.
- [21:27]
Kip
Let's do that.
- [21:28]
Greg Brockman
And you're learning about the permitting process, whatever city you are.
- [21:31]
Kip
And you could ask it questions and interrupt that podcast and be like, what? I don't understand what you mean here.
- [21:36]
Greg Brockman
Yeah. So I think the whole experience and the way in which humans sort of get this type of information is changing, which is, I think, a good thing. I'm curious if you had this perspective on the permanent. I just wouldn't do it, to be honest. Or, like, correct.
- [21:51]
Kip
Exactly.
- [21:51]
Greg Brockman
Or I would pay, like, you end up paying, like, this, like, market inefficient price because you're like, well, I'm not willing to do the research because this is boring, not interesting to me. So, like, therefore, I'm gonna get gouged by somebody and they're gonna charge me $20,000 to take a tree out of my backyard or something crazy like that.
- [22:08]
Kip
But it's also not even the price.
- [22:10]
Phil
Right.
- [22:11]
Kip
It's the time to get to that same outcome would have taken me weeks, and it took, like, five minutes.
- [22:16]
Logan Kilpatrick
Yeah, right.
- [22:17]
Kip
It's like, I can just accomplish so much more in my life, and I don't think any human realizes, like, the rate of progress you can now make.
- [22:31]
Phil
So if you're listening to this podcast, you are more than likely thinking about business ideas and how to market your business. Every entrepreneur has a million ideas vying for attention in their brain. And there's probably that one idea you can't stop thinking about, the one that came to you at 2am and you got up and you wrote it down somewhere. Well, I'm here to tell you that that idea could be worth $5,000. Right now, the Hustle and indie hackers are teaming up to find the next great founder and get him or her started with $5,000. All you have to do is record a 60 second video pitch telling them your business idea and how it makes money and why it's awesome. One winner gets $5,000 cash plus promotion to millions of Hustle readers. Make sure you enter by April 4th at TheHustle Co Bigbreak. That's TheHustle Co Bigbreak. The winner will be announced on April 11th on the hustle Daily show podcast, the Hustle Newsletter, and on the Hustle social channels. Now get to pitching.
- [23:38]
Logan Kilpatrick
It's a combination of there's things you would never have done because you didn't have the time or money and now you can do and just how quickly you can do them. And one of the ones I just don't want to gloss over because you mentioned something that is actually one of my favorite workflows for deep research. We did an entire video of it, which is deep research into Notebook and then turn it into audio, create podcasts and ask podcast questions. That's actually how I've been learning about call options, which this is the worst time to start doing call options. But that's a whole other thing that Kip and I can't see. No financial advice, but that's one of the things we've been playing around with. But when you say that's being integrated, do you mean I can just go to Google Gemini? Like you're integrating an LLM notebook into Gemini or is Deep Research still going to exist separately or are they going to be combined? How is that integrated together?
- [24:23]
Greg Brockman
Yeah, I think for today's launch, when you do a deep research query inside of the Gemini app, you'll see an option when the deep research process is done to like convert the deep research into an audio overview. And I actually think maybe, don't quote me on this. I think there's maybe like a direct way to also just like take the Deep Research artifact and go straight into NotebookLM if you want to like leave. And it's like A subset of the features. So like, actually the like interruption piece wouldn't be included in this. It's just like a single shot. If you want the like full NotebookLM experience of everything, you would go back in a NotebookLM. But it is cool to just like be able to have that single shot version of the overview created for you.
- [25:00]
Logan Kilpatrick
Yeah, maybe. Since we're on NotebookLM, why don't you give the users a little bit of your ways of using NotebookLM? Cause that's actually a favorite tool of this show. Again, I don't think enough people even know that tool or are using it enough.
- [25:15]
Greg Brockman
Yeah, I think for folks who haven't tried NotebookLM, you can think about NotebookLM sort of as the knowledge assistant that you might have, like Imagine or a learning assistant or a tutor that you might have or honestly, just like a way to bring content to life is another version of this. Josh Woodward, who leads the Google Labs team, has this idea of sort of like infinite repurposing of content. And I think NotebookLM is like a really great example of this. You can take imagine you have, you know, something really boring like a onboarding manual to set up, you know, a vacuum cleaner. And for whatever reason you're one of the 10 people who actually reads the onboarding manual and you have to read it, but you're, you're really bored and you don't actually want to go through the whole thing. You could take a PDF version of that, you know, 150 page onboarding manual that says all the different gizmos and gadgets and things that your product does. Stick it in a notebooklm and really quickly generate like a summary, a learning guide, principally, like one of the things that folks are most excited about, a podcast level conversation with like really sort of witty and smart sort of nuance in the conversation about whatever the content is that you're uploading. And it now has features like you can actually interject mid conversation and say things like, you know, hey, this is actually really boring, spice things up a little bit or you know, I don't understand this point that you were just making. Can you sort of re articulate what this is? And one of the most common flows that I actually see, folks who I work with and I work remotely, so I don't commute to the office, even though I probably should go to the office every once in a while. Folks take a bunch of like work documents, they put them in NotebookLM, create audio reviews and then they like listen to them on their drive.
- [27:00]
Logan Kilpatrick
Yes, exactly. I know Kip has a ton of use cases here, but just two really quick ones. So this one here is exactly what you said. It's like a strategic knowledge assistant that's specific to one of the cross functional pods that I run here. And one of the cool things I can do is like add all the documentation. So I add all like meeting transcripts, add all the docs that have been created that month, that week, and then when I'm out walking the dogs, I can actually either listen to it or now you have this interactive way of being able to talk to the podcast directly and ask questions. And so it's a pretty cool way where you can actually take knowledge with you with work and actually conversate with it. Like if you're out of the office, if you're driving, if you're walking. I have knowledge assistants trained on every single project and I can always ask them questions. And so it's like you have a project assistant, slash executive assistant for every project. And as I said, I wasn't lying. Like all the. This is like one example where I had deep research teach me about like call options and then try to pick out like five call options that are underpriced and why it believes those things are underpriced using third party sources. And what I say is like, use the most trusted third party sources to come up with that hypothesis. And then I had a whole conversation with it in this interactive mode when I loaded the podcast and had a whole conversation that lasted like 15, 20 minutes all around its prediction on why meta was a good call option. I kept tell you that was done four weeks ago and actually it was pretty right because I think Kip has probably done that trade. So pretty powerful. I know Kip, you love this tool, but that's pretty powerful stuff.
- [28:30]
Kip
Well, hold on. One of the threads we're going here, and I do think this is true, is that the Google suite of AI products that you all have rolled out I think are the best to help humans learn. Because I'm about to pull a deep cut out of one of my favorite Google tools that nobody talks about because we've talked about deep research, we've talked about NotebookLM. I love learn about.
- [28:51]
Logan Kilpatrick
Yeah, that's right, I love learn about.
- [28:53]
Kip
Which is you can basically just decide you want to learn about something like how do I make pasta from scratch? For example. And what I love about learn about, it's like where deep research is like a deep dive on a topic to gain perspective and then you can kind of ask follow up questions on NotebookLM. This is much more like, almost like a structured learning process and course. And it creates this amazing composable web experience over here. And then you can break down and dive into different components. And basically it solves a lot of the empty box problem of like, hey, I don't know anything about this thing.
- [29:28]
Phil
I don't even know what to ask.
- [29:29]
Kip
And it's prompting me a bunch of different aspects about making pasta. I happen to know how to make pasta, but it's like, if you don't know how to add pasta to boiling water because you've never done it before, like, this is very helpful. Right? And so I think if you look at Deep Research, NotebookLM and Google learn about, like, there are three tools that are really masterclasses in helping people learn.
- [29:50]
Greg Brockman
Yeah, I love that. And I feel like the magic, the real magic and the bow of this is like, how can you bring all that together, understand a user's intent, put the right product in front of them?
- [30:01]
Kip
Yes, yes.
- [30:02]
Logan Kilpatrick
This is going to be my point.
- [30:03]
Greg Brockman
This is the challenge. The sentence that I just said sounds, you know, simple to say when you actually look at it, like, these are the engineering and product problems of the decade. Like, really, these are, like, not a trivial thing to bring together that level of technology. Especially in a world where you're sort of balancing this user context problem, which is like, if folks talk to AI models all the time, like, every time you talk to a new AI model, it has no context of who you are. It doesn't know what you've done before. Which is another thing that sort of just landed in the Gemini app, which is the ability to personalize model responses with your Google search history. And the model can really intelligently say, like, here's what this user has done before, here's what they're interested in. How can we actually use that to sort of prime the model to give you the right stuff? And I think it's like that level of personalization that's going to enable pulling in the right, like, for you, Kip, it'll know, like, here's this. Learn about, like, experience, because Kip's been doing a bunch of stuff over here and this other product versus, you know, search for call options and bringing that stuff. So there's a whole, like, spectrum of different use cases. And I think having the personalized context means that you can get the right product surface or you can get the right product experience in front of the right user Persona, which I think is not something that happens in today's software. 2.0 Product suite software is static. It's like predefined for you in a lot of ways.
- [31:23]
Logan Kilpatrick
Just to touch on this because I think this is a really important point you're making and actually something even internally we are dealing with a little bit. Which is what you're basically saying is instead of at some point going to this kind of dropdown experience, you will have one assistant that you talk to. And in the background the assistant is basically deciphering intent and the assistant can pass you to whatever model it thinks fulfills that intent. Because even in us trying to implement assistance across our go to market, which is like a much more simplistic thing than you all have to do, we have a sales assistant that can actually help sell through chat. We have a support assistant that can help do a bunch of support tickets. We need to actually create a upsell assistant so we can actually upgrade customers to different tiers and sell them on those different tiers. And what we have is like these individual assistants. And the way we decipher intent is somewhat like wherever you are in the go to market we say well this is probably the right assistant, but even us, we need multi bot orchestration. At some point where you have one assistant, the assistant can say you're trying to buy product for the first time, you have a support question, you are an existing customer and you need to upgrade. And that is really complicated. And so I can't even imagine how complicated it is for a Google where the intent is like any single possible thing in the world. How do you ever decipher that intent? But I think what you're saying is even when I think about my AI experience with Google, you all have a ton of great products and I would love to end with one of the ones that I think is maybe transformational for this year, the image generation. But they're disparate. We jumped into NotebookLM, Kip jumped into the Learn product. We jumped into all these different AI models you can choose to the dropdown is the plan for Google to like pull them all together at some point. So you just have like one AI interface as a knowledge worker, maybe one AI interface as a consumer.
- [33:05]
Greg Brockman
Yeah, this goes back to one of the challenges that Google has. It's a company that has lots of product services. I do think more and more the Gemini app is becoming this sort of unified place to get a lot of these experiences and sort of a externalization path for Kip was showing learn about which is one of the Google Labs experiments. I think more and more of those experiences are finding their way into the Gemini app. So I'm hopeful that that trend continues. I'm hopeful that the Gemini app ends up being like the single place to get a lot of these experiences. But there really are just like different user journeys. Like, I think about the four or five different products that I use, like all day, every day, and like, there's different journeys a part of those products for different users, and I think they will continue to be like different products. We didn't talk about AI Studio, but like, AI Studio is another one of these where, like, that's actually how folks access the native image generation capabilities today. And like, I think in the future we're still going to have AI Studio because, like, the user Persona that we care about in AI Studio is a developer who's sort of exploring these models and wants to build something which is very different in a lot of cases than sort of the main experience for users going to the Gemini app who, like, are trying to use the Gemini app really as like an assistant on a daily basis. It's like a daily active user product versus AI Studio is really intended as sort of the portal to the developer world. For folks interested in Gemini, maybe talk.
- [34:23]
Logan Kilpatrick
About the latest image release. When I think about AI and where we might have been a little wrong, I thought video would be further along than it is today. And I mean, AI's ability to go from text to video, right. It's still pretty clunky in a lot of cases. It's not like production ready. And then the other one was images, text to image. And I felt like the first iteration of that were all these great tools where you could go to image. But for image to be really useful for people, there had to be great editing tools as part of that. And so Google's latest release, Kip and I were playing around with it last week is really awesome. And maybe could you just give the context on what that model is and why it's so good?
- [34:59]
Greg Brockman
Yeah, no. 100%. So for folks who haven't been following closely, we launched Gemini 2.0 back in December and we showcased these capabilities actually, and we rolled them out to a small group of trusted testers to get some initial feedback. And then last week we rolled out to every developer the ability to use Gemini's native image generation. And I think the thing that's actually capturing a lot of interest is the native image editing capability. Because the model is natively multimodal, you can pass in an image and you can say, hey, update this image to. In the example we're looking at on screen. Add a little chocolate drizzle to these croissants or add a strawberry drizzle to these croissants or I've seen a bunch of really cool examples which I did not think about of, of taking black and white images and actually asking the model to colorize those images.
- [35:47]
Logan Kilpatrick
And you can see, I saw that.
- [35:48]
Greg Brockman
Yeah, you can bring them back. You can take in two images and you can sort of fuse them together. You know, I saw a funny example of someone posting, you know, a bunch of like hot dogs and stuff like that and merging them together into this comedic image and lots of different random things like this. That is the full spectrum of really useful to really silly. But the thing that I think is capturing folks attention is if you think about how you would have had to do this workflow, pre native image generation and image editing, it's just hard to do. Like the number of people who can do that in whatever the professional tool is, is pretty limited. You can just do it with like a very, very simple text prompt. So like now the opportunity space of people like creating dynamically edited images is now essentially every human on earth is able to do that, which is just. This is the thing that continues to blow my mind is you get a capability and this is what AI is enabling across so many different domains. It's this thing that only a few people could do and then overnight it ends up being this thing that everyone can do. And I think actually coding is like before image generation is also having this like parallel moment with Vibe coding and everything where like before code was an artifact that only, you know, professional developers could create. And now it's like every human on earth could go and create code as an artifact if they want and pretty much get what they were looking for. It's the same thing with image editing today. Before you had to be very good at using one of these tools in order to do this stuff. And now it's like every human on earth can do this and it just changed overnight, which is just such a weird experience to think about.
- [37:20]
Logan Kilpatrick
Yeah, vibe designing.
- [37:22]
Kip
I like that.
- [37:22]
Logan Kilpatrick
This is the kind of dream, I think for AI in general or the transformative mission of AI in general as it unlocks creativity somewhat overnight because it allows people now to unleash their creativity and they're no longer hindered by having to learn the tools. And I know that sounds like a lazy way to think about creation, but I don't think it is because I think the creation part is the important part and the learning the tools shouldn't hinder that part of you, like the ability to create things as this Visual Story. Could you just end on what this product is? Because we discovered this today. I would love to know, like, what is the way we should think about Visual Story, which again, for people who are following along on YouTube or even on RSS, these are all available in AI Studio and you should really go in and play with that product because there's a bunch of great tools in there.
- [38:06]
Greg Brockman
Yeah. AI Studio is, again, it's our service for developers, intended to bring the models to life in a way that ultimately wants to get you to build with them. So here we're trying to. This example that we're looking at, trying to capture developer sort of imagination of what products they could go and build themselves. But principally, AI Studio is not intended to be like the Daily Assistant product. Like, it's a very thin surface on top of the models. We make a bunch of very opinionated decisions to keep the core AI Studio experience the same as the experience you would get in the API. So we don't have a bunch of fancy bells and whistles that you yourself couldn't do as a developer in AI Studio, which makes it limiting as a product. If you want. People are always like, why don't we have Deep Research in AI Studio? I want that. It's like, because Deep Research is something you could build sort of a similar deep research experience using the API, but it's not available to developers today. So we don't want that experience in AI Studio. Yeah. So if you're someone who wants to build stuff, getting your Gemini API key, all that stuff, testing out the latest capabilities of the models happens in AI studio. And our team sits in Google DeepMind right next to the model team physically. So that's oftentimes our product is sort of the fast path to externalize the latest Gemini models, which is a lot of fun and it's cool to be on the frontier and it's cool to see the excitement with native image generation.
- [39:27]
Logan Kilpatrick
Very cool. Kind of like the way ChatGPT was for text. I feel like this is similar for Image. It's really the first time I've seen the ability to kind of get the image crisp and concise to the way you want it. This was awesome. Logan, we really appreciate you coming on and giving us a deep dive into what it's been like to be part of this kind of AI journey. You're working with two of the most transformational companies there are you or you have, and also just going deep into like how people can start using the Google tools straight after this episode.
- [39:56]
Greg Brockman
Yeah, this was a ton of fun. Thank you for having me. Hopefully the call options go well and we'll we'll all be celebrating somewhere together and in a few months that's our.
- [40:04]
Logan Kilpatrick
Litmus test to whether AI is truly good or not. Is is it making us money with call options?
- [40:10]
Greg Brockman
I love it.
- [40:11]
Kip
Thank you so much Logan. This was awesome. Appreciate the time in yeah of.
- [40:34]
Phil
We'Ll be right back to today's show, but first I want to tell you about a podcast I love. It's called Nudge. It's hosted by Phil Agnew, brought to you by the HubSpot Podcast Network. Ever notice that the smallest changes can have a big impact on Nudge? You learn simple evidence backed tips to help you kick bad habits, get a raise, grow a business. Every bite sized 20 minute show comes packed with practical evidence from admired entrepreneurs and behavioral scientists. Nudge is fast paced but still insightful with real world examples that you can apply if you're a marketer. You're going to love this show because it discusses the psychology behind stellar marketing. They just did a great episode titled 99.9% of ads are Genuinely Awful. With Tom Goodwin, you can learn why so many ads today are ineffective and what marketers are getting wrong. Listen to the Nudge wherever you get your podcast.