Inside Google Labs: 3 AI Tools That Will Change How You Create

13:49Megan Li:And I ended up doing that because although the initial images that it had selected were pretty great, there were a few even better ones in Mixboard that I wanted to use. So you can always go back to the board and decide that you want to one-shot download an image. And then when you're going back into the presentation mode, you can replace any images that way. But it's just like such a fun, delightful way to transform memories now at this point, too.

14:15Thomas Iljic:It is. It is. Now for a quick message from our sponsors at MITxPro. The AI landscape has evolved. Success now hinges on moving from concepts to controlled high-impact deployment. Teams that build thoughtful execution plans are gaining efficiency, resilience, and lasting strategic leads. That's why MITxPro and the Computer Science and Artificial Intelligence Laboratory CSAIL created a new course for leaders who want practical guidance and not vague inspiration. It's called Deploying AI for Strategic Impact. This nine-week course, led by 15 MIT professors, teaches rigorous evaluation of AI models, accurate understanding of infrastructure demands like compute data and costs, and practical deployment strategies informed by real-world case studies across manufacturing, healthcare, logistics, cybersecurity, and more.

15:10Thomas Iljic:If you're responsible for AI outcomes and want to build real deployment confidence, check out this newest course from MIT. Link's in the description. Explore deploying AI for strategic impact and enroll today. That's so awesome. So asking a follow-up question on that point, it will maintain the same layout. So if you were to replace that image, you know, scrolling up a couple slides, like let's say. It's not going to regenerate the whole slide, right? It'll just fill it into that one spot.

15:42Megan Li:It does actually regenerate the whole slide, which is the interesting part. So this is powered by Nano Banana Pro. Now it will regenerate the whole slide, but the model is really good at understanding exactly what you want changed in the image. So you could actually do something like just say circle this and select that image to change and everything else should look the same at the end of it. But behind the scenes, it is a regeneration. Okay.

16:07Thomas Iljic:Really cool. Yeah, that's awesome. And I know that that's true because I've used Nano Banana and it's very good at those precise images. It is. Yeah. It is. That is so awesome. So who, I mean, you're the project manager on this tool, right?

16:23Megan Li:Yes. I actually just brought on a new PM who's jumping in and will be helping lead Mixboard into his next chapter in the new year. But it is under my team, which I'm excited to continue to help push it in many directions.

16:37Thomas Iljic:That's very cool. So along those same lines, like when you first started creating this, who did you have in mind? And who do you think will benefit most from this tool? Because I can see so many different use cases. So I'm curious for you as like the person behind creating it, like who you are really targeting with this.

16:55Megan Li:I think that, you know, right now we're really trying to make creativity accessible to everyone. I think one of the best things that's happened with these AI models is there are these amazingly skilled tools that now anybody has the ability to create a like work of art oil painting image or a super cute, you know, origami shaped teddy bear, for example. but you still have this challenge of needing to come up with those ideas needing to know know that you have the tools but like how do you have or how do you spark the creativity in folks that may not have thought that they could build such things previously and so when we think about who is Mixboard for we're really trying to make it accessible for anybody to just like get in and start exploring their ideas we want to keep it easy to use we want to make it fun we think that The act of creation should be enjoyable.

17:50Megan Li:And so Mixport is really meant to just, you know, be a really easy entry point to just get started and explore your ideas and your creativity in any direction and appeal to a large audience as a result of that. This is not meant to be a power user tool, but really it's meant to just bring creativity and make it accessible to everybody by leveraging AI and using that to help fuel and power human creativity.

18:14Thomas Iljic:I can see use cases both for individuals, for graphic design teams. I could see there's a variety of things you could do. I keep going to really nerdy stuff, though, and thinking like a D &D campaign and having a theme board. Or even like, you know, the new wardrobe.

18:44Megan Li:Yeah, I think that that's the power of Mixport is that it just it applies to so many different use cases and so many different users. And we want to make sure that we continue to maintain the ability for many different users and use cases to come in here. I had one one teammate who was using it to help plan a bachelorette party weekend. And then I had another person who's really leaned into the home redecorating side of things. and is using Mixboard to imagine what her backyard would look like, even for themed parties, and what her living room could look like with all this different furniture that she's been considering buying right now.

19:22Megan Li:So it really does kind of span everything from both professional use cases to personal use cases. Totally. And then even, as I mentioned, the creativity side of things, of just what is a cool storybook? My kids love watching me as I'm making different fun creations here. And it's funny, you actually mentioned earlier when you were helping me write that prompt for Ollie the Cat Dragons, because this morning's first thing that I made with my daughter was a dragon made out of clouds. And so you can imagine like a mixed world that's just entirely imagining what are a bunch of different cloud-shaped animals that could exist here and how can I then write a story off of all of that?

19:59Megan Li:It kind of like just lets you explore and tug at these different ideas. And as we're chatting, I'm actually going to try and do that right now.

20:05Thomas Iljic:You could go from an idea to a universe.

20:09Megan Li:Yeah, exactly. You don't always know where you're going to end up when you start here. And that's kind of one of the joys or the magic sparks that I found with it.

20:16Thomas Iljic:Oh, that's fun. I was just listening to, I believe it was the CEO from Linear on TPBN today. And he was talking about design. And one of the things he was saying is that design is such a process of discovery where you kind of have to go out and make a lot of like, just like basically essentially do what Mixport is doing. come up with a bunch of different directions where you have to find the the your way through all of the possibilities so i think this is a perfect tool for designers as well i know i just saw the words flux capacitor on the screen a minute ago you definitely did as i mentioned

20:52Megan Li:we're trying to uh we're trying to keep mixboard fun and quirky um so we do want to have a few of those easter eggs throughout that just make you kind of second guess and say wait did it really just say that um and usually the answer is yes it did if uh if it's in the case of explored it

21:06Thomas Iljic:i love it these are so cool they really are i will say like if i was doing this for my kids

21:11Megan Li:you know maybe i'm going to delete some of these that feel a little bit a little too gnarly yeah then you can also say like okay write a fun poem about this and this is a cool area also where like you can kind of go from both text to image and back and forth and then you could say something like you know um what happens next and just there are no rules with mixed board which is another part of it so now i've like you know what happens next in this poem we've just decided it's this thing and then you could say you know create a cartoon image for this and see what happens based off of that um another thing is you can select multiple images as well we'll give this a second you could go children's book with it even if you want it or something a hundred percent so this is what happens next in this story you could also you know take three or four of these images and say like uh create a um jello themed version of these dragons uh because i was in fact trying to make uh axolotls out of jello last night um via an ai image generator not in real life that would be more challenging i was curious i was about to add

22:29Thomas Iljic:a good a good tray for that it was like that is an activity there but now we have our jello themed dragon that looks delicious i want to eat that i would absolutely eat that dragon candy

22:44Thomas Iljic:how cool on the children's book point actually this is a real use case that i was playing with and i'm still going to probably write about this at some point it's trying to make an entire children's book end to end with nano banana pro uh and i started doing it the other day but i'm wondering now if mix board would be the better choice to do it because i have let's say like three starting frames i have like the cover of the book i have like let's say like the first three pages then i have a list of all of the pages um or like the pages and what needs to happen on each page could i theoretically give that to mix board and have it create the rest of the pages for me I don't know.

23:22Megan Li:You can certainly give it a try. One of the nice things that we have enabled, which actually came like two days after the initial launch, is if we go into... Here, let's... Oh, this was another fun one that I made, actually. And I'll show you one of the tips or tricks that will help with the storybook side is I recently went on a hike with my family as well up in Yosemite. And so we turned that into a really fun presentation here. and for this one I actually gave a bunch of both memories from the trip itself so like tips that made our hike enjoyable with the kids because it's always helpful to remember that for future hikes this one I actually prompted and asked for some interesting facts about the hike here I actually downloaded my Strava entry because I always track the hikes that we do so I uploaded that and then this was a download of a bunch of interesting facts about the Merced Grove hike which is where we went.

24:19Megan Li:So when we go into the presentation mode, you'll notice it just nicely pulls all of this together in these fun little pinecones as bullet points. I wouldn't have thought of that, but I love it now that I've seen it.

24:31Thomas Iljic:Yeah, no, that's amazing. Same.

24:33Megan Li:This is wild. It just generated this based off of the Strava entry that I had in here. A little infographic.

24:40Thomas Iljic:Yeah.

24:41Megan Li:Yeah, exactly. Exactly. Because Nano Banana is great at infographics. And if you do decide to go the route of trying to make your story here, one thing you can do is we allow you to actually download each slide as an image. So you can also start to like pick and choose which ones work the best. I've had some experiences where I actually had to give a keynote talk last week and I used Mixboard. I just dumped a bunch of my previous blog posts, a bunch of interesting other slides I generated previously, and I had it come up with a presentation for me. And I didn't end up wanting to use every single slide, but there were a few really well done slides that I wouldn't have even thought of trying to visualize my concepts in those ways.

25:19Megan Li:And so I downloaded those and I ended up with like this patchwork of a presentation where some of them were entirely like mixed board generated slides that just really added an extra level of like fun and well-designed slides into the entire presentation.

25:34Thomas Iljic:Were you able to like either through prompting it originally or re-prompting it, make it match the aesthetic of the deck that you were creating? Like how did that work?

25:46Megan Li:it's funny what i actually ended up doing is i downloaded an image here from mixboard i then used gemini to help me identify what fonts mixboard had decided to use and i actually changed the visual style of the rest of my slide deck to match the slides that mixboard had helped me produce because those were just better design than what i had been doing off on my own that's

26:08Thomas Iljic:awesome i love that yeah i would have totally done that that way yeah i'm like this is better I'm going to go with this. So is all of the text also generated by Nano Banana or is it coming from like Gemini?

26:22Megan Li:It is generated by Nano Banana, but all of it is powered by our, you know, world class models. So the end output here, though, is a just it's a Nano Banana image is what you're seeing, not like layers of text, which is the interesting sort of like paradigm shift almost in a way that this technology is enabling what used to be multiple layers to just be a flat image. Because at the end of the day, any slide that you're seeing or any like, you know, infographic that you're seeing, it might have been multiple layers to build it. But all that you see is that end image. And now we have models that just go straight to that end image.

26:57Thomas Iljic:Yeah. I have a couple like future focus questions, but is there anything else you want to show us before we get to like where this goes next?

27:04Megan Li:Yeah, I think just I've showed you kind of how to mark up the images when you're in the presentation mode, but also just making sure you know that you can easily click into any of these and it takes you into a similar edit flow. So one of the let me actually go into a different board here because I do actually like the way that that one transformed or that one turned out. So let's let's try something fun here. Let's take this and then let's also say a castle made out of gingerbread. So we're going to generate that right now. And there's some fun ways you can kind of mix and match the images within the board as well that are worth exploring.

27:50Megan Li:So now you could click into here and you could say, okay, I am going to circle this right here.

28:00and put an x we'll give this a try and say put the dragon in the green

28:08Megan Li:circle with an x beside the wasn't called a gingerbread house gingerbread castle and we'll we'll try and see how that one works out but it's this idea of like not only combining images but you can mark up images you can mix and match them in any which way you want wow But definitely worth jumping in and playing around and just seeing where your imagination can take you, knowing that we have this open-ended canvas and we'll continue to add feature requests as folks are using it more and giving us more feedback as well. I think that's one of the other big things to note is just we wanted to get this out as an early experiment.

28:49Megan Li:Oh, look, there it is. That works. That works perfectly.

28:52Thomas Iljic:That works perfectly. it both cut the dragon out and placed it there that's really cool and it seems like it matched the lighting perhaps the lighting just worked well with it but it doesn't look completely out of place no it doesn't it doesn't look like you cut out a dragon and stuck it on the paper you know like it's like it could it looks good but also i love that you said my jello uh dragon

29:22Megan Li:doesn't look completely out of place in my gingerbread castle. That alone is a lot of fun.

29:26Thomas Iljic:Well, they're both booed. Nothing wrong with that sentence at all, is there?

29:31Megan Li:No, it is spot on.

29:33Thomas Iljic:Yeah, it looks perfect. It looks perfectly like it goes perfectly together. It does. Well, what's interesting about this is it's almost like recreating the capabilities of a layer-based system without needing the complications of a layer-based system. Yeah. So if I was a professional graphic designer, I can mix and match things here with just text and drawing and showing things visually where I want it to go. And it's simple to use. No layer management. I know.

30:01Megan Li:It makes you think about what happens next now that we're starting to really revisit old assumptions on how things should work.

30:09Thomas Iljic:But you could still edit it, I assume. You could say, turn the dragon 90 degrees counterclockwise or something, or something silly like that. And I'm pretty sure it would do it. Let's see. I won't hold it against it if it doesn't, but I'll bet it can.

30:29Megan Li:I'm going to fingers crossed and see if it's able to listen to this. You have the right expectation, though. We should be able to do things like that. And as I say often also, this is the dumbest the models will ever be. So if something doesn't work today, just try reprompting it sometimes. But also know that it will get better. Okay, let me try one more time.

30:49Thomas Iljic:Yeah, give it the best of end. Give it like two or three shots. See if it can do it.

30:56Megan Li:Turn the dragon in the red circle to face the other direction. Perfect. All right, let's see if...

31:06Thomas Iljic:There's a lot of assumptions in there, right? Because it has to be smart enough to know what the other direction is. yeah yeah so if it can do it that's pretty amazing i'll be dang what a cool tool what a cool tool i'm already thinking of things i'm going to do

31:23Megan Li:when i sit down with it tonight all right well that one didn't work but uh give it one more

31:28Thomas Iljic:say maybe say like turn it from facing the left to face the right and see if it gets it that way the truth is that dragon is a twisty curly dude it's entirely possible face might be it's complicated yeah it's a complicated image you know yeah it's not like a like a like how many gummy dragons are in gemini's training data we have to ask ourselves it's a very busy image

31:56Megan Li:as well i'll give it that much we we somehow landed uh okay all right it was able to actually make some changes okay it did not turn around but it did breathe fire and i like gold okay yeah and so i think fun i think if you were even more precise with the language it would do it

32:15Thomas Iljic:and from my experience with nano banana i think it will yeah i agree i think there's still a bit

32:21Megan Li:of like prompting work that i could do here which is what i also always say to folks like sometimes it takes a few iterations um and then there's other things you can do like you could always make the dragon in this image face the other way and then put it back yeah and put it on this image here and that might be something it's able to do a little easier the original has a very fruit roll

32:39Thomas Iljic:up astropop vibe uh how cool what i like about mix board is that like normally when i work with the nano banana it's usually in gemini itself so this you know i'll have to start another chat or I'll have to, you know, go a couple rounds back and forth in the chat window. Here's a picture. This is so much more, this fits the visual editing workflow so much better, in my opinion. I agree.

33:05Megan Li:Yeah, we're really trying to lean into both a, like, visual first idea exploration, but then layering in more and more as I've been using it, the text elements become really interesting to think about how to pull those in. And then you can imagine like what other modalities make the most sense to pull in next. Those are the types of ideas that we're starting to explore as well on the team.

33:28Thomas Iljic:That's cool. So that was two of my questions, actually. One, would you ever expand Mixboard to include video?

33:38Megan Li:I think that it depends on where we want to focus. I think that right now, Flow is an amazing video tool. So I'm tempted to say we don't really need to go in that direction. Because if you're trying to do an amazing video, Flow is probably the tool that you want to go to on that front. If, however, you wanted just a quick five-second GIF that animated one of your images, Like maybe that's something that makes sense to bring into Mixboard. Similar to one of the recent launches we actually had last week on my team was Pomelli, which is a tool that allows SMBs to create on-brand creatives. And we allowed one-click animation of any image into like an animated visual creative.

34:24Megan Li:So you could imagine that maybe we would explore at some point in the future the ability to one click animate this and maybe this automatically flies through the sky for a few seconds. Not to say that we're looking into that exactly right now. There's other things on our list. But I think we would want to take an opinionated stance on what are fun, delightful ways we could bring in moments like that rather than lean directly into video, which is where I would say flow is absolutely the tool to go do that.

34:54Thomas Iljic:Yeah, yeah, that makes perfect sense. And then my other question is, would you ever consider new ways of working with a Gemini Pro on Mixboard, i.e. via voice commands, like being able to chat with it with your voice or being able to draw alongside Gemini? And like, because I think there was a tool I saw a while back was like, draw with me, where you can like draw and AI will draw with you with something like that.

35:19Megan Li:Yeah, I think those are very much within the like realm of things that we want to explore. I think getting feedback from users who are also using the product is going to help us sort of calibrate and steer towards what we want to prioritize. I think as much fun as this has been like showing you all the stuff we've been doing, one of the things that we're also very aware of is just how do we get people over the hump of knowing where to even start and making it a lot easier to, you know, solve that blank canvas problem. And so we're thinking a lot about, you know, how does onboarding feel more seamless and really get you into that creative experience to begin with?

35:55Megan Li:And then also, how do we continue to let you expand upon your idea? A lot of Mixboard right now kind of relies on a user having an idea and self-directing on where to go. I think there's interesting new ways we can create experiences or help shape the tool to help a user or to help pull an idea out from a user's head and kind of co-create. So you said draw with me, I think it goes beyond just draw with me to like co-create with me, like get creative with me. What is that type of an experience start to look like?

36:25Thomas Iljic:Ooh, that's cool. I like that. There's so much about it the right way. Yeah, I'm just excited. I mean, what I mean, I guess like let's say, you know, mix board a year from now, two years from now. What what's your dream for for what it does from here? What do you want to want to unlock for people?

36:42Megan Li:Boy, I think one of the most interesting things to me with the pace of AI right now is that it's accelerating so fast. It's opening up a realm of things that are possible that would have almost been impossible to predict six months ago. Like if you told me that the way to reimagine creating compelling presentations was through an image model, no, I would not have been able to predict that. And so I think more than anything, as we think, especially on a one to two year timeline horizon, I just want to be able to keep an open mind. I want to be able to listen to users, build what they want, or at least understand what they wish they could do, and then figure out how to turn that into a compelling product experience for them.

37:22Megan Li:And I think it goes a little hand in hand. Certain obvious things like people wanted a bigger board. Great. That's a pretty easy feature request for us to meet. People want to know, how do I get started when I just stare at a blank canvas? That takes a lot more product thinking on the team, and I want to be able to spend time on that. And then also just let the technology lead us a little bit as well. If a new model capability comes along that just makes you rethink something the way Nano Banana has, I want to be able to jump on that and figure out how do we bring that into the product? How do we bring that to users in a way that gets them excited and unlocks that net new possibility or that net new shape or format that could exist in the future?

38:00Megan Li:So it's an answer and a non-answer, which is to say, I want to keep being creative. And I think there's many directions it could go still.

38:07Thomas Iljic:Yeah. Well, Jacqueline, thanks so much for sharing this with us today. What a cool tool. What's the easiest way for someone to go find it?

38:16Megan Li:Great question. Mixboard.google.com. Check it out. You can also go to labs.google, which has all of the cool experiments coming out of labs these days. Mixboard is on there. And then if you want to give it a try and have any feedback, check out the Google Labs account on X or just ping me directly. We're always looking and the team's always looking for feedback. So please share it if you have any feedback as you're playing around with this tool. We're listening.

38:44Thomas Iljic:Absolutely. There are so many different apps and tools and odds and ends on Google Labs. It's really, it's almost intimidating, but it's laid out in such a simple, clear way. I love that there's a description that says what it is and what it does and click here. It's really laid out beautifully. And that's not something the AI space has been great at, nor are general websites. I regularly go on rants about websites that don't say what they do. and I appreciate that this really does that well.

39:21Megan Li:I'm glad to hear it. It is a very exciting time to be building and there's just so much out there that is worth trying these days.

39:30Thomas Iljic:Well, Jacqueline, thanks so much for joining us. Thanks for joining.

39:33Megan Li:Thanks, bye.

39:36Thomas Iljic:All right, everyone. And now we're here with Tomas Ilgic. Tomas is the Senior Director of Product Management at Google Labs and he oversees Flow, which is a cool product in Google Labs that we're going to talk about today. Tomas, how are you? It's great to meet you, sir. It's great to meet you, too. I'm good. Excited to talk about AI, talk about this product. Awesome. Same. Awesome. So Flow is an AI-powered filmmaking tool designed to show and tell, or designed around a show-and-tell philosophy. Do you want to show and tell us a little bit about it? How funny. Yeah, of course.

40:19Thomas Iljic:Sorry, I had to. I mean, but you're exactly right. Actually, that's part of the starting point is when we started playing with this technology, you know, for the several past years, we realized that like prompting had to go away eventually and it had to become much closer to kind of like where the users are. And what they want is basically just show and tell and just keep iterating things. And so, you know, the mental model is, yes, you should be showing and telling. you should be feeling like you can mold the clay and kind of like, you know, you get an output and then you can continue refining it.

40:46Thomas Iljic:So last year, probably in December, we launched VO2, which was kind of a new model. And we realized like, oh, people are really able to do some stuff that's really interesting now. And so following that this year, we sprinted to build a flow for IO and then it's been launched since then. It is a tool for, as you're right, for an AI filmmaker or an AI filmmaking tool. Right now, we're focused initially on like the niche of people who we thought would be interested and really it's been like interesting to just build it with creatives we're trying to understand like the name is reflective of kind of our ambition here is to figure out like how do we keep these creatives in the flow how do we meet them where they want how we change kind of the ui or even the interaction patterns so that they can you know kind of like put their vision out there and then start kind of crafting and getting to where you want so it's it's a journey um but this is this is where we launch flow that's awesome um and yeah i'm very excited to just show you a little bit maybe i'll just uh you know um go ahead i would say that there's this really cool if you go to flow.google you'll see like there's a really cool video here that we've made with a lot of the quick creatives we've worked with when we built it um wow i encourage you to take a look at it yeah it's uh the technology has done great great steps but i'll just jump into the tool um so i'm going to share this server tab here so when you enter flow actually um maybe i'll just you have all your projects inside.

42:02Thomas Iljic:So you can see my projects, for example, over here. And I'll just open up one that has a bunch of, you know, cool examples I think that I can show you. Right now we support both video and images, but I'll start with video because this has been our starting point. And you can do, you know, there's multiple modes in the tool. So you can start from like text to video, which is something that people, you know, have known and used for a couple of years now. And obviously it's very fun, but it's also kind of like much harder, I would say, to craft a prompt that matches exactly kind of what you want.

42:35Thomas Iljic:But with these latest models, you know, physics are really good. They're incredibly good at photorealism. They have audio and get your characters to speak. Right now you can't control like the voice across multiple clips, but that's obviously something that will be working in the future. And so you end up with things like this with, you know. You see this whole stretch, this used to be nothing

42:54Megan Li:but wildflowers. My father and I, we'd walk it every Sunday.

43:00Thomas Iljic:And so you can see, you know, it's starting to become like, you know, really interesting in terms of the clips that you can get. And, but it requires you to, you know, prompt, try different prompts. Some people come up with very elaborate, you know, super complex, almost like Jason templated ways of calling the model. So still kind of a lot of fiction. So this is why. I love that you could hear his foot on the gravel road. Yeah. I mean, just the fact that there's audio, sound effects, dialogue, I mean, that's all new, relatively new with VO3, right? In 3.1. That's exactly right. Actually, VO3, so I think the big reveal with VO3 was it had audio.

43:33Thomas Iljic:And I think that kind of really just, you know, it gives like a new dimension almost to videos suddenly compared to, you know, the ones without sound. And now everybody expects sounds and expects sounds to work really perfectly. So we'll keep working on it. But yes, it has an incredible dimensionality to the output. But then, as I was saying, obviously, you know, it's still very limiting to be working just with text, particularly when you have a vision in mind. And so you'll see there's all these different modes in the tool, particularly frames to video, which allows you to kind of like almost you keyframe the different pieces of the scene and then you kind of animate them or stitch between two frames.

44:07Thomas Iljic:You can do first and last. And that allows you for, you know, it's almost like another way to just communicate. It's like, actually, I want this environment, but I want you to just like zoom into the environment, for example. So I start from this picture and I just ask, imagine you're a drone going into the scenery. So now I have a little bit more control is I can describe the environment visually and I can get the camera to kind of move into it and describe that a little more. So that's starting from still in. So you have this little button here and everything you generate in flow that allows you to just reuse the prompt.

44:40Thomas Iljic:And as you can see here, I have a first frame, which is this image I uploaded. and the prompt was, you know, FPV camera into the picture and throughout the apartment. Now, what you can do as well is obviously you can also condition on the first and last frame. So, you know, the different models, ours and other companies' models now are getting much better also at first and last frame. And so if you want to do, for example, a looping GIF here for my daughter, I want this little scroll to come in and out. I can actually just specify, like, this is our starting point, this is the end point, and here's what I want to happen in between.

45:10Thomas Iljic:And the thing I'll call out, which is super interesting when you think about these models and what they can do is you can see it matches also, you know, everything that you're trying to insert in those scenes in between those frames, the model will try to match and feel that it's coherent. And so you'll see I'll hammer this theme probably throughout the demo, but I think of these as like little simulators in a way, you know, people think of them still as, you know, just a function that generates something, but I think of them as simulators and you're giving them anchoring points. This is kind of what I mean.

45:38Thomas Iljic:Now simulate this thing I'm asking you to do. You can do so obviously with audio, so you can do the same thing. I can have a first frame with a character and she can be speaking again. She can speak in multiple languages. It's actually supporting, you can ask it in. Wow. Nothing like Coco in the morning. You can ask for the types of background music that you want or not, and you can try different things. Continuing down the interaction path, this is something really cool that we figured out, which we call doodling. So sometimes, you know, you have your first frame and you're like, you know what, like it's that's kind of what I want, but I also want these other modifications.

46:15Thomas Iljic:And again, text is a very poor way to just like describe everything. And so it supports doodling. So what you can see here, I'll just zoom in a little bit so you can see, you know, we added some annotations. You actually, you can edit your image. The way you would do so is, you know, if you had a frame here, add an ingredient, you would just edit it. And then in the edit, you would be able to add some annotations on the frame that you just uploaded. On the image itself. On the image itself. And then you would use this as your starting point. And so what you can see here, I'm just going to zoom in.

46:47Thomas Iljic:So we have this first frame and we're like, you know, this is not exactly what we want. This is kind of directionally like what we care about, but we want to add a window here. We want her to have baggy pants, curly hair, change the bus. So that allows you to also specify, you know, visually where those changes might be happening, which is very hard, obviously, if you're using this through text. and looks like you don't have to be super exact either like i noticed the add window drawing is just just just like that area somewhere exactly somewhere there so you can be as as specific and or as loose as you want and obviously you know it's the hit or miss right there's some cases obviously where the model might not pick up on everything but what that shows you is the model actually understands is able to understand what's within the image and understand what you've written down which is like add window white baggy pants none of this is in the prompt it's all in the context that you're passing visually.

47:37Thomas Iljic:And so then when the model displays the changes, as you see, as I said, it's a simulator. So you see that the bus going through behind the building, because there's two windows in this particular case, this is why I love this example, is it actually figures out, well, if that's happening, then these other things should be happening. That is so cool. I did not realize it was going to happen during the video. I thought it was just going to generate with your changes. No, that's right. It's kind of almost, it applies the changes and you can see them being applied. And then what you could do is you could reuse that.

48:10Thomas Iljic:You know, you could take any frame and then just use that as a starting point if you wanted to or something else. But it's kind of, yeah. The fact that it's the bus throughout the building is always kind of for me a delightful moment where you're like, oh, it's really trying to understand the physics of the scene. It's almost, well, I was going to say is it's almost like a new way of storyboarding where you're drawing the storyboard, but with an actual video. I think that's right. So the idea for us is try to figure out what are the different ways you'd want to input and interact with the models.

48:44Thomas Iljic:A lot of interaction also and back and forth between the still images and the video. And the reason is it's easier sometimes to operate at the image level, to take your time, to annotate, and then sometimes you want to animate. So in the future, obviously, you'll figure out, You'll have things like video references if you want more motion and controlling things that have a temporal effect to them. But really, for you to be able to jump between all these different modes that matter to you as a creative is kind of worth looking for. And so in that vein, one thing we've been working on quite a bit is also with the Google DeepMind team is multiple references.

49:20Thomas Iljic:And so in this case, let's say I want, you know, this is my character. Maybe I'm trying to make an ad reel. This is my brand logo. So, you know, I picked something fairly complicated that I generated on purpose here. But like, you know, you have flowers, a capybara, whatever. You know, you can say the woman is wearing a black leather jacket with the pattern in the back and she turns around. And this is what you get, you know, out of it, which shows you how very quickly you're able to ideate on an idea. And so what we are seeing people do, particularly in the more creative professional space, is like it might not be exactly the perfect thing you wanted.

49:54Thomas Iljic:Hopefully over time we allow you to get there. but for quick ideation and alignment now you have all of these people who like i can do a previs of what you know whatever thing we're trying to achieve in like a few you know a minute an hour whatever with a few people we can align and then we can go into the detail um so really many many ways to input the model and many many ways to kind of like give inputs um are you uh are you going to add or do you already have a feature where you can talk your prompt so like speech I guess I could do it with Whisperflow or something like that. But what do you recommend for that?

50:30Thomas Iljic:Though you're right. I mean, right now we don't have kind of like a little mic button here, but it's been interesting to see how many people have been asking for it. I think as the models get a little faster and latency gets a little smaller, I think that will become very interesting because then you pop this up in Bing in your screen and then you iterate on the output conversationally until you're happy. A lot of people have the most. Yeah, I would much rather as as a person who's, you know, been a writer for years, who's done a number of creative endeavors. I can very much see why the ability to just sit back and talk out loud about what you want would be more conducive to creativity than leaning over on a keyboard.

51:12Thomas Iljic:You know, it depends. Screenwriters might quibble with that. but that's why they might that's why you need a different you need the different modalities for everyone's skill set right like a storyboard artist would want to draw a director would want to talk it out you know that's right cinematographer would want to shoot it you know yeah yeah and over time what I'm hoping is we you know we turn this is to basically more of like it should feel almost like an assistant director kind of you know persona almost where like you're like wherever you are it should meet you there it's like hey you want to brainstorm about you know the idea and just kind of poke at your script maybe we can do that.

51:46Thomas Iljic:Or you're trying to have this output and you want to refine, okay, let's try. We give you a version and then you can say, no, no, this is wrong, change it here. It's all about this, I would say, iterative process. Again, this is where we went with the flow name is how do we get closer working with creative to understand how they operate and how we can bring that up. And speaking of iteration, this is maybe kind of like along those lines of keeping improving the capabilities. Here I generated a clip where I have this leopard I guess walking through the flowers. We've introduced a lot of different ways for you to start kind of like meddling with the video itself.

52:23Thomas Iljic:So inserting, removing, changing the camera angle. So obviously the more we, I'll show you insertion first and then we can dig into for example our camera changes. What's super interesting here is, again, because they're little simulators, or maybe that's the mental model I encourage folks to think about is if you insert something, the model will say, well, okay, lighting is like this, motion is like this, this is how things should behave. Here, just for the sake of, again, speed, I'm showing you an example I generated earlier, but maybe I can insert, if you remember the motion, a Labrador. Here, I just didn't specify, but I just said, you know, the Labrador and the model decided that the Labrador would be in the back.

53:08Thomas Iljic:Or you can also specify where it should be. So for example, let's take this one. I'm just going to delete the previous prompt. And, you know, you can marquee select maybe this area and you can say a cat in a shiny or shiny metal armor is riding the leopard, for example. Hopefully I didn't make any tackle. I love it. I love it. Yeah, that's right up our street. It's generating right now. We can wait. We'll see that particular version in a second. But while, you know, I did a few with like, so this is probably what, this is what my first attempt looked like last time. That's cute. But you see it's bouncing with the leopard.

53:55Thomas Iljic:It tries to keep the other elements constant. And the other part which, you know, I don't know if I have an example here, but if you were introducing kind of like a metal orb in the scene, you would see the reflection on the orb, etc. So really, it feels like almost like a little simulator in a way. I love it. Et cetera. While this is generating, I was going to poke at the image tab, but first, maybe you have some questions. I did have one that just came up. let's say i'm a filmmaker and i've already went out and shot um some elements right uh what could i upload one of my video clips and have it edited for me like take stuff out and or continue it on yeah so right now in the in the flow tool we don't support video upload just yet we support it for the api i believe so you're able to go in eventually we will get there oh there you go you have our cat with like a much bigger armor in this case oh yeah oh yeah that cat looks like He looks like serious business.

54:54Thomas Iljic:He needs to go to the Gladiator arena. And you can see more of the reflections of the scenery. So it's trying to imagine what's on the other side of the camera. Yeah. What I was going to say is, so we don't have video upload yet, but eventually we will. I think we have it right now with the API in Vertex and Cloud for people who want to fiddle with it. Cool. But the idea is you should be able to bring your assets. I think this world where you're fully in the generative space, like basically the generative and non-generative, it's just a tool. those two things should blend. You should be able to bring your images.

55:27Thomas Iljic:You should be able to bring your videos. And then when you need to generate something, you should generate. And probably you're going to use references, which are also anchored in your actual work. So in this case, we inserted a random cat, but maybe I have the reference of the cat or I have the reference of the armor, and we put it. Maybe it's your own cat. Exactly. Maybe it's your own cat. Maybe it's yourself. You see what I mean? I think that... If I'm going to ride a leopard, I want a suit of armor too, to be fair. I was going to say, what did you want to show us on the image tab as well? Oh, yes.

55:57Thomas Iljic:So we support both video and image, and we're going to make it more and more seamless to jump back and forth. So right now it's two tabs. Eventually it might be just like one big canvas, and you can kind of iterate it in any of these modalities. But the reason why image is very interesting to me, A, because the models, I would say, are ahead of the video models. I'm going to explain what that means in a second, but you'll see kind of where they're going. You'll see the same themes, but just like push to the extreme, I would say, in the image land. And then two, because like there's this big thing of like latency.

56:28Thomas Iljic:A lot of people really like to control like the frames. It's like this is where it starts. This is where it goes. I want to storyboard and then I want to animate. So we really want to provide that flexibility to go back and forth. So here are a couple of like fun examples. First, obviously, I use myself in there. As a pirate, I'm on board. I support this. So here, for example, and this is with our latest image model, which we, you know, Nano Banana, in this case is Nano Banana Pro, which released a couple days ago. Yeah. Here, you know, I put a picture of myself and I said, you know, I want this person and I want six poses on a 16 by nine frame.

57:07Thomas Iljic:And you can see how it's able to basically just generate my likeness in the 16 by nine frame. I think we just added 2K and 4K upscaling. Resolution matters a lot, obviously, for creatives. Totally. And, you know, once I have this frame now of me in multiple poses, I can actually ask, you know, very interestingly, hey, come up with, you know, costumes or here's a reference of this jacket that I want to be able to wear. And what I'll call out here is with the image models, you're seeing this this evolution where we're moving from just generating an image. You just ask for something, you prompt directly the generation side to you're actually talking to Gemini.

57:46Thomas Iljic:And so Gemini does the understanding, the reasoning. If you tell it like it's transformed this image into the 1970s, it's going to figure out all the elements of the image that should be updated. And then it's going to generate for you. So you don't have to talk to the generative side. You talk already to this intelligent counterpart who's able to parse this or to take initiative. In this case, I just needed a bunch of costumes that did it. And then because it has a lot of good visual understanding. So let's say I like this pirate here. So this is me as a pirate, I guess. I can say something. So many great costumes.

58:21Thomas Iljic:I know, right? Sadly, my Halloween costume game was not up to what I'm showing here. But you can say things of like, cool, take this reference and now extract the pirate. And he's standing in a Tokyo subway station. So now this is me on a Tokyo subway station. And you can go even further, which is because, again, it's powered by understanding reasoning generative is all together now in Gemini. What that means is you can say things like, you know, go zoom on the shoe and give me a few angles of like the shoe. And you see it respects the scenes, the angles. You can see kind of like the lever, the texture.

59:00Thomas Iljic:So it's consistent from all angles. Consistency has been this really, really hard piece. Yeah, that's amazing. Yeah, the buckles are in the right place, the seam. Exactly. The buckles are in the right place. You can do, people have been playing with things like they use a portrait aspect ratio and they say, here's my starting frame. Show me the frames that come before and after. Or show me this character from multiple angles. And because you're generating these and at the same time, you have like full consistency between the different angles. And so what I would say is, you know, image is kind of a head in terms of like how powerful of these controls are coming together once you have understanding, reasoning, and generative.

59:40Thomas Iljic:As you can tell, they're little simulators. For the people who are very advanced, you can ask for things like a depth map. And it's like, sure, I understand what that is. I'll just give you a depth map. And the final thing I'd say is basically what's super interesting with these models is they should meet you where you are. So in a way, if you're very advanced and you know to ask for this, the model will take a stab at it or Flo will take a stab at it. But if you're like my daughter, taking this tool, she's like, I just want money on a unicorn. And it's not that different from an interaction perspective.

1:00:11Thomas Iljic:It's almost like the depth of the tool gets hidden because now you have this layer of intelligence that can parse the ask. And so that's really exciting for me. I think it's just a new paradigm. So it's hearing what you say and then thinking, okay, that means I need to grab this tool and I need to grab this tool and that's, or whatever maybe not tools, but specifically like here are the techniques I'm going to use to make that happen. Absolutely. And so you'll have, you know, if you think of Gemini as, you know, powering all of this behind the hood, exactly. Gemini can't all tools. So if you're saying I need the reference of that picture of that video, eventually that's where it'll be able to just go and fetch, you know, the information to be accurate, you know, in the same way that, you know, you would ask it to go to search for information.

1:00:53Thomas Iljic:I have a question for you. That red dot at the top there. What happens if I press that? Does that turn it into a video? Oh, I see a red button. What would I do? I guess I would go back to the video side and I would take it. It's good feedback. We should probably change the default color of this. So when you click on an image, it pops up this edit view. So if I wanted, for example, to iterate on this image and I want to say, I don't know, maybe change the camera angle or I wanted to doodle on it. So I wouldn't, you know, I want to add some text here. This is just the color. Yeah. So when you click on an image into the image editing view, when you click on a video, you go to the video editing view.

1:01:34Thomas Iljic:Got it, got it. Wow. So I guess if I wanted to create a storyboard or even lay out all my clips on a timeline, does Flow have a feature for that? I guess for storyboard, I could take it to mixboard, for example. Yeah. But yeah, what's your... In the current version, we have what we call a scene builder, which was, it started from video. So right now it's video-centric, but we're going to update it to make sure it supports both images and video. So for example, if you take this video, I can say add to my scene. I can switch to the scene builder and now it's going to lay down in a few seconds on this timeline.

1:02:11Thomas Iljic:I had this over clip that I'll just delete. And now what you can do is you can say, maybe I'll play it for a second. Now what you can do is you can start trimming, you can add over clips, but you can do interesting things, which is I can also do things like extend. So for example, mid-action, eight seconds, you know, only goes to so far so I can extend the clip while keeping the previous frames as a reference for the model. So, you know, you could say. So we might not have to watch AI videos that are only an eight second clip. No, that's right. I nerd out and like to watch like, you know, AI films on YouTube at night once in a while.

1:02:54Thomas Iljic:And it's funny because like the early ones, there's very much like three seconds, three seconds, three seconds. They're movie trailers. Five, five, five. That's right. And actually, that's a fantastic point. So you can extend the clip. So maybe they continue walking through the field and it will generate some. Maybe like a rival, maybe like a rival animal, like an armored dog riding an elephant arrives. Sure. Yeah, we could do that, too. You know, like an elephant answers the scenes or like they stumble upon an elephant. Yeah. Stumble upon an elephant. We'll see what happens. but the there you go a little bug I'll investigate that it happens the consistency piece is exactly right because you want to do multiclip this is why people use imagery because consistency in imagery is at a much higher level right now than I would say video and so what people do is they'll get the consistency at the frame level and then they'll use these for the different shots So, you know, a different angle and start there or like condition the first and last frame.

1:04:05Thomas Iljic:So what you can do in any clip, by the way, is you can extract, you know, maybe, you know, I'll push. I'll go here. Maybe that's the frame I want. I can use that as an asset. I like that. And now I can keep editing that image or create a new shot from it. So now I can say maybe, cool, this is great. Show me this. This scene from above, right? And so now it's going to generate from above and I could just start from there again. So you can go back and forth between image, video, animation. I'm thinking about that. Yeah. From my experience working on sets, you know, you have a shot list typically, right?

1:04:46Thomas Iljic:The director and the DP. Yeah. So you could generate, you know, according to your shot list, all the different angles and then you can pick the angle that you like the best. Oh, that's cool. Look at that. So now I could... I keep picturing him running toward the camera. Let's do it. Let's say running towards the camera.

1:05:09Thomas Iljic:And so now what you could do is you could use this as a first frame. You could use the other one as a last frame, for example. And you can think of like a transition of like maybe it's a drone camera shot that goes around. So basically image, I think, is ahead in terms of like how controllable and how consistent it is. and also the latency aspect. And when you start pairing it with video, it's very powerful. Okay, so there you go. I think we kind of switched that maybe to some extent compared to the previous ones. We may want to, you know, regenerate that. But this is how you go. But that's still like, yeah, that's still exactly what he asked for.

1:05:40Thomas Iljic:He asked for, you know, they're running towards the camera. Yeah. That's cool. So this is, yeah, I would say this is this exciting space. And yeah, we're excited to just make flow much more controllable. Again, molding clay, figure out how you navigate between all of these different types of inputs. Obviously, multi-clip is going to become very, very important next year as we get to the form. But again, iteration, molding clay, and the different inputs is really key for us. So we're excited to keep working with creatives. Tomas, do you see a time where we're able to get to clips in, you know, minutes in duration, like a film scene?

1:06:25Thomas Iljic:Yes. Across models here, you can keep extending 8 seconds, 8 seconds, 8 seconds. I think you can see there's other models who will also generate like 10 or 12 seconds already at once, maybe more. So we can generate more and more. I think the question I will have is more what's the level of controllability? So is it feasible or is it going to be feasible? Yes. We'll be able to generate longer form clips. How useful it is and how do you iterate on top I think becomes the question. You know, if you look at most shots in most movies or anything, it's really like a few seconds at most. And so I think we're dealing with like, cool, what's the right level?

1:07:03Thomas Iljic:We should probably give you some control over the length, right? Maybe you actually want two seconds and really very precise. One minute, it becomes, what do you feed into it? Do you feed a, you know, multiple scenes? You have an idea. You just want to get a first draft. It will start probably deviating from our vision a bit more, but maybe that's useful for ideation. So I think it's that trade-off. Yeah. But capabilities-wise. I think about it in the context of in film, there's sort of this running competition to try and have the longest one-er, right? The longest one-shot where it's, you know, you're trying to have a whole movie that's one shot.

1:07:41Thomas Iljic:But I think in order to do that with AI, you would almost have to be able to pause the shot at any point. I almost think it'd be better to do it with a world model where you could like almost go around, like you generate this clip. Yeah. And then you could like go all around them and be like, okay, now like let's have the camera go up here and then, okay, clay and have I keep going. Do you get what I'm saying? Where the film's film process began with the creation of the universe within the film. Right, exactly. I love that because I think I can actually, maybe I can give you a quick preview of like, or a view of the genie stuff, which I think is also what you're talking about, at least our Google instantiation of it.

1:08:22Thomas Iljic:But I think for me, if I think of what the future looks like, is you create a world in a way. You specify the world and then you shoot in it. And so the ability to pass time, go around, it remembers what's happened. There's kind of a glimpse of that if you look at the Gini model, which is a different model than the one you have in here. But if I kind of show you this one. It's so cool. You probably saw that.

1:08:49Megan Li:World memory even carries over into your app.

1:08:51Thomas Iljic:Sorry, I'm just going to mute it. What's really cool in this one is, so it's more real time, which again, real time has its advantages and disadvantages, obviously, it's like how do you pause to control. But what's really cool is it has a really long context understanding. And so if you turn back around, it has states. So if you want to do location scouting, for example, you're creating this location, this world, you can keep navigating in it. You can take another shot for the beginning of it. You can modify what's in it. And so I think that back and forth between generating, then re-navigating within what you've generated, changing things, going back in time.

1:09:26Thomas Iljic:Like, I think this is, to me, this is all the super exciting stuff that you could do today. And there's a lot of things that we're still in this incremental phase where, like, it feels like just, like, doing some of the things you could already do just differently. Right. I think we're going to start seeing this year a lot of stuff of, like, you know, I want to change this and I will probably get this change across the entire movie. the jacket is different boom do it like all day before or like I need a location this is the entire definition of my location I'm going to navigate in it I'm going to change this it's a bus here it's a car here and now I can use this as the reference to reshoot the scenes or have the scene being kind of regenerated using that new reference I think it just the toolkit is going to expand quite quickly this year yeah hopefully bridge just new things I think that will open up a lot of people's minds to using this as a creative tool, too.

1:10:20Thomas Iljic:Maybe some people who are more skeptical of AI, because then you have more control than you had before, which is awesome. You know, something that's unique about AI to me and tools like, well, Flow and plenty of other tools for that matter, is that when you look at them, we already know what their future looks like. We know there will be longer videos. We know they will be more manipulative. We know there will be audio controls that can be done and things like that. It's just, when does it get here? And it's really neat that we can see those things ahead of time and know that people are working on that right now.

1:11:01Thomas Iljic:Yeah. And I think what's crazy is it's rare, at least in my field or in tech in general, outside of AI, it was rare that you could be like, yeah, you know, it's going to go this way and this will probably happen and this will also probably happen and it's just a matter of time. And, you know, is it six months, a year, two years? It's kind of crazy that we're able to make these statements fairly confidently, you know, like, sure, I think at some point I should be able to navigate in the video. I will imagine the world I will walk. Surely by K2. It's funny that we're able to kind of like, we've all become fairly comfortable kind of making these statements.

1:11:34Thomas Iljic:But when you think about it, it's kind of crazy. Yeah. It's very much predicting the future. And, you know, it's changed even how I think people build. I think you've got to do, whether it's software, whether it's projects like this, you've always got to be working toward what's available at the end goal of your project. And you probably know what some of those things are. Right. And, you know, the worst case, you get there early. How do you handle that? It's really fascinating. Yeah, how do you handle that? I guess the last question before we go, as leading the product development of this, how do you handle this knowing that the famous line, oh, this is the worst it'll ever be or whatever, you know that the models are going to get better.

1:12:18Thomas Iljic:So are you constantly projecting into the future what you think the capability will be in order to design the features? How do you design? Yeah, it's a great question. It's actually, we've changed quite a bit how we operate, I think, even just in doing product in this space. the way I think about it or with the teams, the way we've kind of worked about it is we try to predict, we have convictions of where the future is heading. And these are like, for example, a year and a half ago, I was like, it's going to be show and tell. Just like, I don't want to hear about prompts. I'm just tired of seeing people just do these things.

1:12:51Thomas Iljic:They're like over-optimizing by prompting just because there's no better alternative. Like, okay, so you have this conviction. Yeah. Then you build kind of like the versions, you build almost for that future, even though the models don't support it really well yet. So it's a little clunky. A good example, you mentioned WISC, for example. So for people who haven't used WISC, WISC is a tool where our idea, it was a premise for show and tells, like you put subjects in style as images and we'll take that as a reference to generate an image. But we didn't have good consistency models. So what was happening is we asked Gemini to actually describe the prompt in the early stages, and then he will rewrite the prompt for you just based on text descriptions.

1:13:26Thomas Iljic:The moment we introduced the proper models that could actually do consistency pixel wise, you saw just adoption just go like straight up, just like immediately. And so the reason I'm mentioning it is like you build, you have that conviction. So you know, there's going to be a window. You build for that window and then you get an early sense from like the early adopters that this is kind of right in terms of what people want. And then the moment the capabilities hit, you have something that's ready on the market to kind of just like ride that wave. It's kind of like, you see what I mean? Yeah. Yeah.

1:13:58Thomas Iljic:You prepare for those windows rather ahead of time. It's not different from what a lot of folks have done in traditional product developments, but the windows are much shorter and they happen much faster and the future is a little bit, it's hard to predict six months to a year down the road maybe. So that's the challenge. Are you willing to? Where do you, while we're here, where do you see the tool a year from now? Yeah. Again, I think my dream for flow in particular is I think it becomes your assistant director where it really feels that you're just like talking to like some you know intelligent counterpart that will and again it's not push button it will just do passes at whatever you're asking and you're able to iterate so you could do everything by yourself if you wanted to we're just saying hey cool we took a stab what do you think and then we iterate on it um and then i think from the model perspective i'm hoping it's more and more to like you build the world you shoot in it at least for something like flow right which is you can already see that that's kind of what this is telling you is like you're showing the characters that are in it you're showing what the scene should look like you're world building in a way and the question is more like what is the right model slash user interaction patterns to really allow for that you know larger context windows things like that behind the scenes will happen wow so cool thank you so much for showing this to us no my pleasure yeah so tomas where can people go to try flow uh yes so they You can go to flow.google.

1:15:22Thomas Iljic:Very easy. I will lead you to this page, and that'll be hopefully your entry point into this experience. What a gorgeous page, too. I said that when we got on, I know, but every time I see it, I'm like, dang, that's good-looking web design. Yeah, we have a team called The Creative Lab. They've been absolutely fantastic, really to kind of channel the creative power of these tools, I think. Thank you. Awesome. Really appreciate you joining, Tomas. It was great to meet you, man. Great to meet you too. Yeah, until next time. All right. So today we are joined by Megan Lee, Senior Product Manager at Google Labs, who's leading Opal, Google's no-code AI app builder that's democratizing AI development by letting anyone create functional mini-apps using just natural language.

1:16:06Thomas Iljic:Megan, welcome to the show.

1:16:07Jaclyn Konzelman:Hi, thanks for having me.

1:16:09Thomas Iljic:So Megan, I wanted to ask your leading Opal, what was the original, I guess, like impetus or problem that you were trying to solve with creating Opal?

1:16:21Jaclyn Konzelman:Yeah, that's a great question. Opal actually started a while back and it didn't start as an original product idea. It actually started with this amazing group of engineers who's still working on Opal now, who are thinking through ways to help people actually chain together multiple AI prompts and calls to different models. So in their exploration, again, this was super early days, they started thinking through maybe this would be an open source library. Maybe this would be targeted towards developers. And they kind of started to build out what became a really robust open source framework for thinking about chaining AI multiple calls together.

1:16:54Jaclyn Konzelman:But as they kind of continued to build, things moved super quickly. And we saw that there were so many tools and different libraries available for developers to do similar things, to be able to create and code these really complex user workflows with AI. but that there kind of wasn't really anything here for folks in between. So how it evolved is that we went from looking at this developer library. I think as a team, we were kind of trying to figure out as the market moved where this might go. And we were trying to think what might this look like as instead of just like a standalone library, an actual product.

1:17:25Jaclyn Konzelman:How might we make this something that can help more people build with AI, not just developers, since we saw so many different solutions in the space. And so where it started to evolve was kind of first saying, all right, what might it look like to have this kind of front end experience? Let's build out this kind of visual editor to help make it easier to think through, instead of just building blocks of code, what would it look like to actually have building blocks that explain what's happening with the code underneath it? And we started to build up things like, okay, maybe there's a step where we talk about generation and that's like how you call a specific AI model.

1:17:55Jaclyn Konzelman:But maybe we also start to have different types of tools that we visualize and kind of abstract away what's happening underneath. And from there, I think we're starting to get closer and closer to, okay, it feels like there might be a product here, but who might this be for? And I think it's an interesting kind of time in AI where traditionally you think of a very specific problem and then you build a solve for that. And I think with AI, because it's moved so quickly, it's kind of almost like reverse in a way where we had the technology to say, hey, it's really cool to chain together multiple AI calls, see people create custom workflows.

1:18:24Jaclyn Konzelman:But how do we actually find the right set of users who maybe don't have access to this capability or it's not so easy to kind of tap into that today? And so from there, I think the biggest unlock, in addition to having that visual editor, was when we started working on this kind of idea of a planning agent and saying, what if we had a way to translate natural language into a completely built AI workflow that already hooked up to certain model calls that could use the best of Google's models and also could access things like tools like web search all in one place? And so I think that kind of became the unlock moment where once we had that ability to let someone enter in a natural language, immediately map out that workflow.

1:19:00Jaclyn Konzelman:And again, it's all built off of that original library of open source code. We started to say, OK, this actually feels like this might be interesting now. And in a way, it also feels like it's really different than what we originally set out to do, right? Build more tools for developers. Instead, we're saying, what if maybe this allows people who don't know how to code or less technical users to actually build something cool and actually useful? And so that's the thesis when we put this out there about five months ago now. But it's kind of a long, winding journey to how we landed on Opal.

1:19:31Thomas Iljic:I'm really excited about Opal. And the reason is because, well, kind of exactly what you said in terms of making it possible for other, for people who are maybe, let's say, not traditional engineers to create basically essentially agents, right? That's basically what we're doing here is we're creating an ability. So me without having to get into an IDE or a command line interface, which is even scarier to actually be able to build an agent. And I think one of the recent things that was just announced is that Opal is actually coming to Gemini. Is that correct?

1:20:01Jaclyn Konzelman:Yeah, we just launched that a few days ago. So it's been a very crazy season for the team. But we're super excited because I think Gemini represents so many users out there who are learning to use AI and today maybe primarily use chat. So we were really excited to bring Opal to that type of experience.

1:20:17Thomas Iljic:That's awesome. So I'm wondering if we could get into it and check it out. my first question is, how would I, if I'm in Gemini, be able to get to Opal and start building with it?

1:20:28Jaclyn Konzelman:So when you actually land in Gemini, if you look at this side panel, you can actually see Gems page. So the reason we brought Opal into the Gems experience is if we scroll to the bottom, the classic Gems are still here. But the thinking is that Gems, for those who aren't familiar, is more of this kind of like chatbot Gemini experience, but with a specific instruction so that you can have models behave a certain way. And in a way, it really all comes down to how do you customize AI? How do you customize Gemini for your specific needs? And we felt like Opal was an extension of that vision. But rather than just having AI talk to you differently or have like a specific custom version of Gemini chat with you differently, we wanted to think about what it might look like to have Gemini act differently.

1:21:07Jaclyn Konzelman:And that's where we felt like Opal really fit within that mission. So when we were looking for a place to kind of bring Opal into that larger Gemini app family, Gems felt like a really natural place to do that. And we really debated around kind of what to call it, how to think about branding. And I think at the end of the day, what we wanted to do was just to move really quickly and bring it some way into the product and let people try it and learn from that experience. That's awesome. That's how we landed here. Yeah, with gems from Google Labs. And I'll actually show you real quick. If we just open up one of these gems or labs gems, if you will, this is in an Opal.

1:21:38Jaclyn Konzelman:The only difference here is that because we knew that certain Gemini app users likely aren't as familiar with Opal, we didn't want to drop them right into that visual editor experience Because even though we wanted to build something kind of simpler, more catered to folks who don't necessarily have a traditional developer background, to your point, we wanted to kind of introduce potentially and play around with an even lighter version of Opal for these types of users. And that's where we came up with what we like to call our tinkerer user experience, where these are folks who are still trying to understand what does it even mean to build a custom AI workflow or build a custom AI mini app.

1:22:13Jaclyn Konzelman:And so here you can see it's a little bit simpler than what you might even see in Opal. And I'll actually show you real quickly. Let's assume, and actually one more thing I'll show you, is you can also start from natural language. So if you're familiar with Opal, yeah, that's a great starting point, right? And app that, and you can actually type something up. Let's actually try something here. That's a prompt. This is a little app idea. Here you can see your original prompt. but we're going to try to build an app that takes a photo of my fridge with ingredients and creates recipes using those.

1:22:45Jaclyn Konzelman:This is a personal thing. Corey needs this. Yeah, it's a personal point, right?

1:22:51Thomas Iljic:Yeah. He's always talking about how he needs this. He's going to be jealous he wasn't here.

1:22:56Jaclyn Konzelman:Yeah, because there's always things that are expiring in your fridge and you need to make the most of it. And that's kind of where we came up with this one was just how do we make it easier to not let so much food go to waste? And you don't always have time to go through everything. So the idea here is we entered in that natural language prompt and kind of what I talked about before, automatically translating that user intent. And what we've done is just broke it down into steps here. And so the first step, if you open this up, all right, question to user. So this means when I go through the app, what am I actually seeing?

1:23:24Jaclyn Konzelman:What are you asking me to do? And then the next step, if we look here, this is a prompt summary. So this must be essentially a step that includes an AI model, but we've kind of abstracted that away a little bit in this kind of lighter tinkerer view. and then we keep going and here you can see the second step now it's actually generating images for that experience and then finally it's putting it all together into this web page so we can actually give it a try and this is kind of the app preview space upload a photo of your fridge i'll go ahead and do that

1:23:53Thomas Iljic:all right i could have used this three days ago we just uh we just uh got rid of a bunch of

1:24:00Jaclyn Konzelman:expired food yeah you could have come up with really creative recipes or gem and i could have done it for you. Totally. And I'll, for those who aren't familiar, this is really similar to the actual Opal experience, except it's a little bit lighter. So while it's doing that, what I might also do is open up the advanced editor so you can see what this app looks like inside of Opal. So we started that journey in Gemini, but if you clicked that open and advanced editor, you can actually open up the Opal experience for even more control. So that's kind of the thinking was how do we make this a seamless experience for people to start either in Opal, if they have a very complex use case or they're very familiar with the kind of visual editor and then be able to use those apps in Gemini.

1:24:43Jaclyn Konzelman:Or if you want to start from kind of the light view in Gemini, you can still go deeper and end up in this visual editor experience. All right. So I opened it up there. And again, we can run this here, but it takes a little bit of time. So instead of rerunning it, I'm just going to show you what this looks like. So I don't know if you got a glimpse of my fridge earlier, but here there was tuna in there. There were some veggies that were about to expire. There was some fruit. And I think there was some deli meat as well. So you can see different types of recipes and it's all put together in one place.

1:25:14Thomas Iljic:That is so cool. That's awesome. And I love the transition. That is genius from a product perspective where you're working with something that is familiar to people who've used gems and then you're slowly onboarding them onto this less familiar user experience with Opal. That's great.

1:25:32Jaclyn Konzelman:Yeah, thanks. So should we start from the beginning? Should we walk through workflow in Opal itself?

1:25:36Thomas Iljic:Let's do it.

1:25:37Jaclyn Konzelman:So one use case that I'd like to talk through, and I'll open it up here, and we won't actually go ahead and create this. I'll just maybe intro it a little bit. So this Opal I really love because it started as this idea where, I don't know if you remember, maybe a few months ago, there were these really fun AI created ads where people were making kind of like branded ads. And you start with this box and you're not quite sure what's happening and it starts to rumble and then it pops open. And then all of this brand's kind of products pop open and it's this really fun, delightful, short clip. And so it started there where I was like, what would it look like to recreate this?

1:26:11Jaclyn Konzelman:And then I saw people circulating the prompt to generate these types of video ads. And then I decided, all right, let's try it for my friend's brand. She just started an interior design company in New York and I wanted to see what it would look like with hers. because when I tried this kind of national consumer brands, it worked super well out of the box, just one prompt. But when I tried it, I realized that the models had no idea what her indie brand was, and it had hallucinated a completely different type of product. And so I was thinking, this seems like a really good use case for Opal, where I could actually try to customize and prod the models a little, give them a little bit more context to work with, to make sure they can actually generate a video that feels true to my friend's brand.

1:26:48Jaclyn Konzelman:So I made this Opal, which I call exploding box video ads. But the idea, and I'll walk you through kind of what this looks like, is it starts by asking the brand name. And then the next step is that it assumes that maybe this brand isn't a well-known, everybody knows this brand yet. Maybe it's a newer brand. Maybe it's a smaller business. So the first thing it does is search the web. So it accesses this tool. And if you're in Opal, you can look at what tools are available by just clicking the add sign. So here, there's a bunch of other things you could pull in, like a maps listing, getting a specific web page, getting the weather.

1:27:22Jaclyn Konzelman:But for this use case, I think searching the web is a good place to start. And then here, you can also edit the prompt. But you can see here, we're taking in that user input of the brand name, coming up with the business name, and then continuing on. And here, as we're kind of going through this example, I'll just talk you through some of the steps in Opal, because I know sometimes it's hard to grok exactly what's happening when you enter this visual editor space. It can look really complex. These are actually just pretty simple prompts. If you break them up separately, each one's just a step that does something slightly different.

1:27:53Thomas Iljic:Right.

1:27:54Jaclyn Konzelman:And so here, this next step, essentially, it mimics the ad format that previously I talked about, that ad, the prompt circulating. So here you can see kind of detailed prompts that it generates. And now it's asking the product, great, go ahead and create something similar to this.

1:28:09Thomas Iljic:Could I ask a question? Is that the output requirements right there under metadata? It's saying create it and put it in this format? Yeah.

1:28:18Jaclyn Konzelman:So this is actually just generating the prompt. Oh, OK. One thing that Opal often does, especially if you use natural language to start, is it breaks up everything into smaller steps. And the reason it does this is it just gives you a lot more control. So one other use case that I really like to use Opal for is testing things. Like if I'm testing this prompt and I'm like, OK, I think I want the model to mimic this longer prompt. And sometimes you run it all the way through and you realize that it didn't quite get something right. It's really easy to just come back and edit the prompt. So maybe you can edit it manually here.

1:28:48Jaclyn Konzelman:In Opal, you can also use natural language by clicking this button. That just edits a specific prompt. And then sometimes you want to edit something about the entire experience. And in that case, you would click here and say, edit these steps. So you could, in a way, rewrite the entire workflow or make edits that impact multiple steps of your workflow.

1:29:07Thomas Iljic:This is great.

1:29:08Jaclyn Konzelman:Yeah. And then this last step is the generation step. That's what you're asking for. So here it just references that generation from previously.

1:29:17Thomas Iljic:Right. Got it.

1:29:18Jaclyn Konzelman:And if you'd like, we can run this through. Yeah, let's try. So, all right, let's give it a shot. So first, this is my friend's brand. And then I'm going to upload an example of the product. So this is an image of a candle holder that I want to be featured. and then as it's running what I really like is this kind of visual experience where you can see where we are in the process because the thing about chaining together multiple calls and different tools like searching the web is sometimes it can take a while so it's nice to be able to track it here and then the other thing I like to do while these mini apps are running is I go to this console tab which you can think of as maybe like a for a developer you might think of this as the place you go to debug but really what's happening is you're just understanding what's happening underneath the hood.

1:30:06Jaclyn Konzelman:So you can see that we're on this step. So let's try to get there. But you can see the interim results. So you can start to kind of see what's happening even before it all gets put together.

1:30:15Thomas Iljic:So you can click into each step and see the actual output of each step. That's great.

1:30:21Jaclyn Konzelman:Yeah. So it kind of gives you some sense of things are happening even if you're waiting here. so yeah that's prompt template yep a completed prompt that's great now it's working on the video which often takes the longest but we also try to show you a progress indicator I like that as well I will say I've used uh sorry to interrupt you but I've used

1:30:46Thomas Iljic:a lot of no code other no code um kind of node-based workflows like this and this is by far the most intuitive, the most that like where it's abstracting away a lot of the other stuff that's perhaps intimidating to people like structured outputs and, and, you know, connecting API keys here and there. And you see this like list of boxes you have to fill in and you get so intimidated here. It's all prompt based seemingly. And there, you can ask questions about it at any step along the way, which is just fantastic.

1:31:20Jaclyn Konzelman:Yeah, I love that there's to your I love that you can think about it as asking questions, right? If you want something to change, you can enter it as a change. And I think where we'd love to see this go is even more of that kind of natural back and forth. So I can definitely see chat also becoming a future part of the Opal experience. All right. It looks like our video is ready. So let's switch to the app tab, which just kind of makes it a little bit bigger. And then we can also full screen.

1:31:49Thomas Iljic:That is awesome. and would you say that this is accurate to the brand yeah and i what you what i should have done

1:31:57Jaclyn Konzelman:is shown you what happens without this opal because the first time i did it my friend's brand is inner child shop but it's very chic as you can see here it's kind of bohemian it's got these nice textures the first time i tried this is just a prompt it thought it was a children's furniture brand so i got unicorns and it was so off-brand and so it's really fun to see something like this where it feels so much more true. It's checked out her website with web search. The Opal workflow kind of is able to take in that product image. So I really wanted that kind of candlestick to be featured because I think it's a really fun, funky representation.

1:32:30Jaclyn Konzelman:So it's really nice that that also is incorporated.

1:32:33Thomas Iljic:With the prompts, were you really specific, like make the candle in the middle of the room or did it just kind of pick up that naturally?

1:32:41Jaclyn Konzelman:That's a great question. What we did was, where is that prompt?

1:32:49Jaclyn Konzelman:The final piece. And so here is where I edited it. And I said, the final piece, the yellow throw blanket. And so what we did in one of these prompts is just specify that this product being uploaded, which was that image of the candlestick, should be the final piece. And so that replaces that last touch.

1:33:07Thomas Iljic:That's awesome. That's so cool. I love that you can get that granular with it.

1:33:11Jaclyn Konzelman:Yeah. Should we walk through a different example too? Sure.

1:33:15Thomas Iljic:Let's do it. Yeah. And perhaps we could maybe go a little bit deeper under the hood. If I'm someone who's doing this for the first time, what are the things that I need to prepare ahead of time in order to do this well or something of that nature?

1:33:31Jaclyn Konzelman:Yeah. Would you like to start from just natural language? Sure. The entire creation flow?

1:33:37Thomas Iljic:Let's do it.

1:33:38Jaclyn Konzelman:All right. let's try so here for someone trying it for the first time you'll enter into this page and so far I think we've seen a lot about natural language so let's start there and then I'll also walk you through for power users what it might look like to build this from scratch but I'll talk you through it instead so let's use a different type of example here I have an app that takes a user provided core ideas input analyzes and expands on it to generate a main headline and then from there I'm having it just create a few different drafts of social posts so I think of this as almost like like a Kickstarter or Brainstormer, and it helps you just come up with one concept translated into different iterations of text and an image as well.

1:34:16Jaclyn Konzelman:And so here, while it's working, we can also talk through how to build this with specific steps here. Yeah, yeah.

1:34:24Thomas Iljic:I see there's user input, generate output, and then add assets.

1:34:29Jaclyn Konzelman:Yeah. All right. And now we have something to work off of, so we can zoom in a little bit. So user input is this yellow box. And so if I wanted to, I could literally enter my own user input and add it to this chain. So here we say enter your core idea or topic. And then here you could imagine if you wanted to add something else like enter in a second topic. You could manually hook this up to the next step as well, right? And so what you do is just add a connection. But then the next step, which is trickier, is you actually need to go in and incorporate your second idea, right? Read the provided core idea carefully.

1:35:07Jaclyn Konzelman:also consider the second core idea. And so you can see quickly that this becomes more manual if you're building it with these manual blocks. So the other thing you can do at any point is use natural language to make that change. And here I'm just going to delete this one. And you can see here you might say, allow the user to enter a second core idea. So it's going to work right now. and what this natural language prompt essentially matches to that adding a second user input we just looked at. But at any time, you can jump into specific steps and change them as well. All right. And while that's working, I'll just talk you through the rest of the steps.

1:35:49Jaclyn Konzelman:Yeah, perfect. So again, generate is just, think of it as just an AI call, right? And I'll jump into one of these once this change is done. But you can actually do a bunch of different types of AI calls in Opal. One of the best parts, I think, is that you can do things like not just generate text, you can generate images, videos, as we saw, you can generate audio, music, all of Google's state-of-the-art models are actually included here.

1:36:12Thomas Iljic:I actually have a question about that. Is it set up through your API or how does it actually, you know, how do you actually work with them back and forth?

1:36:22Jaclyn Konzelman:Yeah, that's a great question. What's really nice is that you don't need to have an API key. You don't need to think about something like, how much quota do I have left? What's really nice is all of this is done for you pretty manually. So we've hooked up all of these models for you to use, and there are some user limits. But for now, kind of we're releasing this in a way where we're trying to be responsible. There are some caps, but we also understand that people are still learning what to build. So we're pretty generous with what you can do here. And whether you're creating an Opal for someone else to use, or you're a user and someone else sent you their mini app to try, you have the same limits.

1:36:57Jaclyn Konzelman:So you don't have to think about paying for anyone else's usage. All Opal users are the same to us, and Google manages everything from hosting and deploying it to easy one-click sharing, which we can also do here. Yeah, but you can see a bunch of different models here. We probably need to clean this up, but there's a lot of cool things you can try, including some of our Gentic models like Deep Research, which is really useful. Yeah, that's awesome. Yeah, so this is the generate step, as I think we talked through earlier. This is an example of an image generation step. You can see Nano Banana and the latest Nano Banana Pro.

1:37:34Jaclyn Konzelman:And then this output is really interesting too. So if you think about this block, most opals are some type of workflow. And then there's usually some last step at the end. And so oftentimes people create this web page, which just means putting it all together in one web page. You can see all of your content from previous steps, including images, any content you created. But the other thing we actually see a lot of users do is this save to Google Docs, slides or sheets. So this is one use case that's a little bit more like a workflow where you've done all these steps and then you dump it into a doc or a sheet or a slide so you can go and use it elsewhere.

1:38:10Thomas Iljic:That's cool. Do you have any plans at some point to connect it to, you know, let's say like other tools outside of Google? You know, like if you want to publish something directly from Opal, for example.

1:38:22Jaclyn Konzelman:Yeah, that's actually one of our top feature requests is people saying, can I embed my Opal in a website? Can I embed my Opal in a slide? So definitely top of mind for us. We would love to do that. That's on the roadmap. The other big question we get is people asking, how can I connect it to other sources? Right. How can I connect Gmail? How can I connect it to third party resources? And that's also top of mind for us. So we are working on an MCP integration. We're starting with first parties. So we are thinking about in 2026, we're hoping it's easier to create custom Opal workflows that connect to your Gmail or your calendar as a starting point.

1:38:57Jaclyn Konzelman:We are also thinking through what it might look like to offer more connections within Opal, but still trying to maintain that simplicity where how do we allow you to do things like connect to a different data source without it feeling like that intense kind of intimidating IDE experience again.

1:39:12Thomas Iljic:Right, right. Yeah. No, I appreciate that. And I understand why people request it, because this is so intuitive. This is so much easier than the other tools. Like, just conceptually, you get it that you want to be able to do even more with it.

1:39:26Jaclyn Konzelman:So compliments to the team. Thanks. Yeah, it's definitely this kind of tension that we're always living through as we kind of think through more feature requests is oftentimes people are asking for more complexity. So they want kind of more advanced features. They want conditional logic. They want looping. And we want those things, too. But it's this weird thing where with AI, I feel like the easy part is building it, the technology. We can absolutely support looping. But the harder part is thinking through how do we present something like looping to people so that it's easy to understand what that is, how to implement it, how do I edit that behavior?

1:40:04Jaclyn Konzelman:So if we think about looping as something where it's like, I want this agent to keep going back and asking questions to the user. I want the agent to continue researching until it's done. but how do you represent it in this type of visual experience where it's not code? And that's an unsolved question that we're always still thinking through. So as we look at new features we're trying to build next year, that's kind of always the tension we're balancing is how do we make Opal more useful for people, more powerful without making it so confusing and complex that we lose a lot of the users who really love Opal today?

1:40:36Thomas Iljic:Yeah, it's funny. It's like the easier you make it, the more people want to use it. And then the more that they want more complicated things on top of it, Because you've made it so easy to do this part where you're like, okay, now let's add other elements to it.

1:40:49Jaclyn Konzelman:Yeah, exactly. So definitely some balance and some self-control we have to have because it's easy to want to race to add all the complexity. Totally. But then we don't want to lose sight of what makes Opal really special for us today.

1:41:01Thomas Iljic:Is there anything we haven't covered here that you can do in Opal or anything that you're like really, really excited about?

1:41:08Jaclyn Konzelman:Yeah, I think one thing that's really fun. And I think also one way when we talk about customization that I find myself using Opal a lot with is the ability to add an asset. So if you click add asset, you can upload a file. You could even do something like draw, which is really fun. So the other thing we're hoping to do next year is bring in more support for tablets and mobile. So you could imagine this being even more interesting if you could draw sketches and then upload it as part of your Opal. But any asset that you upload here can be referenced by the model. So if we take a step here. So this opal is kind of a silly one, but I wanted to make it really easy to make just customized kind of pet comics.

1:41:43Jaclyn Konzelman:So I said, upload a photo of your pet. What's your pet's name? Maybe you can even you can imagine adding a step where you can control what type of story is generated. But it creates a comic strip. And when I first did it.

1:41:56Thomas Iljic:Could we customize this for Ollie, our orange cat? I know the cat there is not orange, but perhaps we say make the cat orange and name him Ollie.

1:42:03Jaclyn Konzelman:That's a great idea.

1:42:05Thomas Iljic:Representing this guy right here.

1:42:06Jaclyn Konzelman:a mascot so here because we actually need to um because we need to upload that let me actually

1:42:13Thomas Iljic:grab a photo of ollie and then we can replace it here oh that's even better yeah if you need one

1:42:20Jaclyn Konzelman:i can send it your way i just took quick screenshots let's see if that works um but the idea here and i'll show you this before we upload ollie is we it says maintain the consistent and art style and overall mood, take inspiration from, and then what I did earlier was add, and then select this photo.

1:42:38Thomas Iljic:I love that. I love the fact that you can just add things in the middle of the prompt. That's great. Yeah, we try to make it easy to know what you're looking for.

1:42:45Jaclyn Konzelman:All right, so now let's upload Ollie. So I'm going to delete this reference, and then I'm going to go ahead and add an asset.

1:42:53Thomas Iljic:I see you can also add YouTube videos, right? Yeah, you can.

1:42:58Jaclyn Konzelman:You can upload YouTube videos as a reference, kind of what we're doing here, or you can add it as a user input too. So there's a lot of use cases we've seen where people are saying, take this music video or take this song and then translate it into like a poem or plan an entire party around this specific song, which is fun too.

1:43:15Thomas Iljic:That is fun. That's cool. People are so creative.

1:43:20Jaclyn Konzelman:I've never thought of that. That's honestly been one of the best parts of building Opal is hearing how people are using it. It's just ways I would not have imagined myself. So it's really cool to take inspiration from that. All right. Take inspiration from, and now we have Ollie. All right. Great. And then if I click away, you can see that an arrow actually automatically got added because this prompt references this asset. That just happens automatically without you needing to do anything here. All right. And let's see if we're lucky that we have limited, limited quota of Nano Banana Pro, but let's see if that works.

1:43:55Thomas Iljic:We don't want people to go too wild with that.

1:43:58Jaclyn Konzelman:Yeah, and we are trying to get more capacity, but as you can imagine, I think there's just a lot of demand right now for AI.

1:44:06Thomas Iljic:Would you ever let people add API keys? I know we're trying to abstract away the horror of API keys, but if they wanted more output and Opal was like their favorite way to interact with AI, would you add that ability?

1:44:20Jaclyn Konzelman:Yeah, definitely. I think we would consider it. I don't know that we would absolutely do it. I think this is, again, a good example of the tension of simplicity and complexity is thinking through how do we best support kind of more complicated use cases without alienating the people who really find Opal special today. And I think that's something that the reason we put it out there so early is because we want to understand, like, what are those killer use cases? Where are people, like, really getting the most out of Opal? And how do we make sure that the roadmap reflects what people actually need for that kind of primary use case?

1:44:54Jaclyn Konzelman:Makes perfect sense. A lot of it is kind of more everyday use cases. It's people saying, I'm building this tool for my team, and I don't have any coding ability, which is really cool. or it's technical users saying, I'm a developer, but this is a really cool place for me to kind of sandbox and prototype workflows and partner with my non-technical teammates to kind of iterate on that. That's actually one way that our team uses Opal as well.

1:45:17Thomas Iljic:Oh, yeah. Say more about that.

1:45:19Jaclyn Konzelman:Yeah. And actually, I will, let's see if I can find one, but I'll talk you through it a little bit. There's a few different ways that our team actually uses Opal, but we use it quite frequently in our own internal workloads. One of them is this use case where if you see all these images that are generated automatically in the back of your Opal apps, when we first launched this app, we just used this prompt. And we tried to kind of vibe e-ballot a little bit, but the images weren't great. So we did not auto-generate them. We actually had a generic kind of like pentagon icon. And then we said users have to go into the theme tab, which actually I will show you quickly, and generate the image themselves or kind of come up with different prompts to generate a good image because we just weren't confident in the prompt quality.

1:46:02Jaclyn Konzelman:but what we did was we built an opal to emulate essentially the experience of automatically creating an image so we did and maybe opening this up is helpful I don't have the exact opal on hand but we just said like okay here's the title of an opal here's a description and then maybe here's also a third step that allows users to enter a customized prompt and then we just iterated on the image prompt that we used and what was cool about this is previously it would have been like in the code and then someone has to take the prompt from code and then it's hard to actually live see how that prompt changes with all the various inputs so someone like me wouldn't necessarily be able to help eng it would be a lot of back and forth just to get us to a good place but with an opal it was really cool because i'm like okay i built the opal i took the prompt that you're using in the code and now i can as the pm iterate a ton on what that prompt looks like until i'm happy with it and i can test the edge cases like what if someone has no title, no description, and then enters a completely random prompt?

1:47:01Jaclyn Konzelman:What might that image generation look like? And be able to really test it and kind of do a lot more banging on that prompt until I was super happy with the quality. And we were able to get there pretty quickly because kind of we were able to work in parallel. And then once we were kind of perfecting that image prompt, we were confident enough to actually say, let's automatically generate these images. So now every Opal you make automatically has that cover image because we're feeling a lot more confident about what comes out there.

1:47:28Thomas Iljic:That is so cool. I love that. You're using your own supply. I don't know the right phrase. Yeah.

1:47:38Jaclyn Konzelman:I'm like, it's very meta, but we love using Opal to build Opal. I think another fun use case is we have this Opal that takes looks at our GitHub repo. And then it asks, like, how many days are you looking at? Like in X previous days, what commits have happened? And so it allows me to actually go in. And, of course, I know what's happening in the code. And, like, we chat about what people are updating. But sometimes there's things that you kind of forget about. It's like even if it's last week, right? It's like, oh, I forgot we fixed that paper cut. So it's been really cool to be able to use an Opal where in the last X days, I can see a summary of everything that's happened in the code.

1:48:12Jaclyn Konzelman:And then I also remixed that opal and said, okay, now turn it into snippets with emojis so that you can actually see, right? It's like in the last week we did. And then it's these like really nice short snippets that I can send on Discord to let our community know, hey, we fixed these things that you've been asking for. And it doesn't require as many steps in just coordination and communication and remembering everything that we did. And it's really, yeah, just a nice way to gut check. So now we just run the opal, ping the team and make sure we didn't hallucinate anything, do a quick verification.

1:48:41Jaclyn Konzelman:But for the most part, it's been a really nice way to expedite kind of closing that loop with our user base.

1:48:48Thomas Iljic:Very cool. Do we want to check on this Ollie comic real fast before we close out?

1:48:54Jaclyn Konzelman:Yeah. I realized, though, that this Ollie opal would require us to upload a photo of the cat instead of the reference because what it does is just mimic the style. So we could run it, but I don't think it'll do exactly what you're hoping it will do unless we change up the workflow.

1:49:08Thomas Iljic:That's fine. We don't have to do it. Well, I did see there was some advanced settings there. Maybe for folks who are a little more tech savvy, what can you do with the advanced settings? Anything interesting?

1:49:19Jaclyn Konzelman:That's a really good question. So the way we think about advanced settings is just going one layer deeper. And so it's the whole idea of simplicity, but complexity on demand. So when we open up advanced settings, you'll see actually different types of features pop up here. So with the classic generate step, you'll see that you can add a system instruction. So this makes it really easy to test not just prompts, but system instructions as well. The default is just telling the model that this is part of an AI system. But if you have a custom prompt here, you can do a typical kind of like, you are an expert comic book writer, right?

1:49:55Jaclyn Konzelman:and then play around with how that system instruction changes the output. And again, when we go into that console and you start running this, you can see all those interim outputs come out. That's a good way to kind of gut check what you've done here. The other thing you can do here, so for image generation, for example, when you open up advanced settings, you see aspect ratios. So you can change that set. If you have video operation, you also have different things. And then even for things like user input, you can say input is required. And that's actually fresh off the press. Um, if you want the user to not be able to continue, then you check that checkbox and then you force them to upload something before you continue with the rest of your workflow.

1:50:34Jaclyn Konzelman:Um, and you can also kind of control here what you're requesting users to upload.

1:50:39Thomas Iljic:Ooh, I like that you have any as an option.

1:50:42Jaclyn Konzelman:Maximum flexibility. And a lot of times we default to that. Yeah, exactly. Cause, uh, you never really know what people want to input. And sometimes AI does a really good job with any modality. So that's actually a really nice way to do it. Like, tell us about your day and you can upload an audio recording. You can type text. You can upload a doc and all of that works.

1:50:59Thomas Iljic:Very cool. Well, I have I have just three more quick questions for you. The first one is, so who who have you found is using Opal the most in terms of like their role or use case? Or is it really broad? Like, would you say that there's anyone in particular who's really, really leaning into it, whether that's, you know, people in marketing or people on engineering, you know, spinning on prototypes? who would you say is like most opal friendly yeah great question and to your point it is pretty

1:51:28Jaclyn Konzelman:broad and we've been a little bit surprised by the main kind of reasons people come to opal one bucket is kind of people just trying to automate repeatable tasks so a lot of times this tends to be like knowledge workers it's not even specific roles though i've been really surprised to see that it's everyone from we've had lawyers use create opals a lot which is really interesting. We've had PMs, which was more expected, kind of creating tools for their team or kind of doing similar workflows to what I've shown and talked about our team using them for. But we've also heard a lot of technical developers prototyping within Opal and using it as a way to kind of test different tools and combination of model calls.

1:52:06Jaclyn Konzelman:So that's kind of one main thing here is being able to create this custom workflow and use it again and again. So that repeatability, it's almost like the use case is more common than the profile of the user, which has been really interesting.

1:52:19Thomas Iljic:Yeah, that makes sense. That makes sense. And where would you like to take Opal from here, you know, into 2026? Do you have any big ideas or what's your long term vision for it, I guess?

1:52:30Jaclyn Konzelman:Yeah, I think for next year and just in general, kind of the future, what we're really excited about as a team is thinking about how Opal allows people to just make things. It really lowers the barrier to being able to create useful things. So not just fun things. I think when we first launched this product, we thought people are going to create like fun GIF creators or like a meme generator. And we were really surprised by people's creativity, but also their utility, the way that they were able to create these things that solve very specific pain points in their own lives or their own workflows.

1:53:02Jaclyn Konzelman:And that's got us really excited to think through how do we scale that and make AI something that feels custom to you or your specific team or your specific problem. Like a good example is even people were creating custom kind of job application opal. like to help them find very specific types of jobs. And they were finding things that they weren't finding when they were just browsing the internet, right? Looking on different job boards.

1:53:25Thomas Iljic:That's interesting. Because yeah, you showed us at the beginning, you can tie in web search as part of this. So you can, and deep research I think is in here too. So you could actually create a pretty powerful workflow to help you search and find niche things like that.

1:53:40Jaclyn Konzelman:Yeah, and I think that's what's really inspiring for us is seeing people kind of push the bounds of what's possible to just make stuff. And I think that's the vision we have is how do we continue to make that easier for people to kind of explore the bounds of what they have in mind and actually build it? And so with that in mind, we're kind of saying, all right, well, what might it look like to be able to use the things you make more easily? And so a big part of where we want to go is saying, all right, right now we know that we're desktop primarily, right, especially with things like a visual editor.

1:54:07Jaclyn Konzelman:But we started by looking at kind of what we built in Gemini and saying this feels a little bit more friendly, a little bit lighter. How do we bring kind of some of that ease of use that our kind of Opal Labs Gems introduces and merge it with our Opal standalone editor and kind of continue to build this product in a way where it's easier to create and use your Opals from mobile, maybe from different parts of your life and kind of make it easy to create custom software for any specific problem or kind of paper cut that you have?

1:54:37Thomas Iljic:I didn't even, I can't believe I didn't even mention this, but I noticed that there's a microphone there. So you could, in theory, do this all with voice prompts as well, right?

1:54:44Jaclyn Konzelman:Yeah, exactly. And what you see here is actually a microphone just for the edit step, but you can also do that with creation. So in this zero state, you could enter in what you're trying to build. It transcribes it. But you could also build, and this is an interesting use case is, I don't know if I have it here, but you can create kind of like podcasts as well. Oh, awesome. generating podcasts or uploading an audio note, just like a rambling audio note, and then creating an app that takes that audio note and then transcribes it or summarizes or transforms it in some way into some other type of asset.

1:55:17Jaclyn Konzelman:So it could be really good for on the go. But that's the kind of thinking is how do we make it easier and easier for people to not just create, but access the things that they're making?

1:55:26Thomas Iljic:Very cool. Last question for you. If I'm using Opal for the first time, what's the first thing I should create?

1:55:32Jaclyn Konzelman:Oh, such a good question. I would say think of the most annoying thing that you experience in your daily life and then try to build something that makes that like slightly less annoying. I mean, we started with that example of that kind of like recipe leftover chef app, but that's a good example because then I think you can actually test it. I think what's really cool about AI is you can build things that are just useful, not just things that are fun. And sometimes they're both like that recipe leftover chef app. Yeah, definitely both. Yeah, exactly. Think about something that's really annoying for you and then see what it might look like to build an AI solution to make it slightly better for you.

1:56:07Thomas Iljic:Love it. Where can people go to try Opal?

1:56:11Jaclyn Konzelman:Go to opal.google. Yes, that's the full URL. You can also add.com. We'll reroute you. But opal.google takes you to this visual editor. And then you can also go to the Gemini app and check out the Gems tab. So Gems Manager, you'll see Opal right at the top there as well.

1:56:26Thomas Iljic:Awesome. Thank you so much for joining, Megan, and showing us Opal. This is really cool. I think people are going to love it.

1:56:32Jaclyn Konzelman:Awesome. Thanks so much for having me.

1:56:35Thomas Iljic:To anyone watching, please take a moment to like and subscribe. If you haven't yet, also make sure you check out the Neuron.ai. That's where you can sign up for our newsletter and join 600 and some odd thousand other people who read it every morning. We'd love for you to be one of them. And on that note, farewell for now, humans. We'll see you next time. Thank you.