Google's Nano Banana 2 vs. Nano Banana Pro: Which One Wins?

Google's Nano Banana 2 delivers Pro-level features at half the cost, with web search capabilities that can generate current event imagery in real-time. Addy and Joey test the model across multiple scenarios and discuss Perplexity and Claude's rapid feature additions in response to OpenClaw.

--

The views and opinions expressed in this podcast are the personal views of the hosts and do not necessarily reflect the views or positions of their respective employers or organizations. This show is independently produced by VP Land without the use of any outside company resources, confidential information, or affiliations.

Transcript

Nano Banana 2. 

Oh wow. I mean, that's literally happening now, I think. Or just happened. 

Yeah. It is Bill Clinton in a congressional hearing with texts in the background saying, house Committee on Oversight and Accountability hearing into the Jeffrey Epstein files, 

which, okay, you've answered my question on how well connected the Nano Banana model is to the internet.

Very well connected. 

Yeah, it was impressive use case to blows everything else outta the water.

Alright, welcome back to Denoised. Addy, how you doing? 

Hey Joey. Saw you at the VES awards. 

I was doing some red carpet interviews, which we'll have a video coming out soon. 'cause I started asking everyone what they think about ai. A lot of expected answers, but a lot of, uh, couple surprising and interesting ones.

Looking forward to that. 

Yeah. All right. Speaking of ai, this week we got another big product line, Nano Banana 2 that is also holding up to two fingers. So this came out. My main question has been, obviously, you know, Nano Banana 2 is gonna be better than Nano Banana. The original, I've been trying to figure out where does this fit in Nano Banana 2 and Nano Banana Pro.

Mm-hmm. 

Because it's called two, not pro. 

It's funny, uh, when I sent you the first text, your first question back is like, yeah, is it better? I was like, I know. Let's, let's figure this out. 

I've been living in a Nano Banana Pro world. Yeah. So like, I'm like, is it better than what I'm currently using? I You've done some tests.

I did a bunch of tests. I think the, the, the quick answer is, it is. Nano Banana Pro Light. So it is, it has a lot of features that Pro had, including it can source the web for current events, which Pro can do, it can, has much better real world knowledge. So if you reference like an actual place, it knows it and I have some demos of that.

Uh, but also PRO has that as well. So I think it's just kind of pro light. It's literally half the price of Pro. So Pro is about 15 cents an image, and now the banana of two is about 8 cents an image. So I think for the most part, like I would run this, I'm gonna test some other workflows on this. Now and see.

'cause it's like if we get the performance, if we get the same output, but Right. For half the cost and 

faster. 

Faster performance and faster inference 

times. Yeah. 

Yeah. My inference test times were actually kind of slow, but it's probably because everyone's using this right now and yes. The servers are backed up, so, yeah.

Uh, I wouldn't put much weight on the speed right now, but it, it should be faster. 

So to recap, you think that this is as good as Nano Banana Pro, but half the cost and perhaps even faster with maybe some additional improvements? 

Yeah, we'll go through some tests. I did. My biggest test is like, 'cause we use Nano Banana Pro mostly for like, give it a bunch of reference images, like location and people, and.

Composite them together into a shot. That's like our main use case. I didn't do a bunch of testing with that with Nano Banana too, so that'll be kinda my next thing. But in the initial test, kind of quality feature wise, really like obviously way better than Nano Banana and on par in a lot of cases with.

Pro, 

let's get into the test and then figure out what it, what some of the things are that it's better ATTN. Then I ha I have a theory on the Google product strategy that I'll share with you. Yeah. So the good old 1970s, busy New York Street. So this was inspired by Joker one, and I've been using this actually since Joker one, like the very early days of ai, uh, non Joker two, not a big fan of that movie.

It's, it's really complicated to do. Right? Uh, so my prompt is very open-ended and I let sort of the, the classifi free guidance fill in the blanks on the model. So I just say busy 1970 Street in New York in daylight, lots of people in lots of cars. That's it. 

Mm-hmm. 

And that's my whole prompt. You can put it through yourself and maybe get this exact same thing.

All right. So the interesting. Thing. I think in general, this is better than Nana Banana Pro in that it is referencing more things from the 1970s than Nana Banana Pro did. 

Yeah. Everything that standing out to me is all of these movie titles that it came, it's like these movies from the 1970s. Yeah. 

It even got the Lugo of Shaft, like the movie.

Right. And then, um, on top of that, the outfits are correct on the people. The ni the New York cab in that era is accurate. I think the, yeah, it reminds me of that same car from Ghostbusters, whatever model that is. 

Oh, the, the, yeah, whatever their, um, that the, yeah. 

Station 

wagon. 

Yeah. It's, so, it's like a car version of that as a cab.

And I think that's accurate. Even that, uh, that um, Greyhound bus like that is from that era. Right. So check, check, check. Mm-hmm. 

It looks like New York. Feels like New York 

feel it's dirty. Like you have graffiti on the subway sign you have a 

newspaper stall. Yeah, we have a subway entrance. And the subway signage is, looks accurate.

The graffiti, everything's graffitied up. Kind of like New York was in the seventies. 

Yeah. 

And then on the other side we have a bunch of like adult theaters and peep show 

stuff. And that's what New York was in the 1970s, right? Yeah. Yeah. Most, most of Manhattan was like this. Well, I will say this. This image still, I don't like it.

Can I be honest? Here's why I don't like it, because it's not the details. It's getting everything right, but the perspective is too damn clean and you don't see an intersection and like it still doesn't look authentic even though it's correct. If I'm making sense. 

I kind of get what you're saying. I mean, also you did give it a generic prompt.

Yeah. 

So. Like if you wanted an intersection, then you could prompt an intersection. I'm looking back in the distance here. So I mean the, the text is falling apart in the signs, in the, in the way back. But yeah, I mean, look, compared to, um, what did we run this on last week? Red. Red, 

uh, 

red, red. Fire, fire, fire.

Red. We ran it on. Yeah. 

Fire red. 

So much better than fire red. 

Yeah. I mean, it's night 

day. 

Yeah. Blows it outta the water. 

I actually ran it on Seedance five light, which also came out and uh, gosh. Oh, right. It was such a joke. Like, I don't even know what that model's for. 

Oh, you did not like Seedance five. 

I was trying to get the same thing out of it, and I couldn't.

Okay. 

Yeah. 

Not for New York. All right. 

Maybe Beijing or something. I don't know. It'll nail it. 

I think, you know, I, I think to what you're saying about it, no. Having better real world knowledge. And possibly surpassing not have been at a pro. I think that, I think that's on point. Well, what's your 

theory on why that is?

I'm guessing it's a newer model that's trained on more current data. While we're, while we're out talking about current events stuff, lemme just jumped to my. The tests that ran that. So basically the model can, has two things. It has better real world knowledge and then it can also, it has a parameter where you could turn on web search and it could access the web and kind of research current events or past events or just like access web knowledge.

How do you, how do you do that through prompting? 

I, good question. 'cause this is through the API, um, I'm guessing if you're in Gemini and you're. And it's making a decision in your prompt based on if it should source the web. It's a, in the API, it's like an actual, like parameter you can say, yeah, true, false.

Like turn it on. Good question about if you're using this in Gemini itself, but it does, it is possible. I'm assuming Gemini will figure out if it needs to access the web and, and utilize that. 

I'm guessing it's ever connected to the web. Like it, it's not disconnected at any point. Right. 

Well, you did though.

Web to make it. I can't run, 

its local. I'm talking about like a, uh, like a, almost like a rag, like a retrieval augmentation system where it's constantly pulling from the web for information. I'm, 

uh, just to guess, I mean, I was guessing it was just querying like some Google results or something and then kind of.

Looking at that for whatever, based on the, uh, inquiry, 

it would be interesting to then generate images based on the current news cycle for the day, every day and see if that changes. 

Well, 

okay, 

hold on. So first up, let's talk, we'll come back to news. Real world knowledge. I had a test prompt that was LA centric, aerial photograph of Randy's donuts, which even if you haven't been to LA you probably are familiar with what that looks like.

This was Nano Banana. One. Mm-hmm. Which did not have real world knowledge. And so this is a very generic looking building that kind of looks like Laish, but it looks nothing like actual Randy Stones. Mm-hmm. It's just a building with a giant donut on top. This is Nano Banana. Two. 

Oh, I was gonna say, is that an actual photo?

I was like, oh, you should, this looks like the actual real Randy. Yeah. Has big Randy's doda. It's on a corner of a street, which really is, I think it's probably cleaner looking version of the street than it really is. But yeah, this looks like it. And this is, uh, not been out a pro, which I, I actually, I don't have, I should probably look up what the real image looks like to compare it.

But these two, I mean, as far as this donut sign and the text. Get it pretty right. 

I thought the donut itself was a dark brown color, but I, I haven't been there in a while. This one? Yeah. Yeah. I think that's more accurate. 

Yeah. Also, didn't it actually change? 

Yeah. Right. I mean, this isn't your part of town, not mine.

So you tell me. Um, I think, 

I wanna say I drove by like a month or two ago and it wasn't Randy's anymore. Some other donut shop, regardless 

got acquired by private equity. It's what I was gonna say was you think it references the Google Maps. Uh, imagery. 

I don't get, I don't, I think this part, I think it's in the model.

I think the model itself has just more world training, uh, up to some certain point. I think that's in the model itself. 'cause also this isn't, this isn't a parameter that you can turn on or off. This is you give the model a prompt. It has, uh, real world under understanding. Mm-hmm. So I think it's just what it's, it's just all the data that it's trained on.

It just knows 

it. I mean, this is so solid for like, um, if you're writing a news article about Randy's Donuts and you forgot to take a photo of the place you just generated and add to your news article, 

score it on. I mean, that doesn't feel like that's open. That's not journalism at that point. Yeah.

That's a slippery slope. 

People are doing it. I'm just saying, Hey, I'm not a 

journalism major. I do not doubt that. Other one was, uh, the fourth Street bridge. This was Nano Banana, which looks like la It's too clean a bridge. 

It's too clean. Yeah. 

But the fourth Street Bridge has the new. Ridges the arches. Oh 

no, my friend.

This is, 

uh, let me school you on la That la stuff, 

is that incorrect? Is 

that incorrect? You were so incorrect. So the, the last one could have been either fourth or third street. Yes. They both look like that. But this one is sixth Street. What's this? And it is completely incorrect. Those arches are actually slightly, uh, tapered and angled like this.

If you're looking at my arms, not straight up perpendicular. Yeah. So it got it. Not this, not 

either. 

Oh, I schooled Joey on la. Yes. 

Yeah, right. I thought, I thought I just took its word for it. 

No, this is not right. 

The prompt was the fourth Street bridge in downtown Los Angeles at Golden Hour show, the specific art deco concrete arches of the bridge spanning the LA River.

Uh, you confused the model. If you had said Sixth Street Bridge with contemporary arches, it probably would've nailed because the fourth Street Bridge is, uh, yeah, it's, it's like a hundred years old. It's old, but this one is brand new. They just finished it like two years ago. 

Right. I thought that was this one.

Okay. 

Yeah, it doesn't look like that. 

All right. Okay. So it all got 

it's user or my friend? Don't blame the model, 

but now lemme go to, um, I'll talk about the current, the, the web search stuff. So prompt was create a photorealistic editorial news photograph to depicting the biggest news story in the United States today.

Nice. February 27th, 2026. Okay. N have been at a pro but had this, which is also a bit weird, it's just kind of generic, like protesting in front of the White House for Vote and Democracy, which is topics that have been relevant in the last year or so. But I wouldn't say like 

this could be used for any day of the week.

Since year? 

Yeah. I would say specific to 

right 

Today. Yeah. Nano Bonta doesn't have that option, and so Nano Banana just made up something else. Also kind of similar topical stuff over the past year or two. Mm-hmm. Nano been added two. 

Oh wow. I mean, that's literally happening now, I think. Or just happened.

Yeah. It is Bill. Bill Clinton. In a, uh, congressional hearing with texts in the background saying House Committee on oversight and accountability hearing into the Jeffrey Epstein files, which, 

okay, you've answered my question on how well connected the Nano Banana model is to the internet. Very well connected.

Yeah. 

Yes. So this has pretty wild, this is what, this has probably been the best, most, most impressive use case to blow everything else outta the water that I've seen. 

Dude, there are so many. Interesting things that can come out of this. If you have an image model that is fully aware of the latest and la like news cycle, you can automate YouTube content for one.

Uh, yeah. I mean, you could do outta the image. That's cool too. I mean, yeah, you could still do that now. I mean, getting to have an image is, is, is pretty cool because you could done something where you have like a, have the an, an LLM source the web and then create a text prompt. 

Yeah, you can do it manually.

But 

feedback, I, 

I'm not creative enough to like think of. Novel use cases, but this is really powerful. I appreciate that. 

Yeah. One shot outta this from one text prompt with the web search turned on, uh, is pretty crazy. 

Yeah. And there's no Bill Clinton in your prompt. It just kind of figured it out. 

So that was, yeah, that was the best, the best use case.

Yeah. Overall, just like these were just kind of like detail style outputs, more of just a comparison. Like this was a macro shot from Nano Banana, the original one. Here's not have been added two. So a lot more detailed on par. This was not have been Banana Pro. 

Oh, Patek. Wow. So, um, I don't think Pateks have a circular shape.

Yeah. I'm, I'm kind of getting into watches I can't afford to protect. Those are like $50,000. But, um, yeah, the, this is pretty impressive that. It got the font right for the most part. And I think the watch dial itself. But um, I don't think it has exposed gears like that least. 

Yeah, I don't know. I don't know if that center thing is real.

Mm-hmm. 

Or like accurate. I mean, it says it here too. Uh, let's see what I got. Fisherman portrait. 

Not this guy again. 

This was not a minute one. This is. Pro this is two 

I prefer pro you. 

Yeah. Pro looks a little bit more realistic. 

Yeah. Realistic. Yeah. This has that tone map look. 

Yeah. It looks like that kind of, yeah.

That expanded HDR look. Mm-hmm. But still way better than Nano Banana one. 

Yes, I agree. So Nano Banana Pro is I, I think still by far quality wise, the one to beat, but Nano Banana 2. Is more agile, connected to the internet and is able to do mm-hmm Like a few has its few tricks up the sleeve that pro does not have.

Yeah. And I think for a lot of stuff, I would probably now try meta a two. 'cause it's like if you could do it. And it's, uh, faster and cheaper than, what do I need to spin up pro for? If two can already handle it? 

I think that's what Google wants you to do too, uh, is because they probably uses less GPU so it's cheaper for them on the inference side.

That's my guess. 

Yeah. Yeah, yeah. Yeah. I mean, I think the official product name is Gemini Image 

Flash 

three flash. Yeah. Light or flash. Yeah. Lemme see. Text rendering. This was original, not a banana. This is to And this is pro. 

Yeah, pro is. Closer, much closer. It has the, the, the chalk. Yeah. Like hand prints from the chalk.

But the, the two light bulbs are totally ruining it. I'm sure you can remove that. Um, but yeah. Okay. 

And then what was this, uh, you did another test where you took an actual image. Okay. To 

do a project. I have a whole backstory and I'm gonna tie it to my theory. All right. So, um, turns out I have a rat problem in my backyard.

Like my, my neighbor's behind. My house. They, um, it's a dirty backyard, let's say. So they're probably coming over to my backyard. So I go to Home Depot, I pick up this rat poison and I was 

like, you didn't watch like oui and have like an, an appreciation for the rats. 

I love the, uh, I, well, okay. I love oui.

Don't like the rats. I love Remy. I saw a lot of examples of Google's latest product, which is called Pom or photo shoot. That is really dialed in for user generated content for marketing and commercials. So instead of, you know, feeding it a really polished image of a product, you just take a iPhone photo of your product and then you could generate campaigns with it, a bunch of things with it.

So this is a J iPhone generated photo that I had in my phone. I ran it through Google. Uh, Nano Banana 2, the new one. And if you go to the next image, uh, my prompt was make a commercial for decon rat poison, where two rats are vacationing and they're drinking margaritas and they have this as a snack.

And then the bottom it says rats like the taste. So this is a fun, cute way look kind of rat esque, like, make the rats look cute. And it did a decent job just based on, yeah, sort of from my initial photo. If you zoom in the, the. Text is all correct. It even put those little bricks onto a bowl, like a snack bowl.

Oh, I guess the, some of my other generations did, but yeah, it's, it's right there. 

But I mean, yeah, the, the bricks are displayed. It's got the little, the, the leg, the lime that goes there. This is off center. But that's probably 

right. Yeah. And it has like a tile top instead of a wood top. And then I, I prompt more of my prompt went into that bottom text if you go, yeah, rats are spot on, dude.

Okay. No, I was looking at the rats because the rats look good and even like their little rat fingers or grab gripping the martini. 

Yeah, 

the, uh, martini, whatever that is. Uh, 

margarita, 

margarita glasses. That stuff looks great. 

Uh, and then at the bottom, the, the rats like the taste I want. I said my prompt was, uh, bold font with a gradient.

And slight transparency, so I kind of like 90% nailed it. But the fact that I don't have to, to do a second pass in Photoshop to add that text, that's a huge plus. So my theory is that the Google Pomp pomp photo shoot product is using Nano Banana 2 under the hood, which is really meant for user generated content, specifically iPhone or phone generated imagery as product reference.

That's my theory. 

That sounds like a good theory. I mean. Makes sense that yeah, maybe they had not have been added two under the hood when they rolled out the Elli product. I'll go to the product page in a second. Mm-hmm. And yeah, I mean, maybe they rolled that out. They rolled that out last week. So maybe it was like a stress test and refined some stuff and then roll out, not have been added two as a the full product.

Yeah. So if I were to kind of segment markets, if I was in Google and I was looking at, you know, go to market, then I would say the non Banana Pro product is still geared for our world for film, television, and high-end usage with high pixel count. Mm-hmm. Versus Nano Banana 2, which is better for, um, high volume of content with, uh, imagery that may not come in Perfect.

Yeah. And at the cost that it's priced at per image, that also makes more sense for e-comm, where you're like, I need to do a bunch of product shots or variety or different markets and stuff. And so I need to like, I just wanna make a lot of images and so if it's more cost aware and cheaper than. That is a more appealing model.

Yeah. 

That can handle this kind of 

stuff. And Nano Banana one, I think is just so far behind now compared to these two models. I think they're eventually gonna sunset that, uh, very soon to free up GPUs. Yeah. 

Yeah. There's really no sense to, I mean, I, it's sure it's slightly cheaper now I, or like half the price of Nano Banana 2, but for the quality upgrade you're getting from Nano Banana 2, it's, I I think what worth it, right.

I would say. But yeah, I mean, also the only issue here too is just like the. As we zoom into the text, it starts falling apart. Yeah. But you're doing this like in the Gemini app? Correct. So you got the Gemini output, or did you use this somewhere else? I think I 

used 

Freepik. Freepik. 

But did you? But yeah, I mean, I wouldn't 

expect it.

The result. Were you, were you doing one K or were you doing 4K? 2K because 2K. Okay. So I could also, two can go up to 4K. So maybe you, maybe if you did 4K, it would, um, I'm curious how it would do like the refillable word it looks like. Yeah, 

yeah, yeah. Garbled it a little bit, but like the shape of the package is spot on, man.

Like the Oh, 

I know. In this top part it got it nailed and it, the, the dead mouse 

nailed. Yeah. I mean, my photo's not the best, but it kind of recovered detail 

from it. Yeah. Your photo is like, it's blown up a weird angle with like glare. Yeah. Your, you tore the pa the top off. 

I was, I was ready. Kill those rats, you know, like crap.

Yeah. Over here. Yeah. This is a crazy demo. I mean, look, we're to get this image from you like a year ago. You're like, here's a crappy iPhone photo of my product. Make me a, uh, you know, a, a good ad and you got this. You'd be like, your head would explode. 

Okay. Exactly. And two years ago from today or three years ago, the only way to make this would've been either in cg, you build the pool and put the rats on a patio chair and da da da da.

Mm-hmm. Or you have a a 2D artist come in and composite the entire thing. 

Mm-hmm. 

And now we can do it in seconds. 

Yeah. It's crazy. Let me switch back to share. I'll share that. Um. Uh, so yeah, here's the Elli Pommel 

page. Yeah, I think it's just po called photo shoots Now. Um, I'm not sure Google's naming convention, man, you guys gotta figure this out.

But, um, yeah, this is, this is the product I'm talking about. It's basically like single image to full campaign automation where it'll generate PDP product detail photography, as well as, um, product in situ in situations. And, um, I think is able to then generate it in multiple formats. So 9, 16, 16, 9 aspect, whatever ratio.

So it, I think it's going up right against Adobe's product called Gen Studio. So it's like two giant tech companies trying to figure out marketing automation. 

That, plus there were like a million startups that were also trying to do this. And I remember seeing all the posts like Google just. Destroyed a bunch of startups.

Yeah. And this, this is really like, now that I'm, I'm in the marketing world now and I, I'm trying to understand how they are operating like this, this is putting a lot of fear into a lot of people's hearts. It's like, oh, do we, do we need to exist as an agency anymore? If somebody can just do this on their own?

If the brand themselves can like, make thousands of assets with every variation of market and stuff that they would need. Yes. And one go. Yeah. 

Yeah. I, I think to counter that, my response would be, of course you still need an agency because it's not just about generating in quantity, right, or generating multiple variations.

It's about, uh, curation of assets. It's about a overall creative arch and having targeted creative per region. And a lot of that expertise only exists in an agency. 

Yeah, I mean, it's just maybe, I think you still need an agency. It's just the agency will probably be a lot smaller. Than it was in the past.

Yeah. And more specialized. 

Alright. Anything else? Not a two or 

That's it for now. 

I'm waiting now. Uh, I mean, if there's a Nano Banana, not a two. There's gotta be a Pro, pro two coming out at some point. So 

you seem like a pro guy to me. You have a MacBook Pro? 

Yes. Nano Banana Pro. 

Banana Pro. You're a pro editor.

Cling Pro. Final Cut Pro. 

Seem like the pro guy. 

Resolve. 

Yeah. 

Do To resolve this studio. Resolve. Resolve. It doesn't do pro. All right, what we got next? Okay. Let's talk about AI agents. My favorite thing. Okay, so yeah, we talked about OpenClaw. Uh, a few weeks ago, and, uh, I'm still in my OpenClaw experiments, but we talked, we talked about how, like, you know, uh, and it was called Quad Bot Philanthropics, like please don't call it that.

And then he calls it OpenClaw and then he gets hired by, uh, open ai. Meanwhile, uh, anthropic has, uh, not set idly by, so they launched three, they've been launching updates like every day this week, but they launched three updates specific that kind of tie into what Open Cloud's doing. One was remote control.

So basically if you spin up cloud code, you know, it runs in terminal and. It's very powerful in what it could do, but it was sort of like you're kind of sandboxed into what you were doing on your computer. Now they have a remote control feature where you can run cloud code on your computer and then log into that session from any other web browser or your phone and continue, or what you were doing on the computer.

So if you're using cloud code as, not for actual coding, but like as an assistant or having giving it access to stuff on your computer. Now that. Leap that, uh, a lot of people liked with OpenClaw, where you could kind of run it on your phone. Now you can do that with cloud code. So that was like one feature that a, they added that was uh, 

so is it like an iOS app?

It's not an app. It's either, I mean, it basically, they already have the cloud iOS app. So you now you could run your cloud code session on your iOS app, but you're basically still controlling the actual session on your computer. Yeah. You're just remotely connecting into it. 

It's just much better than like WhatsApp.

It's uh, basically what the way OpenClaw works is it's running this thing on terminal, on your computer that you're leaving it run all the time and then you are porting in to, could talk to it either via, via terminal or WhatsApp or Discord. However, there's like a million options of how you can set up to talk to it.

And then a lot of people liked that feature of it where you could talk to your ai, your personal AI bot that's running on your computer via phone. So now with cloud code, you could do the same exact thing without. All of the extra OpenClaw setups if like running an AI agent on your phone was like very appealing to you.

Now, uh, Claude added a feature where it's like super easy to do and you don't have to do OpenClaw. If that was one of the main reasons that attracted you to. Open clock was the phone thing. 

So it's your very own agent that you're running at home available to you everywhere you go. 

Yes. 

You are using in a much more elegant package than like messaging.

It 

kinda, yeah, it's, it's still a very like specific use case. If you were using cloud code as an AI agent on your computer, now you could talk to it. You don't have to be on your computer to talk to it. You could talk to it via your phone or another web browser. 

Right, 

that that's one feature. The other features were that they launched were memory inside Cloud Code.

So now there's better memory features, which was another thing with open cloud, whereas you keep using it, it kind of remembers patterns and things that you like. So now that's also built into cloud code. So sort of another feature. Overlap from OpenClaw. And then the last one was, um, schedule Tasks. This one's in Cowork, which is separate product from Claude Code.

It runs in the Claude Desktop app, but it's sort of like Claude Code Light, where it can still do things on your computer, but now you could schedule things to happen. So this is also the main thing that OpenClaw does under the hood that is like very simple, but very powerful. Is it, uh, you could talk to it via sort of any other device, eh, it's just a bunch of like m markdown documents that set up.

It's memory and instructions, but then it could also run crone tasks, which is just basically like a piece of code that runs at a specific time. And the crone thing is like old school and simple, but like actually really useful for AI agents where you could just, every morning have it be like, you know, check my inbox or like, run an analysis on this.

Or just have these recurring tasks, stock 

market 

stuff that, yeah, you can't do in cloud code. You can't set up Crohn's to like run in cloud code. But now Claude. Added this feature in their cowork product. And another, it is just like another feature grab, like all these little things that were attracting people to OpenClaw the trade off being, you're using this highly experimental product that can like go off the rails, is like one by one being added in claw, added into claw added in clawed.

So slowly, but, well, not slowly. This is over a week quickly, but surely they're adding more and more features that kind of put it on par with some of the benefits people were getting out of OpenClaw. And Claude is, you know, already like a much more refined and developed product. 

Let's, let's take a step back.

Correct me if I'm wrong, but from what I'm understanding is now that open AI. Uh, Aqua hired the open cloud guy. Yeah. And most likely, open AI is gonna re-release OpenClaw as a paid product, but way better and da da da da Cloud is anticipating that competition and now they're retrofitting cloud code to compete with this upcoming thing in the future.

Is that right? 

Okay. So yeah, my understanding with them hiring Peter Steinberger, it, it, they didn't acquire Open Cloud. OpenClaw is still an open source product, right? He's still actively developing it. There is like a new update every two days for open Cloud. So it's still actively developed and pushed.

My vibe of that is just like open ai. I wanted to support it, hire him to keep developing it. And it's like if more people are using AI agents and are using open AI models, that's good for open ai. I don't think they acquired that. Or are gonna roll it into like being an open AI only product. I'm sure they're building something separate that has similar features that, you know, would eventually roll out and be more, you know, user friendly for a, for a wider market.

Yeah. So you think just, and it's not like a wall gar garden per se, it's just like a really compelling integration. 

I think the OpenClaw thing that sort of un or just awoken people to that, like people want some sort of like AI agents that are. More integrated in their computer, smarter, kind of learn from them.

Powerful, uh, and not piecemealing, different little chat bot steps along the way, but something that's more holistic and integrated and that the tech is getting there. 

Microsoft missed a boat on this big time, right? Because I think was like pilot stuff. Yeah. Copilot was gonna roll into Windows and it was gonna essentially.

Take over your computer for whatever that means. And OpenClaw, is that right? It, I think it's Linux space. It's it's open source and all that good stuff. 

Yeah. You can kind of run it and integrate it into, into anything, you know, at your own risk right now. 

Yeah, it's interesting. 

Yeah. Back to the other, uh, the initial OpenClaw agent thing, so the other.

Product launch that ties into that space is, uh, from Perplexity, Perplexity Computer. And so this is basically a virtual machine computer that can kind of connect to all like your email, all of your like web platforms and is you can spin up subagent, you know, you can give it a task recurring thing. So sort of as like central virtual AI computer that can automate and connect to a lot of stuff.

Yeah. 

So it's only for the max plan. I haven't really dabbled with it, but um, they're 

all competing for that same. Same market, right? It's the, the local kind of local thing that lives on your machine. 

Well, this is web based, so this like, feels like it kind of lacks the appeal of OpenClaw because it's not on your system.

I don't know if it has a local, it, it did, my impression was it was like a, a virtual machine, so I don't think. Can connect to your local system, it can just connect to all of your cloud services. Mm. So it seemed, it seemed, it felt a bit like, uh, Manus, maybe like a little bit more powerful manice Yeah. That could kind of spin up a virtual machine and do automated tasks, um, that, uh, the mebot.

So yeah, it's still in that same agent space and more powerful space. But, um, I don't think it could do local computer stuff. But yeah, just another product launch into the, into the world of, uh. You know, I think AI agents unlocked a lot of things. 

Yeah. I mean, look, you're, you're way ahead of this stuff than I am.

Like you, you're experimenting with OpenClaw now, obviously. So Yeah. You, you tell me and the viewers where you're ATTN how this world is evolving. Very curious to hear it. 

So the, one of the bigger questions with OpenClaw has been terms of service and if you're allowed to connect OpenClaw. To any of your subscription plans if you have like open AI codex or Claude Max or Google's version, which I forgot what it's called.

So once I connected it to, uh, I switched from paying for API, which is really expensive, from Claude to, um, I tied it into my Cloud max plan. So it's using. My same monthly subscription. So that's made it much more appealing. It has been debated if there have been different stories of some people saying they got like booted off of Claude Max from using it, that it violated terms of service.

But then some other people at Claude have posted on X saying that they just don't want people to use Claude Max for like business purposes. Where like you would be using the API, but if you're using it for personal development or just kind of like app development, uh, or OpenClaw that like, that fits the use case.

So, so far our fingers crossed I have not gotten booted using it that way. And not having to worry about like the daily cost of tokens has made it much more appealing and it's been good for, um, moving and consolidating some stuff that was already set up with, uh, like Zapier automations. Um, it's streamlined that a lot.

More as far as like writing social media posts, a lot of the web articles and newsletter stuff. So it's been handling that and I'm still, the next kind of phase is like something that could, uh, chop up these videos and turn 'em into shorts. That's been my, that's been my next quest. 

Yeah. Speaking of which, and you're no stranger to Eddie ai Premier gal.

Yeah. Who I follow on YouTube. She's probably the biggest premier influencer, if that's the thing. She had a. Big endorsement for Eddie ai. She liked it a lot, so just thought, I'll let you 

know that. Oh yeah. That's a good, oh yeah, that's good. Yeah, I was literally using it yesterday. Uh, they do a lot of updates and it's gotten better.

Mm-hmm. Even it gets better every week. I did ask him, I'm like, oh, do you have API Connections? He's like, it wasn't really on our map right now, because my thought was like, if I could. Plug this into Eddie AI could maybe help cut up some of the shorts. Right. But I, I don't think that tool would work. I'm looking at some other tools that, that might be able to, uh, help with that.

But yeah, that's my, my goal with this is like if there's a AI agent that could kind of monitor social media and knows all of our videos and stuff and can like pull clips and post stuff. And kind of learn from how they're received and then like, like learn the editing style and like improve on what clips it pulls.

That's like my kind of goal project for this of like a AI social media agent that could pull from our library of. Content and stuff. 

That's an interesting goal. I like that. 

Still in progress? Yeah, so far. Yeah. So far. Uh, some stuff's working out Well. Have not had all my emails deleted, so that's good too.

Run, unplug it. 

So he did, in one of the product updates, I think he put in a, uh, an update that says, if you say Stop OpenClaw, then it will actually stop like, 

oh, like a, like a hard reset button or a. Stop button. 

Like a hard, like a hard, uh, like kill switch. Kill switch. Like 

kill switch is still software.

You need a hard work kill switch. 

To explain this, there was someone who posted on X, who was also the meta, one of the 

chief 

AI safety, chief 

safety researcher, said officer, or one of the top dogs at meta. Yeah. 

Gave open a call, access to her emails and was having it delete emails. Which I also think is kind of crazy 'cause I just would have it archive, but then it like forgot the context and just started deleting all of her emails.

Entire inbox. 

Yeah. 

And the only way she could get it to stop was to like unplug her computer. 

She was running to it. Literally. Yeah. 

In the threaded replies it says that like, open clock compacted the conversation and so forgot what it was doing and that was why it went, um, off the wall. That's why this stuff's experimental.

So, uh, you know, be careful. Yeah. All right. Real quick, two new model launches that I saw that were kinda interesting. Uh, one is quiver ai. So, uh, a 16 Z backed company that is focusing on, uh, generative 

vectors, 

sgs, vector graphics, generative vectors, their model called Arrow 1.0. 

Okay. That, that's great. 

I think that's cool.

Yeah, that's great 

because I mean, last week you turned on to reframe and I was messing around with that and, uh, yeah, having. Uh, AI vectors is like nice. 'cause you just don't have to worry about resolution. 

Exactly. You can, uh, and then you solve the print issue. So someone things go into print. Um, there is no, yeah, yeah.

There is no. It's all splines and lines right at the end of the day. Yeah. I think the technical difficulty here was that, um, image models work in the latent space and come out into pixels in the variation of auto encoder, and then it goes to another step where the pixel space stuff is converted into vectors.

Like that's how it's generally done. I think they may have solved that problem by going natively out of late space into vector space. Which is very novel and innovative. 

Yeah. This stuff looks 

great. Yeah, it looks great. Yeah. Yeah. It's good enough for logos and things like that. You don't need like a bonafide Nano Banana pro level image model.

Like you're generating stuff that's abstract and you know, clip already anyway. 

Yeah, yeah. And a lot of times those logos that you spit out of like Nano Banana is so, it's too complex coordinate and crazy complex. Yeah. It's like, doesn't 

really work anyways, and that's, that's why Igram is such a good model.

Dialed in for this kinda stuff. Yeah. Yeah. 

And then other model, there's a new video model called P Video from Prune ai. And, uh, they're, I believe one of the only one of the main European labs to ship a, uh, decent video model. And I mean. I've looked at some of the samples. You know, I'd say it's on par with, I don't even know, like A LTX mm-hmm.

Or, um, uh, one. But yeah, I mean the main kind of things from this, the quality's pretty good, but what you get, it's like pretty fast generation at pretty low cost. So, uh, ten second generation speeds for five seconds of seven 20 p video. 2 cents per second. Hold on. Did you say 10 

seconds for a five second video?

Yeah. That's really impressive. 

That's their post rate. Yeah. Yeah. So it's fast. It's fast and cheap. Yeah. And it has a draft mode for video multimodal, it can do audio outputs and stuff. So overall, you know, it's pretty good. Seems like a pretty solid model. I was trying to look at some of the test clips. Uh, they have a lot of animation stuff with speaking characters, so, yeah.

You know, I always like to throw this stuff on the radar 'cause. Depending on what project you're working on, some models might be great 

for it. Yeah. I, I call this like the preview window model, stuff like this. We absolutely need it. You know, when you're ideating and you wanna do it with videos and small, small animatics and things like that, you don't wanna wait five minutes.

Mm-hmm. 10 seconds is perfect and then just keep rolling the dice and it's cheap and it's gonna 

speed. I, you know, that's. The big trade off with automatic pro. And when I'm thinking back to our, uh, LORA reference showdown, like z image quality wasn't on par, but like seeing the speed of stuff coming out, I'm just like, ah, that's actually pretty nice.

Like how can I incorporate some of that into the workflow? Because just being able to like iterate and not have to wait around 30 seconds or a minute is um, it gets taxing. 

Yeah. And multiple sessions, multiple tabs doesn't really solve the problem. 'cause you're still waiting. Few minutes per. Session. 

Yeah, so that's a good point.

Speed. Speed. Speed is good. 

Speed kills, man. 

Speed. All right. Good place to wrap it up. Uh, links to everything talked about@denopodcast.com. 

Hey, you remember I talked about the YouTube high points? Well give us a couple of high points, actually, I think seven 50 is the minimum that you can give us. That seems to really help, uh, boost our videos on YouTube.

So yeah, it's on the app. If you go to the bottom and scroll to the right, you'll see a little, uh, party icon. I think that's the hat height icon. Give it a go. 

Thanks for explaining that to me. 'cause I don't know what seven 50 hype points are. 

It's a new thing YouTube rolled out. It's like you can like a video or you can hype a video.

Alright. Hype it. Hype it up. Thanks everyone. We catch you in the next episode.