The Next Wave - AI Tool Better Than OpenClaw? + NVIDIA’S $1T Prediction & AI Image Wars

Hey, welcome to the Next Wave Podcast. I'm Matt Wolfe, and I'm joined once again by Joe Fier. And we've got a bit of a special episode today. This is going to be our last episode for a little bit. We're going to be taking a little bit of a break on these episodes, but nonetheless, we're going to dive into all of the crazy AI stuff that's happened in the past week, share our thoughts, break it all down, tell you what it means, and really, really fun episode. A little bit of a bittersweet episode, but fun nonetheless. Let's go ahead and dive right in. Joe, where, where do we want to start today?

I think we start with where you spent most of your week at, right? It was up in San Jose at NVIDIA's GTC conference. Feels like everybody in AI was there for the most part. Wasn't a lot of Twitter action. It seemed like it seemed quieter than normal or X.

Yeah, because everybody that would normally be tweeting was sitting around inside of sessions and walking the expo floor and that sort of thing. But, um, Yeah. I mean, NVIDIA's GTC conference is, I've heard it described as the Super Bowl for AI. I've also heard it described as the Burning Man for AI.

Was it, man? Was it?

What, what you want to call it? I don't really care, but basically like everybody in the AI world is there because it revolves around the one company that AI revolves around, right? Like if, if we look at AI as a solar system, the sun of that solar system pretty much NVIDIA, right? Like every company sort of rotates around that one company these days.

Seems like it. And like even in the presentation, what Jensen was saying, which I know they're hours long. So, you know, that's probably why no one was tweeting either. It's like you're watching him for all day long. But it showed how implanted NVIDIA is in pretty much every part of AI. And I know that's something we'll also talk about, like how the future of NVIDIA doesn't seem to be slowing down at at all.

I don't think so. I, I, I know there's obviously bull cases, bear cases. We're not a financial show. We're not going to get into like, you know, what the stock price is going to do or anything like that. But I don't think we're as bubbly in the AI space as what a lot of people are sort of making it out to be. Right? Like I, I do feel like there's been this narrative for months. I mean, if not over a year now where people have said we're in a bubble, it's about to burst. I mean, I don't know about that. I still feel really, really good about where AI is and where it's heading from like that standpoint.

So I guess let's, let's, I'll ask you because I wasn't at GTC and you were in there as much as, I mean, I know there's a bunch of people and you probably couldn't be everywhere, but what's your feeling? Um, you know, walking away from there, is there a big takeaway or something that maybe changed your perspective?

I don't know if I had like any specific big takeaway from that event that changed my perspective. I definitely had a lot of conversations with people sort of behind the scenes and people that are deeper in the space that there's a lot of. I think during the keynote, Jensen actually mentioned that like the biggest majority of people at that event were in the finance space because, uh, they're, you know, there's a lot of financial analysis analysts there. talking about NVIDIA and trying to figure out like where NVIDIA is going from here. And so I got a chance to talk to a lot of these like people behind the scenes. And, um, you know, a lot of them made some really, really good points for why NVIDIA is still going to continue to grow. But there was also one point in the keynote where Jensen himself made a comment about how in the past year they've done a half a trillion dollars, $500 $1 billion in chip sales alone, right? But between now and, uh, the end of 2027, they expect that to double to $1 trillion. And he actually said that he thinks he's being conservative on that number. He thinks it's actually going to be bigger than that. Hey, if you take a look at my web presence online, it's safe to say that I'm a bit AI obsessed. I even have a podcast all about AI that you're watching right now. I've gone down multiple rabbit holes with AI and done countless hours of research on the newest AI tools every single week. Well, I've done it again, and I just dropped my list of my favorite AI tools. I've done all the research on what's been working for me, my favorite use cases, and more. So if you want to steal my favorite tools and use them for yourself, now you can. You can get it at the link in the description below. Now back to the show. I sat in this like, uh, press Q&A session as well. It was like, there was like, I don't know, maybe 100, 150 people in the room, but it was like a press-only thing where we were able to ask Jensen questions. And somebody asked him about that trillion-dollar number and where it came from. And he said that number was purchase orders. That number is companies actually saying, we want these chips. Like, here is our sort of letter of intent to buy them when they're available. Right? So that he basically said that $1 trillion number is from people saying, we're ready to buy as soon as it's available. So that is insane to me.

Was it, uh, I'm kind of assuming because of what I saw, but like, where is that growth going to come from?

Yeah. So there's a couple of things, right? You've got the agent stuff. Like one of the big, big, big topics of this whole event, uh, Jensen dedicated a good like 20 minutes of his keynote to OpenClaw specifically. Uh, there was a whole bunch of breakout sessions. They had a whole booth called Build a Claw where you could go and have them help you set up an OpenClaw agent. Right? Like this OpenCloth thing was such a big topic of conversation during this whole event. And he basically described it as like, this is as big as like the web browser was to the internet kind of thing, right? Like this is the thing that makes agents accessible to everybody. Like throughout the last like year and a half, agents have kind of existed behind the scenes, but they've been like more enterprise or they've been for like real techie. People that needed to understand how to like do some really complicated stuff to set 'em up. And now with Open Claw, it's become sort of accessible to anybody, right? You, you type one line of code in your terminal, which, you know, some people opening up the terminal is scary, but it's really easy. You open up one app on your computer, whether it's a Windows or a Mac, you type one line and you have Open Claw set up and it just is, is there and ready and to, it will go and do tasks on, on your behalf, right? And so he sees that sort of agentic use case where everybody's going to have like this personal assistant going and doing tasks on your behalf and like double-checking their work and maybe helping manage your email. And like everybody's going to have their own personal assistant as one of the big drivers of growth for AI. Um, you know, another thing is like, so we, we broke down last week that the various phases, or maybe it was 2 weeks ago, but we broke down the various phases of how AI models get trained, right? You've got your pre-training. That's the really, really expensive, um, you know, takes months to do. Companies will spend like $100 million to do this pre-training process. But once that pre-training process is done, then they go to post-training and they do things like, uh, fine-tuning the model to get it to like answer like a chatbot. and then you have your reinforcement learning with human feedback phase where now we're basically telling it what we like to hear and don't like to hear and sort of giving it feedback on its responses so that it gets closer and closer to what we're looking for. Well, going into the future, less and less of the compute power is being spent on this pre-training process, right? There's only so much data on the internet. There's only so much data that we can scrape and put into a model, right? So that pre-training process is becoming less and less important. We've kind of gotten to a point where we've put, we've shoved as much as we possibly can in it. So it understands patterns now. Like that pre-training thing is kind of like, we've got that figured out. So now almost all of the AI, um, like movement that's going to happen, most of the next advancements that we're going to see are going to come in that post-training process through fine-tuning and through reinforcement learning, but also through actual inference, right? Like when you go and type a prompt into an AI model now, you can actually see that thinking process. That's called test-time compute where you give it a prompt and it actually thinks through the response, thinks through the response, like almost like debates itself a little bit and then finally comes out with an answer. And so because a lot of the actual like AI use is being moved to that sort of backend of the process, the, the sort of final part of the process. More of the AI compute use is being moved towards that section. And that is only increasing demand for AI because now you're getting more enterprises wanting to, um, build AI in-house, like on-premises, right? They're, they're wanting to build their own like mini data centers in-house to run local models. Um, you know, so like a lot of that, uh, there's, they're getting more and more demand for these chips because of the agentic use case, because of the fact that everything is moving from pre-training to these post-training processes and the test-time compute processes. So there's like a whole bunch of sort of pieces coming together that make it so like NVIDIA was super relevant during the pre-training phase. And now because we're moving into this next phase, well, they're super relevant in this phase and they're building out new infrastructure and new chips that are more optimized for this post-training phase and for this inference phase. So anyway, that was a long answer to your question.

Well, it goes back to, I think even what we talked about last week, which I forget what you called it. You might remember, I think it's 8 different layers of how computing essentially works.

And remember the layers of abstraction?

Abstraction. Yes. And then we, There was something in Claude when we were going through the visualization, we're showing how AI is affecting, like getting used throughout the layers. Remember it was working from the top down and it was like 3 or 4 layers down at different percentages. But I think it's kind of like what you're talking about right now with like just no more computing power and, you know, AI technology stuff like NVIDIA is basically planted into everything. Is affecting every layer.

Yeah, sort of, sort of like there's some layers which I don't think AI is ever going to really touch. Like it's obviously never going to touch this sort of transistor layer. It's never going to touch the binary code, right? You need layers of like AI itself can't just look at the ones and zeros and then figure out what all of it means, right? You need the layers of abstraction beyond that. And the AI is going to sort of understand, you know, whatever, like 3 or 4 layers deep, but won't really try to go deeper than that, at least not for a long, long time.

Yeah.

Like maybe once you get into like quantum computing and stuff, but not with like any of our current models.

Got it.

But yeah, it's moving further and further down those layers of abstraction. But at the end of the day, like every phase of AI buildout needs compute, right? We had the pre-training phase where we were pre-training these models on like, you know, an entire scraping of the internet where we're getting like trillion parameter models and multi-trillion parameter models. Out of like scraping the internet. But now the new phase is we just need to like better fine-tune these models to respond the way we want them to. We need to do better, more reinforcement learning to get them to respond the way we want them to. We need them to get, the test time compute where you ask it a question and it thinks, well, now they're starting to use these, these chips like the Cerebras and the Groq chips, which are designed for really, really, really fast inference. Meaning that it does that thinking a lot faster. So right now, let's say you ask it a question and it thinks for like 15 minutes, but it gives you a really, really good answer. Well, now with these newer chips like Groq, which is a company that NVIDIA kind of sort of acquired, they have like a, it's a weird thing where they have a licensing deal, but they ended up hiring over the people who built it. It all looked very much like they were trying to get around regulations and antitrust laws, but That being said, basically Nvidia owns Groq and Groq, G-R-O-Q, Groq makes these inference chips. And so it's something that thought for 15 minutes. Groq can probably like think at that same level, the same amount of thinking in like 10 seconds. So now imagine if you're giving these Groq chips 15 minutes to think right now, it's thinking at, it's thinking so much more in that same amount of time. So you're going to get output that's similar to like a model that thought for 2 hours, but in 15 minutes.

Sure.

Right. So like that, because we're speeding up the thinking, we're speeding up the inference, you can get a lot more thinking in less time. And now, now start to imagine where you let a model think for a whole day on like these Groq chips. That's like the equivalent of letting a model think for a month with the previous chips. So like at every phase of AI, whether it's the earlier phases where we're doing the pre-training, whether it's the post-training, whether it's the reinforcement learning, whether it's the test time compute at inference, all of these, every single phase requires compute. And who's the biggest provider of that? Who's the one who bought up all of the RAM chips so we have a shortage? Who's the company that's, that, that makes all of this available to us right now? NVIDIA.

That's a guess. And it seems like they are going for, I'm just, the whole OpenClaw thing we've been talking about for a while now. Yeah. Is, I know they released NemoClaw, or at least they talked a lot about it there. And as you're talking about all this inference, you know, basically getting shorter and quicker and able to think longer with the same, it just goes to think like, okay, everything, now you can have your agents basically run longer in the background for cheaper to do better work on your behalf. And it goes back to your whole thing about personal agents for everybody. And now here's NVIDIA that released an option to basically make it easier and more secure, right?

Yeah.

Yeah.

So real quick, like Nemo Claw is, it's basically OpenClaw, right? So it's not like before GTC, there was a lot of speculation that NVIDIA was about to launch a competitor to OpenClaw. That's not what it is. It's, it's a wrapper. I don't know if I like that word, but it's like a sort of a packaging of OpenClaw. So what they did was they took OpenClaw, they made it really easy to install, like one line of code in your, in your terminal to install. You run that line of code and it installs NemoClaw, which is a bundling of OpenClaw plus a bunch of extra like NVIDIA bolt-ons, right? And these NVIDIA bolt-ons are things that help better make, make it better for security, better for privacy, make it easier to install NVIDIA's open source models. They have a model called NeMoTron 120B, which is a model that's really good for agentic use case. Like on benchmarks, it scores pretty close to Opus and GPT-5.4 and like all the state-of-the-art models, but it's an open weight model. And so NeMoClaw makes it really easy to install that. So Nemo Claw is like this bundling of OpenClaw plus these extra NVIDIA security, privacy, LLM add-ons.

Got it.

Right. So that's what, what Nemo Claw is. And Jensen and NVIDIA are really, really betting on the fact that the whole world is going to sort of jump on board this like OpenClaw, Nemo Claw, like have your own personal assistant, right? Like they look at something like Siri or Alexa Right. And they see like, okay, those are kind of like dumb AIs, right? They'll, they'll Google something for you or they'll look at your calendar for you real quick and tell you when something's coming up, but they won't do much else than that. And people still use these like crazy. And they're betting on the fact that in the future, your iPhone, your Android phone, like your everything around you, all of your devices are going to be attached to an OpenClaw. And you're just going to have an assistant like a Siri or Amazon Alexa, and it's going to be able to control anything in your life, right? It's going to be able to control your toaster, your refrigerator, your TV, your everything. And what, what does all of that stuff need? Compute.

Yeah. So it's, it needs a communication device of some sort right now. It's what Telegram and Slack and all these other places, but what's to say in the future, things like Siri or Alexa or all these other— Spot, shoot, someone might already be hacking that right now. Yeah.

Well, I mean, Alexa or Amazon today announced that they're building an AI phone powered by Alexa and the newest version of Alexa. And I'm sorry to anybody who's listening and we're setting off your Alexas right now, but, um, you know, like Alexa's actually built a— or Amazon's actually building an Alexa-powered phone. And there's also Alexa Plus, which is the most recent version. Where it's got more of these extra features. And I do think Amazon in the future wants to, you know, have more agentic use cases, right? Like they already want you to like control your home by saying, hey Alexa, turn on my lights. Hey Alexa, turn my TV on. Hey Alexa, whatever. Right? So like what enables that even better and better is something like OpenCLAW and NVIDIA does work with Amazon. Like, that's, that's another big piece of this pie is how integrated NVIDIA is into like every company you can imagine.

Yeah. It's not even just AI side. It's literally every company out there and their core technology.

NVIDIA, they work with Google Cloud. NVIDIA works with Amazon AWS. NVIDIA works with Oracle. NVIDIA works with CoreWeave. NVIDIA works like you name it. NVIDIA is like, touching them in some way.

It's wild. He said by the end of 2027, that is less than a year or about a year or so, a little more, that that's doubling the company.

And yeah, I mean, yeah, the other thing they said during their Q&A, he didn't say this in the keynote, but during the Q&A, somebody actually asked, and this is going back to like why I still believe in like the Nvidia stock. It's like somebody asked, are you going to actually do buybacks? And Jensen basically said, you know, we've got a giant war chest. We've got just a ton of cash just sitting there that we need to deploy. And our plan over the next few years is to basically give that money back to investors through A, buybacks and B, dividends. Right. And it's basically the Apple playbook. Like if you look at Apple, Apple stock continues to do well. Even though like, like their revenue numbers haven't been growing, like the iPhone sales are not great. Like Apple hasn't been doing well. I don't want to say Apple hasn't been doing well, but their revenue hasn't been like on this sort of like hockey stick exponential trajectory of sales, kind of like Nvidia's has. But Apple sort of has this policy where they constantly buy back stock from the actual investors, which makes that very, very valuable because the supply of stock continuously squeezes and squeezes and squeezes, making the stock still something really, really valuable to own. Well, Nvidia is basically saying we're going to use that playbook. We're going to go and we're going to buy back stock and we're going to give out more money in dividends to our investors. And we're going to actually take all of this capital that we're sending that we're sitting on, we're going to take this. He literally said we're going to take 50% of our free cash flow and we're going to use it towards buybacks and dividends. So like they're making their stock price look more and more valuable by the day, not only in the fact that they have POs for this many chips, but in the fact that they're going to start taking all of this cash flow and giving it back to investors as well.

That's wild. I know another of the big topics that was discussed, I think it was on the keynote, was the whole like space computing concept. And that's something that I know we've talked about before. You've talked about it on your YouTube channel a lot. Thoughts on that? Because I don't know what the timeline was on this specifically.

They didn't give one.

No.

Okay.

Gotcha.

No. So basically what they said was that they're, they're working on it, right? Here's the summary. It's engineered for size, weight, and power constrained environments. NVIDIA Space One Vera Rubin module, IGX Thor and Jetson Orin, uh, blah, blah, blah, blah, blah. AI inferencing for orbital data centers, uh, geospatial intelligence and autonomous space operations. All right. That's just like a lot of gibberish there. Um, but basically they're saying we're building these GPUs to be able to work in space, but Jensen himself admitted the exact same thing. Like we did an episode, I don't know what, 4 or 5 weeks ago where we actually talked about these data centers in space. And I kind of went off on a rant about like, here's why we're not as close to that as everybody thinks we are. And the main issue being heat dissipation, right? Space is a vacuum. GPUs put off a lot of heat. The sun puts off a lot of heat. The whole idea of putting data centers in space is that you, you can kind of keep them in a spot where the sun is, it's always being seen by the sun. So it's getting solar energy nonstop, 24/7, right? The satellites can sort of move. And stay in contact with the sun. So they're always getting power all the time. Well, that causes additional heat problems, right? It causes problems of, well, the sun is hot. And then, you know, you've got your GPUs, which put off a lot of heat. That's why there's all sorts of discussions about how GPUs and these data centers use so much water. They use the water to cool off the data centers, to cool off the GPUs, right? So, so these GPUs put off a lot of heat. The sun puts off a lot of heat. And in space it's a vacuum. Like heat doesn't just sort of like, you know, move away from them in space because it's a vacuum. It's, there's no way for the heat to just like flow off of the GPUs, you know? So they haven't figured out that problem. And in Jensen's keynote, he said as much. He said, you know, we're trying to figure out how to do this. Like, you know, we've obviously got the rocket ships to put them there. We've got the GPUs to do it. We've got the solar panels. We've got all the pieces. The one missing piece is how do we dissipate the heat off of these GPUs? Uh, and he, at the end of his like little, you know, this was a very, very small portion of the keynote, but at the very end of the keynote, he said, you know, we have our best engineers working on it and we'll update you on that guy on that soon. But he didn't give a timeline because they literally don't know a timeline because they don't have a solution yet.

Well, it seems like, yeah, because I know Elon talks about the same kind of thing too. And my, guess is they're probably working together somehow. I'm not sure.

Yeah, I don't know.

I mean, I'm looking it up, but I don't have a definitive answer on it.

They definitely work together in some ways. I know Elon buys chips from NVIDIA. That's definitely true. And Jensen and Elon are friends. They're not adversaries like Sam Altman and Elon are. But like, What's interesting is Tesla makes like their own chips. They're like one of these few companies that when it comes to like their vehicles and stuff, they're actually designing their own hardware. They're not actually using Nvidia stuff inside of like Teslas. So I don't know. I don't know if Elon is actually trying to build his own GPUs to put up into space or, you know, most likely what will end up happening is if Nvidia develops them, they're going to work with SpaceX to get them up there. Right. So they're going to be working on a partnership, at least in that sense.

We got to figure out who's, who's going to break the code, you know, of the whole heat dissipation thing. That's going to be, I don't know if it's one of them or if it's some other company we haven't even heard of yet. So yeah.

Yeah. I don't know.

I don't know.

Maybe, maybe they got to put giant, maybe they got to immerse them in giant fish tanks and then put this watery fish tank up into space with the GPUs inside of it.

I would take the rest rest of the glaciers that are on, on planet Earth and somehow get up. No, I don't know.

I've also heard, um, you know, a more, uh, like a potentially more viable route is actually putting data centers in the ocean because then the cooling is fairly solved, right? Um, but that would also likely heat up the ocean and cause other environmental issues. And yeah, I'm sure there's a whole, all this stuff. That'll just be its own new set of problems, I'm sure.

Yeah. All right.

TBD.

Well, check the news somewhere.

But yeah, I mean, like if I was to point out like overarching themes of GTC and we can go ahead and wrap this up because we've talked about this quite a bit already. I would say like I was blown away by how many companies NVIDIA is like sort of has their hands on, right? Like every single cloud computing company pretty much uses them. Every single robotics company pretty much uses them. Um, Jensen actually won the Motor Trends Man of the Year award this year. I actually saw them hand them the trophy in person while I was at this private Q&A, um, because of all the work he's doing in autonomous vehicles. Um, you know, they're doing stuff in life science, they're doing stuff in networking and telecom. Like their telecom business alone was something like $35 billion last year. Everybody looks at NVIDIA as like, they make the compute, they make the GPUs, but like they have so many other businesses that like, yes, the GPUs like overshadow that. And that's what everybody talks about because AI is the biggest thing on the planet right now. But like NVIDIA is more than just GPUs. They're networking and telecom and they actually have software. They're building these NeMoTron models that like they have like their whole Omniverse. They do all the sort of like AI training inside of virtual environments. Like they are just like this massive, massive ecosystem, but everybody just sees them as the compute company, which, you know, is, it's understandable because that's like 93% of their business. But, you know, they still have like billions and billions of dollars coming in from all of these other revenue sources.

It's just wild.

So, well, cool.

Um, NVIDIA GTC. Uh, yeah, I think it's pretty cool. And all the keynotes are online too, which is awesome. So check it out.

Yeah. And like, if, if, uh, again, um, if you want to play with OpenClaw, if you've seen like all the buzz about OpenClaw all over social media and YouTube and stuff, but you're like, oh, that looks really unsecure and I'm worried about it. Look into the NemoClaw version because supposedly it patches a lot of the security and vulnerability issues and is, you know, it handles a lot of that kind of stuff for you.

Cool. I mean, I'm in on that because, yeah, I've been a little nervous to jump in, to be very honest. And that's available right now. Yeah, I could see you could try it right now. So that's cool.

All right. What are we up to next, Joe? Where do we go? Where do we take this from here?

We take it into more visual ecosystems, realms. And let's go into, I know what, you used to use a lot years ago when you first got into AI, which is Midjourney, right? They're back on the map. They're back, at least in the news.

I don't know if they're on the map, but I don't know if I mean, I'm not going to let the cat out of the bag, but yeah. So this week, so last week Midjourney released V8, right? So they're on their 8th iteration now. And it's, yeah, it's very fantasy, very avant-garde at the bottom left. Uh, yeah, I mean, it, to me, it seems like where Midjourney excels now is for like really sort of like out there, weird, creative sort of stuff, but not realism. Um, like you can't really see it on this one, but if you zoom in, the girl's only got 3 fingers and it's like, Didn't we kind of figure out the fingers issue a while ago now?

And this is their announcement photos, right? Like these are the ones they—

Yeah, this is like their announcement page here.

Yeah.

Um, but basically they upgraded the, uh, let's see, it's much better at following detailed directions, uh, ability to understand your aesthetics, mood boards are amazing. There's a slightly different web interface. You can access it at alpha.midjourney.com. But again, I believe Midjourney, you have to be on a paid plan to make it work. And I mean, you could just like scroll through some of the images that people have generated with Midjourney. And I just look at this stuff and I go, yeah, I mean, it's, it's all right. But I feel like we're getting a lot better out of like Nano Banana and some of the other platforms we've gotten access to. I mean, this one's looking pretty realistic.

I was going to say, like her, I saw that there's a guy up higher up. That looked more realistic.

Uh, this one?

Uh, not that one, but that's a little freaky. I didn't even notice that one. Left guy.

Oh, this guy maybe?

Yeah. So I mean, yeah, but you're right. I feel like Nanabanana has been making the biggest splash lately with realism. And I mean, really any style, right?

How many fingers do you count there?

1, 2, 3, 4. Unless one of it's That's a claw hand. That's a claw. Can we do a quick generation, Matt?

Like just, yeah, we can, we can. This model is actually very fast.

Well, let's generate some hands and just put it the hand test.

Yeah.

Okay.

Um, let's see a closeup of a hand.

No, no, no. Actually do like a, a bunch of like, I don't know, a circle of people in a circle holding their hands out. Like just test it really hardcore.

A group of people in a circle, all putting their hands in.

Sure.

Sure. Let's see what that does.

So it's fast, you said?

Yeah, I see it's almost done here. Holy moly. All right. So this one is done.

Um, how many can we zoom in?

That's as far as it's letting me zoom in.

Let me see.

Maybe I can go a little further.

Uh, oh, that one's a little weird.

Yeah. Like this guy's got a big old middle finger. That's a, that's an elephant trunk for a finger.

Carrot. Yeah.

So I mean, look, you get a bunch of hands in the picture and, uh, not, not amazing. Not amazing. Here, let's look at another one.

I mean, I don't know. Ooh, that one's even worse. I feel like on the top right, those hands are blurring in.

What's going on there?

Yeah.

Yeah. What's going on?

Melding into each other.

What is going on? This guy's got a finger growing out of the middle of two fingers.

Wait a minute. All right. Do the same test. Go to Nana Banana. I'm just curious. Same exact prompt. Let's just, we got to do a side by side.

Okay. Um, all right. So I'll go ahead and create an image and I will paste that same prompt in a group of people in a circle, all putting their hands in.

Comparing on speed and quality, even though I feel like we should more compare on the quality. If they're big enough, it was speed. I mean, it's like, that's great, but it was fast.

Yeah. This one is definitely not quite as fast, which is interesting because Nano Banana 2 is usually pretty fast. Focusing on hand details. It's good to know that they're focusing there because that's what they will be judged on.

What's a little pull-down if it can, uh, okay.

Okay, so here's our image.

It looks normal to me so far, even though those two girls seem to be twins. They seem cloned.

It's okay. Twins exist in the real world.

They do.

You're right.

Yeah. I mean, I think it passes the hand test.

This one worked. Yeah. So Nano Banana, they're fine with hands, it seems like. Midjourney, on the other hand, I gotta hand it to them.

Got it. There we go.

All right.

So let's just move quick unless there's anything else on the Midjourney. I mean, they, yeah.

So Midjourney, it's, I think it's really good at just creating weird random stuff. Like I gave it this thing to make it like a see-through clear elephant with a galaxy inside the elephant walking on a sea of books kind of thing. And you know, it, it gets the weird random stuff good. But like, yeah, to me, the reason I wanted to bring up Midjourney was less about like, this is impressive and more about the fall from grace. Like if you look at Midjourney when it first came out, like when you and I were doing like our Hard Fork shows back in 2022-ish, right? We were talking about Midjourney towards the end there. And like that was like what was available and it did okay. And then pretty, pretty quickly it got to the point where everybody's like, holy crap, Midjourney is proof that we're not going to be able to tell the difference between AI and real images anymore. Like Midjourney is just that good. Holy crap. Right. That was around Midjourney version 4 where people started to go, oh my God, this is blowing my mind. And now you fast forward today and like, pretty much every other model out there, in my opinion, is doing a better job. And that to me just blows my mind that they like sort of fumbled the ball that badly.

Yeah. I mean, yeah, because you, I don't think, I don't remember the last time you personally talked about on your videos, but you used to talk about it a lot.

No, I mean, every once in a while there'll be like a new model release or a new feature and I'll just kind of quickly talk about it, but I don't go too deep into it because it's just not one of my preferred models to go use anymore. And I think it's Midjourney's own doing. Um, their, their, their founder, like, you know, he's like a brilliant, like genius of a founder, but he does refuse to take any sort of venture capital. They've never raised any money. The whole business has been bootstrapped. So even though companies have come to him and tried to give him money, they basically said no. Um, they've never released an API. So, uh, you know, the tools that are out there like Leonardo and Kria and, um, You know, Runway and all of these platforms, Photoshop, all of these platforms that let you choose which AI model you use inside of the platform. None of, none of those have been able to use Midjourney either because Midjourney just refuses to put out an API where other tools can tap into it. Right. So like they, they're not taking outside money. They're not being able to like sort of capitalize on getting their model inside of other platforms. Which could be a giant revenue stream for them as well, which they can then put into more model training, more GPUs, whatever they need to sort of get this model better. But like the things that people have been like trying to get out of them, they're just like, no, we do things our way. Screw you. Like, this is our, this is how we do it. And I think that is starting to catch up with them.

Yeah, man. I mean, things are rolling out all the time. It is. That's kind of an interesting way of thinking. Maybe he's got a bigger play, you know, the founder there. But how about the, 'cause there was another image model release as well this week and this one from Microsoft, which.

Yeah, so Microsoft, they had an image model already called MAI-Image, but this is MAI-Image 2 and it's actually pretty good.

Pretty good.

This one does text. I would say it's like fairly on par with like Nano Banana, right? It's like, it's about as good as Nano Banana. From, from most of my testing. But this one, I believe anybody could go and use for free right now. Let's see, where's the link to it?

Where do you find it?

Arena? No, Arena is where you could see the leaderboard, right? So they're actually showing that this is like a blind test leaderboard. You give this, you give it a prompt. It generates an image. It doesn't tell you which model generated the image. You pick your favorite of the images. And basically based on enough of those sort of blind tests, tests. It ranks them and people like this one, the third best behind Nano Banana and the, you know, the ChatGPT image model.

Nice.

Okay.

Yeah, it looks like there's a playground to test this thing out.

Yeah. So let's go ahead to MAI playground. And so here we are. Let's go ahead and give it the same prompt just because, you know, might as well. A group of people in a circle all putting their hands in.

The hand test.

This one's not quite as fast as Midjourney or Nanobanana.

Okay. Okay. From your tests, what have you found here? Like, have you tested text and realism?

I did. I tested text. It got all the text perfect. I tested like realism. It's worked really well with realism. All right. So here's our hand test and Looks good to me.

Looks good.

Yeah. I don't see any issues there.

No, that's cool.

So I mean, this brand new model from Microsoft, Microsoft's only been making an image model for like a year. Midjourney was years ahead of them in image models and Microsoft is already better than them at the hands.

Interesting. Okay. Do they have an API? Do you know?

They should. Cause if it's on Arena, that has to, Arena has to tap into an API to be able to pull in images. Okay, so here we go. MAI Image 2 is beginning to roll out on Copilot and Bing Image Creator. API access is available today for select Microsoft customers. So they do have an API, but it's, you know, you've got to fill out an application and get access to it. It's not like a publicly available API yet, but yes, they do have an API.

Got it. Okay.

All right. We can test text. Let's see. Neon. Sign over a dark Tokyo cityscape that says Joe Fier loves tacos.

In Tokyo. Okay. I do love tacos.

Hey, if you leave me to my own devices of typing random stuff, stream of consciousness, it usually ends in tacos or Padres for some reason.

Reason. I was betting the latter, but I'm okay with that. We'll test that.

We'll test a Padres player too and see how well it, like, see if it pulls in logos and stuff.

Likeness and logos. Yeah. That's always a good test. Remember we were, we, I think we were testing Nano Banana 2 the other week and it was pulling in all sorts of likeness. Yeah. That's pretty good. Look at the reflection too.

Yeah. Yeah. And would you say it's a true statement?

Absolutely. It is. Is that the Tokyo skyline? I'm not savvy enough on.

I doubt that part's very accurate.

If I had to guess, but hey, it nailed the text and the truth behind it. So cool. Yeah.

The text and the truth behind it. All right. Should I try a specific player's name or?

I think so, because I'm feeling the likeness and the logo. Let's try to combine those.

Manny Machado on the San Diego Padres hitting a home run. Oh.

I was— oh, sorry, I was gonna give it more.

I could not be generated. It's got— I think it won't generate people with like real people's names, it looks like.

Okay, okay.

So they got some guardrails on it. Um, or maybe it's the Padres. Maybe the Padres is the brand they won't do. Um, let's see, a player on the San Diego Padres hitting a home run in No, home run in Peco Park. Was that what you were going to say?

Just to add a little bit more.

Yeah, I won't generate it. Oh, oh, you've reached the usage limit. Try again in 27 seconds. So my usage limit is 2 and a half images.

2 and a half. Let's do the countdown, I guess.

So it doesn't seem like we're super limited on how many we can generate. It's just like they want you to give it a little bit more spacing in between. And the only reason there was no spacing in between was because it refused to generate this one. You would think it would sort of waive that requirement. If, uh, you know, it fails on a generation, it should let you generate another one quickly immediately after a failed generation.

You'd think so. It's not kind to Microsoft.

I'll talk to Mustafa about it. Hey, there you go. That is a weird looking Apeko Park, but he's also aiming in the complete wrong direction.

And he's swaying with one arm.

He's throwing with one arm and he's, he's hitting it out into like the, the third baseline crowd. Look, you can actually see the ball here too. It's like, and look, he's like, screw the fans. I'm just going to hit it to them.

Well, look at the baseball diamond. The chalk is going behind him to first base where it should. You have the, the, the foul poles right in the middle and center field.

Yeah.

Yeah.

Yeah. It doesn't understand the design of a baseball stadium. Hey, but it got the Padres like jersey and logo and everything, right?

That's true. That is true.

So, but he just like turned and apparently like nobody pitched the ball at him. There's not even a pitcher's mound there, but he just like turned and hit it into the crowd.

This is like one of those games you could probably, it's almost like what Magic, I don't know, like where you like spot things that are a little off in images. I mean, like so much, there should be a good published AI video or sorry, book. Yeah.

That'd be fun. Well, it's interesting because you could tell this is like probably supposed to be the Western Meadows building out here. So it kind of knew like some of the elements of Petco Park. It just didn't know where the things are supposed to go.

Yeah. Yeah.

Yeah.

All right. Well, these are fun tests.

Yeah. Yeah. So those are the two new image models that came out. Let's be real. We're probably still both going to Nano Banana when we need images.

Absolutely. And surprisingly, ChatGPT, like I've been is still good too, but Nana Banana, number one in my book.

Yeah.

All right. Let's change the, turn the tables a little bit. And now let's talk about kind of the real world implications. We were kind of playing around with like, hey, what are some of the things that maybe when we're not staring at a computer, we're like, how do our lives change with AI in the middle of things? And there was a job market visualizer. That was pretty interesting. Here it is.

What do you know about the guy who created this job market visualizer, Joe?

He came from OpenAI.

Am I wrong? No, you're right.

I am right. Yeah. He was one of the co-founders very early on.

Uh, he might've been one of the early founders. Uh, he might've been a founding member. I don't remember, but he left. He, uh, so Andrej Karpathy is the guy who created this US job market visualizer. And he actually helped design a lot of the like autopilot system for Tesla and then ended up going over to OpenAI and was like, you know, you know, one of their brilliant engineers over at OpenAI and eventually left OpenAI to kind of pursue his own projects and his own projects are seemingly, you know, making stuff like this and putting with OpenClaw. I think he's building some sort of like educational system as well.

Okay. So this one I thought, I mean, we're both looking at this and we're kind of surprised by some of these careers or these occupations, but it looks like what it visualized 300 42 occupations.

Yeah. So basically the, if we look at this, it's this giant grid and these are various occupations and the green shows occupations that are fairly safe and growing job markets. And then the red and orange are declining jobs. Right. So a negative outlook for those jobs. Um, you know, so some interesting stuff here. Yeah. Cooks, you know, positive outlook. It's actually growing faster than average. That makes sense. Home health and personal care aides, much faster than average.

We were talking about yesterday, aging baby boomers, you know, it's like, yeah, that's, there you go. Prove that.

Yeah. I mean, that makes sense. And I mean, like some of these are like, make a lot of sense to construction laborers, carpenters, electricians, plumbers, painters, construction equipment operators. Apparently construction and building inspectors are in decline though.

Cameras. Maybe just like you can have anybody just kind of be the inspector at this point.

Boilermakers?

Boilermakers.

What is a boilermaker? I'm actually not— what does that mean?

I don't know. That's one of the tiniest boxes on there, it looks like.

What boilermakers do? Boilermakers assemble, install, maintain, and repair boilers. Clothes vats and other large vessels or containers that hold liquids and gases.

Okay. Okay. Yeah.

Well, apparently those are going away.

Those aren't great.

General office clerks, bookkeeping. That makes, that makes sense to me. Customer service representatives, because like obviously things like Delphi exist. Yeah.

They're not customer service though. It's like, actually I called my plumber yesterday because I was supposed to have some stuff done right now. and I was rescheduling and her name was Bonnie. And I was like, Bonnie, I was tricked. I was duped. And then it took me about 30 seconds and I'm like, this is an AI. That was a really good AI. And I even asked her, are you an AI? And she said, yes, I'm an AI person. I was like, Bonnie, you almost like duped me. It was really cool.

But customer service, I think I find really interesting too though. Like kindergarten and elementary school teachers are declining.

I don't know.

Teacher's assistant. Now, I don't know if this, I don't specifically know. And I don't believe that this is just declining because of AI.

Well, if you look just in general. Yeah. And their color graph, I mean, maybe it's more slow, even though the decline, the orange, I guess, would be declining.

Well, there's a digital AI exposure layer. So let's click on this layer here.

Here we go.

And now things start to change a little bit. Wow. All right.

We didn't look at this. Top executives were even on the decline. You see that in the very top? Software developers.

Top executives.

Yeah.

But here's what's interesting is like digital AI exposure, software developers are declining like a lot, right? But if you flip to this one, the Bureau of Labor and Statistics outlook, it's pretty— and you look at software developers, it's growing much faster than average.

Can we click into any of these boxes? Yeah.

If you click into it, it takes you to the Bureau of Labor Statistics here.

I got you the source.

So, I mean, based on the Bureau of Labor. So like, this is what I'm taking away. And I don't know if I'm interpreting this correctly is this is like actual job growth and job loss based on Bureau of Labor Statistics, right? Like this is the, this is the sort of ground truth as they say, right? Um, if we look at this, this is AI exposure. This isn't reality. This isn't like jobs are declining fast because of software. This is showing this is very exposed as a result of AI.

Yeah. Right.

So like software developers are really exposed as of AI. AI is sort of something that can take over that role. But based on the, the statistics, software is still growing.

And that, I guess, makes sense back to what we were talking about with like agents coming around and now there's just a different flavor of where these developers are going to go work. But it's interesting going over to that AI exposure layer. Yeah, you can now kind of start to maybe project. I mean, you just see where the red is, is— I mean, you can see janitors, construction, all the very hand-oriented jobs, you know, like physical stuff. Not much AI there. Yeah.

Like down here, lawyers, very, very exposed. But the reality of it Lawyers are actually growing 4%.

But then you have paralegals. So in the law space, it's more of the, yeah, those are, yeah.

Well, it's saying little or no change. So it's remained unchanged, but paralegals are one that probably have a lot of exposure. Yeah. Look at that. Tons of exposure if you're a paralegal.

Yeah. Interesting. Okay.

So it's fascinating to, like, to me, what's so fascinating about this is it's showing that all of these roles because of AI have a lot of exposure. Like AI is able to do more and more and more of this stuff. That's what these red boxes are showing. Going, but the real-world data based on the statistics is that it's not having as big of an impact as what everybody's worried about so far.

Yeah. I'm sure the data, as new data comes in, it'll change quickly. So I don't know, maybe he'll keep this up to date so we can come back to it later.

Well, yeah. I mean, I'm sure that pulling it automatically.

Yeah. Well, and it looked like I saw some of the stats from 2024.

So from the bureau, if we click into one of these here, at least I think the pay stats, it says 2024 median pay.

Yeah, I bet there's more layers of data that are, you know.

Oh, but it does say last modified August 28th, 2025.

I'm sure that's probably the best stats we can find.

So yeah, well, one of the other things we were talking about actually before we hit record was like SaaS, right?

Yeah.

And, um, like software developers and how that world is changing. And I think, I think there's a lot of like interesting nuance there because we were talking about how Jensen was essentially saying that he doesn't see like SaaS going away, right? He thinks that, um, you know, essentially SaaS isn't going anywhere. It's just your agents might be using those SaaS tools as opposed to you using those SaaS tools. And if you're not familiar, SaaS is software as a service. It's basically like, you know, HubSpot is a SaaS, right? It's a, it's a, a service you pay for monthly, but it's like a software online that you pay for as a service monthly. And so like Jensen's take is you're just going to have your agents go and use the SaaS. The SaaS is, the software companies aren't going anywhere, but like your Claude, your Open Claw is going to go and actually use the tools on your behalf. And I do think that's sort of like the positive outlook from somebody who is very incentivized to calm people down and say, don't worry, these jobs aren't going anywhere. Like these software companies aren't going anywhere because I mean, NVIDIA has how many clients that are software companies? Probably a lot. He doesn't want to say like, hey, I'm building the technology that's going to like obsolete you. So he's very incentivized. Let's just put that out there. To, to say SaaS isn't going anywhere.

Absolutely.

Um, saying that, I do think there's different levels to this. I think when it comes to like big companies that have like big infrastructure built out already, they're still going to have a lot longer of a lifespan, right? The, the companies that are like simple apps, like one of the examples that we were talking about offline is one we used to promote as an affiliate called Thrivecart, which is like a Online shopping cart platform. It's sort of like a wrapper around Stripe, essentially, right? Um, like companies like that, where you make a software that's like a one-trick pony software that doesn't have any data centers, you don't have a ton of infrastructure built out, you know, you don't have to deal with like warehousing. You're just a wrapper on top of an API like Stripe. Those types of SaaSes are probably going to be in trouble. Like I don't see a world where those companies continue to grow and make a ton of money at like a rapid rate in the future, because I can, like Jensen says, I can have my agent go and do things for me. And if I need a shopping cart, well, I can go design a shopping cart in Stitch, pull it over into AI Studio and connect my Stripe API. I don't need a tool like Thrivecart anymore. And right now it does take somebody who's slightly technical, right? Somebody that kind of knows that Stitch and AI Studio or cursor exist and know how to use them to some level. But that's only like a transient problem, right? Like that's a problem that's going away in the future when more and more people have these AI assistants. They're just going to be like, hey, my website needs a shopping cart tool for checkout. Go build it for me. And it's just going to be built. You don't need a tool like a shopping cart wrapper anymore, right? Like you just don't need that. But the companies that do have the big infrastructure, the data centers, the like that kind of stuff, you can't vibe code a data center, right? You can't vibe code TribeCode, all of that additional infrastructure that was, that's built around some of these big tools. Um, so, you know, my prediction is SaaS is still going to exist, but like where the money is flowing in the SaaS world is going to shift, right? And I think it's going to shift to like a lot of the backend stuff, you know, your, your database companies, um, your, uh, like Vercel type stuff that, that, like the, the places where you're actually like hosting the front ends of your website. It's going to be like all of the plumbing type of stuff is still going to exist long into the future. But anything that's like taking an existing API and just putting like a front end in front of it, which, you know, um, I, I don't have a relationship with the new founders or the new owners of ThriveCart. So I don't mind throwing them under the bus, but like that is, doesn't, that's not a thing that's going to work in the future. That's, it's at this point, it's just a wrapper around Stripe.

Yeah, and I looked something up, like what Jensen was saying on the All In podcast, because he talked about this, is it's the SaaS incumbent SaaSes, which are probably like HubSpot and, and, you know, bigger companies that have been around for a while. People are, you know, these companies who have strong workflows and data moats, like you said, and easy access to robust APIs. And then they were talking about, or he was talking about first-party agents. As well that can be essentially wrapped in with these SaaS companies to support users. So yeah, I think back to, um, yeah, the ThriveCarts of the world, which it's still around. I did see some, uh, some usage of it the other day. Um, yeah, I, I, I think that's more of a wrapper rather than something like a, like a incumbent SaaS with a whole ton of workflows and data moats, kind of like what you were saying. Yeah.

Well, I mean, I don't think even like that type of SaaS is gonna be like an overnight death. Death, right? It's not going to just like that kind of thing's going to stop existing overnight. I think it's going to be a sort of like slow fading out of those kinds of software companies. Um, yeah, but I think it's going to require further ease of use, further sort of democratization of a lot of these AI tools, right? It's going to require, um, it like, I think a lot of people don't want to go and try to learn Cursor or learn AI Studio or learn Claude Code yet. That even though like once you get into them, they're really not that scary. They're really simple to use and intuitive. I think a lot of people still see them as like, ooh, that's, that's scary. That's opening a terminal. That's, you know, that's working with code. I don't want to get into that. And I think there's still like a giant portion of the population that's not as nerdy as us. That's just like not even willing to touch that stuff yet. But once it gets to a point where you're just like, hey Siri, I need a checkout for my website. And then 10 minutes later, you get a text message that says, hey, your checkout's ready. That's a different world.

Yeah. Yeah.

All right.

So moving on, let's talk about something else that Anthropic put out where they interviewed, was it almost 81,000 people about what they want from AI?

What do you want from me?

What do you want? Geez. Haven't I done everything? I'm thinking for you.

You're talking like me now.

What else? I made your life easier. You don't have to work anymore. No.

Yeah. So yeah, like you said, they did, I think AI did the interview. It says we did, we used an AI interviewer and they interviewed 81,000 people. So let's actually see what some of the findings are, or did you already pull up some of the findings?

I did pull up findings and thank you Perplexity for summarizing things for me. But it basically said the goal of this was to understand what AI, what AI going well looks like in real people's lives, not just abstract debates about risks and benefits. So what it said that the top hopes for AI, basically what people want, 18.8% said professional excellence. So offloading any kind of routine work. Then you went into personal transformation. That was like 13.7%. So personal growth, essentially. Uh, life management was next, 13%. So helping with schedules, organizations, things like that. And you have it on the screen now. Oh, great. Time freedom. Yeah. Financial independence, societal transformation. I mean, like, this makes sense. Yeah. This is really cool.

Entrepreneurship, learning and growth, creative expression. So interesting to me that what people actually hope for is professional excellence.

Like, yeah, I don't know why that's— I mean, I think that's the obvious thing for AI though, right? Basically what it's— these perplexities are saying it's offloading that routine stuff. Any of the— So any repetitive action so you can focus on more of the high-level stuff.

Which, yeah, I think for me, the, what I would have guessed if I was just like, if, if, if I was on the Family Feud and they were like, we ranked the top 8 responses, like my number, my guess for number 1 response would have been time freedom. Like for me, that's what I feel like an AI should enable for people, right? Like, you know, I, I want to get my work done. I want to do it with excellence. but I also want to spend less time on it. True. Um, you know, I, I want that time freedom. I want to go and, uh, you know, hang out with the family, go camping, play video games, play, you know, touch grass, that kind of stuff.

Hang out with people.

Um, and I want to get the same amount of work done in less time. That's what I want AI to create in my life. And so like, if I was to guess, I would've thought that would've been like the number one answer.

Mm-hmm. What's interesting, it says where AI has already delivered, it says 81%. Said AI has taken at least one step towards their vision. 32% have said around productivity, dramatic speedups and automation and repetitive tasks. So that's what's interesting because I think you've talked about this, maybe it was on X or we talked about it, I don't know, but like, or maybe it was one of your YouTube videos where sure, you might be more productive on some of these things you're already doing, but then you're actually doing more things because you've created more time and you're already in the task. You're like, screw it. I'm not just going to complete this one thing. I'm not going to go code a bunch of stuff that I didn't even know it was possible about a week ago. So I'm going to just hack all my time with this new stuff at this point.

Yeah, no, there's been this weird thing that's happened where because we have access to AI and we're able to like accomplish things quicker, well, we just stack more things, right? We go, oh, now I, because I can make my own, I could code my own website. And because I can, I can, you know, respond to all these emails faster. And because I can do this and I can do that, because AI makes all this stuff easier, we just try to load our day with more stuff. And what ends up happening is, although AI should be making our lives easier, everybody feels actually more stressed out than they ever did because they're trying to do so much more than they ever did because AI makes it easier to do these individual things faster. Right. And so there's been this, this weird sort of phenomenon that's happened where AI is actually making people more productive, but they're also putting in more hours as a result.

It's so, so interesting. Interesting. Yeah. I think this is a cool report and I bet we can all, if you threw it, I'd probably task people that do this, just throw this into Perplexity or Claude or whatever and just like start asking it different questions around whatever you're I don't know what you're doing in business. I feel like this shines a really cool layer of insight into what people are thinking about AI, how they're using it.

Yeah. Well, I also, I also find it super interesting that the number 2 response was AI hasn't delivered.

Yeah. What is it? How's it?

So number 1, the 32% of people said it's helped with productivity. The number 2 response was AI hasn't delivered. Number 3 is cognitive partnership. Which makes sense, right? Like a lot of people use it for brainstorming and like, hey, I've got an idea for a video, help me flesh it out and that kind of thing. Learning, that makes a lot of sense to me. Like, hey, I hear a lot about quantum computing, break it down for me, whatever, right? Technical accessibility, research synthesis, and emotional support. Emotional support is the slippery slope one that I would like caution people of using AI for emotional support. Oh yeah. Yeah, that just seems like a slippery slope. But number 2 being AI hasn't delivered, to me, that's people like either you haven't used it enough for, or tried to use it for what you're doing, or you must work in a field where it's just not there yet in that field. Cause that to me is like, I don't know.

I still feel like we're so early enough in AI, even though it feels like talking about it for a while, that there's still a big education curve that this moment that we're in where people just don't know to apply it to aspects of their lives.

Yeah. Here's another like little chart, what people worry about. Unreliability is number one, which makes sense when you see that most people want it for professional excellence is number one. Like that computes, right? If professional excellence is your number one goal, your number one worry is going to be unreliability, right? Like you've probably heard the stories about like the lawyers who used AI and then went to court and said something to the judge that turned out to be a case that never actually existed.

Right?

Like they were trying to use existing precedent from like a case that like AI just made up. So like, if you're looking for professional excellence, unreliability would be your number one fear.

Look at number four, cognitive atrophy.

Yeah. That's probably for me. That's probably my, like, that would be my number one at the moment.

And we talked about it.

Because sometimes I find I lean on AI too much to help me with stuff that I get to the point where when I'm not using AI, I'm like, ah, how do I think through this problem? Oh, I'm just going to go ask AI. Like my, like, like it does actually impact your ability to like think through problems and things.

I'm surprised that existential risk is at the very bottom. And when that study came out about the whole, okay, well, if AIs are in control of, you know, wars and whatnot, they all revert to nuclear war. Yeah.

Well, I think, you know, a lot of people have seen these models hallucinate. They've seen them make mistakes. They've seen them like not live up to expectations in certain areas. And they're like, yeah, that whole Terminator scenario, I don't think we're that close yet.

I hope that's true. Well, I think again, this is a deep, you know, report. There's way more here than we can go through. And well, cool visualizations too. But yeah, yeah, I think a plugin, at least this is what I'm going to do, is, is take this and go deep into Perplexor or whatever and just like find meaning out of here. I don't know if we can get to the raw data as well that they publish, but I also find it fascinating too that like they break it down by like regions.

You can see like what North America is most interested in versus what East Asia is interested in versus what, you know, Central Asia is interested in. And it's like they don't line up. It's interesting.

Yeah. So let's, let's transition to something that's interesting.

So, um, Are you saying that wasn't interesting?

Cut that. So check this out. We have new opportunities now for work in the, in the real world, right? No, but I think we were talking about maybe this is where some new positions or new jobs will be created, for better or for worse. But let's, let's talk about what DoorDash is now doing with DoorDash Task Yeah.

So with this DoorDash Tasks, um, basically it's like a, a way to just like earn some side hustle income by doing like little things. Like it seems like it was designed to like, you go to a restaurant and you take pictures of like what you ordered so that they have, you know, pictures of that food item online so people can look it up. Um, or like you go to a grocery store and you take pictures of like the shelves to make sure like the items are stocked or whatever. Right. Like, that's kind of like the initial sort of pitch of this product. Like, that's how they angle it in like the first few paragraphs. But if you scroll down on this article a little bit here, this paragraph is where it gets particularly interesting. We're also piloting a new standalone app where Dashers can complete activities like filming everyday tasks or recording themselves speaking in another language. This data helps AI and robotic systems understand the physical world. Pay is shown upfront and determined based on effort and complexity of the activity. So basically DoorDash is saying, we will pay you to train our AI systems. Like that to me is, um, implicative of like where I think the world is going. And this is something that, you know, you and I, Joe, we've had conversations about this offline in the past. Um, but there's this whole talk about like, okay, if if AI does get to AGI, well then that theoretically means AI can do like all meaningful work for humans, or at least it's getting close to it. We have to obviously be able to like embody it inside of robots and stuff if we wanted to get to like plumbing and construction and stuff like that. But, um, at least from like a knowledge work perspective, uh, it can do like almost all knowledge work. And so where does that leave the economy if we can't make money anymore? And a lot of people speculate, well, we need some sort of like UBI, universal basic income, where the government just gives everybody a sort of stipend every month. Like, here's, here's a check that you get and this is your spending money every month, right? Like, that's one of the theories or one of the sort of suggestions of, of what direction this could potentially go in. And I've always had this thought that I don't really see it going in that direction, at least not anytime really soon. I think more likely what is going to happen is you're going to see more and more stuff like this where the big corporations are going to figure out how to pay people for their data. Um, you know, when, when you think about it, humans are getting more and more sort of stingy with their data, for lack of a better word right now. Like we've seen all the like Cambridge Analytica stuff from Facebook, right? We've seen like data leaks from all sorts of companies where like your emails and passwords get leaked all the time and stuff. And so people have gotten more and more like, I don't want companies to be having my data. They're also using this data to put targeted ads in front of you, right? Like it's collecting information on you so that it can better sell you stuff.

Stuff.

And so because people don't really like giving out their data, there's more and more sort of systems to protect your data, right? Like I, I actually literally use a service that goes around the internet and scrapes my data from like any list that I'm on automatically in the background so that, you know, companies stop calling my phone and stop sending me junk mail and all of that kind of stuff. There's like literally services out there that will go and try to help you to remove your data from the internet. And so it's going to make it harder and harder and harder for companies to collect data to further train their models on in the future. Well, how do they get the data then? Well, they can incentivize humans by, I'll give you my data if you give me money. And so I actually think that that is a direction that a lot of these companies could end up going in, right? Like the money from AI is all going to go towards the top, right? It's all going to go to the Googles and the Microsofts and you know, OpenAI, XAI, NVIDIA is probably the very tippy top of the pyramid, right?

Yeah.

Um, it's all going to go up to these companies. And if these companies don't need workers anymore because AI is doing so much of the work, how does that wealth get distributed? You know, if these companies can have less and less and less employees, the, the money, it's going to go to like the, the inside the company, but not get redistributed to humans in some way. Well, this kind of thing seems like a potential way to redistribute it. Okay. Since you don't want to just freely give us your data, why don't you go out and collect data for us that we can use and we will pay you for that. And that seems like a, like a very incentivized way for these companies to continue to collect data, to continue to train their models, but also to get humans paid Yeah, it's, it's, it's a whole industry of, uh, what is it?

AI data collection, getting paid for that. Essentially it looks like, and I'm looking up, I'm like, okay, what other companies are doing this? Uber looks like they announced in like October last year, they're doing similar stuff where it's like uploading photos, recording audios in their own language, submitting documents to train AI. Obviously that's more driving. There's, there's different market. I'm looking at some other stuff here too. But you know, this is going to be a growing thing because I'm just thinking of training data that has not reached its way online somehow, even though a lot of people have data online, probably on their websites. But it's, it's the personal stuff. It's what individuals are walking around and doing and seeing and touching and feeling, you know, and like, yeah, yeah.

Well, like, I don't really think there's a data problem in the sense of text, right? Like, these companies have already scraped the internet for all of the text they can get. They've got enough text data to pattern recognize how people speak, right? Like, how people type. Like, the text is no longer the issue. The issue now is like the next modalities other than text. It's getting these AIs to understand the physical world around them through through filming videos or through your, your, your, your car that has its cameras and sensors all over it, right? It's, um, you know, I just saw an AI, uh, announcement about a company that's like trying to teach AI how to smell, right? Like, like obviously translating things into other languages. One of the things this DoorDash tasks here is like doing translations. Um, you know, like we were even talking about how like this is, this is like a system of kind of like how CAPTCHA did it, like all over the internet. Like they, they have these little CAPTCHAs, right? Where it used to be like, they took a picture of like a sign and then you would enter the text that was on the sign and it would let you in. Well, that served more than one purpose. The purpose wasn't just to prove that you were a human because you can read the text that was in that picture. The purpose was also you were labeling all of that data for those companies, right? Duolingo has a very similar thing when you're actually using Duolingo. Like it's training on how you're using Duolingo as well and trying to improve Duolingo's system. Um, yeah, I'm pretty sure the people who created CAPTCHA were also the same people that created Duolingo. I could be wrong on that, but I think it's actually the same people. Um, I know Joe's gonna fact-check me. I am.

Keep talking about it.

But like, they, these companies need data and text data isn't really all they need anymore. They need other modalities. And so that's why DoorDash is saying, look, we'll get you to translate stuff. We'll get you to go film things in the real world. And, um, you know, I, I think that is a model that could potentially work. The only problem I have with that model is that, well, I don't want to say only problem. There's definitely a few problems. One is like you become reliant on these companies, right? And now all of a sudden, like these companies control your income. And if you're not, you know, following their terms of service and whatever, they can just like cut off your income. That's not good either.

No. Privacy. How about that? Like you're literally selling.

Yeah. Privacy is also an issue. So there are a lot of problems with it, but the other problem is I also think it could be a race to the bottom. Like, because people will be willing to do this for cheaper and cheaper and cheaper and cheaper. It's not a skilled labor. So like how much money can this really generate for people?

Yeah. I don't think it's the solution, but it's definitely something that we'll see more of to capture the data. The stuff that has not made its way into text and to LLMs already. And we know there's things out there. Um, just to fact check you, there was a single person that, uh, a key founder, Louis von Ahn, uh, he, it was, so it wasn't the full team, but he came from CAPTCHA, it looks like. And later was later co-founded by that same guy.

Duolingo was?

Yeah. Duolingo.

Um, I knew there was a, I knew there was a tie-in between CAPTCHA and Duolingo because like the guy behind both of them, like a big point of what he was trying to do was like this dual purpose system of A, data collection and B, sort of trying to create something useful out of that data collection.

Right. Yeah. No, I, this is definitely dystopian and I don't think it would happen, at least not anytime soon, but think of Neuralink. Not, not everybody is having a Neuralink done. Obviously there hasn't been that many people. But it's like, hey, you want to start tapping into the intuition, like we were talking about Jensen Huang last week and like, you know, the way that people are thinking without— there's all sorts of ways that maybe we can tap into data that hasn't been trained into AI yet. It's very dystopian and probably far off and never possible.

Or maybe we should have a chatbot that just taps straight into Jensen's brain. So I go to like JensenGPT and I just ask questions, but it gets the response directly from Jensen's brain.

Yeah.

Yeah.

It's very possible. There's—

let's sell our brain power. Once we're out of CPUs, we'll start— and GPUs, we'll just sell direct access to the brain.

Someone's going to want that. We'll figure it out. You know what? I think it's time to end on the thing that we always like to bookend these episodes with, which are the weird mechanical things that are starting to look more human-like by the day. Robots. And Matt, you're queuing up a video of one that I found pretty interesting. What's it called? It's Galabot? No, that's who reported it. This is a university. I have the source in the tab here, but it's called Latent Learning Athletic. Latent. Yeah. Learning Athletic Humanoid Tennis Skills. From imperfect human motion data.

Yeah, I've actually seen some of these kinds of demos. They had, uh, like robots playing ping pong when I was at CES. I don't know if it was the same, like the same company, but like, yeah, it's, it's kind of crazy that they can do this kind of stuff.

I think this is a university that did this out of China, I believe, but the video is cool.

So let's just probably out of China. China is definitely the leader in robotics right now for sure. We've talked about that in past episodes too.

Oh yes.

All right, let's go ahead and take a look.

Among all athletic scenarios, tennis is one of the most challenging tasks for humanoid robots. In fast-paced rallies, robots must respond to unpredictable landing points and constantly changing ball trajectories. GalBot introduces the world's first real-time whole-body planning and control algorithm. For athletic humanoid tennis, it enables humanoid robots to perceive dynamic inputs.

I think the human's going easy on the robot.

Autonomous decisions within milliseconds.

Still cool.

Making a fundamental transition from mechanically imitating motions to intelligent decision-driven responses. Our algorithm enables humanoid robots to sustain continuous rallies without interruption, marking a leap from reacting to individual shots to handling long horizon interactive gameplay.

Look at that footwork. Nice.

The learned policy also demonstrates strong robustness, performing reliably be against opponents of different ages and playing styles while maintaining high stability in real-world gameplay.

Oh, good shot.

Interesting.

So Matt, um, I know we're going to start playing pickleball next week, so we need a pickleball robot to practice.

A pickleball robot? Yeah, no, that's, that's pretty cool. Like, um, it's definitely blowing my mind where we're coming with robots. Like I've said this on past episodes about robots too. To me, it feels like the mechanical aspects are there, right? Like it's like these robots, we've seen videos where they're doing backflips and they're doing like dances and they look like ninjas and they can play tennis now. And like the mechanics of these robots, like it feels like we can get them to do any sort of movement you can imagine now. but the brains haven't caught up yet. Like the large language models and the AI models that are baked into these robots aren't quite at that point yet where like it can go and do your plumbing for you or, you know, whatever those physical tasks are that our chart earlier showed you're still kind of safe with, you know?

Yeah. This one. Yeah. The fact that it's, I don't know how, and I was trying to figure out, but I don't know if it's learning in real time. It seems like it's collecting data is what it said of the game and like what's happening. So, you know, it's maybe a step into, cause like the backflips and all the stuff you were talking about was all pre-programmed cause they're like super synchronized.

Yeah. Yeah. That wasn't like an AI model, like thinking, oh, I'm going to go do a backflip and show off right now. It was like, it was either pre-programmed or somebody had a controller behind the scene, you know?

Bingo. And this one's just literally, there we go. All right, let's play a game of tennis and see where this thing goes. Goes.

But yeah, yeah. So, so that's like, that's an example of narrow AI, right? Like, as an example of like, they've trained it specifically on tennis, but if you took that same robot and you tried to get it to play chess with you, it'd probably really suck at that.

Yeah. So keep an eye on the robotics. It's quite interesting. Yes, sir.

Yes, sir.

Good way to wrap this episode.

And, uh, this, this is a good way to wrap this show for a while.

Yes.

For a while. Yeah. So again, as we mentioned in the beginning of this show, this is going to be the last one for a bit. Um, we're going on a bit of an indefinite hiatus and, uh, what better way to show robots than to show robots playing tennis to wrap this up.

Great way to end cap this thing.

And so, you know, everybody who's stuck with us and hung out for this show and tuned into past episodes and nerded out with us for, for the last, however long it's been, year and a half, two years that we've been doing this show. We really, really appreciate you hanging out and learning with us. It's been a super, super fun journey. And, you know, hopefully we will pop up in some other way, shape, or form in the future and continue to do some awesome long-form nerding out with you. But thank you so much to everybody for tuning in and hanging out with us. It's been a fun ride. Joe, thanks for being a great co-host for the last several weeks.

It's been fun. And thank you everyone for watching and having fun with us and this time with us.

So yeah, well, on that note, we're going to go ahead and sign off. Thanks again.

Adios.