Gavin (2:15)

Well, let's. There's two parts to the story. One is, let's talk about what is in this Wall Street Journal story and kind of why it came out. Second of all, I think you and I both kind of feel this very strongly that there's two very different conversations happening here on one is the financial bubble that exists around AI probably the other is the capabilities of these AI tools that our audience and we know are getting better all the time. But first, let's start with the Wall Street Journal story. This is a big kind of exclusive story. It came out this morning. You know, it was big because OpenAI themselves has got a lot of like, people coming out and saying, like, this isn't that big a deal, trust us, blah, blah, blah. The basics here are that they are saying Wall Street Journal is saying that OpenAI's growth has slowed down. And you know as well as I do, every startup is kind of judged on its growth before it gets to the public markets. And one of the things that OpenAI has been trying to do this year is get ready for an ipo. And for those of you who are in our audience who are not like financial prototypes, that means the idea that it is going to sell its stocks to the world at large. We also know that OpenAI has raised more private money than almost any company ever. In fact, that that was like they raised $125 billion before they went to the stock market. So there's a lot of pressure on this particular company to grow and grow and grow. And the only way that this funding makes sense and this kind of represents the overall AI business at large is if this company continues to grow to the size of billions of users in the same way that Facebook is. And the idea that it would be slowing down and that it would not hit a billion users is putting a big kind of like, oh, flag in this for lots of people. There's some important context here. Even in this story. They talk about anthropics growth taking away some of OpenAI's momentum, which we know.

Kevin (3:55)

Right. A lot of other people online.

Gavin (3:56)

Yeah, and a lot of other people online have talked about the idea that, like, look, OpenAI has pivoted now to this kind of new slimmed down, no Sora, more codex, more, more programming, more business oriented approach. So there's a couple things going on, but I do think it's important to hear first your take and then we should try to get into that, like, kind of deeper conversation about what this means for AI at large.

Kevin (4:16)

Well, you every week scream at me, you're either growing or you're showing. And I never understood what that mean. And I think I'm getting what it is now. It means if open AI isn't growing, that they're showing weak. Right.

Gavin (4:28)

And I think this is about that,

Kevin (4:31)

it's fair to say, wow, look, they're

Gavin (4:34)

on the hook for like, much faster than I expected, but keep going.

Kevin (4:37)

They're on the hook for 600 billion, 600 billion in future spending while having only raised 122 billion. Right. And they're going to burn that 122 projected in three years now with Sora gone. Yes. Maybe they get 3.25 years out of it. I don't know. They were bleeding out with Sora. But there's some that would say that OpenAI isn't even really a tech company. Or they're like, in fact, they're like the WeWork of AI. They're like a leveraged infrastructure bet masquerading as an AI company. I don't particularly side with that because a lot of the people screaming that about OpenAI are the same ones making memes about how Dario and Anthropic don't have enough compute to serve the models that they're announcing. So which is it? Did Sam Altman get his greedy, greedy eyes and reach for the infrastructure ring a little too soon, or is he a genius that looked out on the horizon and thought, oh, we're going to need all this. It's going to be like electricity. We have to be able to serve it. Or is it somewhere in between?

Gavin (5:37)

I think there's a really interesting moment we're sitting in right now. And by the way, this, these companies and this idea of AI being this big is affected by more than just like our use of AI, the private funding of AI. Obviously there's things like geopolitical issues, there's all sorts of other things happening. The most important thing right now, I think, is Whether or not OpenAI and Anthropic, the two kind of juggernauts in this space, will continue to prove useful and continue to kind of allow people to do more and more of their actual work with it so that people will pay for it. And I think that's a big question right now. Right. We, you and I have talked about on the show many times. We personally find it useful, although I'm not sure how economically valuable some of the stuff we make with it is. But the idea that business wise, that's a different argument. We're not really the right example that. But business wise, it does feel like a lot of businesses are finding use cases out of this. Now the biggest question I think ultimately will be can that turn that wheel turn fast enough? Right. Because the other thing that's tricky about OpenAI, and I believe that There are lots of naysayers that have been out there and wanting this sort of story and this sort of thing to come true for a while is that there are so much, there's so much money now tied up in all of these giant tech companies in this AI build out that if this does fail in some form or another or it does start to slow down, there is a little bit of a house of cards effect that could affect everything. A little bit, Yeah, a little bit, yeah. I mean, Oracle is the biggest one, right?

Kevin (7:03)

Well, here's like you've seen the memes and I think we've discussed them on the show of the circular trading that happens between these companies, right. Oracle writes a big check to OpenAI to serve their models. OpenAI writes a check to Nvidia, Nvidia writes a check to Oracle. Like it's like six companies all handing money around. And not only are the retail investors like highly leveraged in all of this, if you throw your money into the S&P 500 or anything like that, but your papa or your grandpapa, if you're one of the younger listeners, their pension is tied up in all this. So there is like, there is a bit of a. Oh God, what's the shape, Gavin? Where it's like a three. It's like three sides. It's like a triangle. Gets really tall and there's a few companies and then everybody else goes back to that.

Gavin (7:44)

A pyramid scheme.

Kevin (7:45)

No, it's right, you know, it's like a. More like a circular, like a rattler thing and it's like the snake eating its own tail.

Gavin (7:50)

Yeah,

Kevin (7:52)

like there's a little bit of that going on. But I do just want to say, like, for all those saying that like this particular story and these, you know, like the six or eight companies, if you will, that are really pushing all this stuff, that them slowing down at all is a sign that the bubble is bursting. I don't think that's going to be the first sign. I think it's the million plus companies that you and I have talked about time and time again that went out and raised $30 million because they put a skin on one of these companies and released an app or claim that they were an enterprise solution for something. I think there's a lot of the dam is going to break there. That's where the holes are going to be first before like OpenAI's slightly slowing growth right now is, means that this bubble half burst.

Gavin (8:32)

Well, and also the number I want to point out here is the three year Number which, you know, for a normal person, a normal business might be like not that long at all, but three years in the AI space is a very long time. And I guess the question I continually come back to, and we're going to get into like, why, I don't think either of us believes this affects like AI, the actual technology at this point. But the thing I just keep on coming back to is this idea that who knows what this stuff will do three years from now. Like, I think a year from now the vast majority of coders would have said there's no way I'm going to end up doing most of my coding with AI. And now I think a lot of them, I would say a majority at this point now are coding mostly with AI. Right. And that is like one year ago. So imagine a year, two years, three years from now when these things get way more capable. And I do think they're going to. And maybe that's a good transition, Kevin, to kind of think about this separation of these conversations. Right, Because I think a lot of people in the mainstream continually say like, AI is going to fail, it's going to fail. And that yes, if, if the money side of AI fails, it will be harder for these companies to kind of make the bigger and newer models. But again, we just said it's not going to fail money wise. Even if OpenAI doesn't make more than they're making right now, they have three years of Runway left. That is a lot of time for these things to get better over time. And you and I have seen in the last three months the acceleration of these things just go bonkers. Right, Right.

Kevin (9:54)

Yeah. I would say though, I mean, two, two real shiny, blinking red weak spots. If you're a, you know, a final boss PlayStation gamer, like the weakest points that are blinking right now is one is like, as we know, a lot of times this infrastructure is a massively depreciating asset. You spend billions of dollars to have all these servers and these high end chips and then the next year Nvidia releases something new and now your chips are half as capable, let's say. So there's, there's that massive unknown which is happening right now. And the other one is this open source thing that we keep talking about where it seem every few months there's a new model that gets released, usually by China, that is a fraction of the cost and it's almost as capable as the best things that these companies which are raising billions of dollars. Yes. Are churning out. So yes, three years is a long time for OpenAI to figure it out. They could also like cut their costs and do ads and do a whole bunch more to boost their revenue.

Gavin (10:46)

Sure.

Kevin (10:47)

But these depreciating data centers, the moment they're fired up, coupled with open source, oh my God, it is a bubble. Gavin, we have to get out.

Gavin (10:54)

I have two, I have two counterpoints. I have to. Please hit counterpoint number one. Recently there's been a lot of talk and semi analysis put out something about this, about the idea that older chips can actually serve the newer models better like than they thought they could. So like even though those chips might depreciate, they will still be able to.

Kevin (11:13)

They're still very capable.

Gavin (11:14)

They're still very capable. And they may not be as capable as the brand new chips, but they're as capable. In fact, they're better serving the new models than they were serving the old models. That's one thing. The second thing, Kevin, is that so the new Deepseek did come out and if you didn't hear about that, I don't blame you. It was because it didn't make the impact that the last one did because it was actually about kind of a generation behind right now. It's equal to around opus 4, 5, maybe 4, 6 level in that kind of era. And what's interesting about open source is open source might just be like a distillation so far of like what these frontier models are. I think to your open source point though, the bigger question is at some point you run out of the ability to want to use this like super high end model. Right? Because right now I would say there's a lot of stuff that we can be doing with the current models that we're not doing. Right? So maybe there is a world in a year or two where every open source model on the market is more than enough to serve every person's needs. And maybe that also makes the inference cheaper for these companies too. Right? So there's all sorts of things that can happen from here on out. I do want to shout out, I tweeted something about this and I Quindla is a good friend of ours in the open source world who runs a company called Daily and then also helped make pipecat the open source AI audio software, actually replied to me and said something really interesting, put the whole script thing on screen here. But his basic take here is just that he thinks that LLMs and AI at large are going to be bigger and more important than the Internet and that we're just kind of now scratching the surface as to what the use cases are. And I think this was in response to me asking about like are. Is. Is this whole kind of bubble conversation around the fact that whether AIs are as big as the Internet or they may be something less. And Quinla's argument, which I think is a good argument, is that actually they might be a lot more. And it's just a matter of finding ways to integrate all of these things and new capabilities into the sorts of stuff we do. And obviously Quinn does a lot of stuff with voice, but like it does feel like that's what we have this kind of moment of time right now where like this new magical thing has come into our lives and we have to figure out like, well, what the hell do we even do with this, right? Like where do we put this and what does it go into?

Kevin (13:20)

There was a time in my life, Gavin, and then I'm going to assume in yours as well.

Gavin (13:25)

Where

Kevin (13:27)

I remember it was a transition. Sorry for the, for the old heads out there. Stay with me. For the youngins, just try to follow along. There was a thing called a486Gavin, and it was a processor. And there were many before it, a whole lot after it. But around the time that the 486 became the Pentium, intel had this crazy new chip, all these new capabilities. Then they were plugging RAM into these computers. The amounts were insane. The hard drives went up from four gigs to like 616 gigs. The numbers got bigger. But something really interesting happened which is where you could offer somebody a $6,000 high end Pentium rig with all the RAM in the world and blah, blah, blah, blah, blah. If it didn't have a modem in it and it couldn't connect to something called the Internet. I would take the 486 with a monochrome screen and barely any ramp because that was clearly it. And I wonder if we're going to have that moment. I don't wonder. I have my opinion. I want yours though. I wonder if we're going to have that moment where we, we look at intelligence similarly. I'm not going to be interested in a device if it either doesn't natively have some sort of AI, some intelligence built in, or if it isn't readily easily accessible by AI models to be interacted with or, you know, in some capacity. Do you think we're going to see it become that, that present?

Gavin (14:40)

I think, I think that's true. I honestly think that that's the thing that will prove the use case of this and what's interesting is like, on the other hand, it also might be where AI becomes the commodity, right? Where it's just in everything. And then what AI is special versus not special? I don't know. And that becomes another question of, like, are these companies worth as much as they're worth? There's no shortage of major things happening, especially if you as a viewer at home push that like and subscribe button. Because you know what? The amazing stuff all begins here. We are the origin point, the big bang of everything interesting comes out of us. Our two little heads explode out. Thank you.

Kevin (15:15)

And I do want to say, like, someone did say that we are the most impactful podcast ever, ever released in the history of ever. Others said that we are the tastemakers for the community, that without us, AI would be a bubble that bursts. We're the only ones, like, holding the door, so to speak. And I just want to say thank you and I agree. So please, like subscribe. Leave a comment down below. Juice that algo. And if you want, you can back us on Patreon. Go to aiforhumans show. That's our official website. You can sign up for the newsletter, which still drops twice a week. And you can buy us a coffee. How good is that? We like to stay caffeinated.

Gavin (15:44)

Sip SIP newsletter drops once a week. Unfortunately, the twice a week just couldn't, couldn't, couldn't keep going.

Kevin (15:49)

Well, here's the thing. I get it on two different inbox inboxes. Gavin. I only check one of those later in the week. So it's really like a second treat for me.

Gavin (15:56)

You can do that too. Let's talk about all the cool stuff that people are actually using AI to make. And Kevin, I do want to start with this clip that you sent me, which is just a perfect example of how money should be spent in AI. This is putting Tom Cruise into famous movies running in them. Tell me a little bit about what this is and who made it.

Kevin (16:13)

There is a YouTube channel called Alternate Alternate reality movies. There's still a sleeper channel. So shout out to them, give them a sub. They've got a bunch of like kick ass cameos starring Chuck Norris where he just shows up unannounced in a bunch of movies. They're mashing up different genres, but this one was very simply Tom Cruise runs Faster. And it is clips from everything from, you know, Forrest Gump to Back to the Future where Tom Cruise shows up unannounced. He'll even face palm a T1000 and run faster. He runs faster Than buses in movies. He'll run faster than Arnold Schwarzenegger on a horse. It doesn't matter. It's just Tom Cruise bolting at high speed and it even kind of has the Tom Cruise run, which is.

Gavin (16:57)

Yeah, it's unbelievable. And I think it goes to point like, you know, obviously you can't do this with a lot of current AI video models. This is probably being done with open source tools. So it just goes to our open source point before this idea that you can face swap in Tom Cruise and make this stuff stuff. It's just a very, very fun use case of technology.

Kevin (17:15)

Watching him throw the boys from Stand By Me away or like hip checking Sylvester Stallone as he's running in Rocky to go. This is just so good.

Gavin (17:24)

Tom Cruise has become like the AI of Hollywood in a lot of ways. He's just going to be all actors in the future. Kevin, we should keep going here. There is a cool voice interaction from OpenAI that they're calling Chappie, which I kind of found interesting. The coolest thing about this is that it actually allows you to use real time voice, OpenAI's real time voice to fill out forms, which is a real good use case of OpenAI's technology here.

Kevin (17:48)

This is like okay to, to the thing that I said earlier, not to pat myself on the back, Gavin, but this, what's really cool about this demo is that it's someone using their voice to interface with like a web browser. But it's not doing the traditional agent thing where the AI takes a screenshot or, or scrapes the DOM of the page or whatever else. It's like the, the, the, the, the tech stack gives the agent direct access and information about what's on the page.

Gavin (18:13)

Right.

Kevin (18:14)

So filling out the form like you said, or engaging in dark mode versus light mode. This is to me like the signal of like.

Gavin (18:19)

Oh right.

Kevin (18:20)

I like. There were rumors of the open AI wearable. The big device with Johnny I've is going to be.

Gavin (18:25)

This would be a use case. This would be an actual.

Kevin (18:27)

Yes case.

Gavin (18:28)

Yes. Do you want to play this in audio? It's got an audio demo. Let's just take a listen to a brief moment of this.

Demo Voice (18:33)

In this demo, I've even added a cursor just to show the user what's happening on the screen. Let's start with a practical example here. I've defined a simple application with a wake word and the ability to go from light mode to dark mode. Hey Chappy, can you go to dark mode?

Kevin (18:47)

It's done so it just.

Gavin (18:48)

Yeah, so it just seems to be.

Kevin (18:49)

I had to say it's done because it happened so quick. Go ahead, Gav.

Gavin (18:52)

Yeah, no, and what it's interesting here is you're. We realize that the thing's not talking back. This is just you saying something and then it does the stuff automatically within the actual application.

Demo Voice (19:02)

Can we do night to F3?

Kevin (19:06)

This is a chess demo. The piece moved already, by the way. Like that, quickly. The piece moved because the AI didn't have to read the screen. See what he's talking about, who's trying to. Whatever. It just knows what's happening on the, on the page. And so when you extrapolate this to all of your apps across your entire ecosystem, suddenly a device where you don't have to go and hunt for the thing that you want to load to then ask it to do something, you just tell the AI and it inherently has access to all those things and can manipulate it. I do get excited for that device.

Gavin (19:35)

Yeah. I mean, it goes to show you how important really good computer use is. Right? Because when you can plug a very smart agent into a very good computer use, suddenly you're able to do all this stuff. It also made me think of all of the people who are paralyzed or who have disabilities, like the ability for them to. Instead of having to like have a speech to text and try to find the actual, actual form. That's a massive level up too. Again, this is just a cool thing that came out from OpenAI's developer channel. And like, this is the future of it all.

Kevin (20:04)

That is huge. When you talk about, like the bcis or brain computer interfaces, something like a neural link is what comes to mind, no pun intended, for a lot of people. Usually that is think about where you need the mouse to go. Think about the click you need it to make. Think like that will be gone. That's. It's a huge, huge boon for accessibility.

Gavin (20:23)

You're totally right. Yeah. Another really cool thing that just came out is Blender is now being plugged directly into Claude. This is a direct combination of within cloud itself, where you can talk to Blender. If you're not familiar, Blender is an open source 3D modeling tool that I have actually spent a little bit of time lately with. And guess what, Kevin Claude does Blender really well. Like I made. Oh, does it kind of raccoon gangster thing sort of scenario, which is very cool. We're going to talk about that a little bit more on Friday's show. But it is a very cool integration and Again, this is another example of a very complicated system. Blender is a nightmare. I tried to learn it myself. I shouldn't say it's a nightmare. I am. I'm just not built to learn complicated systems in that particular way. But Claude is a translator in this way. Like it translates your ideas directly into Blender and it gives you 3D models and it's not really difficult to do.

Kevin (21:09)

So.

Gavin (21:09)

I was also very excited about this.

Kevin (21:11)

This is like a. I won't get too super weedsy about it, it literally just came out, but. Nemotron 3 nano omni. Oh boy, that's a fun mouthful. But Nvidia released this. It's a 30 billion parameter model with a decent context window, so it can, it can cram a lot. But what I really think is interesting about this is that it is an Omni model, which means it's one tiny model. Think of this running on your phone or running on a small robot in the near future. And it natively handles audio, video and text. So it does not need to call a tool or connect to something else to understand what it is seeing or what it is hearing. It can process that natively and across modalities. So you can feed it a video, ask a question, and it could answer you with audio. Like there's. It's a really, really interesting space. I love, like Nvidia has a vested interest in powering these things because again, future of robots, they're making things that plug directly into robots. I just thought it was interesting. Perhaps not as interesting as the talkie LLM though.

Gavin (22:10)

Well, this is the thing, like all those things we just mentioned are things that are important to understand about, like how AI is, is advancing in different pathways. There's AI video or there's other stuff. AI is also going backwards, Kevin. And this is actually a very interesting scientific project, what this is. First of all, it was created by Alec Rad and David. Sorry, David, I'm going to butcher your last name.

Kevin (22:31)

Duchovny.

Gavin (22:32)

Not David Duchovny. A very smart man. David Devunod. Sorry, I did butcher it. And Alec Rad, if you're not familiar, is actually the person that worked at OpenAI and kind of was the engineer behind the first iteration of ChatGPT, GPT2 and other ones like that. This is a model that essentially is only trained on text from 1931 or earlier. So what you're getting out of this is an LLM that is trained to answer, but nothing past 1931. And what's really fascinating about this is it just gives you a sense that like these LLMs work on being trained on all this data. But we do have a lot of data pre1931 and it speaks like an older a person from that era. And like when they ask a question of like who's the president right now? It thinks of somebody in 1931. And it's just another good example of playing with the idea of what AIs are capable of and then like maybe what smaller focused data sets could make AIs into. Right. Because to me I immediately think about like character stories, right? Or like oh, what would it be like to do a, have this AI write something versus an AI that we have right now Write something in general. I think it's a very cool idea.

Kevin (23:42)

Yeah. And the important distinction again is that it was trained off of that data exclusively. So it's not like a newer model that's being told role play like you're an old timey something. A lot of that data came from patents and came from documentation that had to be scanned. So there was like a lot of OCR involved because there was no yield Reddit for it to scrape. So I think that's cool. I think having a model like that and seeing if it can like investigate or, or invent new science, like modern science. You know, how would it react to being told that there is a device called a pager or a cellular phone or even a phone. How would it react to that? And then it got me thinking like what sort of, what sort of other LLMs would you want it to train exclusively on? Like, like I was thinking like a, like a, a guardians LLM trained only on Groot and it can only answer with Groot would be interesting to me.

Gavin (24:33)

That's important. You could fake that one pretty simply because it would just.

Kevin (24:35)

What about Fruit B movie LLM or Cookie Monster?

Gavin (24:39)

Well, see, this is what I'm getting at is this idea that you could create an LLM that would essentially be a much more diverse, much more interesting like very specific character. And you can think in the future of if you have, if everybody has this collection of AI agents, if one of your AI agents is like B movie hero and it's only B movie dialogue and about what a B movie hero would gonna do. That would be a lot of fun, right? And like maybe useful for something maybe that maybe you give that B movie hero whatever the most recent technology development is and they find a way to kind of open it up to a whole different world. But I don't know this is just a cool way of looking at.

Kevin (25:13)

Literally, this makes me realize that we're never going to be responsible for revenue at any of these companies. And I love it.

Gavin (25:18)

No one should. No one should ever give us responsibility for revenue, except for you when you go to our Patreon. Thank you, everybody. We will see you on Friday. Great show. And we'll talk to you soon. Bye. Bye.