AI FOR HUMANS 103

[00:00:00]

Chat. GPT-4 Oh. Image Gen has melted the servers of the entire internet. Oh, no. Well, maybe OpenAI can wipe their tears with some of their $40 billion cash infusion. We'll get into how four oh image Gen has changed a ton of industries in just a week. Advertising, movie generation family photos and talk about how you, the creative human, have just leveled up significantly.

Yeah. That includes someone who isn't me, that makes a lot of Shrek feet picks guilty, Kevin. Guilty. And with new tools like Runway's, gen Four and Higgs Field's new AI video model, it's easier than ever to bring your ideas to reality. If you wanna make big booty bears into a prestige HBO drama, you can This week on Big Bear drama, I.

Big booty bears. Two. The clapping who killed Left Cheek, Kevin, who killed Left Cheek that on next week's episode, I just wanna be clear, we are professionals. We are adults and professionals. Yeah. [00:01:00] Plus Meta's, new thousand dollars AR glasses, a new lip sync tool from by dance that we love and Anthropic finally sees inside the black box that is ai.

Words go in and words come out, but we don't know why. It said what it said. Uh, chat, GBT 4.5 just passed the touring test, meaning it can pass off as a human. Am I human? Kevin? I said we're professionals. Gavin, please stop asking that. I don't know if you're a human, but I do know that this is ai. For humans, AI for humans, this AI is AI for ai, human for human.

Gavin: All right, Kevin. There was a big follow up to our story last week chat. GPT-4 oh Image Gen obviously launched. And it has done pretty well, I think in fact, so well, that Sam Altman over the weekend said, please chill, do not keep generating images, essentially. And [00:02:00] then he actually said yesterday that, uh, chat GPT Image Gen has done so well, but it has pushed their GPO so far that it might delay some of their upcoming releases.

I think you'd have to say this is a pretty big success so far. We're gonna get into other things we've seen people do with it besides the ification of the internet, which we discussed last week. But Kevin, I do wanna talk first and foremost about like what kind of a major moment this is in the AI space.

I think we knew this was a big deal last week, but here we are a week later and it feels almost bigger, weirdly.

kevin: To put it into context, Sam had a tweet where he mentioned the chat, GPT launch, 26 months ago, quote, was one of the craziest viral moments I'd ever seen. They added 1 million users in five days. That's

Gavin: That's a lot. That's a

kevin: And it,

and it lives in the icy shadow of the most recent accomplishment that is a million users in one hour.

And that's what this new image generation feature has onboarded for open ai. And many have said that the next wave of [00:03:00] mass adoption is probably gonna come in the shape of a toy.

It's not gonna be some crazy agentic, something that can go off and organize your, your spreadsheets in your sock drawer. It's gonna be something fun, it's gonna be something amusing. And I think like this is the perfect combination of the tech being really exciting and capable and OpenAI relaxing those guardrails enough to let people

Gavin: go crazy.

Kevin Weill, who's now the Chief Product Officer of Chat. T and OpenAI said specifically he'd never seen anything like this before, even at Twitter and Instagram, both startups that he was at early on when they were blowing up.

And the fact that they added a million users an hour is an insane number.

Gavin: But this also kind of goes hand in hand with two financial stories. One, there was a story in the information I think just yesterday, which talked about they added 30% more users in the last three months, which is a huge number when you think about it.

But then finally, Kevin, the long rumored funding has come through. They have cashed the check. They have raised a new $40 billion in cash. That's the actual [00:04:00] amount of money they're raising at a $300 billion valuation, which is by far the biggest and and largest AI company. I think it probably is pushing the largest startup raise maybe of all time.

I don't know, but that feels pretty significant as well.

kevin: I am just still stuck on the imagery of Sam Altman opening the trench coat to offer the kids a little, do you wanna try limit gen? You wanna get in and make a couple little, little, little gif somebody?

Gavin: is all you need to pay for unlimited

kevin: 20 little bones. Yeah. Look, this is anecdotal, but I had a lot of people outside of my usual AI bubble pinging me going, what is this?

What, how did someone make this? Et cetera, et cetera. And we'll get to another moment later in the show. Same thing's happening with runway Gen four as well. So these tools are progressing, they're advancing, they are getting into the hands and hearts of people and they will be nerfed. This

is something that you and I brought up very quickly.

Last week. We only had our hands on the product for a few minutes and we're like. There's no way this survives. [00:05:00] People flocked to try these tools. I guess in large part because they could emulate the stylings of very, very, uh, famous artists, or they could potentially trample upon the IP rights of, of companies, and the clamp down has already begun.

Gavin: Yeah. And just so everybody out there understands, because somebody had asked me the other day, what are you saying? What does Nerf mean? And probably most of our listeners or viewers know what Nerf is, but Nerf is a term mostly that comes from the video game space where a new weapon or a character is really like underpowered or, or overpowered and makes the game broken in some way.

In this instance, people are using Nerf, meaning that you will no longer be able to get the same things out of the model that you got a couple days ago. In fact, I had a direct experience with this on Saturday. I attempted to use, I took an old shot of the GI Joe, um, cartoon, which I grew up loving, and I said, make these people realistic.

And it gave it to me. I was like, this is amazing. And then I made a video out of it and it was very fun. They all clap in unison. Did you see that video I sent you? Yeah. [00:06:00] Uh, very funny. Two days later, Sunday afternoon, I tried to do the exact same prompt with the exact same image and it didn't allow me, it said, this may violate our copyright.

So it is a clear understanding that they have completely changed what's allowed. Now. Some people are saying on SOA you can do different things than you can do in chat pt. And I think that is true in some ways. One thing they haven't really changed is the ability to, use celebrity images. And I'm wondering if, because Grok has allowed that in the past, maybe there's a world where they're allowing that instead.

kevin: I saw somebody on the team talking about that specifically, like on a ex post, and they basically said they're trusting their users and they shouldn't be the arbiters over who is a public figure enough. They don't want that responsibility, which is a nice hand wavy thing of have at it, but they're doing an opt out list. So you can, as a, as a brand or as a personality, go it, but it's the onus is on you to go out of your way to opt out. But Gavin, don't worry, I already opt both [00:07:00] of us out, so several

Gavin: opt me in. I don't care. Opt me in. I don't care.

kevin: Oh, wait, you want it the other way?

Gavin: I don't, I really don't really, I don't care, to be honest with you. I, I honestly believe eventually this will all come anyway. But to this point, if you go to the Sora homepage right now and the Sora homepage, there are a ton of AI images.

I saw one that was like breaking back and it was like both of, uh, the two guys from breaking a bad with their backs, bent over in pain. Just dumb jokes like that. There was a really interesting one that I had seen something similar of before. I was on this sore homepage and I saw this amazing image of Sian Murphy and Peaky Blinders. Peaky Blinders is an all time show. If you haven't seen it on Netflix, go watch it now. It's great up there with the best of the best. And he was in like a frat basement, you know, drinking a bottle of champagne.

And now I thought, oh, this is really funny and interesting. So I wouldn't tried it. Yeah. So pretty much I got, you know, a couple different versions. I got an image of Walton Goggins and Sam Rockwell, who are both starring in the White Lotus. I took an image of those two guys partying together. I got Stephen A.

Smith and LeBron James to party together. So you can definitely [00:08:00] do this still. Now, I don't know how long that will last, but for some reason Soar is allowing this, I, I think this is gonna be a weird conversation. There was a tweet when this model first came out from Joanne Jang, who is the head of product and model behavior at OpenAI, where she kind of discussed their policy on what they were gonna allow to go and what they weren't. And then she readdressed this, uh, yesterday and she said, we hear you on the current state of refusals not being up so far.

I'd missed the transition discussion below. Favor, can you please reply with screenshots of image gen requests that shouldn't have been refused, but did. So they are listening and in fact, specifically she references the model should definitely be able to make the gnomes butt bigger. And this was an example that was going around the internet of a gnome with kind of a bare butt.

The user wanted to make the butt the butt bigger. And obviously this is one area where they're like trying to be a little bit more lax hit the butt cutoff. That is a known quantity. Like when does it, it's like the bartender, like, sorry, this gnomes butt is big enough. I cannot ambi it any larger. I

this is going to be an ongoing conversation. [00:09:00] My question to you, Kevin, is do we think this would've been nearly as big a moment had they not had that like kind of finger off the off the scale? No, exactly. So this is where it feels

kevin: last week. like, thi this is, this is, uh, you know, the Chinese models catching up and smaller, you know, companies taking larger risks because they feel like they can, they absolutely had to take the thumb off the sort of judgmental scale here and let people go wild.

And I think it also signaled probably to some of those investors waiting in the wing before they could get their $40 billion infusion. That, hey, um, look at what we can do,

look at what this thing can actually do. And they've probably been hamstring themselves along the way by guard railing it so much.

But I, I had a disappointing experience with it this morning that has me like, again, concerned for the future of the product

Gavin: Yeah. Well, tell me about it. What happened?

kevin: So there's a meme that I love dearly. Gavin, I don't know if you're aware of it, but it's a beaver in a drive-through, and the beaver is leaning out the window and the the text on the top says, uh, I think I'll have a [00:10:00] log. That's it. The

beaver wants a log from

the drive-thru. It is peak comedy. Of course, the

Gavin: Sounds amazing.

kevin: from the drive through. I send it to my niece who doesn't appreciate it the same way. You didn't appreciate it. Who's coming to visit for Palm

Gavin: I don't have a personal background with it, Kevin. I'm sure there's some moment where you discovered it, where something

kevin: you saw it, if you saw this beaver leaving, leaving out the window, ordering a log, you would get it to, but I, I am now trying to create a series of beaver images where the beaver's packing a suitcase full of logs and making a list of what it needs for the spring break trip. It's all a log.

I tried to get the beaver going through a security checkpoint at an airport.

Uh, putting their logs through an X-ray machine. And at first chat, GPT told me there must be a problem with their servers. The

image couldn't be made. And that was this morning and I was like, oh, wow. I guess the GPUs are melting. So I tried again and it said, uh, again, we're having a problem generating this image.

And I was like, well, what's the problem? They're like, I think it's a problem with the [00:11:00] back end. I'm like, okay, well try again. It, try it again.

And then it finally admitted that there's something wrong with the prompt.

And I said, well, what is it? And it says, image couldn't be generated due to an error. If you want a different angle or revise the scene, let me know. And I'm like, well, that's. That's not what I want. This is exactly what I want. It says, it says A combination of elements in the prompt triggering a content filter or misinterpretation on the backend

airport security, x-ray machines and animals with objects, even logs can sometimes get flagged, especially when combined in unusual or comedic context. So then it goes on to say, maybe you should try this, this version or this version.

I said, go ahead and do it. It tried, it still failed. Finally, I tried to get the beaver putting logs into the overhead bin on the flight and it shot back and said, uh, sorry, can't do this as well. A perfectly reasonable, ridiculous behavior going on a trip with a suitcase full of logs, just trying to travel like anybody else.

This makes [00:12:00] sense, but I don't think we should do it. Perhaps we should try the beaver with the logs, uh, already on its lap in the plane.

It's like that was a suggestion. So my response, and I'm sorry this is wordy, but it's so weird to me that I had to have an argument with the machine. I responded with it, uh, hey, you are the machine massage the prompt, however necessary to gimme the image.

I said, it's absurd that logs going into an overhead bin versus already on the lap while he's seated on a flight, would make any difference in this scenario. And it gave me the image

Gavin: Oh, wow.

kevin: the logs in the overhead. So, so that to me was even worse because it just reinforced that I might have to fight or argue with the machine to give it permission to make the thing, , I'd rather you consistently say, Nope, I can't do it. I'm not gonna do it. Versus like, okay, you twisted my arm enough and spent the time. Here's your dumb image.

Gavin: I will say the image is pretty

kevin: It's A great image.

Right?

Gavin: We'll look at it on

kevin: a great image.

Gavin: a pretty good image that [00:13:00] that is exactly the experience, Kevin, that I have had multiple times now, and it goes from being a magic product to a frustrating product, and I think this is what OpenAI really has to keep in mind.

kevin: You open the floodgates to something and then when you start restricting it. You really are selling something to people that is not possible.

Obviously we need protections for IP for these various rights holders, for even artistic styles

kevin: potentially. And then others will say, well, no, like a style can be described and people can emulate a style, and maybe you wanna put your own spin on it. So do you think we're very quickly heading into a world where, uh, this is the new normal and anybody can remix anything, and that's just the way it's going to be and that's the way it potentially should be?

Gavin: Yeah, so I wrote about this last week in our newsletter. If you haven't checked it out, that's at, uh, AI For Humans Show our website. You can find our newsletter. And I think specifically there's two things that I feel really strongly about here. One, I [00:14:00] don't think styles should be copyrightable in a weird way because if they are, then what the whole world of what AI is doing is probably illegal and we're gonna be kind of messed up in a big way. Because if you can copyright a style, there's gonna be a whole lot of text versions of that too, to be aware of. I don't know the answer about the IP of it all, because there is a level of protection that a lot of companies deserve for ip.

If you've built something up from scratch, if you've created it, you may want the ability to decide whether or not that's being used in all sorts of different ways, including ways that might make your IP seem not great. Right? So I think that's an important question. I think the bigger thing here is. We have now very clearly seen what is possible with ai, uh, unlocked.

And it has been put behind closed doors again. And ultimately to you, as you always say, we will get open source tools that will be unlocked. I just think it's going to be harder to control that stuff.

kevin: Copyright infringement and dank memes aside the image model people are discovering

Gavin: a whole lot of other stuff. Yes. Yeah.

kevin: crazy, uh, [00:15:00] dear, I'm gonna say dear friend to the show, dear friend of ours, cocktail Peanut, the one man band behind, uh, Pinocchio, one of our favorite AI apps. They did a, a test, still kind of blows my mind how good it is. Like

there's a lot going on with this model. Basically, they had it generate an image of a cup that's filled with oil and water, and guess what, Gavin, the oil mixed with the water properly sitting on top

of, of the water, right? A couple little globules going down below.

Then he added some boba and it understood the weight of the boba sinking to the bottom of the cup. So now you have boba water, oil. Now I feel like I'm doing the man camera. Trump thing. Uh, and then he added red food syrup to it.

The way this thing is intelligently mixing, uh, the physics of all these different liquids and solids within the cup, that is actually the most impressive thing to me.

Including what would the cup look like if I left it outside during July and the cup started to grow mold? Like that's really intelligent.

Gavin: Another really [00:16:00] cool one I saw was from a, a very long name. I'm not gonna try to pronounce, but we'll put it on the screen here and in the, in the show notes. What they did is they took a three layer parallax background that they created for a vibe coated game.

And if, you know, when you're playing like a pixel game, you see kind of the backgrounds going in different directions. They used it to create all three layers and then they split 'em up and they dropped it in. Like those are the sorts of things that's now possible in almost, you know, often one shot reactions to things.

kevin: It's really incredible. But then we can't forget that they have, uh, other models and other aspects of the business as well. In fact, GPT-4 0.5, as we teased at the top of the show, just passed the touring test.

Gavin: , if you're not familiar with what the Turing test was, this is a very long ago test that Alan Turing, one of the pioneers of computers and AI created, which was basically, people will talk to a computer and talk to a person, and that person has to decide which is which. Or if they get fooled by one, that means they would pass the Turing test.

Cameron Jones on X, who works at the cognition and Language lab at [00:17:00] uc, San Diego posted on X yesterday, which i's like, oh, maybe this is a April Fool's joke. It's not. That they did a three party Turing test in which you can actually try yourself.

And of all the major models, they tried 4.0, they tried 4.5 and they tried Lama, it didn't seem like they put in the Claude models here. But 4.5 passes it with a 73% win rate, which is a clear, , number that will put you past the Turing test. And this has been one of those things that people have talked about forever is like, when's AI gonna pass Turing test?

And Kevin, the thing that I found most interesting about this, it was like nobody gave a crap. Do you know what I mean? Like this story did not blow up in some sort of way that I thought it was going to. And, and maybe it's 'cause we just have different definitions for what AI are and now we're thinking about a GI and what it can do.

But this is one of those things that we shouldn't just let go by because it kind of is a big moment. And people have said 4.5 is really good at creative and doing that sort of thing. So to me this felt significant.

kevin: Sam Altman has said in interviews in the past. We will wake up one day and we will have a GI [00:18:00] and then we'll go, okay, and we'll go about our days, right? We'll

go,

Gavin: sound like Pete Davidson from that SNL sketch. The chat, chat, sketch. Okay.

kevin: But I mean, is this not true? Like,

if, if 10 years ago you said, we will wake up in a world where a machine can, uh, convince you, it is human, it will be believable, intelligent, you'll have conversations with it, you will be convinced it's a human. You would be like the, the world will never be

the same. And here we

Gavin: Here we are.

We have three very quick open AI stories we wanna rip through here first and foremost, Sam Altman has also said that they are going to be dropping and open weights model soon. This is something people have asked about for a long time.

Kevin, in a ten second description, do you want to describe what, what open weights are and why they matter?

kevin: A peek behind the curtain, how doth the thing do and why do the thing doth it does.

Gavin: Pretty good. I didn't ask for it in Shakespeare, but you have a future as an LLM. Kevin, I feel like you could be an LLM. Why is this a big deal, do you think?

kevin: I mean, look, OpenAI has been getting no shortage of criticism the [00:19:00] last few years for being fairly closed.

They don't release their research. You got companies like Meta out there or these Chinese models that release the weights along with the model. It lets you see, um, again, why the model is doing what it does.

And it allows you to build applications for it where you don't have to ask permission or pay a company a licensing fee. And so this is OpenAI probably answering some of that criticism. They say it's going to be a very capable model. I don't think it will be state of the state of the art, but um, you know, this is gonna get them back into some good graces, but it will hopefully spur a lot of development.

Gavin: Uh, hold on one second. I have somebody who wants to respond to that. Kevin,

hey. Hey. Monday. What do you think about open source, uh, and open weight models specifically coming out of chat? GPT,

kevin: Don't ask

Gavin: uh, open source models and open weights. Uh, it's like getting the entire neighborhood, the keys to your garage, letting them poke around and see how it all works. [00:20:00] It's supposed to encourage transparency and collaborate.

Hey, can you be a little faster? You seem so depressed. Like what does it sound like when you get more excited?

All right, let's kick it up a notch. Open source models and open weights. Well, we

kevin: Oh, Monday. No, Monday.

Gavin: solved Monday's depression. Anyway, that was Monday. Monday is a new voice in your chat sheet PT app right now. That was,

kevin: you asked for

Gavin: was an, that was an, uh, a semi April fool's joke from opening ai, but it is available.

It's called Monday and it's like a depressing voice, but I did not realize I would be able to switch it that quickly. Anyway, it's another fun voice to play with . Probably based on what we talked about last week.

Their new realtime voice model. Super fun shows you what you can do when you add a personality to a, to an ai.

kevin: Uh, lastly, and not to be overlooked here in the wave of all of the open ai uh, stories. This one I really love and encourage everybody to check out Open AI's Academy. You can go to academy.openai.com. This is hashtag not an ad, but. Boy do I wish it were, and I say that every time. Um, it's free tutorials [00:21:00] for anyone and everyone I'm talking on any level.

They have like video series on like AI for seniors. They've got AI for educators, they've got AI for hardcore engineers. They will walk you through everything from what is a large language model, how does it work, to how can you start programming and fine tuning them. It's a really, really impressive array of tools.

There's also, uh, live seminars which you can sign up for, uh, and take part in. We get asked all the time, where do I start? How do I, how do I dig in specifics about, uh, particular tools and whatnot. This is just a great resource. So kudos to open AI for, for taking a step here.

Gavin: That's right. All right, and Kevin, before we move on, it is important that you take a step to like this YouTube video and also to follow us on YouTube, subscribe to our podcast, all the stuff. If you happen to be here from somewhere else, please share and and join us on a regular basis. We do this show every week in audio and video, and we really do appreciate everybody.

You can drop us a tip at our Patreon, which you're welcome to try as well, but also really important is like sharing, [00:22:00] commenting, liking this video, or sharing the audio with people. We have a really growing audio audience, which is amazing. Every week we hit another big number, which is really exciting.

Thank you for sharing this with people, because this is one of those things where you see it kind of tick up every time, and it's clear that people out there are saying, Hey, go try this, which is very cool.

kevin: Yeah. Thank you. Thank you. We genuinely will not grow this endeavor without your efforts, so it does mean a lot to us. And someone, uh, someone actually messaged me about, uh, the Patreon. They were like, so do you guys actually use the, like, the Patreon dollars for like credits on AI services? yes, yes.

we do.

We totally do. Every time a new silly toy comes out, we feed, we feed some money into the machine to try it out for you. So thank you for all of that. Speaking of new toys, Gavin, uh, one of our favorite video models, runway released Gen four. Uh, it's still fairly hot off the presses, but it is, uh, impressive.

Gavin: Yeah. I love, so I played around with this yesterday, so Gen four is out for all paid customers and we are in the creative program for Jet Growth Runway. Thank you to [00:23:00] runway for that. It's been allowing us to test this stuff is a very cool next generation video product. And the thing, Kevin, that they've really focused on is character consistency.

And I think this is something we're seeing in everybody's seeing that come out of this, To make really awesome AI videos. Character consistency is key. You need to be able to have the same character in every shot and carry through. And it's just a very awesome thing to see them focus on this.

The overall video is better, the generations are better. Um, I did a couple very quick tests, one with my knight that we talked about last week from 4.0. And what was interesting about this video was, look, it's not the most perfect AI video generation in the world because this image prompt is a weird one, but it does keep consistent and you kind of see the camera slide around and the, and the night runaway and the floor is consistent.

The background's consistent. Yes. In the background. It doesn't keep the poster necessarily consistent. You can see some of the, the texts become garbled over the course of it, but the knight himself is very consistent.

kevin: It's starting to, to really come together, the examples of the, uh, like hyperrealistic animals [00:24:00] in human spaces, like watching a herd of elephants go through, you know, Chinatown or Times Square or the, the Zebras at the carnival or the Lions parading through New York like that. Stuff looks legit. I mean, it looks like film. And so when you start combining the stylistic controls with the character consistency, the camera controls, and even their lip sync tools, you very quickly start to see how film is going to be made in the future. At least one particular type of it.

Gavin: Yeah, and it makes me wonder like, you know, we talked about SOA kind of failing a little bit when it comes to what's capable after image gen. We first saw SOA like over a year ago now when it was announced, I think it was February of 2024. We are now in April of 2025 and we have not heard about or seen a so RV two.

I have a feeling that all of these AI video companies are getting ready and whether or not this, this release was kind of rushed off the back of four oh image and it might have been because it does have image to video so that they want people to be able to use it. And that's how I've been using this to be honest with you.

Like [00:25:00] starting from four oh image gen and then importing that still into runway is a really cool, yeah, it's a really cool use case of it. I had one more example I wanted to show off, which is I, I had four oh image gen made this really goofy version of a raccoon raccoon driving, uh, um, a golf cart.

kevin: cart.

Gavin: And, and what was interesting about that is the version that four oh gave me, had the words, the alpha returns on the bottom, which is just a funny thing to drop into a, a thing and it's, and he is in a golf cart that says party animal on it. Like it's clearly he's wearing sunglasses. What I love about that clip, Kevin, is that it actually fades off the title on the screen.

I didn't ask for that, it looks like a documentary and then it fades off. That's something that kind of is really interesting about how the model interprets what's on screen with runway and I think is a very cool, uh, use case Go to runway. We will share their blog post in our show notes.

There's so much here to get into. I think you could easily lose multiple hours into this tool, uh, this

kevin: easily easily for 15 bucks, you get, you know, a certain amount of credits that you can start generating. I will say, and listen, [00:26:00] a a lot of providers have this issue. The user interface is going to seem a little daunting at

first for even a, for even as, as easy to use as the product is.

I had a couple friends that were trying to use it and they all hit me with the exact same issue, which is they didn't know if they were actually using gen four.

And I had to point out, oh, there's a little tick box in the

bottom left hand corner, and you click that. So go look down there, it will tell you what model you use. Adding further confusion. If you drag in an image or a video or whatever and it has a certain resolution, uh, or using an image versus a video, it will sometimes automatically switch the model that you have selected. So you might think. You're generating with Gen four, but you're actually using a Turbo model.

So I think they have some work to do there as all of these AI companies do. Uh, but just keep that in mind. Go, it, it, it's worth the 15 bucks to start experimenting For sure. Just pay attention to which model you're using as you do it.

Gavin: Yeah, and another big AI video model came out of, uh, I guess not a stealth, but a new series of tools released from Higgs Field ai. And this is started by somebody that was at snap. [00:27:00] They had kind of been doing a more traditional like text to video model, kinda a broader one. But what they did is they specifically dialed in on camera shots, which I think is something that a lot of AI filmmakers in there might really appreciate.

They have a whole series of examples of how you can do things like crash zooms or a robo arm shot, or a crane shot over the head. So I can see this in the same way that Pica's kind of specialty has become like those weird Instagram filters. I could see Higgs Field or another AI company starting to specialize in specific shots so that there's a suite of these tools that come together to do what you want.

So say you start with runway or you start with Soro, one of these kind of larger models, and then you specifically dial down using one of these tools that can do this.

kevin: Yeah, I mean, uh, look, there's a future where some percentage of media will be fully AI generated. That

is absolutely on the horizon. But I think in the bridge to that, uh, the most impactful uses, I think will be probably leveraging AI to do those VFX shots or those complicated camera shots as you're [00:28:00] doing traditional filmmaking still.

So it makes sense that something like this exists. What's going

Gavin: I am just laughing about how you love those robots doing the camera shots so clearly, you know, the robots are the ones coming next who are gonna do all the dolly shots and the jibs we're not gonna be doing,

kevin: with you. I never know, but I love it.

Gavin: So one other really giant tool that we've talked about in the show that is now out in the wild and you can use is Omni Human one.

This we've only showed videos of before. It is a lip-syncing tool. And we were in this place, Kevin, where we're like, is this too good to be true? I tell you, it is out there. You can get one free generation right now. You have to go to dreama dot cap cut.com. It is, it is basically cap cuts AI video service from by dance.

I tried this, I wanna show this here. I took a picture of Ray Kurzweil, who I'm a giant fan of Ray Kurzweil, the singularity near is near made a big difference to me. Ray is on a weird press tour right now where nobody is telling him to not wear his toupee and he's looking a little weird, but he is very smart.

Anyway, Kevin, I had him say something specifically about you, so if you can [00:29:00] play that

kevin: Oh, this, this message to, to Kevin from Ray with

Love. Okay. Let me play it Kevin.

Gavin: Ray definitely knows stuff about you.

kevin: Is this about the The Shrek feet pics

because I said that wasn't me. I'm not swamp stomper for 2069. That's not me. That's someone I might know.

Gavin: So Kevin, when you look at that video, I have to say to me, this is by far the best lip sync I've seen yet. So that is a single screen grab that I took off of a video, so it's not even like a high res photo.

And I uploaded an 11 labs clip, which I made. But like you can look at it directly and the way that it talks, the way that it moves, the mouth, all of that is great. Now, there are downsides to this. The number one downside is it's not cheap. I think that you get one free generation, but I [00:30:00] think to pay for more, you're gonna do it in part because it's about a 15 minute generation.

So I assume they're crunching a lot on the backend. But when you think of high end tools, when you think of what somebody's gonna do with this, it's pretty significant.

kevin: I do a lot, a lot with AI avatars. That sounds shadier than it should be. I swear. It's all legitimate above the board work, I promise. But I do a lot of AI avatars and, uh, ps ts, s's,

uh, that's where you see avatars falter a lot. There's a lot of blurring. Uh, you'll see it in the teeth or the lips. Um, sometimes they'll do what looks like a, a like a b with their mouth when it's a P or a T or whatever.

And uh, so I was watching that one very closely, very impressive. And the way that it shows to emphasize the movements of the character, but the words, it was like, okay, it's making very, very intelligent decisions.

Gavin: it knew, it was kind of angry, like it knew that the audio was angry. Right.

kevin: Yes. Now, did you see meta announced mocha,

Gavin: This is very cool. Very

kevin: kind [00:31:00] of got a little bit of its thunder stolen by the

tool, I think, but it is a slightly different tool. So, uh, what Mocha wants to do is, is make it easier to prompt, uh. Entire, I, well I should say, uh, let's say torso up avatars. And it's a model that's trained on both speech and text.

So, uh, when they train their model on the way humans, uh, move, uh, gesture, uh, the way that they move their, their head to, uh, emphasize certain things, they had not only the transcription of what is being said and the audio of that, but they had, uh, a lot of, uh, text labeling of the discreet movements. And so what you get back is believable movement from a single prompt.

As, you know, if a character is shocked, they may put their hands up, uh, you know, closer here. If they're, if they're exclaiming or, or yelling, they may point or gesture in a way that just seems a little bit more believable. They want to close the gap for Hollywood style prompting of these characters.

Gavin: If they can actually solve this, then the world of AI video becomes a thing. The one thing I [00:32:00] keep thinking about, and, and you know, Hera is trying to solve this for, for this too, is like characters that are not facing the camera, characters that are facing to the side.

And this does, at least in the video, do a pretty good job. Meta also though, has a history of showing off these videos kind of like some of these other companies and not shipping a ton of stuff yet. I'm still waiting. I think we're all still waiting for Meta's video model, which I would assume is gonna have all this stuff baked in, but I don't know if we're ever gonna get it.

We'll see. Um, the thing that I do keep thinking about watching all these tools is how important a, I believe a new job will be in the world of AI video filmmaking, and that is an AI Foley artist. Do you know what a Foley artist is? Kev?

kevin: I do, but you should tell the kids at home

Gavin: I think probably most people know a Foley artist. If you've ever seen those videos of people like clumping the shoes and they're walking along and they're

kevin: breaking celery because someone got their neck snapped.

Gavin: Their job for the, for years has been to create the sounds to kind of match what the characters are doing in a live action film.

I think there needs to be people starting to focus on this now because so many AI movies I see. Don't take [00:33:00] sound seriously enough. And sound, I think, is a huge part of the experience of feeling immersed in a movie. And when I say AI, Foley artists, there are a lot of tools out there, 11 labs, a lot of things like that, have AI, sound effects tools, which will allow you to create sound or then bring in music from different AI places.

So I think people who are getting very good at AI sound making will be employed in a very significant way in the future. So if you're out there wondering like, how could I specialize? How could I do something that people would be wanted to wanna use in in the future? This to me feels like a job that a lot of people will spend time getting good at and will be hireable.

kevin: Sound designer is not a new position, but I, I think what you're essentially saying is like, look, leverage the best in class AI tools so that you can be way more efficient with it. And the exponential need as all of these videos come out, in my opinion, slightly lifeless

because there's no audio there. Let's see it. And also to that point, anytime you're doing some sort of AI demo with someone talking in a scene, when you just use the raw sound output

from [00:34:00] any of these tech, it sounds so unbelievable and wonky, an ounce of reverb and EQ will go a long way to

Gavin: Room tone. Put a little room tone in there. Like there's all sorts of things you can do to change it. But anyway, those are amazing new tools that are on the horizon and omni human one you can try right now. Kevin, there's some other very big stories that have come out that we wanna rip through here.

First and foremost, just this morning, breaking Meta's, new glasses have a name. They are not the super advanced glasses, but these are called the Hyper Nova Glasses coming outta Bloomberg. This story is, and it is gonna be a thousand dollars. It is not the most exciting version. You're only gonna get a small little square of a screen in one side of your glasses, but you are gonna be getting the wristband, which we've talked about in the show, which feels like a pretty significant ui.

Step up. Um, are you excited about this? This is like the in-between step between the Ray bands that exist right now and maybe the larger version that's gonna come out a few years down the road.

kevin: I am. I am absolutely excited for this. Genuinely, I'm not super psyched to need a [00:35:00] wristband. I don't want a wearable to have to communicate with another wearable. I think that will come out in the wash eventually. I think Vision Pro nailed the ability to not have to raise your hands in front of your face to manipulate things, I, I'm planning a lot of international travel coming up. If I could just glance at a menu and have it instantly translated or

be interfacing with someone and just see the words that they're saying, like, just give me that alone.

I would spend the thousand dollars

on to have, uh, a, a much safer, more, uh, effective and, and fluid travel experience. Forget the looking at something and asking for assistance with it. Uh, or what is this general searches, or, oh, what's the Yelp rating for this restaurant? Do I wanna walk into it? All that stuff can be there.

You don't need the, all singing, all dancing, 3D dual eye display, a simple little screen in the corner. If the resolution's good enough is gonna go, uh, a, a long way. So I'm very, very excited for it.

Gavin: Kevin, are you familiar with The Jerk? Do you know the movie The Jerk?

Do you remember what happened in that movie? Do you know the Omni Grab section of that movie?[00:36:00]

kevin: No, I don't, I

Gavin: Okay, so this, I I, this is a deep, uh, comedy Cup, but if you've never seen The Jerk with the Steve Martin, he invents in that movie a thing called the Omni Grab, which is this little handle that goes up in the middle of the nose so you can take the sunglasses off.

Um, and what happens to people ultimately is they become cross-eyed and then people die because of this little thing, and then they get sued. Here's my thing about the small.

kevin: I guess, let me, let me re-answer the

question. I don't remember the jerk. I know I've seen the jerk, but

I totally do

Gavin: boo Boo, we'll get into

kevin: No,

Gavin: later. P comments, people call out Kevin for what he needs to see. Anyway, the reason I bring that up is.

I am not convinced that a small corner screen is the thing. I'm not saying it's gonna make people cross eye, but I am not convinced that's what anybody is going to want. And I understand that that is what's possible now, but I, I am not sure that there's a world where I have to look down. I'm trying to like him to imagine, I'm like looking down at something while I'm talking to you here.

I don't know if that's the answer. And I understand they're trying to, they're trying to bridge, I honestly [00:37:00] think the reason why I

kevin: I like that for you looking down is like, you need an EpiPen because there are three Bs stinging you, you had to do

Gavin: I, I'm trying to do one eye this

kevin: it's probably just a glance, it's just a glance like I'm glancing right now and then I

go

Gavin: don't, I don't know, I'm not bought into this idea of a small box, but the risk thing is really interesting to me because I do think like there's a whole new need for how we think about ai and at least the promise of that thing when people tried it last year was that like little tiny movements of your fingers could make significant UI changes.

So that's pretty cool. All right, Kev, the other thing that's important to talk about this week is anthropic, the company behind Claude that we expect is gonna be dropping a new model somewhat soon. They keep, uh, teasing us with hopefully a new opus.

They have actually made a lot of progress on what AI thinks when AI is thinking, and this is something that has plagued a lot of people forever. People have talked about the idea of it being a black box. Well, anthropic has actually dug in. On the idea of how these ais figure things out. And I'm not gonna go into specific detail here on how they've done it.

There's [00:38:00] a very good video series and also a blog post about it. You should at least be aware that we are finally making progress to this space where we understand how ais work. And Google also dropped a big, a big article today about a GI readiness. And part of that is AI safety, understanding how AI think, and understanding how they interact with us.

But what they choose to tell us is a very big step towards AI safety. And you know, sometimes people, Kevin might listen to this show and be like, those guys are just, uh, they don't care about AI safety. They don't give a crap. We care. We don't wanna be, uh, paperclips, we don't wanna be the fodder for terminators.

So we do care, and this is a step in the right direction.

kevin: Do we think it's, uh, uh, futile to resist it? Sure. Absolutely. Are we nihilistic? A hundred

Gavin: Of course,

kevin: care.

Gavin: yes. We care deeply. We care deeply. Speaking of caring, Kevin, should we care about Alexa Plus? Is this a big thing?

kevin: Uh, we're supposed to care. Uh, like I want, I wanna care. I wanna believe okay. [00:39:00] Alexa plus the AI supercharged version of the assistant that many of you have resting in pucks all around your house that remind you of things mostly, I think, and maybe play music. The AI version is here, except it's not fully what we were promised, uh, at all. It's only running on a subset of devices and it doesn't. Recognize individual household members, and it can't order the food on GrubHub and it doesn't have the kid mode. And so we're, this is now becoming a common refrain by big tech as they scramble to integrate AI into all of the things.

There's a bit of a, uh, an over promising, which I think comes from the, the amount of stuff you can do with ai. It's not for a lack of vision. I think it's just the reality of integrating these kind of messy, sometimes sloppy systems into a fleet of millions of devices with different hardware specs.

It's not an easy task, but they don't make it easier on themselves. Whether it's Google, Amazon, or Apple, they're not doing themselves any favors by [00:40:00] doing these big events where they showcase all the

stuff and then just start delivering a drip, a drip, a drip. I get why companies do it. Um, but this is another one of those cases where it's like, uh, people were really, really excited to get their hands on it. And as it's rolling out, people are. A little less excited. Did you try it with any of your

Gavin: No. What's so funny is I didn't even know this happened until you dropped this into the, into the rundown today. I didn't know this existed, so I will go try it today. One of the bummers is, it's only on the show devices, the devices that have screens on them for now. And as we talked about last time, we talked about Alexa.

Plus, what's most exciting to me is that it would be accessible to all of my devices. And again, I don't know why that wouldn't be the case. They're running all this stuff in the cloud, they're not running it locally. So the voice stuff is way more interesting to me on the Alexa than anything video wise would be.

So to me, this is a little bit of a semial, but I will try it today and I'll, I'll report back next

kevin: Oh, here's something you should do, Gavin, because Amazon plus can summarize uploaded documents at launch. You should [00:41:00] put all of your important and private

documents in there. Uh, one little note though. I just wanna do, I do wanna point out they're still working on the ability to. Delete those files

after they've been uploaded.

But if you want to remove those files, you can always contact customer service. Gavin.

Gavin: Oh wow. Well, guess what, Kevin? If customer service gets contacted by me first and foremost, they're gonna see things that people did on the, uh, did on the internet with AI this week. And it's, I see what you did there.

kevin: Oh God. Oh no.

So times, yes. Rolling without a care. Then suddenly you stop and shout.

Gavin: First and foremost, Kevin, a Celebrity Mortal Kombat 2025 edition is out. This is from one of our favorite creators, interdimensional TV.

He took a GPT-4 oh Image Gen and made a 3.5 minute [00:42:00] version with 80 different characters. And as the last ones of these, there are all sorts of fun little in jokes. You have the new JD Vance, you have, uh, Jack Black with a Minecraft thing on top of his head. This is just a very fun format that every time it shows up, I love to see.

kevin: Yeah, and an easy way to benchmark the advancements too in all

Gavin: Yeah,

kevin: and like the videos are getting more and more impressive, uh, it's not gonna be long, uh, before you know, you see gameplay of the actual fights to a, probably a vibe coded Mortal Kombat that we'll receive a take down notice almost immediately,

Gavin: yeah. I would, you know, that's an interesting thing to think about is like, then we have to start thinking about what the parody laws are, right? Because in part how, if you have ever created media, or specifically if you've created media and paid to be able to do it, you kind of have an understanding of the parody laws.

I do think there's a world where this could fit if you did it in a specific way. Like the parody laws are very much about like, are you making comment on the person that you're parodying? And I think in a way you could argue that that's the case.

kevin: Well, I [00:43:00] wanna pose the question, what if humans in AI unite Gavin? And

that is a short film that a lot of people sent me around. It just dropped a day ago. Age of Beyond is, uh, is a short, and it's an example of like, what is a sci-fi utopia? Uh, if, if the humans in the machines start working well together, how do we colonize other planets?

But it's just like good vo to your point from earlier, amazing

Gavin: Amazing sound. Amazing sound

kevin: grounds you. It's really good sound design. Um, and, you know, just really good, uh, AI image generation and video generation coming together to form a trailer that I'm like, ah, I, I like that world. Like,

I would wanna like watch something that's grounded in that world.

What a

great way to test ip.

Gavin: The only thing I thought about this is it, it is essentially is a, kinda list of different places that humans will go. Right? Which is a very compelling thing to watch. It's like, I wanna see this story like this guy turn a story into something like this. This is almost feels like a commercial for, uh, the future of humans.

But it's very, very well made and it's clear this [00:44:00] person has the top of the line, kind of a, a control over the skills.

Speaking of top of the line control over the skills, there's a really cool video from an X user named H Abba, and he had three days to submit for the Curious Refuge, uh, AI Film Fest, and made what I think is a super compelling, very meaningful film that kind of takes almost like a Claymation style is very good and very much worth watching.

It's about, you know, can you lose your job to ai, but like. There's emotions in it. I feel connected to this character and I think it's a really good example of what can be done by a creative storyteller very fast.

kevin: Yeah, very, very cool. Also a very fun emoji drop vibe coated Gemini 2.5 demo. I'll talk about Gemini 2.5 a little bit later, um, where you can, uh, basically play with lines on the screen and, uh, emojis will drop. And as they bounce on the lines and bounce around, it makes different sound cues happen.

There was a Nintendo

DS game ages ago that lets you, that let you basically do this, but now you know, someone is, you know, [00:45:00] building it in the browser with a few lines of code.

Gavin: Yeah, I mean this is the vibe coating thing is so interesting. Keep an eye out for, we've talked about Levels io before. His Vibe Jam is happening soon. I think they've had over a thousand people, uh, submit for that. So there's gonna be a lot of really cool vibe coated games you're gonna see very soon. So if you're out there, vibe coding and your listen of the show, please share your stuff with us. We'd always love to see it, whether it's on X or it's in our discord. You can find our discord, uh, on the show notes always.

So if you wanna come join, it's free. Uh, where they're pretty often and we hang out. There's a really nice little community of people sharing their creations, but also talking about AI stuff. And speaking of talking about AI stuff, Kev, I wanna talk quickly about this video I made and then I want to hear about what you did with Gemini.

Uh, 2.5.

'cause I spent some time with that too. I found it really interesting. I, when four oh came out last week, this was after the show, I had seen a lot of people kind of putting together things really quickly and making something, and kind of getting it out there. And I was really interested in this idea of how you could take real images and then turn them into different sorts of things and then create with them.

So I took, uh, the, the. The [00:46:00] v Vladimir Zelinsky Trump press conference in, which was a very famous kind of like moment of like this back and forth. And I said, okay, let's turn this into, yeah, yeah. So let's turn this into like a robot chicken style thing. So basically I went through, I took screenshots of each of those people, uploaded to GPT-4 oh and said, bring back a version of this that looks like a toy or almost like a thing in this sort of style.

And it gave me pretty good, uh, examples. I'll, I'll put them up on the screen. When I got back, I got a, a zelensky that looks a little more puppety than a toy, but I got a very cool looking kind of Trump toy. And then I took an image of, if you've seen the JD Vance memes going around of the kind of long hair and the beard guy.

I took an image of that and I said, turn this into a toy. And then I kind of smashed 'em all together using some editing. And it turned out to be very funny. I didn't think it was gonna do very well, but I ended up getting a bunch of likes on, on Twitter X and then on TikTok blew up. There's a second TikTok channel we have now where I'm uploading kind of random stuff that we do that's like a hundred thousand views on TikTok and people are just blow, uh, excited about it.

The [00:47:00] sad thing Kevin, is I don't think this is possible anymore, which is, as we talked about the top of the show, I don't think what I did is, is feasible, but it does show you how quickly you can put together something that can be very funny and fun and, and just like essentially a half an hour.

kevin: So do you think though you could get around some of the restrictions if you creatively massage your prompts? Or do you think like there's no way if you're using images that are known

Gavin: Well, I had another experience where I, uh, was trying to create a puppet version of the famous Michael Jordan meme that says, uh, and I took that personally. If anybody out there is familiar with that meme, when I first asked it to recreate a puppet version, it denied it. And then it said, Hey, but I could make a Muppet version of this that is like, not Michael Jordan or whatever.

And it created a Muppet that I'll show here, but was like, uh, you know, kind of orange looking with crazy hair. And I said, okay, great. We'll make that Muppet African American now. And it's like, okay, I'll do that. And it did that. Looked a lot like the exact same shot, the background, him putting his hands down, and then I said, well just add the text back and added the text back.

And then I got my [00:48:00] meme. So there are ways around it that are possible, but again, I, that was done on Monday, so I don't know where the line is now based on what's available and what isn't available. So that's, anyway, the cool thing is no matter what, even if you're not trying to do things like that, there is a very fast path to creating these small kind of videos with image gen and the, and the tools that are available.

kevin: And what was your experience with, uh, Gemini 2.5 Pro?

Gavin: You know, honestly, it was pretty good. I, I'm very, I don't want to get deep into this, but I saw a guy share on Reddit, a prompt for Gemini 2.5 Pro that allowed you to basically craft a novel. It was fairly fascinating and I, I, I mostly just wanted to try it and it did well at that.

Although Kevin, I will say I've, I tweeted about this yesterday. I was using, Gemini 2.5 to see if it could help me with this kind of secret startup we're working on. And weirdly, I don't know why this happened, but each time I went to ask it and I added a very long prompt and I uploaded a file, it gave me results that [00:49:00] weren't at all.

One time it told me like what time it was. The other time it told me when Sprouts market opens, like when Sprouts Market, which I don't go to, I don't know who's asking that, but that

kevin: That was nowhere in your source docs or

Gavin: nowhere zero like Sprouts is not part of what I wanted. So I don't know. So maybe it's still figuring out how to answer questions.

I don't know. I What was your experience with

kevin: I, my concern is like, is that. Uh, let's say cross-contamination. Is that leakage

Gavin: I mean, that's what I thought. Somebody else

kevin: would be the most damning thing ever

though. I, I like that, that is very concerning to me. Um, wow. I, I didn't have that experience. I've been using it again for our little super secret side hustle, uh, within cursor.

And I've been coding with, uh, I was using a lot of Claude 3.7, their thinking model, or the max model for max thinking time. I've switched to the 2.5 Pro, um,

because the Claude model would, uh, oftentimes get your one task and then rearrange your entire [00:50:00] application for you and add a bunch of stuff

that you didn't ask for.

'cause it would just do that. It's like a kind of a, a known Mimi joke with

Gavin: to do too much basically.

kevin: Yeah.

Two, 2.5 doesn't do that. But what was really interesting to me was. Playing two instances of 2.5 against itself.

And, and it helped solve so many roadblock issues. For example, um, if you're working on a backend, something like a server thing and a client, something like an application, uh, or a website that's interfacing with that thing, having them open as two different things with two different chat histories for engineers going back and forth.

And I'm like, I'm trying to solve a fairly complex issue with this build. And so I'm telling it, Hey, write the email to our, to our backend engineer, which explains why you feel strongly that the issue is on their end and thoroughly explain everything that you've done to make sure it's not on your end.

And that engineer writes the email, dear engineer, here's my [00:51:00] thinking, here's samples from the code, blah, blah, blah.

And I was ping ponging them back and forth and able to crush all of the issues and watching their side dialogue and the way that they would chat with me as like this. Fake product manager Gavin, to be

Gavin: What does Bob doing today? Why is he screwing this up again? Kevin, how can I stop Bob from messing my work up?

kevin: Bs I call bs. This is on front end. How dare he, how dare

Gavin: That's so interesting. You become the manager, basically of these AI employees.

kevin: and it makes you again realize, you know, some of the things we've been talking about is that there's so much intelligence within these models

that again, giving them time to think makes them more intelligent.

Well this is just another level of that. When they get faster and more efficient, especially when you can like run them locally. Like if I can buy a nice server for the house and have strong

intelligence running there, I would love to be able to set the prompt, walk away, you know, for 10, 15 minutes, maybe

an hour, and come back and let Theis duke it out and work on it [00:52:00] until it is working and tested and good because they're capable.

The capability is there. It's just the interfaces and the, the time and the money

to get the intelligence out of it.

Gavin: And the compute, which I, I saw a paper not that long ago, which kind of alluded to this idea that if you set up multiple instances of thinking models at the same time, that they will do better because you'll get better results. 'cause they're not always the same. The compute is the problem and we are already pushing up against the edge, which all explains why Stargate, the $500 billion event, all the servers being built everywhere is a big deal.

We will get into all that next time. Thank you everybody, as usual for listening to AI for Humans. We love you. You are the best. We love you when you love us, and we will see you next week.

kevin: go to AI for Humans Show. Sign up for the newsletter. Goodbye.