AIForHumansOpenAISora2Spicy

Kevin Pereira: [00:00:00] ChatGPT is adding erotica capabilities. This is not a drill. Nope. And verified adults might be drilling GPT five by Christmas. Christmas is about to get spicy in here. However, some other open AI features have come early. Hey,

Gavin Purcell: I see what you did there. Kevin Sora two has now opened up a storyboard feature in longer generations.

Guess it's okay.

Kevin Pereira: Okay. We will unpack that glowing recommendation. Plus Google updated their VO video model. We're talking 3.1. It allows users to extend scenes and add old men to anything.

Gavin Purcell: And early testers are saying that Gemini three. Point oh is incredibly capable. Plus Google might have developed a new cancer treatment with ai.

Wait, wait, I'm sorry. Did AI actually do something good Gavin? Yes, Kevin, you were just too busy riding hot dog erotica to notice.

Kevin Pereira: It's gonna get so much better in December. Draw me like one of your French dogs, Jack. And Ander will unveiled their eagle eye system. We talked [00:01:00] about it before, but the videos look insane.

This is the future of warfare, at least while humans are still fighting it.

Gavin Purcell: Kevin, I'm not worried about terminators. I am more worried about Bob Barker coming back from the grave and punching me in the face.

Game Show Host: I'm sorry, but you lose. You lose real big.

Gavin Purcell: This is AI for humans. Everybody,

welcome everybody to AI for Humans. This is your weekly guide to what's happening in the world of ai. And Kevin, this week we have some crazy things happening. First and foremost, something is that I like to call this the shot heard round the world. Sam Alman. Sam Alman has said, come what kind of shot, Gavin?

Yeah. Dot, I wasn't gonna get into that, but you can keep going and imagine. Sam Altman himself has said that they are now going to loosen restrictions and let adults be adults on Chad CPT and open AI will allow erotica. Kevin the, if haven't seen the clip, it's great. He takes

Kevin Pereira: a massive drag off a menthol in the corner of a [00:02:00] hotel room and he's in that very special chair.

Just sit. Let the adults be adults,

Gavin Purcell: baby. I'll, we should have had like boogie night sort of, uh, outputs from Sora for this. But the funny thing about this to me was, so first of all, like this comes from a post where he basically said, look, we have figured out that like we're better about understanding how to put parental guidance on chat, GPT.

We feel like we can loosen up a little bit for people. Wait,

Kevin Pereira: hold on. I gotta make all these noises.

Gavin Purcell: Okay, go

Kevin Pereira: ahead. Go ahead. This was on the heels of solving like mental health issues with chat GPT, right? Yes, yes. Like so like the, the, the, the thing that blows my mind about this post, and I'm, I'm genuinely sorry to interrupt you here, Gavin, 'cause you were driving to actual information, but that the one, two punches, it starts with, you know, we at OpenAI, we've taken this mental health thing seriously.

We believe we've finally established our guardrails. So now you can put your che GPT. Hey,

Gavin Purcell: we're bleeping that part. We can't say that out loud, but hey, here's the thing. Their, their freedoms are good for people to explore stuff. Yes. The other part of this is they are gonna [00:03:00] be bringing back some version of.

The four oh personality, which we have also heard a lot about. Yeah, and let me just tell you, there are people that are very mad about this idea for a couple reasons. One, a lot of people are continuing this conversation of like, Hey, OpenAI is supposed to be getting us to a GI and they're getting us now to erotica.

Mark Cuban himself came out who has been a SOA two promoter. We talked about this last week. Mark Cuban came out and said that this could be a very bad thing for kids and everybody else when they get their hands on it. The thing, I think it is important for people to understand about this is I actually think it's not the worst thing for to let people use chat PT as they wanna use it.

No. And I think that there won't be that many people making erotica. I mean, sure there might be some out there, but like it's okay to have better guardrails on this. And the other thing to remember is, Sam has said this before, but also I believe this. It does not have to. It does not mean the entire company is pivoting towards erotica.

Right. This is a massive company now, we just saw SOA two come out, which is a separate entity within OpenAI. Like [00:04:00] this is a whole team that works only on SOA two. There is a massive amount of people who are working on the state-of-the-art models, I'm sure, whatever the next model is, we've talked about how it's beaten all these competition records.

This feels to me a little bit like the narrative of like how people see Chachi PT and how everybody is on it. And that way when you see a headline like this, you think, oh my God, I can't believe this. I don't, I don't know. Do you think this feels overblown?

Kevin Pereira: I look, I thought the timing of the announcement and the specific mention of erotica as one of the things was odd.

I mean, yes. Sam had to even come out recently. When did he do it? It was, uh oh, less than 24 hours ago. As of, yeah, of this recording. He came out and said, all right. My tweet blew up on the erotica point much more than I thought it was going to. Dude, you said erotica like, and we know that, you know, X AI is over here with yfu bots that will riz you up in ways that made you blush when I showed you Gavin, like right off the rip.

So it's not. It's not hard to imagine that that is a market [00:05:00] sector that they may be going after. Right. Or at least relaxing things enough so that if people want that experience, they can have it. We'll see if it makes its way into apps, I don't know. Yeah. But I, I, I side with you on the. Let people use the models the way they want.

I understand like Mark Cuban's objections to this. Sure. When it comes to like children and teens, but I'm also on the parental responsibility side of things as well. So if OpenAI says it's for verified adults and your child or teen has an account where they're a verified adult, that's, I mean. That's a, it's not a failing on open ai.

It's a problem. Yes, it's a problem, but like at some point it's not Sam Altman's problem, right? Yes. And I think that's what

Gavin Purcell: he's trying to get across, right? Yeah. Like he's, he's not the, they're not trying to be the moral police. I think the bigger question here is. It's a, a tool that is in 800 million people's hands weekly.

Mm-hmm. Right? And right now it is a chat interface. So there'd be a lot more

Kevin Pereira: tools in people's hands, if you know what I'm [00:06:00] talking about. Gavin, have you ever thought about that one for a

Gavin Purcell: second? Yes, I have thought about that one now. Um, anyway, I think this is one of those things that like. It is about the size of what Jack GPT is and I think it's one those, oh, it's so easy to do.

It's so easy. So you can't finish your thought.

Kevin Pereira: Let's pull back. Let's pull back. Oh yeah, let's pull. Pull out of it. We're not gonna pull out. Alright,

Gavin Purcell: let's talk about another big announcement they made this week, which is an update to Sora two. Bill Peebles. Mr. Pees himself has come out and said that there are two new futures.

They just dropped. SOA two pro users can now generate up to 25 seconds in one go with these storyboard features. So this is a little bit like Google Flow's feature, which allows you to create multiple scenes in one generation. They're also letting SOA two non-pro users get 15 second outputs instead of 10.

And this is cool. I have to say like it's always cool to see a company like this, kind of like immediately kind of try to improve the product. The other kind of side note is supposedly, um, Gabriel, I can't remember what his last name is, but the guy that works at opening, I said they're gonna have less. Uh, content blocks soon.

Yeah. [00:07:00] So we'll see how that works. I spent some time with this, Kev, and you know, it's interesting the things that they've shared have always been, you know, kind of picked out and interesting in some forms. Mm-hmm. I actually just signed up for GPT Pro to si to try this. You've got, we've been on pro before I got, I gave them $200 to the machine and I wanna say, um.

It was okay. Um, if you have a second. Yeah. So have a, you just dropped

Kevin Pereira: $200 on something. Yes. And you seemed underwhelmed, like just, well,

Gavin Purcell: okay. There's, there's two parts of this. Let's talk about the two parts. First of all, if you, we talked about this briefly last week. If you're on pro, you can get better results in general for the ten second clips.

And those are improvements. I've seen those improvements and it's much better in general. I wouldn't say like, it's like the difference between, you know, like SOAR one and Sore two, but it does give you better outputs. It's the, it's the additional time in creating the additional scenes that is kind of an issue if you go into our drive.

So if you remember a while back, I created this video for vo, I think it was VO three with an ORC on the city, [00:08:00] A city sidewalk, like kind of like on a bench with a pigeon. I use that as my driving image. And I created four separate scenes. And just so you know, before we watch this, the four scenes that I've created were the, or basically looks at the pigeon and says like something like, I gotta get up outta here.

He jumps outta this scene, he's supposed to land on a cab that is moving, crunch the cab down, turn to the cab driver and say something like, get us outta here. And at the end I gave it a voiceover to do a scroll that talks about us. So play this clip and let's see what it looked like.

Game Show Host: It might be time for me to go.

Uptown. Good, sir.

Kevin Pereira: Okay. Okay.

Sora 2 Announcer: No humans or orcs were harmed in this video. Also, don't blame open AI for this. You can blame Kevin and Gavin. Goodbye for now.

Gavin Purcell: Okay, so for people who are just listening, explain to explain [00:09:00] what you just saw. And then maybe explain some of the things that are a little funky about a video wise.

Kevin Pereira: Uh, I mean, yeah. Okay. So started off very strong. It looked like a blizzard. Cinematic Yeah. Of an or on a park bench who kind of leaps into frame and then onto. A taxi cab and if you're just watching it like I was in the first go and not pausing it, you don't realize that there's other destroyed wonky vehicles in the background of this scene.

So it looks like this, or head, this wasn't the first take. There are other crushed vehicles around and then the camera sort of pans about and moves around. Then you realize there's like a, a weird jeep. Uh, taxi Cab Hybrid. Yeah. Yeah. Like it looks like it's a Rubicon edition of a Ford, uh, uh, navigator. Then it, he leaps onto a second taxi cab.

There's a weird Mad Max video, like, or a weird Mad Max vehicle in the background. And then you get a scroll, which is made up of 16 languages and Yeah, some, some hieroglyphics.

Gavin Purcell: So, so of text. So the interesting, so this is [00:10:00] a one, I, I wanted to try it. One shot output. I didn't make a second one of these. I just wanted to give it a go to see what it was like, and.

You know, again, like it got the scene changes basically. But each one of these scenes, I would say is much less interesting to me than if I had generated them individually and tried to kind of move it along. And I just think it goes to show you like, this is like early software, right? Like we are not in anywhere sort of sense where this happens.

I think the thing at the end was the most interesting thing to me is like. That shouldn't be that hard in some ways to create a scroll. Uh, obviously it got the words right when it was hearing the voiceover. And by the way, I gave this pretty significant prompts. These are prompts that were pretty significant on their own, but it does have this like garbled kind of gibberish thing that felt like an old school, like, um, output from an AI image model.

Yeah. All of this stuff just points out like, yes, we should always be careful. And Kevin's good to remind me that this is like, you know, we're at the very edge of where this stuff is happening. This was a one shot prompt for 25 seconds of AI video and yeah. It's interesting, right? It's not there yet, but it is [00:11:00] interesting.

Kevin Pereira: The 25 second mark is the most intriguing one because now you're talking like a traditional commercial spot. Yeah. So if you can truly prompt a, an almost 32nd spot on with a one shot that is very impactful, and you could throw a thousand dollars at the machine and get a bunch of ads, let it write the script, let it direct the scene.

And if, if. 10 of those are usable. Yes. That's amazing. That's, that's, that's incredible. Um, the storyboard feature was one that I really wanted to get my hands on and haven't had time yet because, you know, OpenAI was touting, I think this feature is something that they're using in their motion picture that they're doing, like this is one of their internal tools.

And I go like, oh, wow. Amazing. I can, in my head, I would go and describe my scene. Right. Or, or, or give it my script. It would generate the frames for the storyboard. I could re regenerate those easily, give some notes, and then I hit the button and it just makes it, and you're shaking your head no, Gavin.

'cause you can't believe that I thought about how great this tool is going to be. And that's how it works, right?

Gavin Purcell: That that is not what we got. Right. I will say, like we talked [00:12:00] about that last week. We showed off that video that was a custom created storyboard for their critters movie. That is not what this is, what this is, and, and we'll show a little bit of it here, is there's a shot.

You get, you basically multiple boxes. You get write your prompts per scene and then you can actually associate how much of how many seconds you want that scene to be. So you have a total of 25 if you want. And you can say the first shot is four seconds. The second shot is this, second shot is this. Each one is its own prompt.

And then you kind of send it away to go generate. Long story short is like. We'll probably get to a place where this makes sense. But the thing I think that's tricky here is the consistency between them. Right now, you're much better off like generating individual scenes and the kind of stitching them together yourself.

Mm-hmm. If you know some editing things. Um, I do wanna, very quickly before we move on from soa, there's, there's a really interesting shot of last week's open, which if you remember last week's open, you watched. Kevin and I were in chairs and we took off up above into the rockets and it doesn't look amazing.

This version of it, at least halfway through, is much better. This is from Sora Pro, so if you wanna play that here, you [00:13:00] can kind of get a sense of it. This is using your and my cameos as an example. Ready?

Game Show Host: It's AI for humans

on. Woo.

Gavin Purcell: Great. Okay. And then so, so what was, yeah, A pair attached legs just slammed us down onto the, okay, so what was cool about this is the much better than last week. You see our faces are much clearer. You see the rockets go up much clearer. We kind of get stuck in the roof, which is kind of funny, right?

But then it cuts to the two chairs empty and then like a weird pair of legs. So the first seven or six seconds of that is much better than what we created last week. And then there's kind of this weird extra thing, whereas last week it sent us. Further out and into the moon. Yeah. All of this just goes back to like the kind of slot, uh, machine sort of style of this.

It is way better, it is started state of the art. And we're gonna get to VO three in a second and talk about how it's better than VO three still, I believe, but, but 3.1. But it is still AI video and there are still, there is still funkiness. You're still gonna have to pull the, the lever a bunch of [00:14:00] times to get what you want.

Kevin Pereira: I, um. Uh, totally anecdotal, but, uh, the amount of videos I've been receiving from friends asking, wait, is this ai Yeah. Has gone through the roof? Yes. And like there was a slow trickle. And I'm assuming that as more SOA invites get released and more people get onboarded to the platform and more people go, Hey, I could take this and grind Instagram reels or tiktoks with it.

I don't know if you're receiving this Gavin, but there's like a new trend of, it's a weird one of like. Animals having, um, uh, some, uh, bowel issues, let's say, and human beings. I, I'm dead. So I saw the elephant one. Is that you sent that to me, right? No, I didn't. I didn't. There's an el You did send me the elephant one.

No, but you did see the

Gavin Purcell: elephant one. It was crazy. Im sure

Kevin Pereira: it's very similar. There's a camel one with, uh, two influencers and a convertible. There's another one with Yes, that's the one I saw. Yeah. There's another one with horses at like a central park and like it's. It's very clear that there is like, oh, this is a prompt that is working.

It's jarring. It's believable enough. Yeah. That people go, oh, [00:15:00] whoa, whoa, wait. Is that, and you know, now that there's watermark, removers, I I, I even received videos from people that, uh, like they're, and they're not senior citizens. Let me flag that. 'cause a lot of my friends are Sure. I'm receiving. I received a video from someone going like, is this fake?

And there was a glimpse of the Sora watermark. In it at one point and it's just like, that's, it's just the, it's the lane is now flooded. Yes. And I really do think, like, this wasn't one of our stories, but Alexis Hanian recently made some headlines 'cause he was saying dead internet theory is basically here.

Sure. Meaning that, you know, you and I are real. We know each other, but we have to assume all the comments beneath our video player. And, um, anybody outside of our dear patrons thank you is probably a robic. Like he, he's now saying that's the case. I'm fully there with video now. Like when I see a video, the first thing I do is assume that it's not real, and then if I care to know if it is real, I'll go a little further.

Gavin Purcell: Sure.

Kevin Pereira: I guess

Gavin Purcell: that there's a slight difference there in that the video was still created by somebody as of right now, right? There's not like a lot of. Bots out there that are creating these videos on their own. But I've [00:16:00] seen a lot of

Kevin Pereira: n innate and grind set hooks. Sure. That will just generate videos. Sure.

So, I mean, there is a, there is a, a possibility that it's a lot of bots generating media that could be wrong. Yeah,

Gavin Purcell: no, that's for sure. Uh, we should keep going here really quickly. Um, speaking of OpenAI. They are showing pretty incredible growth in revenue. There's a chart showing how quickly that they're growing from Epic AI versus other companies of the past.

But at the same time, Kevin, there's a very interesting story that's going around and we're gonna get to like how OpenAI compares to Google in just a second. That OpenAI is burning through money faster than ever before. Yeah. And one of the things we talked about last week and we continue to talk about is how OpenAI productizes their stuff.

And productizing, meaning how much money they get people to pay. Obviously we just paid another $200 to the machine this week. Um, but I think that their big question over the next like six months to a year is gonna be how do we turn the corner from losing upwards of $20 billion a year? Right now, when you think about how much money that is, that's massive, and how do we turn it into a moneymaking business?

And the promise of this [00:17:00] has always been, Hey, we'll do new science or we'll create new products that people will wanna use and pay extra for. Yes. We talked about agents last week in the new agent kit situation and how people might pay for more, but we should turn briefly now to Google, which is doing actual new science and is kind of a big deal.

Sundar Phai just came out and said, this an exciting milestone for AI and science. Our CS or C two S scale, 27 billion foundation model. Built with Yale and based on Gemma, that's one of their smaller models, generated a novel hypothesis about cancer, cellular behavior, which scientists experimentally validated in living cells.

So Kev, this is from Google, at least AI, doing real new science, which is very, very exciting for us as people.

Kevin Pereira: Yeah, um, there was a lot there. I, I announced that the internet, as you know, it is likely dead and you cannot trust it. And Gavin immediately lept into earnings for open ai. So that's, that's a new normal.

And now AI [00:18:00] is crushing science, which is a dear. Bright spot of the future, the, the light at the end of the tunnel, which was the brilliant sunshine and not, uh, the headlights of a Mack truck hitting us all. It's doing it. And this is the early stages of that. So they, they basically created a model, um, that could understand.

Uh, the way single cell organisms, uh, react. Um, and they, they have this immune context, neutral phase and this immune context, positive phase because some tumors, uh, in the human body, they will trigger, uh, an autoimmune response and that's good. 'cause then the body will attack it. Others hang out and don't trigger that response, but are still very active.

I'm distilling this in a very, very, yeah. Terrible way. Wow. Dr. I feel, I feel like, no, I'm doing a, you've

Gavin Purcell: level up. A really terrible

Kevin Pereira: job of distilling this. But I read, I read it and reread it and totally understood it in the moment. And now I'm like, ah, it's a let the AI do this. This is vi, I mean, you wanna talk about reasoning up chat, G-P-T-I-I got you all day.

You wanna talk about [00:19:00] this, this groundbreaking science. I'll leave it to the pros, but this is the promise of the, the future.

Gavin Purcell: Yeah. And long story short, what I wanted to kind of, uh, show here is that OpenAI, there's been a lot of conversation around like getting productized and doing stuff and need to make money.

Google is kind of at the same point of doing VO three point ones. They're also working on these things and tangentially showing actual new science. There's also a story that DeepMind is working, uh, using AI to help bring fusion to reality. So when we talk about this idea of like AI driving actual things that will be impactful for human beings, I just think Google feels like they're ahead of the game a little bit right now.

And again, just from a money standpoint to the money point real fast. Google has a business like meta that prints dollars for now, so they can actually forward, are they gonna afford to put all this money into science? Whereas OpenAI is gonna have to turn the corner at some point?

Kevin Pereira: Again, I, I like. We've been doing this podcast for a while and um, you know, some people are very supportive and that's great.

Other peoples are like you. You're [00:20:00] celebrating AI slop. How dare you. Yeah, there's deforestation. We're using all the drinking water and what are we getting, like Bob Barker and, and Stephen Hawking and an MMA match videos. And I'm like, yes. And they are entertaining. Yes. But this is the other side, the other side that is.

Much harder to explain. That tends to move slower and that when the results happen, like the results here could be groundbreaking. Cancer treatment, right? Taking this cold tumor and turning it hot with this previously undiscovered thing, that's amazing. But when it happens, it's not as shareable. It's not as visually impactful.

It's not as memeable. It's a white paper and a dense model. That could be incredibly beneficial to each and every one of us right now receiving this message, but it doesn't get. The attention that doesn't get erotic, get headlines, doesn't get erotica headlines. It sure doesn't, Gavin.

Gavin Purcell: That's the funniest thing about all this, which I think is important to know, is you always, as a person in the news today, have to be very clear about the stuff that's being delivered to you is being delivered in part [00:21:00] because they believe you will read it and they believe you will click on it, and it is much harder for them to get erotica as a word when it is associated with a.

CEOA $500 billion company. That is almost like instant click. Yes. Cancer treatment from AI may not be as much. So yeah. Before we move on to some other amazing stuff that Google is doing, including, uh, Gemini 3.0 and VO 3.1, we have to ask you at home to be a friend of ours, Kevin, be a friend and actually like, and subscribe this video.

We do this show every week. We have a lot of fun doing it. We are also doing a thousand other things, including our startup and then, which is actually going very well. We're in the middle of a bunch of conversations, but you are here and you are our friend, so like, and subscribe and then share something else.

Hong honk

Kevin Pereira: in the comments. If you're an algo clown, where am I? Algo Clowns at? The algo's been weird. Gavin, it's been rough. These algo has been weird. Not just for us, but for like actually successful people on YouTube. Yeah. Yeah. Everybody got a little gripe. Gripe going on. Hey, algo clowns, big Honk Kong to you.

We've tried juice in it. We tried, I think, what did we do last week? We punched, I don't know. [00:22:00] Ho. Big, big red shoes. Big silly nose. Whoa, that's a good idea where my algo clowns honk. Honk because leaving a comment actually helps pop our video and get us notice. So please do that if you listen to it, only five star review on whatever platform you do.

And sharing is caring like gen. Genuine sharing is caring. We do not have the dollars to advertise for this podcast, so any growth we have is off the backs of your effort. So thank you again. Please make someone aware. Of this podcast and rant over Hong Kong.

Gavin Purcell: Yeah. I keep getting really cool messages from people where they're like, Hey, we watch you guys every week and we tell people to watch you because you help us understand what's going on.

And like that is the mission of what we're doing. So if that's you, first of all, thank you. But second of all, get that person you shared it to, to share somebody else so we can keep getting to new people as well. Alright, moving on to more Google stuff. So Kev, there's a very exciting thing. I know you're excited about this.

Yeah. I saw some of this as well. It is a, it is a kind of rumored out there, but I think people are confirming it. Gemini 3.0 has been having some [00:23:00] incredible outputs from these, uh, Ella Marina type of situations. Yeah, it is not available for people to use, but some of these outputs are really big. The rumor around this is, is that it is a.

Big step up from Gemini 2.5, and maybe even from GPT five. So like, let's talk through some of the stuff you saw here and what makes it interesting. Yeah,

Kevin Pereira: it looks like it. So the first mention of Gemini 3.0 Pro was spotted within Gemini. Uh, someone noticed it within a developer console. It the, the rumors as of our recording, which is October 16th, in the afternoon-ish, October 22nd, this thing might actually be released, but I'll digress.

Early results as people get served up. This model in these arenas that you're mentioning where it puts models head to head, you give it the same prompt and a human being rates the output. It is night and day what this model is capable of doing. If these early rumors are true. Um, you, if you're, uh, watching the video of this podcast, you'll see on your screen some examples.

Um, I cherry picked a couple gaming ones. [00:24:00] Gavin? Yeah. These are one shot prompts for making full clones of games where the model is sourcing or generating graphics. Sound, um, inventory management, scoring, uh, movement controls, attack, defense, you name it. Things like Zelda or Binding of Isaac. There was a great example of, um, which was a two shot, meaning they had to give it a second prompt to add more features.

They're making simulators. Yeah, I saw the we one. There's a, there's a whole

Gavin Purcell: we two simulator or whatever that is like the, no, not we two. The, what is the new again? What is the name of the new, uh, the switch two. The switch two, switch two, sorry. The Switch two. Yeah. Pretty incredible. But

Kevin Pereira: I think there, what I, I recall there was a we one as well, but, but basically they're saying, hey, make not only make an emulator for the system, so they're, uh, it's emulating the looks of the controls and the console or the handheld itself, but it's generating the games that go inside of it.

Granted, they all end up kind of looking like Atari games from back in the day. Yeah. But like it's, it's Crocking the task. There was a Game Boy one that was really impressive where it's like a game boy [00:25:00] color. And when you actually fired up with the power switch on top, it takes you into a selection screen.

There's Tetris, Zelda, Pokemon, um, there was a Mario game. That's crazy. I haven't seen this one yet, but that's nuts. It looked like Mario running about. And these are again, one shot prompts. Now, on the one hand, there's gonna be some engineer out there. Well, first of all, there can be someone out there claiming this is all stolen and Okay.

Yes. Or fake.

Gavin Purcell: Or fake,

Kevin Pereira: right. Possibly. Uh, it could be fake, but a lot of the, uh, the folks that are doing this, Gavin are making their prompts and the code that gets spat out available for review. Yeah. Okay. So you can kind of go insanity check. Now, granted, yes, they could be using Claude code or GPT five Codex or something and trying to fake this, but yeah, that's a long way for a lot of people to go that have a semblance of a reputation in this space.

Sure. So I'm, I'm choosing to believe, but I also, I want to believe. Yeah, of course. Um, the, the, yes, it, it could be, uh, stolen valor fine. Uh, yes. A lot of the code for these things might exist in other places, and it's cobbling together. Sure. I'll give you that. [00:26:00] The fact that in one shot, the model is able to understand the task.

Yeah. Contextualize it. Grab the assets in, make it work and interactable, um, make decisions for how you're gonna navigate that app and get in, and get out like this. This is a, a, this is a next level. This is like the next level.

Gavin Purcell: And this is, and this to your point, like this is next level of coding, right? We will have to see when this comes out.

Like what else? It can do really well, right? Yeah. Like it's a, it's a question of where we go from here. We have been saying for the last, I don't know, six months to a year kind of pre the SOA two announcement, which felt like it kind of leapfrog VO three. Um, that Google was in the driver's seat and that Gemini 2.5 Pro was very good.

Yeah, this is 3.0 supposedly, and to Kevin's point, it may be coming in a week or so, so we're excited to follow up on that. There's another big update from Google, which is VO 3.1. This is a step change update to their text to video model, which has audio built in. And, you know, we've talked about VO three at lunch here, and now we're at 3.1.

There's [00:27:00] some cool updates to it, like basically it's better with character consistency. It will keep your shot stronger. They actually have better textures. There's a really interesting video from the Google team themselves showing off the textures of like horse mains flowing and all sorts of stuff like this.

And then Kev, there's a new Will Smith eating spaghetti. We got it. Video the benchmark.

Kevin Pereira: Kevin, do you wanna describe it in detail for the audio only users? Do you wanna Yeah. So this

Gavin Purcell: is, you know, it's, it's Demis ab by the way, shared this himself, which is always interesting. He called it the touring Test for AI video, which makes me laugh.

This is Will Smith, like really enjoying a plate of spaghetti. And as somebody, Kevin and I both know this from people who have been watching this stuff forever. I can't get enough of this. This is magic. Yeah. He's saying, I can't get enough of this. We hear slurping. Long story short, it's very good. Also, you wouldn't be able to do this in SOAR two because you can upload pictures of real people.

You can upload from a picture to start. Um, there's a many people who are playing around with this so far. I also spend some time with it, Kevin, and as I am want to do, I took, uh, one take experiences to make sure we can give a sense of what it [00:28:00] looks like. There are two clips using that same photo that I used for SOA two that are available and the exact same prompt that I use for the SOA two output only.

The first shot of the guy of the ORC jumping off the bench onto a cab. Both of them are not amazing. I will say like, and I think something that VO O three has struggled with is physics. Mm-hmm. And one of the things that Sora too did much better is physics. So if you look at both of these, they're kind of in slow motion of points and the physics of them are not amazing.

You all right?

Kevin Pereira: Yeah, I'm good. Did you hear that? Were you able to hear any of that? Okay. Nah. Sorry, I'm allergic to the bad physics in this VO video. My God. No, it's not a, it wasn't a bit, I don't know. I, I, I'm choking on something here. Um. I'm gonna go a step further and say, these look bad. Um, yeah. Now the, the visual quality of the individual, uh, elements of the scene, yes.

But the, you're, I mean, you're underselling how bad the physics are in this. [00:29:00] Yeah. Um, the, the then, and there's, there's a scene where the cab kind of. Comes in, the orc kind of jumps, but it doesn't feel really connected to like, jumping onto the cab. He kind of lands near the cab. Yeah. Now he's on top of it and Oh, the back of the cab becomes the front.

Yes. What a weird warping. And now it's driving in a completely different direction. That is, um, just a lack of coherence that I, I don't expect from a foundational video model

Gavin Purcell: now. And, and those were in the, not the fast model, the regular model. I will also say as a second sider note for vo, that kind of made me mad this week.

I was subscribing to Google Ultra for three months. Mm-hmm. So we could do these tests with VO three when it first came out. And by the way, I will say like VO three in general has been a great model and this is first shot. Yeah. I always wanna show people what it likes. First shot, Kevin. I had about 10,000, maybe even more, 12,000 credits that were kind of stored up that I was like expecting to use whenever they updated the model again, or when I wanted to go back and do a project.

I did switch back from Ultra to pro because I didn't wanna pay the $250 a month that they were gonna charge. It was a discount to [00:30:00] start. I went in, uh, yesterday and my credits are back down to a thousand. So they basically took away 10,000 credits for me, and I just think that's a bad consumer sort of scenario.

Like 11 Labs actually does a really good job of this. They kind of let you. You know your rollover minutes from night and

Kevin Pereira: weekend? Yes. Yeah. Yes. Yeah, the whole, so anyway, that was a

Gavin Purcell: bummer. That was a bummer to me. So each of those outputs cost a hundred credits on VO 3.1 regular. So now I have 800 credits left.

Again, we're always happy to see new models and, and you should see what people are making with it. But again, always. Keep people's like examples. Yeah. In the back of your head is like, they might've taken 10 to get to this, they might've taken 20 to get

Kevin Pereira: to this. Right. You don't know for sure. One of the things that I, I, I do wanna point out, like, again, I, I, I think those two examples were tough.

They were rough. Uh, if you pulled the slot machine handled 10 more times, maybe they would be amazing. But Yeah. Off of your one shots, uh, a little underwhelming. They did show off the ability though, to use flow. Where you can, uh, use elements, which is always impactful, right? Yes. You can give a, a, a person or a, a, a character, [00:31:00] a setting and maybe like an item to either hold or interact with, so you can give it the individual elements.

You can use start, end, end frames with it now, which is great for seeing consistency and control. Um. But they showed an example of a drone flying over like um, a football field, Gavin, and there was like a cow standing at the 50 yard line or whatever. They wind it back, they draw over a portion of the video and they say, add an elderly man.

Yeah. And then the same video sort of plays out. The drone comes down and you see the elderly person there just kind of standing there in the like, that was really impactful. Yes. Like that can become, that part's very cool. I wanna get my hands on that. I admittedly have not had a chance to play with that yet.

But so many times when using these tools for promotional purposes or to make proof of concepts for things, you just wanna tweak one element of a thing and you don't wanna re-render the entire video around it. So that. That sounds pretty powerful.

Gavin Purcell: Yeah. And that is like something that happened with Nano Banana, right?

Like they opened the idea. Basically. It's a much better layering system it feels like, where you can kind of take specific things [00:32:00] out and kind of manipulate them. So I will very much admit like that is not something I had the time to try. But I will try it before uh, we have our next show or maybe some point talk about in our newsletter as well.

Kevin Pereira: OpenAI has some big announcements. Okay. Erotica on the horizon. That's great. Google maybe curing cancer, giving us some new video tools. Poor anthropic. Nowhere to be found. This, this poor little, but wait,

Gavin Purcell: there is somewhere to be found. Kevin. They just introduced 4.5 haiku. That's right. That is a, their new small model, they, they have 4.5 sauna and this is 4.5 haiku.

So this is. Very cheap. This is their small, tiny model. And Kev, the, the kind of interesting thing here is that benchmark wise, it is almost as good as sonnet 4.5. Yeah. And compares pretty closely to GPT five, which is a really interesting thing. So for all the coders out there. Um, this might be your new driver.

It's funny, I was just talking to people that are working with us on in then, and I was saying, Hey, I'm gonna try to sign up for, uh, chat GBT Pro to test a bunch of stuff. Yeah. Does anybody on our team wanna use [00:33:00] um, GPT five PRO as a coding tool? And they're like, ah, most of us are just cloud code and we're happy with it.

They're so, like, they own the coder world in a big way and like, so this is a big step up to make sure that like. You don't burn through your credits as much and maybe there's things you don't need that larger models for at some point. Yeah, uh,

Kevin Pereira: yeah. One of the biggest things there is just the double the speed.

I mean, like if you're looking at the, the capabilities of the model, like yes, it's on par with some of the best, which is phenomenal considering it's the smaller model, but it's the speed. You know, if you're not doing a massive architectural, something that requires deep thinking, you've already got your plan in place, now you just wanna execute it.

The fact that you can iterate, because again, slot machine pulls, sometimes it takes like three or four tries. With a model to vibe the code successfully. Well, now you're, you can be literally twice as productive as you were before. Yeah. For a fraction of the price. So I, I think that's super exciting and I wanna use it to make AI slop in articles, Gavin, because as we've said, the internet is dead.

Don't believe anything you read, because it's probably written or [00:34:00] created

Gavin Purcell: by aerobic. Well, this is a story that I wanted to touch on briefly, and it does, it does point to your, what earlier stage about the Dead, dead Internet sort of theory, that there is a big article going around. I think this originally came from Axios, but the share of articles that are written by humans are generated by, I ai.

AI has now surpassed humans. It is about 52% to 48%, which if you're out there reading and you've seen a lot of stuff that feels similar, this is partly why Yeah, I, I just wanted to point this out because like. One of the things I have found and, and I write the AI for Humans newsletter every week, and I find it really enjoyable to do at the beginning stages.

When we talked about making a newsletter, you and I were always like, Hey, how can we leverage AI to help write this? And how can we leverage AI to kind of make this into something that. We'll make it really easy for us to do, not that we didn't wanna put out a good product, but that might be something interesting that we could start with a a, an AI draft of some idea.

Yeah. I have transitioned to fully writing by hand again, and when I not, I'm not literally like scribing by candlelight with a feather, but like I type them all out myself. I [00:35:00] don't really use AI for anything other than research because. What I have found is that if I use AI to generate drafts or if I use it to do stuff, it doesn't come out like what I really want it to say, and then I'm kind of fighting the AI to kind of say something of of importance, whereas I know what I wanna say.

I do think there's huge value, obviously, in AI writing for people that need the help or who wanna get something across, but I think ultimately this is going to be a separator. I think people will have a sense in their heads of like, what's AI written and what's not? And I think if you're out there working for a company or anywhere that's kind of like really leaning into like publishing a lot of AI writing, that's only ai.

I don't think that's gonna end well eventually. That doesn't mean that you can't have a human assisted AI writing, but I think this is going to be something that like eventually will get called out very significantly.

Kevin Pereira: Yeah, I mean, I still do a lot of creative writing, um, and technical writing as well for, uh, for, I mean, pretty much every day I have to do it.

Sure. You know, as a jumpstart, you know, as a, as a get [00:36:00] 60% of the way there, pretty good still requires a lot of cleanup and then a lot of creativity to make. Interesting. And that's like. There was another article, I dunno if you saw a couple weeks back, Gavin, that talked about the rise of like sort of AI slop in the workplace.

Sure. And then people, people are just handing off and passing off things and how it doesn't actually do the job. It kind of kicks the can. Mm-hmm. 'cause eventually along the way, a human being has to touch it and make it real or make it better or else your company. Is shipping, you know, cs basically, you know, you're shipping like a 60 or 70%.

And what it's, what's happening is that um, people are starting to think less of certain coworkers that ship ai, right, that just ship it out. They're starting to scrutinize the formatting of things before they even read the content. Yeah. Which is. Also fascinating to me, but I, I'm guilty of that as well.

I'll receive emails and I look for, not just for M dashes, but I look for like structure and emoji and bullet points and whatever. Especially if I've already communicated with someone a handful of times in the past and you can see, wow, they're suddenly communicating [00:37:00] dramatically different. Did they get Yeah, hit on the head with a lead pipe that makes 'em love emojis and bullet points, or is there something else going on here?

And I just like. All of this adds up to, again, yet human in the loop being incredibly important and becoming more so, like, use ai, leverage it, but do not remove yourself from the process because the, the, the, the numbers have already been run. It is going to damage, uh, your relationships and your reputation.

Gavin Purcell: And that's also, by the way, a tension is the way everybody's gonna be going forward. I mean, I think if we talk about the way AI is coming for many jobs, I think what's gonna happen for a lot of people is they're gonna be in the business of expressing themselves in some way. And when you express yourself truly and uniquely, that means you are doing you right?

Yeah. Like, and that doesn't mean, again, you can't use ai, but like that's what will separate you from other people in interesting ways. All right. Let's move on to, uh, a different way of expressing ai, which is AI in the military, Andro and Palmer Lucky have introduced something we have talked about in this show very much before.

Yeah. But it is out. Yeah, this is Eagle Eye. Kev, do you want jump in [00:38:00] and tell people what this is?

Kevin Pereira: Oh, listen. And I know, like, look, I know the, the war, you sound

Gavin Purcell: like you're getting too excited about this. This is war. No, dude, lemme tell

Kevin Pereira: you about this stuff. This is like a Call of Duty, add-on, come to Life, baby.

Which is

Gavin Purcell: kind of frightening, which

Kevin Pereira: is no, it really, if you've ever played like a Tom Clancy game, like this is, uh, splinter Cell ghost Recon times 10 because it's real. This is a, uh, a military grade helmet that provides protection for soldiers on the front line or maybe the second line, whatever line you wanna be.

On, you can wear the helmet and it's packed with sensors, um, and, uh, communications devices. And so basically the gist is Gavin, if you're wearing one of these headsets and I'm also wearing one, you can tap in essentially to what I am seeing and have your real time view your goggles augmented with my data or with data from a command center.

And again, we talked about this before, but when you look at the videos, I'll tell you. I mean, obviously I'm very excited for this. Like I, this is like science fiction, uh, come true. Um, yeah, you know, you want your team [00:39:00] for not, I'm not saying that the, the, the, the battle that you're fighting is the right one or the righteous one or anything else.

But if someone's gonna be out there, you want your team to have the best capabilities, right? You want, you want your side. Sure. You want your side to be alive. Um, and this, I

Gavin Purcell: don't wanna, I didn't moral No, because you know what I'm doing moral. The moral quandaries are turning in front of your eyes. It's not,

Kevin Pereira: it's not because like that, that is such like a nuanced discussion that someone's gonna be like, someone's gonna take a clip outta context and be like, look at how excited you're getting.

And I, I hate that It's disingenuous. Think. Yes. That's disingenuous. So, so lemme just talk about the tech here for a second and then we can punt the, the, the morality discussion. Sure, sure. Way down the line. 'cause that's not what this is. So you can augment your view with data that's enriched by other soldiers.

So if you have line of sight on a target, uh, or something, I can have the wire frame of that pop up in my visor so I can essentially see through objects. I can receive way points like you would in the video game if someone pings. Very cool. You know, epic Loot, it's [00:40:00] on your, your dynamic compass. There are cameras in the back of the headset so you can have like a digital rear view mirror to see what's coming up behind you.

There's some sort of control system which they're showing where you can use to like, uh, request for additional support or for something to be like, you know, detailed, there's a real time mini map tracking you and the friendlies that it has visibility on. So it's like, it really is. A video game, heads up display coming to life through a partnership from, uh, and roll, uh, meta Oakley, uh, for the optics as well as, uh, I think it's Qualcomm for some of the comm chips.

So even in a situation where like your wifi or your satellite connectivity gets knocked out, these things can still communicate with each other. Yeah. And again, the warfare of it all makes a lot of people get upset. I get it. Let's take a step back and go, how great would this technology be for first responders?

Like if you need to change Yeah. Or for

Gavin Purcell: paintball, right? Imagine in a

Kevin Pereira: paintball game. Sure. Yeah. Yeah. Yeah, for dog walkers, you know, you never know who's gonna need this [00:41:00] kind of military grade technique. Oh, infrared. I

Gavin Purcell: can see my dog's butt warming up better get the poop bag ready. Yes. Deploy the

Kevin Pereira: bags.

Deploy the bags. Now deploy the bags. Tactical scooping,

Gavin Purcell: but yes,

Kevin Pereira: not, but no like, like change the gun in the videos to a fire hose and suddenly being able to have. Infrared vision or get visibility on where someone might be trapped behind. Absolutely. It's, it's super impactful. And so I, I don't wanna pretend like there isn't a greater discussion to be had around this, but the discussion I'm hoping to have on this podcast with you is about how frigging cool the technology is.

Yes.

Gavin Purcell: And it, and it very much is. And by the way, the other side of this is like when the war begins with the terminators, like we have to make sure that this is unhackable by the machines. Because I think it would be good to be able to have a communication that's human only, right. Because it's AI derived and started, that might be a hard thing.

Well, I, I don't know what the meat version of this will look like when we start to have to fight the machines in the, in the long run. Well, that's for right now. Say now. It's very cool.

Kevin Pereira: Well, it's, it's almost, it's laughable in a sense that it's like we're gonna be [00:42:00] upgrading human soldiers and you go like, well, that's gonna work for like a minute.

But arguably, we're already like at a point where we should not be developing anything for humans. Yes, with the future of warfare, yes, but I digress. And

Gavin Purcell: fingers crossed by the way, like we're entering a world. I know there's a lot of stuff happening in the world right now, especially with what's going on in Russia.

But like I really do hope that humans at some point can progress past this. Like in, when I say this, I mean like war at large, and hopefully we'll come together, but you and I both know that there's conflicts across the board. It's not gonna happen before OpenAI

Kevin Pereira: allows erotica in December.

Gavin Purcell: It's

Kevin Pereira: not, well, maybe not.

That will be the thing

Gavin Purcell: that solves world, world peace. That's what I think's gonna bring us all together. It gonna be a world peace. That's fantastic. All right everybody, it's time to look at what you did with the AI this week and say, ah, I see what you did there.

Game Show Host: So times, yes, rolling without a care. Then sudden you stop.

Fun shout.

Gavin Purcell: Alright, Kev, there's a [00:43:00] couple fun things this week. Um, Matt Schumer, who, if you remember from a long back when had a controversial moment of the ice space, but has continued to do interesting things. Has been working on a new SOA two kind of mod that he created, which is like, he's calling it interactive SOA two.

This is a piece of code you can download, but it, and it's expensive to run, but it will allow you to generate new SOA clips based on the, where the last one took off. And it kind of looks like a video game. And yeah, it's just a cool way of playing with the SOA two API and and seeing different things you could do now that you can put it into programming and not just through the SOA front end.

Kevin Pereira: Yeah, I mean this is like, you know, it's a cool tech demo. It, it, it does like some pre-K of the possible paths you can do to try to make it seem like, 'cause these Sora generations can sometimes take, you know, 30 seconds a, a minute time. Yeah. Depending upon the quality and the style. Like, so he's, he's doing some backend work to, to render these things quicker, but.

Man, you, you and I have talked about how AI is gonna disrupt future of game making and playing and like, this is just another example of that. I think like you can imagine the next [00:44:00] telltale game or whatever, being powered by something like this, where truly whatever branch you want to take is completely different than any branch that someone else is taking.

That's. It's pretty wild.

Gavin Purcell: Yeah. You know the interesting thing when you say that, I always come back to the idea of like, well, people mostly want to play stories and the telltale thing that they always did so well was like figuring out that branching in some direction or another. Yeah. It is this interesting idea of like what games will look like in a true sandbox when you can kind of go anywhere and do anything like how much people actually want that there will be like.

An amazing AI Gary's mod that I can't wait to get my hands on in a couple years Yes. For, but for right now, it's a very cool thing.

Kevin Pereira: It'll also only cost the price of an Xbox to play it for 10 minutes, which is also very cool. Maybe in two years it'll be cheaper. Maybe it'll be 10 years. It'll be cheaper, I'm sure.

Um, I did put in this next story because, uh, I love a good bit something, um, uh, Boston Dynamics has their spot robot, which is, uh, often imitated, sometimes for much cheaper. But it is the four legged robot that moves about and has like a. A little claw, a little pincer on it that it can move [00:45:00] about. But, uh, uh, this is a post by, uh, it's math.

It's max X on X, so it's uh, max, ooh. Um, this is their work. I'm assuming. It says that we taught spot two stack 15 kilogram tires autonomously. It uses its whole body and shows some dynamic manipulation. And when you watch the video of this thing struggle, like me trying to get off the couch, but it is just, it is taking these.

Fairly heavy tires considering the size of the spot and nudging it. And muscling it Yeah. And reacting in real time to get this thing up on its side and move it into the goal, basically like the target area, and then stack another tire on top of it. It's, you, you like, you root for this thing until you imagine the near future where you're hiding under blankets and canned food in a, in a post-apocalyptic supermarket.

And this thing is hunting down. It's it turning

Gavin Purcell: over the, it's turning over all the uh Yep. Shelves as it comes to find you. Yeah. It's pulling them off. And sniffing you out, but it is very cool. What is cool about this, you're right, is [00:46:00] to watch how it figures out to get the leverage on certain things, and this goes back to that, the fact that it's autonomous.

It is doing this on its own. It's that reinforcement learning in the real world, trying to figure out how to manipulate stuff. Super cool and interesting. All right, there's one last thing. Uh, LK Campbell, who's one of our favorite Discord users. If you're not in our discord, feel free. Join the, join Our Discord.

We're in there often and you can pop in there. Found a very cool way using SOA two of how to determine who is a content violation or not. And I will say, before we get into this, Kevin, one thing that's really interesting about this whole conversation is it does seem like if someone has passed away. They are not a content violation, which is a very interesting thing.

So LK Campbell basically created a format in SOA two, where he basically has a game show. And if you play the one I made one here, just, just play this and you'll kind of see what it is.

Game Show Host: 10, 20, 25 in Liberace. Come on down. You are not a content violation. Did somebody order sparkle? Hello darling.

Gavin Purcell: Look at this legend, folks.

Straight from the, so basically he, it's a, it's a prompt where you give the date of the [00:47:00] time and then you say like, yeah. A name of a celebrity and they say, come on down, LK Campbell did a bunch of these and it turned in one of those remixes. But one of the most fascinating things about this is you put in Bob R, you put in Fred Rogers.

All the people that we have seen come up, show up, and then all these other people, a pagoda. I try, but there does seem to be this interesting, weird thing and I don't know, maybe there's a legal scenario with this for if a person has passed away or not, where they sure are not asking for permission ahead of time in some form.

Interesting. Yeah.

Kevin Pereira: Do you think, um. Do you think in December when OpenAI relaxes the guardrails on erotica Gavin, that Sora is gonna get a similar treatment?

Gavin Purcell: Well, you know, that's a really interesting thing that I was thinking about while we were talking about earlier in the show. I would argue probably not, but if you look at what ROC has been making, maybe that's where we're headed because that worries me more than erotica.

The audience for erotica is probably this big. The audience for video is much, much bigger. So like if we are suddenly entering into a world where like. Spicy content comes to soar too. We may be in trouble, but for now, [00:48:00] Kevin, that is everything. Thank you everybody for tuning in. We really appreciate it. We will see you all next week.

Kevin Pereira: Hong.