← The Artificial Intelligence Show
The Artificial Intelligence Show
#142: ChatGPT’s New Image Generator, Studio Ghibli Craze and Backlash, Gemini 2.5, OpenAI Academy, 4o Updates, Vibe Marketing & xAI Acquires X
01:13:00
Loading summary
Transcript73 lines
- [00:00]
Paul Raitzer
It's like this future of work, like, what does it even look like? And this definitely makes my brain start to hurt a little bit trying to visualize that. But the idea of a single interface for all of your communications and strategy seems like a logical target for them. Welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raitzer. I'm the founder and CEO of Marketing AI Institute and I'm your host. Each week I'm joined by my co host and Marketing AI Institute Chief Content Officer Mike Caput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all. Welcome to episode 142 of the Artificial Intelligence Show. I'm your host, Paul Raitzer, back again with my co host Mike Put after my solo talking into the screen for an hour and 40 minutes. Session 141, the Road to AGI. So if you didn't catch that one, 141 was the first in our Road to AGI series and I kind of walked through a theoretical AI timeline of sort of what's happening now, what's coming next and what it means. So you can go check that out. That is available now. And we are back to our regular weekly format today. This episode is brought to us by the State of Marketing AI Survey soon to be report. This is your last chance. So if you haven't, if you're a marketer or business leader, want to be a part of our State of Marketing AI report for 2025. This is our fifth year, Mike. Fourth.
- [01:40]
Mike Caput
Fifth year. Fifth year.
- [01:41]
Paul Raitzer
It's the fifth year we've done it. So we've got a ton of fascinating historical benchmarks and data that will build into the report. More than 1800 people have already been a part of the survey, so we'd love to have you join in. You can go to stateofmarketingai.com and just click on the link to take part in the survey. You can download last year's while you're there, check that out. But the new one will be coming out. When, Mike, when are we thinking we are shooting for?
- [02:06]
Mike Caput
The end of April for the come out or mid April rather for.
- [02:10]
Paul Raitzer
It's gonna be a fast turnaround.
- [02:11]
Mike Caput
Yeah.
- [02:11]
Paul Raitzer
All right. So we're gonna turn this thing around fast and so you will have some fresh data. We will definitely talk about that data on the podcast once that comes out. But again, go to stateofmarketingai.com it takes what, three minutes? Mike, go through and take the survey, Give us your feedback. We would love to hear. It just talks about piloting, scaling, AI within organizations and your own perspectives on AI and your career and a little bit into your life. So we'd love to have you be a part of that survey. All right. With that, I mean, we had a couple of ads this morning, so we're recording this Monday, March 31, 11:40am and just this morning there was like two major things that got added. So yeah, it's, things are moving fast. But we, we had big week last week with some new models and new capabilities and other models. So let's, let's dive right into that.
- [03:00]
Mike Caput
Mike yeah, it was a huge week. Paul Our first main topic, OpenAI has introduced 4O image generation. So this is a new image generation capability directly within the GPT4O model, meaning you can now generate stunning imagery right within ChatGPT. Much, much more advanced than the previous Dall E image generation capability, according to OpenAI in the launch announcement. QUOTE gpt4o Image generation excels at accurately rendering text precisely following prompts and leveraging 4O's inherent knowledge base and chat context, including transforming uploaded images or using them as visual inspiration. These capabilities make it easier to create exactly the image you envision, helping you communicate more effectively through visuals and advancing image generation into a practical tool with precision and power. Now you might get from that quote this is not just a better image generation image generation tool, it is a fundamentally different one. So this is actually baked right into the 4.0 model. So that model is truly multimodal and as a result it gives it new capabilities so it can produce much better images because the full intelligence of 4.0 is brought to bear on the prompt, which didn't used to happen with image generation in ChatGPT. It's better at generating text and images way, way better, which was a weakness of past image generation models. And you can also upload and edit or manipulate images using the tool. So you can change any element or apply any style to an existing photo or picture. Now what's really cool here is it has quite the capacity for quote in context visual refinement. So you can kind of progressively prompt and shape the image through conversation with GPT4O and still maintain visual consistency across multiple iterations. So if you start a new iteration, it's not going to look totally different than the image you started with. So right now this is available in ChatGPT+Pro and Team Sam Altman posted that because of insanely high demand their GPUs are quote melting and they roll out to the free tier is going to be delayed for some time. So Paul, there is actually a whole piece of this we're going to address has its own topic in the next topic which is the, the fact that this image generation has gone viral with people generating images in a popular animation style. But before we get to all that, I want to first get your initial impressions of this. Like I know you've used it, like have you found it as impressive as everyone's claiming it is?
- [05:49]
Paul Raitzer
Yeah. So I, I did have a chance to finally experiment with it on Wednesday. I think it was, I think it came out on Tuesday so I guess it wasn't too long after. I thought it was interesting timing always OpenAI likes to drop things right after Google drops things. So we'll talk about 2.5 from Google which came out at like 11am or 10am that morning. And then at 1pm OpenAI drops their image generation thing and sort of like steals the thunder. But we will come back on the Gemini model. It's quite impressive on its own. So I wanted a use case to test the text part of it because that has been a massive flaw of these models previously. And so I actually make my kids birthday signs. My mom used to do this for me when I was a kid. It was like one of my favorite things, my birthday. You wake up in the morning and there's like signs made around the house. And so I, I carry on that tradition and so every year for my kids and my, my wife I make them birthday signs and hang em up around the house. And so this year I thought well let me see if I can make these with ChatGPT image generation. And so I just went in and gave my son's name is that he's, you know, turned 12. And I want to do like some fun and clever sayings on signs. I'm going to give you some themes and like let's develop some stuff. And so it was really fascinating because the first thing I did was Pokemon and so it would create it and I asked for specific characters and it would create it and then it would just disappear. And I was like what the hell was that? Like it was there. And then it would say oh, due to copyright or whatever, we can't generate this image. I'm like yeah you can. You just did like do it again. And it's like well I can't do that one. But I can make one that looks like this. And I was like, fine, do that. And it would come out looking almost exactly like the actual character, but it wasn't the character. And so I immediately realized like, okay, there's some filter like classifier here that's not rooting out your initial request for copyrighted images, but it's, it's like extracting it. So I don't know. Anyway, we'll come back, we'll come back to the copyright thing in a minute. But it was able to start creating these things. So I asked for like a 8 bit baseball one. Cause he loves video games. I asked for like Minecraft related things and it was able to roughly do those and that it nailed the text. They had like one typo in, in like 12 signs that I made. It was like one. It smelled his name wrong and I was like, hey, you did that wrong. And it fixed it. So it definitely is quite impressive. You know, I think that I immediately started, you know, when you were online on Twitter, you just see all of these, not only the studio Ghibli things that we'll talk about, but you were seeing like ads being made, people taking like Coke signs and dropping them into backgrounds. And. And you start realizing like if people weren't aware of the impact these tools are going to have on creative workers, creative firms, movie production, stock photography, it is quite apparent when you spend some time looking at the samples of what people are building or when you just start building things yourself, these capabilities are significant. And you can definitely start to imagine a world where you're using AI more and more in creative work. And then the other thing I thought about when I tested this is Sora's next like the video generation stuff is this is the prelude to that. And so imagine this level of control and consistency. But applied to 10, 15, 20 second videos, I got to imagine when the GPU shortage sort of goes away and they have more capacity, that capability is probably already sitting in there is my guess. They just don't have enough GPUs to roll it out. So image and video, yeah, just like whole nother world. And that's. You can go look at VO2 from Google, their video gen model and imagine three from Google, like they're similar certainly, you know, probably on par with some of this stuff. So yeah, we're taking some leaps I think this year in image and video generation for sure.
- [09:47]
Mike Caput
I want to just talk for a quick second about maybe some of the bigger implications here that you alluded to for say creatives or like business Use cases. Because like, I've always found it pretty fun and creative, no doubt, to like generate images. It's impressive. Like, it's really cool, the stuff we had before. But honestly now with the text being accurate, I was generating literally mockups of ads in old styles of like really good ads with their logos, the exact font and stuff. It was crazy how accurate it was. Like you could see yourself now actually using this to create high quality ads, business visuals. I could see infographics, charts, things like that at some point.
- [10:27]
Paul Raitzer
Yeah. And I think, you know, I. So if anybody's watching this on YouTube, in my background, there's a extra size copy of the COVID of our book, the artificial intelligence book. There's a logo of Macon. Like there's, there's different things back there. And I start thinking like as a non designer, I wouldn't say non creative, but like I have no design capabilities at all. And when we want to do projects like that, you are reliant on the designer to like get the vision out of your head. But all I have is words like I can't sketch it. I'm not just not good at that. And now to think about Mike, like to your point, whether it's a logo, a webpage design, interior design of your home, the design of a, of a book or a digital asset, all of it, you can now just use words or a driving image like, hey, I love these three book covers. Develop some in this, this theme. But here's the words for the COVID And then say, oh, no, no, no, like, oh, that's awesome, let's do that in blue. And all of a sudden non designers have these abilities and I don't, I don't know what that means, honestly. And I don't, I don't think OpenAI knows what it means. I don't think Google knows what it means. But I think it's really important that we have these conversations because I just feel like these tools are starting to truly creep in to democratize the ability to build things. And I don't know what that means to the people who do that work daily for a living. I think some, you know, some of them obviously are just going to take these tools and have superpowers and continue on. And a lot of the world's going to just be ignorant to the fact that these are even possible like that, you know, you can have business leaders who don't know. You could go do some of this stuff yourself or at least mock things up yourself. But I think that's more and more what's going to happen in all the knowledge work is whether you're working with an attorney, an accountant, a graphic designer, you're going to have the ability to do the first drafts yourself. Now for anything, basically, and you still may rely on the experts to do the final products and bring it home. But some of that early work might just be done by the AI. Here's a draft of my talk. And then give it to the speechwriter and let the speechwriter fine tune it. Here's a financial analysis of the business. And then let the financial analysts do their final work. I don't know. Like, again, I don't understand what the implications are, but it's a reality. Like, these tools are there. They're beyond first draft capability in most of these things.
- [12:55]
Mike Caput
Yeah. And just quickly thinking out loud, I wonder even if you accept that your designer or creative professional has superpowers with these tools, I wonder how that'll change expectations. Like, I'd be like, why can't you give me a hundred variations today? Two hours. Yeah, maybe.
- [13:11]
Paul Raitzer
Why is this taking two weeks?
- [13:12]
Mike Caput
Maybe I'm a jerk for saying that, but I feel like the expectations of what's possible, possible if you assume someone is enabled or empowered with these tools.
- [13:20]
Paul Raitzer
Has just changed 100%. Yeah. And you, I mean, we worked in an agency. I 100% could see that, like the expectations just become way faster, way cheaper, way better. And I think that's going to be a reality for service firms and internal, you know, creatives and writers and things is like, once everyone catches on to what these things can do, the expectations for what you do is going to change. And I think the faster you get there and be proactive about this, the better prepared you'll be. You don't want to like sit or wait around until all your clients figured out that you could do things way faster.
- [13:59]
Mike Caput
All right. Our second main topic is very closely linked to the first one. So the new four zero image generation capabilities have gone truly, like, insanely viral, primarily due to a single use case. And that is people are using 4O to turn their personal photos into animated illustrations in the style of Studio Ghibli. And Studio Ghibli, if you don't know, is a legendary Japanese animation studio. It is famous for producing beautiful, often hand drawn animated films that appeal to both children and adults. Like, one good way of thinking about it, it's like kind of the Pixar of Japan, but with a much more dreamlike and like, poetic vibe. I would say it's very calm, peaceful, Thoughtful, distinct animation style that is really, really well known. And it also happens to be the animation style that users latched onto when experimenting with 4o image generation. Now it seems like X in particular, though I've seen Elsewhere now, like LinkedIn is just flooded with everyone, not just AI, you know, early adopters using tools to apply like a Studio Ghibli filter to all their photos. Now a lot of people have found this really fun, wholesome and creative. The style is really kind of joyful to look at, I think, but it's also generated a ton of backlash because people are wondering just how much of Studio Ghibli's copyrighted work may have been used to train this model. So, Paul, I want to kind of frame this through one post which we saw that kind of really illustrated a. I don't know if it's surprising, but like a deep well of anger about this issue. So a former Googler and a leading voice in AI, Cassie Kozakov, who posted about this trend on LinkedIn, she showed off some Ghibli photos that she had done of herself and she basically kind of commented like, hey, this might be some great marketing for Studio Ghibli. Everyone's talking about it now. The comments, though, kind of disagreed. They were almost 300 of them. They are very majority negative. People are just kind of raging and extremely upset about how Ghibli's work is just being essentially ripped off for this use case. Again, it's a really distinctive style that has been around for decades. So I guess I wanted to start this off by asking, like, are we about to see a wider backlash here, at least among creatives towards AI?
- [16:35]
Paul Raitzer
Yeah, I think it's coming in addition to like the books being stolen from for training, which we'll touch on as well. So I tweeted sometime last week when it became pretty apparent that OpenAI was steering into this Studio Ghibli thing and Sam himself changed his icon on Twitter. And I said that the AI model companies had entered the quote, don't give a fake phase of IP investment. Meaning like, we're just gonna do it. And then interestingly, cause I was referring earlier to how these things work, where they, it's. It's very, very obvious. They were trained on copyright material. Like ask it to do something for the Simpsons or Marvel or Disney. It'll do it. At least as of like Friday, it would do it. It'll. It takes about 20 seconds to create the image and it sort of appears from top down. It's using this auto aggressive model to like build these things. And so you'll like say, turn me into Homer Simpson and it'll do it. And then it just goes away. And so obviously it knows who Homer Simpson is. It was trained on Homer Simpson. It has the ability to output Homer Simpson or any of these other copyrighted characters and material, but by disappearing it, they're basically like, not hopefully going to get sued. But in the case of Studio Ghibli based in Japan, which doesn't actually have, you're not allowed to sue for copyright infringement, they're allowed to train the model. So the belief, and I don't think OpenAI has confirmed this yet, is that the reason that Sam and others allowed this, like steered into this and even like joked about this character or this studio being used this style is because they can't get sued by this company. Now, again, I don't know that a hundred percent to be fact, but I do know that that is the rule in Japan. So I think what they're doing is basically like they trained on everything and the ones where they won't get sued, they're going to kind of just let it go. Now, interestingly, there was a couple people who started saying, like, hey, this thing's getting nerfed already by the weekend. Meaning they're making it safer and not like allowing it to do all these other things, because it can do a lot of things, not just copyrighted things. And he said, quote, we are going to do the opposite of nerfing it. Meaning they have every intention of like pushing the limits here. They're going to let this thing go. So I don't know if they've just become convinced they're going to win these lawsuits or they just have enough billions set aside for the lawsuits. They just don't care. But it's obvious that they're just full steam ahead. Xai is going to be full steam ahead. I got to mention, Meta is going to do the same thing. I don't, I wouldn't have thought Google, but they've definitely been more lenient, I would suppose, with the things that their models are creating. So I feel like we're just kind of pushing the limit here. And then society is just going to kind of get used to that limit having been pushed. And then all of a sudden it's like, ah, you can just make anything you want. So, you know, I, I don't know. I think that there is this definite frustration with people related to the impact it has on creative work. And then the other side of this was last week, Mike, there was a lot about like the. The lib gen books thing. You want to give us a rundown of what that was because it's in the same vein?
- [19:51]
Mike Caput
Yeah, absolutely. So on March 20, the Atlantic, the publication, published a database of all the books that Meta may have used to train its models. Books which it doesn't have the rights to because Meta was proven to have trained on books from a database called Library Genesis or Libgen. And it's a pirated book database. You can go on it and get books that you have to typically pay for for free. Now, another post about this has gone even much more vital than the one from Cassie. And this is from marketer and writer Ann Hanley, who we know she posted on LinkedIn about how all three of her books she found from this database were used without permission to train Meta's llama models. And this post has almost 850 comments, 438 reposts on LinkedIn, which is insane. That might be like one or crazier posts I've ever seen. And the comments are just, as far as I could tell, just all very, very negative towards Meta, but towards AI companies. So I guess maybe unpack this a bit more for us because it just really feels like these two things, maybe it's just bad timing for them, but it really feels like creatives don't have much of a leg to stand on here in terms of doing something about this.
- [21:04]
Paul Raitzer
Yeah. So I think, and I'm just kind of thinking out loud here, I think two things are happening. One, people are becoming aware of how this has been working for years. This is not a secret that this is how this has been done. Your books have been being stolen for years and used to train models for years, as has your creative outputs, your designs, your photography, your painting, all of it's been stolen for years. That is not new. People's awareness of it is new. And then the second component that I think is only going to throw some fuel on the fire here is go back to the last section where we talked about the impact on creatives where I think this is the year where people actually start to feel it, where, you know, maybe I'm not making what I used to make to do logo designs, or I'm not getting paid what I used to get paid to do writing, or the client's kind of doing their own writing now, or the CEO is able to write his or own scripts because they're just using ChatGPT and they build a co CEO that writes their speeches for them. Like, I think this is the year where the rest of the world starts realizing what these things can do and starts doing things themselves that they used to use other people to do. And so I think combined with an awareness and understanding of how these models work with an actual impact on people's livelihood or perceptions of value and fulfillment, that is a recipe for a lot of backlash. And I would not be surprised at all to see these sorts of posts continue. And it just takes a few well placed influencers who decide to make this a talking point for all of their followers to now realize what's happening too. And you know, Ann is wonderful. Anne is a great friend. She has been on, she was our keynote for last year's AI for Writers summit. I believe she and I had a wonderful talk about the impact of AI in writing. So, you know, and Anne is one of the most trustworthy and honorable people I've ever met. And if Ann has a problem, Anne's followers are going to have a problem with what Ann has a problem. So, yeah, I think it's, it's fascinating to see. And I, I, I, I don't know, I, like, part of me is involved in this because I am a writer. You're a writer, Mike. My wife's an artist. You know, I, I think I live this personally and then I observe it and comment on it for, for our podcast. So I live in this weird world where I actually, I feel both sides of this. Like, I, I'm inspired by what you can build now, the democratization of the abilities to do these things. And I like using the tools. And then the other part of me is like, but I know how they're trained and I know the impact they're going to have on people. And sometimes I'm not really sure how to feel about it all.
- [23:49]
Mike Caput
So our third main topic this week, I almost hate to say it, they kind of flew under the radar, which is crazy. This is a huge topic. But, you know, ChatGPT's image generation, like, sucked the oxygen out of the room and kind of overshadowed the fact that Google unveiled Gemini 2.5, which they are calling their, quote, most intelligent AI model to date. So this first release in this new line is an experimental version called Gemini 2.5 Pro experimental. It is making waves in the AI community because it is topping industry benchmarks with significant margins. So Gemini 2.5 pro experimental is in a category of models Google calls thinking models. These are AI systems designed to reason through their thoughts before they respond, which results in better performance, better Accuracy. This means that they do much more than just classification and prediction. They have the ability to analyze information, draw logical conclusions, incorporate context and nuance, and make informed decisions. So right now, the new Gemini 2.5 Pro sits at number one on the LM arena leaderboard, which measures human preferences for AI responses. It shows particularly strong capabilities in reasoning and code generation. It leads in common coding, math and science benchmarks. And in practical terms, it is demonstrating really impressive reasoning skills across a range of challenging tests. So it's achieved state of the art results on math and science benchmarks. It scored 18.8% on humanity's last exam, which we've talked about before, which is a data set designed by hundreds of experts to capture the frontier of human knowledge and reasoning. 18.8% is a very high score so far on that for all the models out there. And it is now available if you want to try it in Google AI Studio and in the Gemini app, if you're a Gemini Advanced subscriber. So Paul, what do we need to pay attention to here? This is clearly a really, really powerful model like 4.0. It has multimodal image generation, it has a long context window, so that's a million tokens right now, which is about 750,000 words it can hold in its memory and pay attention to at any given time. Just seems like these models Google is putting out are just getting really, really robust.
- [26:18]
Paul Raitzer
Yeah, and this is sort of a preview of the next generation of models. So the, that there, if we go back a year or so, it was text in, text out, so you could put text into your prompt, it could generate text back to you. Now you could do image generation and you could do some reasoning last year and things like that, but they were through separate models. Usually it wasn't all baked into the same PI, I guess, for lack of a better analogy. And so what's going to happen now is like Claude 4, GPT 5, Gemini 3, Llama 4, all these next generation models, which I assume we will see all of them this year, they will all be multimodal from the ground up. Right now that means text and image and I guess voice and it'll, it'll eventually also include video and audio in there. So you can imagine like Sora from OpenAI being baked right into ChatGPT or you know, VO2 I mentioned earlier, from Google being baked right into Gemini. So you're gonna have these multimodal models that it's able to input and output modalities, multiple modalities, and then you're gonna have reasoning on top of it and then you'll have some sort of classifier that actually knows which function to use for you. So if you go in and you're having a conversation, it knows whether you use reasoning and think more, it knows whether to create an image or a video. We don't have to pick from our 17 models in the dropdown like we've talked about many times. And then the context window, which is where Google has a pretty clear advantage at the moment. The reason for a non developer like you and I Mike, that that's important, the average business user is imagine it having access to your CRM system or to your Google Drive where all of your knowledge is, all of your documents are. The context window is basically what goes into the prompt, like the back end of the prompt and what happens is within that window it dramatically improves accuracy, reliability, reduces hallucinations. So if it can remember that that information, those tokens, then it, it becomes way, way better and more practical for use in businesses. Otherwise it just kind of forgets things and it can make errors. And so the bigger the context window, the more accurate it becomes. It's why NotebookLM works so well. Like if you build a Notebook LM and you put, you put like five PDFs in there and a video script, it basically talks to you based on that context that those documents you've put into it. And so the bigger that context, and we know Google last year they talked about 10 million tokens being tested and working. They've, I think Sundar's been on record is talking about basically infinite tokens. So the goal is to be able to stuff as much information as you want into this system and it's like insanely accurate with what it outputs and recommends to you and the decisions it makes. So context window matters a lot to the average user. It's just sort of an abstract concept.
- [29:15]
Mike Caput
Yeah, And I wonder too, with, I mean 750,000 words already is an insane amount and with higher limits you have to think I, I mean that could comprise every document your company has created if you're relatively a small company. Totally.
- [29:29]
Paul Raitzer
I mean our, yeah, like an average business books like 50,000 words.
- [29:33]
Mike Caput
Right.
- [29:33]
Paul Raitzer
So do the math. I mean it's a lot of books, it's a lot of content just at the million tokens. So yeah, it's. And again they can do multimodal so you can put video and you can put different kinds of documents. So yeah, hard to comprehend.
- [29:50]
Mike Caput
All right, let's dive into some rapid fire this week. So this first one literally hot off the presses here. So we just found out before recording, OpenAI has launched OpenAI Academy, which is a new initiative aimed at democratizing AI literacy for people from all backgrounds. So this is right now a free community powered learning hub that features bite sized video tutorials that cover Everything from basic ChatGPT usage to more advanced applications like creating videos in Sora right now. The current offerings include content specifically tailored for educators, students, job seekers, nonprofit leaders, small business owners and a couple other groups. Now what makes this kind of interesting is the community focused model. So rather than simply building just a repository of content, they're creating an interactive ecosystem with both virtual and in person events. The platform hosts regular workshops, discussions, collaborative sessions led by both OpenAI experts and external innovators. And according to social media posts from OpenAI's VP of Education, who apparently worked at Coursera, this launch represents just the first phase. The academy is designed to be globally accessible, though it's currently only available in English. They're going to expand additional languages soon. They also indicate they're looking for motivated hosts across the world to help scale their in person events globally. Now, right now you don't get any type of certificate or accreditation through the academy, though they say they've got at least one other big announcement about this coming soon. So we'll see kind of how that works out. Paul, this definitely validates the need we've seen for widespread AI literacy. What do you make of their approach to achieving that goal?
- [31:44]
Paul Raitzer
Yeah, I was, I was actually really excited to see this. I think it definitely validates what we've been saying, this need for AI literacy. It's nice to see them, you know, pushing that. You see similar things like HubSpot has AI classes, Salesforce has AI classes, Google Cloud, Microsoft, like a lot of these AI model companies, AI software companies have and are moving into the eye literacy space. And I think it's really important. There's a lot of value that can be created from these model and software companies. Now the challenge sometimes they face is that they can't be brand agnostic. So like OpenAI is not going to have courses on here about Claude and you know, co pilots and Gemini and things like that. But if you look at how they've got it structured because I joined as soon as I saw it, they have collections so they have like chatgpt on campus. Awesome. Like I, I think that's huge. That's, I assume that's mostly for like students and maybe teachers. They have ChatGPT at work which gets into like some specifics about how to use different productivity components. They have Sora tutorials for the video, they have an A AI for K to 12 educators which is fantastic. And there's four items in there now. But like I think that's great. And then they've got a lot on the developer side. I would expect they would go pretty hard on the developer side but I did notice they've got like just some like AI for older adults. Like that's awesome. And so some of these are functioned as live stream and so it's, it's real similar to how I'm actually envisioning building out. Our AI Academy is a mix of on demand courses, live stream, in person events. So you're truly kind of creating all this. So just in my initial couple minutes of looking through this, there's actually a lot of it I could see where we, we may actually be recommending components of this as part of our mastery program. We say hey, for additional learning, here's some great stuff on OpenAI, here's some stuff on Coursera, here's some stuff on LinkedIn Learning. So yeah, I think that it's fantastic to see. I expect we're going to see a lot more of this from these companies because at the end of the day they need AI literate buyers. And so for people to use ChatGPT to the level they want to grow, they need to educate them on how to use it. So it makes total sense that they would make a play like this. And it's interesting they announced this last fall, but this is not what they announced. Like they pivoted what AI Academy was going to be I think, which is great.
- [34:07]
Mike Caput
Our next rapid fire topic consists of a few more important OpenAI updates. So in addition to the image generation update, the company has also released some other significant updates to 4.0. So according to OpenAI, the updated model reportedly feels more intuitive and collaborative. There's particularly particular improvements in STEM and coding tasks. GPT4.0 now generates cleaner front end code. It more accurately analyzes existing code to identify necessary changes and consistently produces outputs that compile and run successfully. Now I would say if you are not a developer, I personally encourage you to check it out. I found it actually to be significantly better. Maybe that's just kind of the vibe I'm getting. But for a lot of non developer tasks it also seems to have improved significantly. And for business customers, OpenAI is rolling out one of their most requested features, the ability to connect ChatGPT to internal knowledge sources. So this is in beta for ChatGPT customers. And it allows the AI to access and pull information from an organization's Google Drive workspace in real time, which is allowing it to provide more personalized and contextually relevant responses. OpenAI says this is just the beginning. They have plans to support additional connectors for collaboration tools, project management systems and CRMs. On the business front, OpenAI is projecting pretty extraordinary revenue growth, according to sources familiar with the company's internal communications. OpenAI expects to more than triple its revenue in the coming year to 12.7 billion, up from 3.7 billion last year. And that growth trajectory is expected to continue to with projections of 29.4 billion for the following year. And this all comes as Bloomberg reports that they are getting closer to finalizing a 40 billion funding round led by Softmark. So Paul, do any of these updates seem particularly notable to you? I mean, I'm personally interested to see what becomes possible when you can connect this to internal knowledge sources.
- [36:24]
Paul Raitzer
I'm interested to see what happens to the knowledge sources themselves when I can just use Chat GPT.
- [36:30]
Mike Caput
Right.
- [36:31]
Paul Raitzer
So I think of like Asana, we use for project management and Asana's got some baked in AI stuff now but like if I could just connect ChatGPT to it, would I use Asana's AI tools? HubSpot has some AI capabilities, like even some new things that I've started seeing that I really like with within like auto summaries of companies and things like that. Yeah, with their breeze intelligence which I'm pretty sure is actually built on OpenAI APIs. So I, I, I like, I, I wonder Google, same thing. I've got Gemini right in Google Drive. Like would I use chat GPT's integrator instead of Gemini's? So as a user slash CEO buyer, I don't know what actually any of this means. Like I start to think ahead, it's like wait, so am I just gonna, will we centralize all of this into ChatGPT and just connect it to all of our tech stack, or am I gonna use the AI native within each core piece of our tech stack? And I don't know the answer to that, but I do think this idea of being able to connect in is makes a ton of sense. I can see that being valuable.
- [37:32]
Mike Caput
So you're saying you could see a future where ChatGPT is just so known and intuitive you end up just using that as the interface with these tools?
- [37:40]
Paul Raitzer
Yeah, like you just log into ChatGPT in the morning and it's connected to your project management system, your CRM, your Google Drive, and you just live in ChatGPT. Like you're just talking to it all day long and it has access to everything you need. It's like, oh, what are my top three tasks for the day? And it goes to Asana and it grabs them and all right, what was the conversation Mike and I had last week about AI Academy? And it goes into Google Drive and it grabs it and it's like, okay, great, like, draw me an email to follow up with Mike on that. And I never leave that thread and maybe I just have a thread each day and it's like, I don't. I don't know. Again, it's like this future of work, like, what does it even look like? And this definitely makes my brain start to hurt a little bit trying to visualize that. But the idea of a single interface for all of your communications and strategy sure seems like a logical target for them. I would imagine they'd be trying to build that.
- [38:35]
Mike Caput
Next up in rapid fire. In February, AI leader Andrej Karpathy posted on X about a concept he called, quote, vibe coding. This is a new kind of coding. He kind of invented a term for where you, quote, fully give in to the vibes when you're coding, basically by just talking to AI over and over and having it do all the coding to complete your projects, he notes. For instance, quote, I'm building a project or web app, but it's not really coding. I just see stuff, say stuff, run stuff, and copy paste stuff. And it mostly works along those lines. There's now the introduction of this term called quote, vibe marketing. So this is now gaining steam in some marketing circles, based on a number of posts online that say the era of vibe marketing may be here. So here's their argument. The current marketing landscape typically involves AI tools being used in isolated ways for specific outputs on individual channels. However, in the coming year, we're expected to see interconnected AI systems working together with shared context. These systems, so the argument goes, will feature multiple AI agents and workflows managed by quote manager agents that are trained by human experts. As a result, this transformation will fundamentally change the role of marketers. They'll basically evolve from individual executors to orchestrators of complex AI systems. So basically, marketers will start operating on vibes focused on strategy, storytelling, creative direction, while AI handles all the messy execution. The proponents of this argument say it could create incredible efficiency gains. A savvy solo vibe marketer backed by an orchestra of agents could outperform a typical Entire agency. So, Paul, this is definitely an interesting idea. I don't disagree with marketers becoming orchestrators of AI tools and agents, but good luck trusting so much marketing at major brands to agents and vibes. I mean, at least today. Do you think?
- [40:46]
Paul Raitzer
Yeah, I don't know. So. So my friend Ali K. Miller was sending me some of these links. She and I were catching up last week and she was sharing some of these links with me. And so I was, I was diving in a little bit to it. And the way I think about this, because honestly, like, I went back and re read Andre's original tweet like 5 times a week or two ago and I was like, I don't know if I get it. Like, this is kind of a confusing topic. And so let me. I'm going to think out loud here, Mike, and tell me if this makes sense. What I'm envisioning for Vibe marketing is, all right, we're doing a product launch in 30 days. I want to go in and I want to build a campaign. I want to build all the components of it. Like, let's go. And I'm just talking to Gemini or ChatGPT and it's like, great. Like, let's start with a plan, you know, the really excited helper assistant. And it's like, okay, yeah, build, build out the plan. And it builds. And it's like, okay, that looks great. Let's go ahead and write that first email. And it writes it and it's like, okay, that's actually really good. Build me a nurturing sequence now for like when people open, don't open. And it just builds it. It's like, okay, let's get to the landing page. Can you design a concept of a landing page? And it has image generation capability now with text. So it like builds the landing page. How's this? Like, that looks great. Like, write me the code for that. We're gonna jump. And I, I'm just envisioning you're basically just sitting there and just doing the campaign. And I think that's the spirit of the concept here, is that you're just kind of feeling it as you go and you're like, you have an idea and you visualize it and it's like, that's cool. And like maybe three months from now we can do video with it. Like, hey, knock me out. Like a, a 30 second trailer for this idea that I can use to put on X or LinkedIn and it creates a trailer. And I, I don't know that that isn't a Thing, like, I, I do think that I could see people who are at the frontiers here and really understand the capabilities of these tools. You could see doing this where you used to need five people like, you just, you just do it like that. I was listening to a podcast with Sam Altman, a YC podcast he did with. Oh, shoot, what's the guy's name? Gary Tan. The president. Yeah. And he. It was in November and he's like, with AI, you can, you can just do things. And I, I think that's the spirit here, is like, when you know what these things can do and you want to do a campaign, you can just do things. Like, you, you can just go in and design it and develop videos and write copy and create landing pages and build paid ad copy and like social media shares. Like, you can do all of it. And again, this goes back to. What does that mean? I don't know. I just know you can. And like, if I had to do something tomorrow to launch something and our team was like stretched and they couldn't do it, I could sit down for an hour tonight and do everything. I just outlined for people so someone on my team could do it. But, like, as the CEO, if I just need to do something, I just go and do it. I did it last night. I did this crazy research project. Like that, no joke, would have taken me, I don't know, like three days. I did it while I was like getting my teeth brushed. I started the research project in the Chat GPT app. I said, here's what I need to do. I need to prep for this meeting tomorrow. Here's everything I need to know and like, go. And I brushed my teeth, went and put my daughter to bed, Came back, it wasn't done yet. Went and checked on my son, came back, laid down, Boom. I. I have this research support. It was crazy. So I think that's the spirit here. I don't know that I'm like, in love with the vibe marketing, like, name, but I get what they're saying and I guess it gives a name to this thing.
- [44:19]
Mike Caput
Yeah. It feels like almost like the ultimate triumph of the stereotypical idea guy. Right?
- [44:26]
Paul Raitzer
Yeah.
- [44:26]
Mike Caput
It's like, hey, let's come up with a bunch of ideas and then suddenly they can actually all get done.
- [44:31]
Paul Raitzer
Yeah. The idea people are now the creators and the developers and the. Yeah.
- [44:37]
Mike Caput
Next up, Elon Musk's X AI has acquired X, the social media site in an all stock transaction that values Xai at $80 billion and the social media platform at 33 billion. This merger officially combines two companies that were already pretty deeply interconnected behind the scenes. So the rationale behind this focuses on the overlap between the two companies. So X provides a massive stream of conversation data that can be used to train AI models. It also is a built in distribution network for xai's GROK chatbot. And the combination creates one of the few foundation model companies with a widely used consumer facing product. Though analysts note that GROK still lags behind competitors OpenAI, Anthropic and Google in certain areas of state of the art performance. So this move effectively transforms investors in Musk's original Twitter acquisition into shareholders of xai. And it also formalizes what was already happening informally, the sharing of data talent resources between the two companies In Musk's push to remain a leader in AI. Some people point out this may also not be the final consolidation that Musk is looking at. Tesla, with its fleet of approximately 5 million vehicles collecting multimodal data represents an even more valuable data source that could eventually in some way be integrated with xai's operations. So Paul, like you followed all these companies closely, credit where credit is due. I feel like you predicted this like two years ago, that this was something like this was the overall play, like what do we need to know here?
- [46:21]
Paul Raitzer
Yeah, so I, I, I think this is the most predictable outcome in business ever. So when he bought Twitter, recall he didn't, he tried to back out. So like when, when Musk bought Twitter, he kind of half jokingly, you know, made the offer and then he tried to back out claiming it was like because of bots and all this stuff and so he had to buy them. He was forced to buy Twitter and then proceeded to tank it like it was, I don't know what the valuation was before this event, but it was under 10 billion. So you go from 44 billion to 10. Now he borrowed a lot of that money, so he owed people those billions or tens of billions of dollars. Well how do you get out of it? You create an AI lab. Because what is the most valuable asset in the world right now, maybe besides Nvidia chips and data centers, it's to own an AI lab. And so they're worth 20 billion more than Anthropic now based on the, this data with no revenue. Like so this xai, it's, there's no revenue, they have Grok, but like they don't have what OpenAI has in terms of the growth. And so the only way out of this was to do this exact thing. It's, we Always knew that X became the training source for xai. But to do that, I would imagine legally you need them to be the same company, otherwise you're just, I don't, I don't know how it was working before. Maybe they were licensing the data to them. So yeah, it's, it makes total sense. I, it's probably a smart move. I think you're just basically making up a number to like make the investors and people you owe money to whole and you just roll on. And so it's just, it's such a weird world where like tens of billions can just get like thrown around and put on paper. It's like, it's worth, it's worth 44 billion. It's like, okay, like Xai, I guess it is because of the data source, but pretty wild. So yeah, this, and this was like, what was this? Like Saturday night or Sunday night? This was like a late night thing. And he just announced it on Twitter like, hey, by the way, bought X from xai. Bought X. And yeah, crazy. So I'll see. I. For you, the average person at the moment, it just means that if, if you're all your data from X wasn't being fed to X, it is now. That's pretty much anything you've ever said publicly or I assume in dms. Like it's all training data now for xai.
- [48:43]
Mike Caput
Anthropic has just released some research that gives us a peek under the hood of large language models like Claude, offering some insights into how these AI systems actually quote, unquote, think so. The company published two new papers that are focused on what they call interpretability, which essentially creates what they call like an AI microscope to examine the billions of computations happening inside these models. This research actually addresses fundamental questions about AI cognition that have remained mysterious until now. So some examples of how it is working kind of under the hood. When Quad communicates in multiple languages, is it using separate language systems or thinking in some universal mental space? We didn't know the answer to that question. When it writes poetry that rhymes, is it planning ahead or just making it up word by word? When it explains its reasoning, is it showing its actual thought process or sometimes constructing a plausible sounding explanation after the fact? The findings are surprising even to the researchers themselves. It turns out Claude often thinks in a shared conceptual space across languages, suggesting it has developed a kind of universal quote language of thought. When writing poetry, for instance, it actively plans ahead, thinking of potential rhyming words before crafting lines that lead to those endings that contradicted the researcher's initial hypothesis that it would simply proceed word by word. And perhaps most intriguingly, the research confirms that AI models can sometimes engage in what you might call BSing, providing plausible sounding explanations that don't represent their actual internal processes. In one example, when given an incorrect hint for a math problem, Claude was caught in the act of fabricating reasoning to match the expected answer rather than working through the problem logically. So Paul, we've been saying for years, but it bears repeating often this is your regular reminder that even the people building these models do not fully understand how they operate. So honestly, it does seem like this research should be a bit of a big deal if we're able to better start understanding what goes on inside of it.
- [50:56]
Paul Raitzer
Yeah, this research totally went under the radar. If you think 2.5 from Google went under the radar, like, this is wild stuff now they've been working on this, we've covered this stuff before. I think the term was mechanistic. Interpretability, I think is like the technical term they use for this stuff. So anthropic has been pushing on this. I know Google does research like this. I'm sure OpenAI does this. Like everyone's trying to figure out how these things think, why they do what they do. Interestingly, the Sam Altman podcast I mentioned with Gary Tam was he told the story of like why they built GPT1 and it was actually because this internal researcher was looking at like it was like Amazon product reviews and a neuron in, in the system was flipping like on and off related to sentiment and they couldn't figure out why it was doing it. Like it was it understanding sentiment even though it hadn't been trained on it, I think was like the concept and that led to them actually pursuing the path of building GPT1. That wasn't what they set out to do originally. And so this whole idea that like these models just do things that we don't understand and sometimes it leads to an entire research path. And so I think that that's what this research is demonstrating for people who haven't been following. Is this reaffirming the fact that we don't know how they do what they do? And while there's people like Yann Lecun who think these things, probabilistic machines just making token predictions and that's all they do. You look at research like this and you're like, are we sure that's all they're doing? Because it sure seems like there's something else going on in these models. And so I Think it's fascinating to like follow along with this research. And I, I, I mean I, I love these kinds of papers because it does give us a window into how it works. And the other one that goes back to that Golden Gate Bridge thing I think we talked about in the fall where they found a way to like get the thing to just tie everything back to the Golden Gate Bridge. Like they found the neuron that was firing basically that was causing it to do this thing. And so yeah, it's, it's such a, like a open ended research area where there's just so few answers right now.
- [53:04]
Mike Caput
Next up in rapid fire, the CEO of Replit, Amjad Massad has made some waves in AI circles by saying in a recent interview, quote, I no longer think you should learn to code. So Replit software uses AI to automate and augment coding work. And Mossad has long been a proponent of using AI to massively increase the leverage that great programmers have. And he advocated for a while learning at least how to do some coding in order to build even more with AI's help. Now in this interview he says he's really starting to believe in agents and a path where they optimize to get better and better. And that in turn has altered his opinion from even a year ago when he was recommending that people learn to code even a bit. Now, not anymore. He says instead you should, quote, learn how to think, learn how to break down problems, learn how to communicate clearly. He then said in a follow up post to the interview, quote, I understand all the cope. It was hard to arrive at this conclusion. There are obvious domain exceptions, but the trend is hard to miss. In my work, I've popularized learning to code more than anyone else. A good chunk of my life's work. Bittersweet. Okay, Paul, you and I are not programmers, but we're talking about this because Amjad is the CEO of a major AI company. He said he spent a long time like his whole career arguing this opposing view. But now it seems like he's convinced of a future where something like learning to code doesn't make as much sense as maybe prioritizing other skills.
- [54:40]
Paul Raitzer
It sounds like, yeah, this is one of the great unknowns. I mean, just because this is his opinion doesn't mean he's right. He is someone who's very thoughtful about this and has a company where, you know, his goal is to build like a billion developers. So he wants everyone to be a developer. I, I, I mean I've, I've met him, I, I've talked with him. I, I, I think he's as unbiased as one can be when this is your business. So I don't think he's doing this for hype. I don't think he's doing it to sell more subscriptions to Replit. I would imagine he truly believes this. And there are a lot of people who don't, like, there's a lot of people on the other side who still see the value in coding. And I think it's just representative of where we find ourselves. You're going to find experts on any side. So, like I do the AGI podcast, you're going to have some people are like, you're crazy. AGI is ridiculous. It's not a thing. It's not going to happen for 10 years, if ever. And then you're gonna have other people who say two years and they actually think two months. Like, there's, it's, it's all over the board. And like. Yann Lecun is so strong in his beliefs that language models are not the path to intelligence. But Yan also was very strong in his beliefs. Yann Lecun, the chief scientist of meta, chief AI scientist at Meta. He was also extremely strong in his beliefs that back in 2016, that AI couldn't win at the game of Go. And he was wrong. Like, people are wrong sometimes. Jeff Hinton is, like, so convinced that AI is going to destroy the world that he left Google and is like, making his life's work to dismiss his previous life's work and say, we went the wrong way, we shouldn't have done what we did. Like, people have opinions. And I think that's the whole goal of our show, is to share those opinions with you, share those perspectives so you can figure out your own perspective on this. I have no idea if he's right or not. Like, if my son was a senior in high school right now and he wanted to go into coding, right. I don't know enough to say, don't do it. I would just have this perspective in the back of my mind and make sure we're thinking, thinking about that as we're making these decisions.
- [56:37]
Mike Caput
McKinsey has released its latest State of AI report, examining how organizations are restructuring to capture value from AI. So this research is based on a global survey of nearly 1500 participants across 101 countries. And it reveals that companies are beginning to make organizational changes designed to generate future value from generative AI. So they find that the adoption of AI continues to accelerate dramatically, with 78% of respondents now reporting their organizations use AI in at least one business function. That's up from 72% in the previous survey and just 55% the year before that. Generative AI usage has similarly jumped to 71%, with companies most frequently deploying it in marketing and sales, product development, service operations and software engineering. Despite this rapid adoption, the survey finds we're still in the early stages of organizational transformation. Only 21% of companies say they have fundamentally redesigned workflows as they deploy AI, and less than one in five say they are tracking KPIs for gen AI solutions. Interestingly, larger organizations appear based on their data to be moving more quickly than smaller ones, with companies exceeding 500 million in annual revenue, more than twice as likely to have dedicated roadmaps to drive adoption of Genai solutions and dedicated teams to help drive that adoption. So Paul, I know we wanted to talk about first the recency of this data, but also maybe touch on that for us and the overall takeaways you found in this research.
- [58:19]
Paul Raitzer
Yeah, I think there's a lot of information here that supports a lot of the things we talk about, you know, just in terms of the early stage of adoption, the lack of education and understanding within companies. So I think it's a worthwhile report for people to read, give it a download, check it out. They do a nice job of summarizing the findings. We'll put the a link in the show notes. It is interesting. Like I, I've always said on the show anytime we talk about research it's always like go to see how it was done. Who, who, who did they interview, when did they interview them, that kind of stuff. And I did find it interesting when I went to this I thought like oh great, this is like brand new study, like it's going to be super relevant. And the survey was from a two week period in July of 2024.
- [58:58]
Mike Caput
Right.
- [58:58]
Paul Raitzer
And I thought that's odd, like why would you wait eight months to release a state of AI report report? Which honestly like made me think about the role AI will play in research reports in the future.
- [59:10]
Mike Caput
No kidding.
- [59:10]
Paul Raitzer
Because like why, why wouldn't you just take all the data and, and either train a model to, to do this analysis so you don't wait eight months to release it or at least like accelerate the review of the data. Like that's how we're doing it. Like the way we're going to turn around a survey in two weeks instead of eight months is by infusing AI into the process and helping Mike and I do this way faster so we get more relevant data out. So yeah, I don't know again, I guess. Great report. Read it.
- [59:40]
Mike Caput
Yeah.
- [59:41]
Paul Raitzer
Secondary note, maybe an example of how AI is going to speed some things up in organizations.
- [59:47]
Mike Caput
Yeah. And we, and we'll share more once we publish our report, but we're doing even more this year with that, which would be really cool. I mean, just looking even at the last year when we published the report, we used AI quite a bit to accelerate it. But it's night and day now. What we can do.
- [60:01]
Paul Raitzer
Yeah.
- [60:03]
Mike Caput
Next up is a fascinating inside account has just been published by the Wall Street Journal, revealing some more details behind the dramatic November 2023 firing and rapid reinstatement of OpenAI CEO Sam Altman. So this article is adapted from an upcoming book by Wall Street Journal reporter Keech Hagee, and it sheds new light on what happened during those chaotic five days that briefly upended the AI industry's most influential company. So apparently, just days before his sudden ouster, Altman was warned by, of all people, venture capitalist Peter Thiel over dinner in LA that the, quote, AI safety people at OpenAI would, quote, destroy the company, echoing concerns about effective altruism advocates who worry about AI risk. But ironically, according to the article, it wasn't ideological differences that led to Altman's firing. It was something much more mundane, governance issues and management style. The real Trouble began when OpenAI's nonprofit board started discovering what they perceived as a pattern of deception by Altman. Some of the most damaging testimony came from within Altman's executive team. CTO Miro Murati had privately shared concerns about Altman's toxic management style, documenting instances where he allegedly misrepresented safety approvals and pitted seniors employees against each other. These complaints, combined with board members catching Altman in what they said were direct lies, ultimately led them to vote to remove him. Now, what they didn't anticipate, as we've covered before, was the massive employee backlash. Within days, almost the entire company had threatened to quit unless he returned. And even more, both Murati and Sutskever Ilya Sutskever, formerly at OpenAI, who had both provided evidence against Altman, ended up signing the letter supporting his reinstatement. So this book should be an interesting read. Paul. It's not the only book coming out about the inside story at OpenAI either. We also learned that journalist Karen Howe, who we know well, has announced pre sales of her book Empire of AI Dreams and Nightmares in Sam Altman's OpenAI, which relies on seven years of her reporting to tell that story. We always knew there was all this, like, deception and drama going on. We covered Gosh has a ton of it. Does what we're learning now about it surprise you at all?
- [62:33]
Paul Raitzer
No. And, yeah, I think obviously there's just a lot more coming. I think Sam is. Every podcast that I have listened to Sam on, which is probably over a dozen since this all happened, he gets asked this question about his fire.
- [62:46]
Mike Caput
Right.
- [62:47]
Paul Raitzer
It's always the same response, like, okay, I. I will answer these questions again. Every once in a while, he, like, lets the guard down and provides a little bit more perspective on it. I. I mean, my general take is that there. There is how Sam has viewed these things, and then there's how others viewed these things. And he's been pretty consistent that, you know, there's probably things he could have done different or better, which may be the things that are being highlighted here. You know, if he looks back, does he think they were worthy of him being fired and humiliated and going through that craziness? Probably not. But he also generally just takes the high road, and it's like, man, we've learned a ton, and I gotta keep moving. That's the thing. So I'll read the books. Like, I mean, it's fascinating to see and to hear more about what happened, but I don't. I don't think it, like, changes anything moving forward. I think they're pretty focused on the future. And, Yeah, I don't know. It's always. It's always intriguing, though, to get a few new insights. Like the Peter Thiel dinner thing I had not heard. There was definitely pieces of this I was not aware of.
- [63:49]
Mike Caput
Well, it could be one of those things, too. Almost like the Ghibli thing, where it just becomes wider knowledge of kind of everything that's been happening within these companies. Almost like the Social Network movie or something about Meta as well, you know?
- [64:00]
Paul Raitzer
Yep.
- [64:01]
Mike Caput
All right, next up, Runway has introduced Gen4, its latest AI video generation model that Bloomberg says, quote, challenges OpenAI Sora with more cohesive videos. So Gen4 is a next generation AI video creation system. It represents a significant leap forward, addressing one of the most persistent challenges in AI generated video, which is consistency across scenes. This new model introduces what Runway calls, quote, world consistency, allowing creators to maintain coherent characters, locations, and objects through an entire project. So, typically, past video generation systems have struggled with maintaining that kind of visual continuity from one clip to the next. It's also able to work with minimal reference materials. According to Runway, the system can generate consistent characters across multiple scenes using just a single reference image. They also offer some impressive capabilities for object consistency, allowing creators to place any subject in various environments while maintaining its core visual characteristics. So you can start thinking of this as applying to things like film and storyboarding to kind of capture multiple angles of the same scene by having all these references be consistent across each frame. So Paul, definitely seems like we've alluded to things are moving really fast in visual AI.
- [65:27]
Paul Raitzer
Yeah, video is one of the things I talked about on the AGI podcast last week is just you're going to see these rapid improvements in this space and consistency, length, things like that. And you know, I think for Runway, we've talked a lot about them at least last year we covered them quite a bit. Their, their CEOs on record is saying like their goal is to do a feature length film from a single prompt. Like they're, they're not stopping at like 10, 15, 20 second clips here. And you know, I think that they play an interesting role in the creative space and the impact on creatives. They've tried a lot to integrate creatives into what they're developing and let them use those tools. I do think, I mean, if I had to put some money on who's going to get acquired this year, I would put Runway pretty high on that list because I think what's going to happen is video is going to become so integrated into the AI models platforms themselves that sustaining a standalone video gen tool, like, I don't know that they can build the kind of customer base they're going to want to build once I just have SORA baked right into my thing. I've VO2 baker into Gemini. So I, I don't know. I. Good company. We've been following them for six years. I, I could see them being an acquisition target for sure for somebody.
- [66:42]
Mike Caput
Yeah, I have no idea as to who that would be. But it does occur to me there are two AI companies that also have TV studios and film studios, which are Apple and Amazon.
- [66:52]
Paul Raitzer
So those are interesting ones. And, and two that don't have video yet, which is XAI and Claw and Anthropic.
- [67:00]
Mike Caput
Right?
- [67:01]
Paul Raitzer
Yeah. That's interesting. We could probably do a whole episode thinking about that one.
- [67:07]
Mike Caput
All right, next up in rapid fire, Microsoft has unveiled two powerful new AI reasoning agents for their Microsoft 365 copilot platform. These are called Researcher and Analyst. Researcher acts as an on demand research assistant. It tackles complex multi step projects with improved accuracy and insights. It's built on OpenAI's deep research and enhanced with Microsoft's orchestration and search capabilities. So it can do things like develop market strategies by synthesizing internal company data with competitive information from across the web. Researcher can also integrate data from third party sources and the second agent analyst functions like a skilled data scientist. It transforms raw data into actionable insights within minutes. It is powered by OpenAI's O3 many reasoning model, and it uses chain of thought reasoning to work through problems incrementally, similar to how humans do analytical thinking. It can run Python code to handle complex data queries in real time, allowing users to verify its work as it processes. Now both of these agents will be rolling out to Copilot license holders in April through a New Frontier program designed to give customers early access to developing innovations. So Paul, this like seems like a really cool update to Copilot. I guess I immediately think of like the many, many knowledge workers I talk to or work with who use Copilot, and I just hope these kind of come with adequate education because I don't know if out of the gate, if you show me this announcement and I'm a Copilot user, I'm going to immediately kind of get how do I use these tools?
- [68:50]
Paul Raitzer
Yeah, they probably won't make a history of all these companies as anything. It's like, here's some really powerful tools figured out. My mind immediately went to like, when are they going to launch the accountant and the wealth manager and the marketer and the writer? Yeah, it's a slippery slope. It's a hard position for these model companies to be in where, you know, you have the ability to build these tools that do jobs. You know, collection of tasks, a large collection of tasks that make up a job and how they'll be received. They can be nice complimentary tools that help you do your job. They can also be viewed as replacements. I don't know. I think we're going to see a lot more of these this year and even go back to that vibe marketing. If you just like lumped everything I said in that example into like a marketer co pilot, like couldn't, couldn't you just bundle it and know it as those capabilities? Yeah, I don't know. More questions than answers I have.
- [69:53]
Mike Caput
All right, so our last topic today is our recurring segment we're doing on listener questions where we are answering all the questions that come up from listeners or from audience members in different contexts on different webinars. We're just kind of cherry picking some that jump out as really helpful, possibly for the audience to get an answer to so this week's question, someone said, I want to master prompt engineering, but now that models are able to create prompts for you, is this even going to be important in 12 months?
- [70:28]
Paul Raitzer
Yeah. So in 2023, pretty early on, I was trying to look out and say, like, dude, is prompting like a thing? Like, isn't the model just gonna like write the prompts or improve your prompts? And I, all I can say is we're like a couple years into this and prompting matters still. Like, you know, Mike does demos on this all the time, runs classes on it where you're showing like new prompting techniques for reasoning models, for example, or prompting techniques for image generation models or video generation models. Like, yes, the model companies are probably taking and improving your prompt and not showing it to you. They're like rewriting it, making it better behind the scenes. But your ability to know what the system's capable of and convey what you want to convey, like, what is the goal? What is the output I'm looking for, that stuff still matters. Like it, it definitely I think of it as a skill. And like when we're interviewing people, you know, for, for roles in our company, I want to know their prompting skills. Like, I want to know these things. So I, I, I would, I would encourage universities, high schools, I would be teaching prompting as a skill. I, I don't think that that's going to go away. I think the systems will get better and better at helping you. But I do think that knowing how to talk to these systems is going to be a required part of everybody's job moving forward.
- [71:38]
Mike Caput
Yeah, absolutely. All right, Paul, that's a wrap on a busy week. Just a quick reminder for everyone. Again, go to state of marketingai.com to take the survey that takes just a few minutes to help us with this year's State of Marketing AI report. You can find the link right on that page along with a copy to download download of last year's report and check out the Marketing AI Newsletter, marketingai institute.com forward/newsletter where we wrap up all of this news from today's episode as well as all the stuff that didn't make the list, which is always lots of really interesting news we just didn't have time for. Paul, thanks again.
- [72:17]
Paul Raitzer
Thanks, Mike. It was good to be back together after my solo session and we will be back next week with another regular weekly episode. So thanks everyone for joining us. Thanks for listening to the AI show. Visit MarketingAI institute.com to continue your AI learning journey and join more than 60,000 professionals and business leaders who have subscribed to the weekly newsletter, downloaded the AI blueprints, attended virtual and in person events, take in our online AI courses, and engaged in the Slack community. Until next time, stay curious and explore AI.