← The Artificial Intelligence Show
The Artificial Intelligence Show
#125: GPT-4o Update, Sora Leak, Perplexity Shopping, Elon Musk vs. OpenAI & US Government’s AI Manhattan Project
01:34:55
Loading summary
Transcript107 lines
- [00:00]
Paul Raitzer
I do think that one to two years out, this capability of the AI having access to all the apps and your devices and being able to connect and take actions in them is going to become extremely commonplace. I was trying to think of an analogy for it and it's almost like people take for granted the AI capabilities within your photo apps, the ability to just filter things. And I feel like that's kind of how agentic AI is going to be on your devices. Like you're just, it's just going to be there and you're really not going to think about it. 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 125 of the Artificial Intelligence Show. I am your host, Paul Raitzer, along with my co host Mike Put. We are back after what feels like a month away, but it was really only a week. So back from the Thanksgiving holiday, rested, but also slightly overwhelmed by all the AI news we have to cover. So Mike and I decided we're going to do sort of a special episode here rather than our traditional three main topics and seven to ten rapid fire items. We're just going to go straight rapid fire through this whole thing. We've got. I don't know the, the. So I, I've mentioned before, the way we do this prep for this podcast is we keep a zoom chat that, that Mike and I can post in throughout the week and in a normal week we'll have, I don't know, like 25 on average items. I think we were at about 50 coming into this one. We were, yeah, yeah. So Mike did the, the work on Sunday of trying to organize this into something manageable to get it done within a reasonable time period. So there is a lot to catch everybody up on, including ourselves. And so we're going to do rapid fire only. We're going to move pretty quick through this, but there's a lot to, to catch up on and a lot of it I think is sort of laying the groundwork for some big things that are maybe still coming in December from some of the frontier model companies and definitely some look ahead into what's happening in 2025. So this episode is brought to us by our upcoming webinar, this is happening on December 17th, called the AI forward CEO unlock the power and intelligence of a CO CEO custom GPT. So again, this is happening December 17th. It will be at noon Eastern time. CO CEO. I've mentioned this on the podcast a number of times. This is something I built for myself. I'm actually working on a project, I'll explain more during the webinar called the AI Forward CEO. This was actually an internal project for me that I've decided to just make public and see if it can help other people. But in essence, I started assessing my job as a CEO. This goes back probably three or four months ago. And I was trying to figure out how to free myself up to focus more on the big things I wanted to work on like special projects, vision strategy, performance management, and get out of all the other things that were sucking my time. And so I started building a list of all the things I do. I was actually using Jobs GPT to help me with this initially. And in the process I started setting out my own vision for how do I become an AI Forward CEO. What is the process? And so in developing this approach, again, this is not anything I ever planned to even talk about publicly. It was just something I was going to do myself. I built this co CEO a custom GPT for me that was helping me start to manage my life as the CEO of two companies. And so I built it as a strategic advisor and I designed it to do five specific things. Analyze data, execute tasks, solve problems, build plans and innovate. And as I started using, I realized how valuable it was becoming to me. Like daily I was having conversations with it about different things. I was trying to think through business planning. Like I was working on different things for our online education, for our events. And I would just go in and talk to it. But in the custom instructions, I told it like you're, you're have expertise in marketing, sales, customer service, product operations, hr, finance, legal. Because as a CEO you touch all of these things. And so I just wanted something, I want to say someone, but like something that when I'm laying in bed at 11 o'clock at night trying to figure things out and I don't have anybody to talk to. I wanted to be able to talk to something. And so that's how I designed this thing. And so I'd mentioned it a couple of times on the podcast and we've had dozens of people reach out to us asking like, hey, is this something you're going to make available? So what I have decided to do is I'm going to create a public version of this. So the private version is trained on a bunch of, like, private proprietary information about our company, our financial performance, all that kind of stuff. So I'm going to release a public version that doesn't have all that in there, but that gives people a sense of how this thing works. So I'll preview that as part of this webinar and then I'm going to actually give people a prompt template for the custom instructions that I use to build mine. And it'll show you how to build your own and kind of inject your own proprietary information. So in the webinar, we're going to go through what it is, how it works, go through example use cases. I'll show you how I'm using it every day, and then teach you how to build your own. And this isn't just for CEOs. I would say this is helpful for anyone that's involved in transforming their business, trying to overcome business obstacles or that just wants a CEO mindset. So, like, we're going to share this. The internal version I created, we're going to share with our team, and then that way they can actually, like, interact with it in, in some ways, like how they would approach things with me and so kind of know how I might think about things. So anyway, you can go to SmartRx AI again. That's our AI research consulting firm, sort of the sister company to Marketing AI Institute, if you're more familiar with Marketing Institute. So SmartRx AI. And at the very top there is a ribbon or a banner that has the webinar link to it. You can click Register now. It'll take you to the landing page for that if you're a Smarter X AI newsletter subscriber. So the Exec AI Insider newsletter that I send every Sunday, there will be a link for the webinar in the upcoming issues of that and then the show notes will have a link to the webinar as well. So again, Smarter X AI, click at the top of the page to join it. December 17th at noon Eastern time. It is a free webinar teaching you all about co CEO and how to build your own. All right, Mike, so let's get into the rapid fire, man. We got it. We got a ton to get through.
- [07:07]
Mike Caput
All right, we're gonna hustle here, Paul, but we've got a ton of topics. First, one, big news from OpenAI they announced a big upgrade to their flagship model GPT4O, which of course powers ChatGPT. A post on X from the company on November 20th. Again, we're going through a couple weeks of updates here. Made this announcement saying, quote, GPT4O got an update. The model's creative writing ability has leveled up. More natural, engaging and tailored writing to improve relevance and readability. It's also better at working with uploaded files, providing deeper insights and more thorough responses. So as of our recording today, the new version is tied on the popular AI leaderboard maintained by lmarena AI. It is tied with Google's latest experimental version of Gemini for the number one spot. So it seems like the upgrade, at least in the short term, did some good. Kicking it up the rankings now, Paul. It's kind of interesting to me to see this model getting specifically better at creative writing and just contending with Google's experiment experimental model directly.
- [08:21]
Paul Raitzer
Yeah. So there's very few details about this. It is interesting that they explicitly said this is for creative writing is the main improvement. They did one post on Twitter or X that I I could find there is. So if you ever want to follow along with what OpenAI is doing, they do have a model release notes section on their site. So we'll put the link to the show notes for this one. It was very direct. It just says November 20th I think is the date. So it's kind of like journal entries almost. And then you had highlighted it Mike, but all it says is we've updated it for all paid users. This update includes improved writing abilities that are now more natural, audience aware and tailored to improve relevance and readability. And then the uploading files providing deeper insights kind of thing. Now, it's interesting to note this comes on the heels of the Canvas launch. So on October 3rd they introduced Canvas, which in the post they introduced it, they said, we're introducing Canvas, a new interface for working with ChatGPT on writing and coding projects that go beyond simple chat. Canvas opens in a separate window, allowing you and ChatGPT to collaborate on a project. This early beta introduces a new way of working together, not just through conversation, but by creating and refining ideas side by side. It suggests edits, adjusts lengths, adds final polish, adds emojis. I've played around with Canvas a little bit, but it's very obvious that all of these things together are very much coming right at the writing profession. It can be seen as complimentary to for sure these are just tools, but there's definitely something deeper going on here. Now this likely has something to do. And this is just me like trying to perceive what's going on. This isn't, I don't have a source to cite for this, but it's likely it's something to do with the publishing deals that they've been entering into. Because what happens in these models is you do your initial training when you build the model, but then you do the post training where you're going through and actually improving it or making it, giving it specific expertise and capabilities. And so in the post training you can kind of tune the model based on expert data. So if they, if they come out with their usual 4o model, but then they go in and they take all these licensing deals and they fine tune the model to actually write more specifically, like, I don't know who they, off the top of my head, I can't think of who they have deals with. But think of like publishing houses, media companies, where there is access to licensed creative writing. Plus you hire expert creative writers who then go through and tune the model to have it right in in their style more that's likely. What's going on here is they're just fine tuning the base model with licensed data and expert trainers. Now the big question to me is the implications to writing profession and writers, which Mike, you and I both are, Jenny, is fundamentally changing the way we all create. The thing is, like, we don't know the impact this is going to have on authors, journalists, copywriters, but it seems like it's going to be significant. And as we've talked about many times on this podcast, one of the things this does is it democratizes creativity. And in this case, creative writing, like poor writers, can become good writers instantly by using these tools, good writers can become great writers and great writers can have superpowers. Now, what does that mean to journalism schools and media companies and publishers? I have no idea. But it's, it's very obvious that OpenAI is coming for that profession, for better or for worse. I don't like as a writer, like, what, what's your reaction to this? Like, what do you think when you see this?
- [12:06]
Mike Caput
Yeah, I think Chat GPT, at least from what I've heard people gripe about, at least people that don't use it deeply with a lot of examples, they gripe about like, oh, it's not that good a writer. Well, A, it can be if you give it the right examples. B, something like Claude is already where you want these tools to be. So as we improve ChatGPT, I think a lot of People are going to kind of take a step back and say, wait a second, it can write like this out of the box now. And that's going to be quite disruptive, I would guess.
- [12:37]
Paul Raitzer
What are you, Are you using Claud or chatgpt more in like writing assistance, editing?
- [12:42]
Mike Caput
I would say Claude. Usually it depends on the use case, but in terms of just like really quick off the top of my head prompting to write something, especially again based on examples. So what you provide does really matter here. I find Claude just more reliable at more nuanced writing that could change. I have to give this a test run now.
- [13:03]
Paul Raitzer
Yeah, I just don't use Claude that much. I mean, I have the account and I'll go in there and like test some prompts in there. But I would say I'm definitely more of a power user for chat GPT. I. I find it just easier. But it's also just what I'm used to.
- [13:14]
Mike Caput
Well, especially too with how much we've talked about and expanded our use of GPTs. This balance is quickly changing because if I'm doing the same thing over and over again, there's just no point into me continuing to like copy and paste stuff into quad.
- [13:27]
Paul Raitzer
Yep.
- [13:29]
Mike Caput
All right. Some other OpenAI news. OpenAI's new Sora video generation model appears to have been at least temporarily leaked on Tuesday, November 26, so a group calling themselves the Sora PR Puppets published a project on the AI platform hugging face that seemingly provided access to Sora's API. This had not yet been officially released to the public, and the leak actually allowed users to generate 10 second videos in 1080p resolution simply by typing text descriptions. Several users appear to have created and shared videos bearing OpenAI's watermark before access was revoked just a couple hours later. What's really interesting here though, is the motivation behind this leak. The group kind of released this almost like manifesto claiming that they acted in protest of OpenAI's treatment of the early testers. These are the red teamers and the creative partners given early access to the API. They allege that OpenAI is pressuring testers to present an overly positive narrative about Sora. They are failing to fairly compensate them for their work, describing it as, quote, unpaid labor for a company that is now valued at $150 billion. OpenAI responded saying that Sora remains in a research preview phase and they emphasized that artist participation in the testing is voluntary and comes with free access to the tool, along with support through grants and events. So Paul, what struck me about this were kind of two big things. First, at least one commentator who we take pretty seriously, Ethan Malik, said that the leaked videos are quote, in order of magnitude better than competing tools with the caveat if we are seeing true representations of what this can do out of the box. And second, it was pretty interesting to me that the motivation here was not was a moral stand against OpenAI's practices but, but they explicitly stated they had no problem with the technology itself.
- [15:35]
Paul Raitzer
Yeah, so it does seem like it's real like that the, you know, the videos we're seeing, the leaks we were seeing are real. The thing I find interesting is SO SORA was introduced in February 2024. So we're coming up on almost a year since they first previewed this and they released a technical report at that time. And so in the February 2024 blog post from OpenAI, they said we're teaching AI to understand and simulate the physical world in motion with the goal of training models that help people solve problems that require real world interaction. Soar as a text to video model can generate videos up to a minute long while maintaining visual quality and adherence to the user's prompt. Today, this is in February of 24. Sora is becoming available to red teamers to assess critical areas for harms and risks. We're also granting access to a number of visual artists, designers and filmmakers to gain feedback on how to advance the model to become most helpful for creative professionals. We're sharing our research progress early to start working with and getting feedback from people outside of OpenAI and give the public a sense of what AI capabilities are on the horizon. There was, you know, we talked about Mir Murati, the CTO who left that Sora, it was rumored, was part of the friction of her leaving. Whether she was against the release of it or, you know, things just weren't going as planned. But that was supposedly one of the friction points. You know, I, I don't know, like I think this is going to be a really big deal. I still think that there's. When you see these amazing previews, like even Runway, you know, we big fan of Runway and their Gen 3 technology with, you know, their video production, I just think it's, it's all misleading. Like I think people see these demos and think, oh my gosh, we're there, we're going to be able to create 1 minute long videos with a simple, single, simple text prompt. But it's just not where the text seems to be. Like there's just a lot of barriers to this working and consistently generating those kinds of outputs. Like even for me, when I go in and play around with Runway and see where it's at, I'll put a prompt in and it's like a useless output. Like it looks impressive for two seconds and then the other seven, it's like we can't do anything with that. And so it's like it's not very controllable yet. I think you have to still have deep expertise in probably video production to get value from these tools or you have to spend a whole bunch of time working with them. I don't know that Sora in 2025 or maybe still in 24 is going to have its chat GPT moment where it just works and you put in a prompt and you do get 10 seconds of incredible quality. Because honestly for me, like I feel like even their image generation tool isn't quite doing that yet and video is so much harder than image. So I don't know, I'll be really fascinated to see what kind of progress is made. Google's working on the same kind of technology stability AI is working on this kind of technology. So a lot of people are working on text to video. Meta is doing it. And I think 2020, 25 we're gonna see some advancements, but I just don't think we're gonna see like all of a sudden anyone could create a 60 second video in high definition that maintains consistency from frame to frame for 60 seconds. I don't think we're there yet from a technological perspective.
- [19:08]
Mike Caput
So next up, Perplexity, the AI powered search engine, just released a big new shopping feature. And this allows you to research and buy products right within Perplexity. This feature is called Buy with Pro, which lets Perplexity's paid users in the United States complete purchases directly through the platform. So the way this works is let's say you ask Perplexity a question that could be related to shopping. You still get all your natural language search responses like normal, but you may also now see product cards that show the most relevant items available for purchase and a bunch of details about the products. Now Perplexity actually says these cards, quote, aren't sponsored, they're unbiased recommendations tailored to your search by rai. Now, when you see a product you like, you can then use a one click checkout option right in Perplexity. To buy the product, you click a button that says Buy with Pro. Now right now you actually get free shipping on all Buy with Pro orders courtesy of Perplexity. The company is also rolling out Snap to Shop which is a visual search tool that lets users take photos of items they're interested in and find similar products. And they're also launching a merchant program. So this lets large retailers share their product specifications directly with. With the platform. Paul, I'm curious, kind of what you make of Perplexity's play here on the Surface. Seems like they're coming at a little bit both Google and Amazon.
- [20:44]
Paul Raitzer
Yeah, I very much thought of like Amazon Rufus when I was looking through this. So if anybody hasn't tried it yet, I think most Amazon users have this now. But you can actually chat with Rufus within the Amazon app. And I think on the website I do most of my Amazon shopping on my phone. So, um, but you can talk to it and like find products and have conversations with Rufus. It's like their Gen AI integration. So that immediately thought of that. Certainly, you know, it comes at Google. Perplexity appears to have just a very aggressive product roadmap. They're releasing stuff like crazy. The one thing I found was weird is like, I don't know, maybe I just scanned their blog post real fast and I missed something. But I couldn't figure out how to buy with Pro. Like I went and did a test in Perplexity and I was like planning a trip to Scotland. I'm. I'm not. I'd like to be. Click on my, my list and I want to see like what would pop up. And so it gave me its usual Perplexity responses and I said, okay, what gear would I need for this trip? And then it popped up with a bunch of shopping carts and there was like, I free with the. It was like, take me right to the vendor though. Or there was a Buy with shop pay and I was like, well, that's not Buy with Pro. My where's the Buy with Pro button. And so I had to go do searches. And again, I may just be missing some obvious thing here, but the way you do this is you go to your account, then you're going to have some links, you're going to click purchases and then you're going to click Get Started. And then once you click that, you can add your shipping details and your billing details.
- [22:13]
Mike Caput
Okay.
- [22:14]
Paul Raitzer
And I think that then turns on the Buy with Pro button. So again, Perplexity is moving really fast. Maybe they haven't built out their communications and marketing team really well yet.
- [22:25]
Mike Caput
Yeah.
- [22:25]
Paul Raitzer
But I would imagine if, you know, you're launching something like this, you want to make that really obvious how to actually do the thing. You're. You're proposing. So that's how you do it. Yeah, he said. Arvin said in his tweet the buy with Pro experience is currently available to us Perplexity Pro users to begin with. We will expand internationally soon. User author authorizes Perplexity to transact on their behalf and the checkout flow is accomplished with a mix of AI and human in the loop. And if they don't have a relationship with the vendor, then I think it just takes right to the vendor site and I assume they're going to get some affiliate kickbacks on those purchases, which would be a nice revenue stream. So yeah, I don't know. I mean, again, Perplexity is doing a lot of stuff. This was an interesting one. I'll be curious to see what kind of uptake they get on it.
- [23:12]
Mike Caput
All right, next up, Elon Musk. His AI company XAI, has finally confirmed that it has indeed raised $5 billion at a 50 billion valuation. This is something we've been hearing as a rumor for a while and the fundraise has some significant implications both for Musk's operations at X, formerly Twitter, and his increasingly escalating beef with OpenAI. So first up, according to the Wall Street Journal, the company told investors it had raised 5 billion at a $50 billion valuation. That's more than double its valuation earlier this year. You know, some well known names are involved including Sequoia Capital, Andreessen Horowitz, and this brings Xai's total fundraising to 11 billion this year alone. Now we're also interestingly seeing some reports that Musk is using XAI to pay back investors who funded his takeover of Twitter, which is of course now X. There's another report from the Financial Times saying that investors who backed Musk's $44 billion Twitter acquisition have been given a 25% stake in Xai. Basically this arrangement could help make these investors whole after they suffered significant losses on their Twitter investment. Twitter is dramatically declined as of right now in value by what Musk paid for it. Some estimates say by as much as 80%. And all of this comes as Elon appears to be, I guess, ramping up his war against OpenAI. He has filed for a preliminary injunction against OpenAI, Microsoft and several key figures at the company. It seeks this filing to prevent OpenAI from transitioning to a for profit structure. Paul, this funding's been rumored for a while, seems to be confirmed. I think the use of XAI to make Twitter backers whole is maybe a new piece of information, but if I recall correctly, this is something you predicted, Right, Yeah.
- [25:17]
Paul Raitzer
Is exactly what we think was going to happen. So the. I don't remember which episode it was. It probably would have been back in May when I think XAI launched that. I said this is what was going to happen. But it was very obvious because so Basically Musk pays 44 billion for Twitter. It's now worth less than 10 billion is the current estimate. So his AI startup is worth four times what Twitter is. So what I assumed back then was that the AI Frontier model company would be worth way more than a social network. So when he first announced it and they raised 6 billion, I think was the first round at a 24 billion dollar valuation, meaning it was already worth more than Twitter was, but that Twitter X was essential to that valuation because what they had, he had shut off access through the API to Twitter's data when he bought it, which meant he was hoarding the proprietary data that lived within Twitter. Which is the value in part of XAI is that they now had access, they exclusively had access to Twitter's data, plus they had Twitter's distribution. So anything they built with GROK they could immediately put in the hands of hundreds of millions of Twitter users. And so that was why I assumed that he was going to make people who invested in Twitter whole. He was going to basically run the value of Twitter into the ground. It didn't matter anymore because it was XAI that was the future of everything. And Twitter was just a feeder of data and distribution for it. So it ends up that does appear exactly what they're doing and why he doesn't care that they're losing advertisers on Twitter and he was able to leverage it to influence the, you know, the elections and, and get leverage with the Trump administration and he can use it to increase the value of X. So again, seemed obvious to me and May that's exactly he was going to do and it does appear that's what he's doing. Regarding the OpenAI injunction, the thing I keep taking away from this is it is just very apparent he intends to make life miserable for Sam Altman and now Microsoft and other OpenAI leaders and that he's going to use his newfound influence with the incoming administration to do just that. Now, I have no idea the legal grounds of the injunction and if it could be granted. Certainly, again, like without getting political here, depending on the judges who are involved or the judge who is involved and who put that judge in office and the leanings of that judge, he, he may get a favorable ruling. I have no idea But I think the whole point is he's going to spend as much money as he needs to and he's going to force SAM, OpenAI and Microsoft to spend a ton of money and time and just make things insanely complicated for them. So whether he ends up winning this or gets the injunction or whatever, I don't even know that that's the point of it. I think the point of it all is just to make life miserable for them and make this really complicated and give Xai more time to build what they're building and catch up and then surpass OpenAI. I think that's his whole intention is just muddy the water so he can do what he's doing and slow them down from being successful.
- [28:30]
Mike Caput
I have no doubt there's a huge competitive, I guess, play here to, you know, muddle up what they're trying to do. But also, I just increasingly feel this is very, very personal. It feels like Peter Thiel destroying Gawker. Yeah, they leaked details about his personal life. I'm just like, I view Elon Musk right now, I feel like as the biggest threat opinion.
- [28:54]
Paul Raitzer
Yeah, yeah, I, I, I do too. I mean, I, and I really think that OpenAI needs the government's support for what they're intending to do. And that's what I said. Like, you know, with the election and stuff, like when I put on who were the losers in the election, I was like, Sam Altman's at the top of the list. And OpenAI, like, it's, they're gonna have, they're not gonna have as clear a path to dominance as they, they had previously, that's for sure.
- [29:18]
Mike Caput
All right, so Microsoft just wrapped up its Ignite 2024 event and they made AI announcements at this event. They positioned Copilot as what they call the ui, the user interface for AI. And they talked a lot about the ability to create AI agents in Copilot. So CEO Satya Nadella shared this kind of UI for AI vision. And with that he kind of introduced a bunch of updates to Microsoft's AI capabilities. And these include things like something called Copilot actions, which are customizable prompt templates that automate repetitive tasks. There are new purpose built AI agents in Copilot that take on different roles like agent for hr and it questions one for meeting collaboration. Another is a project manager agent. Copilot Studio is apparently getting enhanced with autonomous capabilities. Again with the caveat, like we talked about in the previous episode, you got to take this all with a grain of salt. But Microsoft says that it will allow agents to take actions in the background without human prompting. And interestingly, that keynote actually concluded with Microsoft's commitment to AI education. They noted that they've helped train over 23 million people in AI and digital skills over the past year. One other interesting announcement to me, actually, that came out of the event is something called Interpreter for Teams, which apparently lets teams users clone their voices so they can have their sound alikes speak to others in meetings in different languages. So, Paul, what did you kind of make of the updates here? Like, definitely seems like Microsoft leads to maybe generate some excitement. In episode 124, we talked about the debacle of Business Insider. Basically doing a deep dive on all the ways that Copilot today appears to be disappointing people.
- [31:15]
Paul Raitzer
Yeah, I, I don't know. Like, I, I, I keep struggling with this race to keep releasing more and more features and capabilities because you can. And the tech is there. And on, on the surface, like, we take in any individual one, they sound awesome. But we still have like an adoption and integration issue in corporations. And so I don't know if like, the intended audience is just developers to go build more stuff or if they're actually trying to talk to like, corporate people who have to do this. Like a Chief Marketing officer. Yeah, and I feel like the people you and I interact with every day are overwhelmed by just the basic premise of generative AI. Like, yep. And, and when you start releasing all these other things, people are just overwhelmed by it all. And I don't know if Microsoft's missing that or if they're, maybe they're messaging that somewhere else in their sales cycle. But again, I've talked to plenty of corporations who have hundreds or thousands of Copilot licenses who have no idea what to do with it. And I don't know, like, I feel like all these features are great, but moving into 2025, we need a far greater focus from these companies on actual adoption and value creation. Like, we always talk about it. Find three to five use cases per person. Forget all these hundreds of other AI features. Just get them a custom GPT they can use like twice a day to save themselves an hour. Like, I just feel like that's the real opportunity in the enterprise that these big tech companies are completely missing.
- [32:54]
Mike Caput
Well, I think if I recall correctly in that Business Insider report, someone complained like, hey, I had this really useful feature in teams that like, transcribed dollar meetings. It was awesome. Legal wouldn't let us use it. Like, that's a really tangible, simple use case that's still running into major adoption issues.
- [33:11]
Paul Raitzer
Yeah. And I think like we'll talk about cohere. I believe in one of the upcoming rapid fire items here and I feel like cohere is taking a much more strategic approach. Writer comes to mind. Like I know we talked about them recently on their funding. Like I think these tech companies that are trying to get more laser focused on summarization, transcription writing assistant, like very like almost like the like the five to ten things that everyone in every company does and forget about all these long tail use cases like let's just nail these five to 10 things and you're going to cover 80% of the value creation from Gen AI. And I feel like we're just getting from all these big tech companies like Salesforce and Google to a degree and Microsoft like hey, here's a thousand other things you could maybe do this. And it's like we don't need the long tail yet. We haven't even solved the head of this yet.
- [34:05]
Mike Caput
All right, we've got some next step, some Google Gemini updates. So Google Gemini first has reportedly been rolling out a memory feature to at least some users, according to TechCrunch. So this feature, much like the one in Chat GPT, allows it to remember personal information. Think of like hey, I mentioned I had a food preference or details about my work so that it can provide more contextual responses. This feature apparently is available only if you are a Premium subscriber. At 20 bucks a month, it can be turned off at any time. It currently works only in English on the Gemini web clients. Second, a report from the Verge indicates that Google may be preparing Gemini to take action within apps. They are reporting that hidden in the Android 16 developer preview is a new thing called Quote App, the App Functions API that could give Gemini the ability to take direct actions within apps. So think about like for example, if you order food through doordash without ever touching the app, you would just tell Gemini what you want. Now Paul, we don't know if this rubric about app functions is the vehicle by which Google is making Gemini agentic, but we do know agents from Google are coming, don't we?
- [35:24]
Paul Raitzer
Yeah. Yeah. And I think it does remind me a lot of the Apple intelligence, like the exact same thing they're trying to do on iPhones, which they haven't obviously achieved yet if you've used Apple Intelligence. But I do think that one to two years out, this capability of the AI having access to all the apps and your devices and being able to connect and take actions in them is going to become extremely commonplace. I was trying to think of an analogy for it and it's almost like people take for granted the AI capabilities within your photo apps, the ability to just filter things. And I feel like that's kind of how agentic AI is going to be on your devices. Like you're just, it's just going to be there and you're really not going to think about it. But it's not there yet and it probably won't be next year in terms of reliability and scale. But yeah, definitely the same thing you're going to see on Apple devices in terms of memory. This is a really critical thing. We've talked quite a bit about memory on some recent episodes. Mustafa Solomon, the head of AI Consumer AI at Microsoft, I think is his title. He recently said in an interview that we're on the cusp of infinite memory, meaning these things are just going to be able to remember everything. I think the people, the thing that people have to keep in mind about memory is memory is going to be a choice, most likely, meaning you can get a very personalized version of Copilot or ChatGPT or Gemini. It's going to learn your history, your preferences, your searches, your prompts, everything. But you're going to have to allow that to happen because you are giving up a lot of privacy of data to enable memory that then enables the personalization of these gen AI experiences. And this gets us back to which companies do you trust with that? So perplexity is going to try and do it. Everybody's going to try and do it. If you're allowing it and you're letting it remember all of these things, you're. You're trusting that company with those memories. And that's something I don't think many people are talking about yet. We keep hearing about this and like, I know with OpenAI, you can go in and see the things it remembers and you can actually tell it to forget something. So like imagine real quick, like example, so if I worked at an agency, like I used to own a marketing agency, if I was using my Chat GPT account to do work for seven different clients in all these different industries, it's going to remember all these different prompts and it might get confused as to what industry I even work in because I'm constantly talking to it about all these different things, or if it's my personal life and I'm searching for something for my wife or my daughter, whatever, and it remembers things, but it doesn't know that that's not my personal Life, it's actually somebody else's. It creates all this confusion. And so the theory is you need to be able to control those memories, which just opens up a whole other world of things. So. And so memory. I think we'll talk a lot about memory moving into next year on the podcast because there's a whole bunch of layers to it that aren't being talked about yet. It's just being talked about as a technical thing right now. But it has far greater implications than the technical side.
- [38:27]
Mike Caput
I'm experiencing a version of this right now because my ChatGPT memory is full.
- [38:32]
Paul Raitzer
So really won't learn. I didn't know it gets full.
- [38:35]
Mike Caput
I didn't either. I just got a message the other week, and so I dived in because you can look at literally every line of memory it's saved. Half of it's really useful. Half of it's like thought experiments I was doing that aren't actually. I was like, oh, imagine we were gonna solve for this. And then it's like, memorizes that about my life. It's like, no, this isn't something. Or it'll be like really mundane stuff. Like, hey, Mike wanted to go cook that recipe last week. It's like, do we need to remember that in preference to other things? Now I'm like, okay, this is increasingly a useless feature because I have to go curate.
- [39:10]
Paul Raitzer
And. Yeah, and so you imagine, like, the research side of this, when you hear Mustafa Salman saying, we're gonna have infinite memory, what they mean is, technically, we're gonna have the ability to remember everything Mike does, and we're gonna build AI that then classifies those memories, prioritizes those memories, figures out which ones are relevant so Mike doesn't have to curate his own memories. Like, it's a whole field of exploration that they think is critical to AGI and true agentic behavior. But to the average user, we have no idea because it's not being talked about. But, yeah, that's. That's real stuff where you're. You're. Your experiences are now being personalized based on things you would never choose to personalize them on.
- [39:53]
Mike Caput
Right.
- [39:55]
Paul Raitzer
Wild.
- [39:57]
Mike Caput
All right, so another cool news item this week is that we just got kind of an inside look at Google's very popular AI tool, Notebook lm. So we actually got some inside info on how it works. Example, use cases, what's on the product roadmap. And this all came straight from the team building it. So there's a new episode of the Google DeepMind podcast out where host Hannah Fry interviews Raisa Martin and Steven Johnson, both at Google, who work on Notebook lm. And the interview gives you kind of this look at all this cool stuff about NotebookLM. Like, for instance, its audio overview feature is able to achieve, like, voice quality because it deliberately adds what linguists call disfluencies, which is like stammers, pauses, verbal quirks that make speech sound more natural. It incorporates sophisticated voice modulation. The hosts raise their pitch when they're uncertain or slow down for emphasis. So there's all these real subtle, like, human speech patterns. They also said, interestingly, this technology is not meant to compete with traditional podcasting. Even though you can create a podcast episode of your sources, it's designed for content that would never justify a full production, like summarizing meeting notes, creating an audio journal for your personal life, et cetera. They also, I thought was really notable, talked about a ton of different possible use cases, including a few in business specifically that jumped out at me. So, for instance, sales teams can use it to share and digest complex documentation. Teams can use it for weekly meeting summaries. You can use it on technical manuals and documentation to put that into more accessible formats. You can search through quotes and writing, whether it's your own or others, to inform new work. And I actually like this one as well. I didn't think of this. You can use it to analyze and review your own resume. So, Paul, we love NotebookLM. Getting a behind the scenes look here seems really valuable to me. Like, what were some of the most important takeaways for you here?
- [42:06]
Paul Raitzer
Yeah, I found myself being inspired to, like, spend more time with NotebookLM and really think more deeply about ways to integrate it into, like, my daily routine. Almost like the co CEO idea. I think there's a lot of ways I can do it. Integration. Quick note. So this podcast, if you've never listened to it, it's phenomenal. So Professor Hannah Fry is a British mathematician and author. She actually wrote a book I love called hello World, how to be Human in the Age of the Machine. That was from 2019. So this is the third season of the DeepMind podcast. I. I would suggest go back and listen to them and think about the time period with which these interviews are happening because they have incredible insights from the DeepMind team. So I thought the episode was great. You know, it talked about how they integrated the voice technology from Google Labs. They. They did. They were very direct and like, we can't take credit for the voice capability that came from a different Google Labs team. And then they figured out how to Integrate it into Notebook lm. They talked about use cases, whether or not it's a threat to podcasters and creators. And then they talked about their product roadmap. I thought one of the most interesting segments was when Johnson was talking about interestingness. And so he was explaining why the voice, the audio overviews work so well. And he said one of the key aspects is they, they program interestingness into it. And what, you know, Hannah Fry asked like, what exactly is that? And he said it's controlled surprise. So it's a. He went on to say, it's a great example of a convergence of three different technologies or breakthroughs that make something magical happen. He talks about Gemini itself and how it can do text, obviously extremely well, but that doesn't do voice. And so then they realized, like, you could integrate this and it would give the computers the capability to now do different things. So underlying is, is Google. And then they have these voices that, you know, in partnership with the other Google Labs team that are trained to extract the interesting components of anything. And they go through some really wild examples of, you know, being uploaded a two word document. I forget what the one. I think the one was just like chicken. It just said chicken like a hundred times, whatever. But it was structured like a research paper and they did like a 10 minute podcast on chicken. And like, what were these authors thinking, only using chicken? And they actually like find the interesting aspect of this. So I don't know, just really cool to hear them talk about it. You follow Rise on X. She's very forthright in sharing what she's thinking about the product and where they're going to take it. So I would definitely follow her on X as well. And then they talked about one of the features that's coming up that Steven Johnson was very excited about is customized hosts, where you're going to be able to, and this is there like 2025 roadmap, say, okay, I want one host to be an economist and I want the other to be a government regulatory person. And I want them to actually debate this topic about AI's impact on the economy or something like that. So you're going to be able to steer the overviews based on the Personas you give the hosts, which is a kind of a wild thing to start thinking about. But I do think that the, the Notebook Om, as I said when we first talked about, like this could be a standalone thing on its own. Like they could build this into a massive company if it was a standalone company. So as a product, I hope that they keep pushing on this as not only integrated into Google Workspace, which I think is great, but I think they're on to something here that that can become a very big product and if they do it right, they could get wide scale adoption within enterprises. But they've got to think deeply about that, that market. And right now I don't think they're really doing that yet. But I think if they handhold people to like again, those three to five to 10 use cases that are prevalent across every enterprise, I think they could get rapid adoption of this within enterprises.
- [46:01]
Mike Caput
So next up, Amazon just doubled down on its relationship with Anthropic. They just invested another $4 billion in the company, bringing their total investment to $8 billion. Now this announcement came on November 22nd. Anthropic announced the funding and said that AWS would be the company's primary cloud and training partner. A key part of this actually involves close collaboration on AWS Trainium hardware. Trainium is this purpose built machine learning accelerator that enables high performance model training. So apparently Anthropic's engineers are working directly with Amazon on optimizing future generations of Trainium. But what's really interesting here is despite expanding their partnership, Amazon also appears to be trying to actively reduce its reliance on Anthropic. Another report at the same time from the information reveals that Amazon is developing a new AI model that can process images and video in order to make it less dependent on Anthropic. According to this story, quote, in developing the new model, Amazon is showing that it still hopes its internally developed AI can gain traction among its cloud customers, making it less dependent on AI for Anthropic. Paul, can you maybe walk us through that contradiction here?
- [47:22]
Paul Raitzer
I feel like you could just place replace Microsoft and OpenAI with Amazon and this entire story and it would be the exact same thing. Like, it's eerie how similar the relationships have become between Microsoft and OpenAI. Microsoft put what, 13 billion into OpenAI but now they're, you know, there's some friction and Microsoft's trying to do their own thing and build their own model so they're not relying on open AI. And now we've got Amazon pouring 8 billion into anthropic but doesn't want to become dependent. I don't know, it's wild. Like it's literally the exact same scenario. I'm trying to think like Google doesn't really have this scenario because they've, they have relationships with a lot of the other AI model companies, but there isn't one. I don't think that they've put billions into that they're similar relationship, but I have no idea. I don't know where this goes. Again, I, it seems like such a natural acquisition target, but I don't know that that's going to be allowed. And so I think we're just going to keep having these sort of arm's length relationships with, you know, eventually tens of billions of dollars being put into these companies, but then they're going to compete with each other too. It's, it's so weird, like the whole ecosystem. So I don't know, I'll just be interested to kind of follow along. I think there's gonna be a lot more to this story as we move into next year.
- [48:41]
Mike Caput
All right, and some other news. Meta has poached Salesforce's CEO of AI, Clara Shi, who is going to lead a new business AI group. She post an announcement on X with the news and she wrote, our vision for this new product group is to make cutting edge AI accessible to every business, empowering all to find success and own their future AI era. Now, she actually backed this up by providing some interesting stats saying things like 200 million businesses each month use Meta products to connect with billions of consumers. Meta's llama models have over 600 million downloads to date and Meta AI apparently now has more than 500 million monthly active users. All of this seems to imply that Meta is well positioned to make a big AI play with business users. And interestingly, Paul, I think you had posted this in our weekly podcast chat with the comment quote, meta is coming for business users. Can you walk me through that?
- [49:45]
Paul Raitzer
I would be fascinated to dig into like how Meta defines a Meta AI user. Like you don't have a choice, like if you're, if you're in Instagram or WhatsApp or Facebook, whatever. Like if you search anything, you're using Meta AI. It's not like I'm choosing to use Meta AI, it's just the thing that's in the search function.
- [50:02]
Mike Caput
Right.
- [50:02]
Paul Raitzer
Though I don't know, their numbers are questionable. So real quick on Clara. So she started a career Software engineer intern at Microsoft 2002 to 2003. Business operations analyst at Google 2004 to 2006. She was actually at Salesforce GM of their App Exchange platform and ecosystem. 2006 to 2009. She was the author of the Facebook Era, which some of you may have read. That came out in 2009 and then a second edition came out in 2010. Interestingly enough, Marc Benioff wrote the forward for the 2009 edition, not the 2010 edition. She then was the founder and CEO of Hearsay Systems, her own company from 2009 to 2020. Then she was the CEO of the Service Cloud at Salesforce from 2020 to 2023. And then most recently as you highlighted Mike, the CEO of Salesforce AI 2023 to 2024. Now on LinkedIn in her job description as the CEO of Salesforce AI, I think think we get a indicator of where she's going to take Meta. So it reads led AI efforts across Salesforce including product applied research, domain specific model development, go to market, revenue adoption, partnerships, acquisitions and responsible AI. The four years I spent as a boomerang at Salesforce have been some of the most rewarding of my career. Okay, it all started in 2022. So this is when she was CEO of Service Cloud with Gucci GPT, a large language model application prototype we developed with Salesforce Research to suggest accurate on brand response to customer service inquiries. Moved at startup speed rapidly build a prototype. Then ChatGPT comes out which led to them then shipping Einstein GPT for service and Sales Cloud. Many of you remember Einstein in probably early 2023 and then she was moved into the new role to build out their AI platform and organization to support sales, marketing, commerce developer, use cases beyond service cloud. And then she goes into what their team built with Einstein Trust layer and prompt builder and model builder and all these things. So I basically think like you could pretty much take her job description for those four years and move that over to Meta and say okay, now do the same thing at Meta. But we have you know, 5 billion users or whatever that number is. So yeah, that'll be really interesting to watch. But I've been following her for a long time. When I saw this I had to go back and like what I can remember, like her actual bio, like all the different things she's done. So yeah, it'll be fascinating. But I don't to my knowledge like I don't, I can't think of like Meta's play in enterprise. Do they my missing something like do they have a play in enterprise?
- [52:38]
Mike Caput
I literally wrote a comment in this. I was like what is the play here? Is it open source? Is it.
- [52:44]
Paul Raitzer
I assume it's going to be, yeah, like turning Llama into an enterprise play.
- [52:48]
Mike Caput
My guess is that's what it is. But I don't know what that even looks like from it.
- [52:53]
Paul Raitzer
Yeah and I like Zuckerberg was at Mar a Lago last week, so he went and had dinner with Trump. So, you know, there's efforts being made. I think they're going to try and make Llama. I know they're working with government already to try and infuse llama into government, so I don't know. I. Yeah, Zuckerberg's doing what Zuckerberg does. I think he's going to just make a massive play here. And yeah, maybe they try and go and build enterprise solutions around the AI and I don't think they're going to try and play in the cloud, but I don't know, maybe they do that too, who knows?
- [53:26]
Mike Caput
We have some big kind of developments on the AI for education front broadly. So first up, OpenAI is making a pretty big push into education. The latest move is a free course that they are offering for K to 12 teachers. And basically it's called ChatGPT Foundations for Educators. It's designed to help teachers incorporate ChatGPT into their classrooms. It is one hour long. It has apparently already been deployed in, quote, dozens of schools, according to OpenAI. However, TechCrunch has reported some skepticism from a couple teachers they interviewed about this course. Some of the common critiques include the fact the course offers contradictory and limited advice on things like privacy and safety, using this data appropriately to get results, and there's some mistrust of OpenAI's promises that user data is actually protected now. At the same time, while teachers are learning how to use ChatGPT, we're also seeing some more advice on how students can AI proof their futures. We just saw a new report in the Wall Street Journal that offered some tips from experts on how students can prepare their careers to thrive in the age of AI. So the key recommendation they provide is focus on skills that machines can't easily replicate. Things like human communication, emotional intelligence and complex project management. They also recommend students avoid hyper specialization, instead developing a portfolio of diverse skills. That might mean combining things like technical expertise with business knowledge or adding strategic minors to complement a major. So, Paul, this is like a hugely important topic. We keep coming back to again and again that need for both teachers and students to urgently figure out AI. Like first, kind of two parts here, what do you think of OpenAI's guide? And second, how about the advice for students? How good is it?
- [55:29]
Paul Raitzer
I think it's, it's positive that OpenAI is providing some education. I don't know what you can learn in an hour when it comes to like a, you know, if this is like a, an introductory level in terms of, you know, let's assume the teacher students don't, don't have any real knowledge of how to do this. Like, one hour's not going to get you very far. I think the other issue is there's just always going to be trust issues when the tech company selling the solutions is the one providing the education. Because in reality, at the end of the day, they're trying to sell you stuff is the assumption I think a lot of companies make regarding the AI proofing, you know, education and careers. This is a tough one. Obviously this could be a main topic. We could talk all day about this. I think there's lots of uncertainty and there's a lot of mixed signals. So I listen to a lot of podcasts and depending on who you're listening to, some say programming is useless. Like, you won't need. There won't be, you know, coding five years from now. It just won't exist as a thing, you know. You know, computer science majors are, have no value. Others say it's essential that, like, computer science is still the future. Some say focus on stem, others say focus on liberal arts. So right now I think the most important thing is infusing AI into all areas of education and teaching students to work with AI as it evolves. Because the thing we do know is every profession, especially in knowledge work, is going to have AI integrated into it, but we don't know the actual implications to each industry. And I think what's going to need to happen is you're going to need domain experts in healthcare and finance and, and legal and education, all these other areas. They need to become highly competent with AI to then figure out the impacts it has on their profession.
- [57:08]
Mike Caput
Right.
- [57:08]
Paul Raitzer
So go back to the creative writing example from OpenAI. Really, like, what does that mean for writers? I don't, I don't know. Like, writers have to figure that out. OpenAI isn't going to solve that for them. The creative writers, the experts, the publishers, the authors, they need to get in and use these tools and figure out, how does this change our profession? The tech companies are just building the technology they can't solve for the thousands of professions and majors that are out there and say, here's what this means to you. The only thing I would pull this back to is like, I have an 11 year old and a 12 year old and am I doing anything different yet? Based on my knowledge of this, the answer is no. So I'm not doing anything to try and influence their decisions or the direction. Now they're still in grade school, so it's not like you know, they're heading into college and I'm trying to advise them on what major. But I talk to enough school leaders where I don't feel like I, I believe anything with enough conviction that I would change my advice to people on what major to pursue in college or what career path to pursue after college. And so my main guidance is we just have to accelerate AI literacy and capabilities at all levels, specifically to the teachers who are integrating into the classrooms and the experiences. And so, like, at this moment, I would most likely guide my children to pursue a liberal arts background with heavy stem. Like, I, I don't know, I would hedge on both sides. I think the liberal arts matters. I think a diversity of understanding and knowledge matters. And I think understanding how these machines work and like, what goes into building them matters. Like, so I don't know, I don't even writing it. Obviously AI is coming for writing and like, I, I don't know that I would discourage my kids from going into writing. Like, you just gotta understand the implications of what's coming and adapt what you do, your career based on that. I mean, you and I are writers, Mike. We're still doing okay. Like, we're still making a living doing stuff and we still write. And so I don't know. Yeah, it's again, topic we'll talk plenty about next year.
- [59:12]
Mike Caput
All right, next up, a major US Congressional commission is calling for a Manhattan Project style initiative for the US to develop artificial general intelligence. So this comes from a bipartisan group called the US China Economic and Security Review Commission. And they basically argue that America needs this type of large scale public private partnership to stay competitive with China in the race to develop AGI. Now, of course, the original Manhattan Project was a collaboration between the US government and the private sector during the Second World War. It led to the development of the first atomic bombs. The commission is kind of advocating for a similar scale of that effort, though they have not specified exactly how much money should be invested in this or how this would look. One suggestion that came out of this from one of the commissioners was proposing streamlining the permitting process for data centers, noting that energy infrastructure is currently a major bottleneck in training large AI models. Now, Paul, back on episode 120, you made a plea for a national Apollo level mission for both AI development, but also AI literacy and upskilling across every sector of the economy. Now, it's basically a similar idea. You just picked a more positive metaphor, which I think was good. The Manhattan Project's a little.
- [60:36]
Paul Raitzer
Yeah, I think that was the key like when I was, you know, trying to make the plea here, the, the Apollo mission obviously put humans on the moon. The Manhattan Project obviously built a bomb. So. Yeah, and unfortunately, like, I'm not so sure that the effort here wouldn't be more along the lines of a Manhattan Project than it would an Apollo project. What I mean by that is the Department of Defense would likely have a massive play in this and DARPA would have a massive play. And those are not. Those aren't put the people on, on the moon missions. Those are defend the and protect the United States missions that usually have defense projects. So, long story short, I think that there's going to be a massive acceleration of AI investments and projects starting January 20th. I think the new administration has a massive focus on this. There'll be a reduction in regulation on AI at a federal level. Climate issues are out the door. They don't care. Any impact that new data centers and all those things would have on the climate is not going to be a concern for this administration. There's going to be acceleration of open source technologies, so players that are building open source are going to benefit. There's going to be major investments in infrastructure, energy and data centers. Like this is all kind of a given. There will be a new aizar in town, I believe. Kamala Harris is currently. I don't think this is the official title, but I think she's in charge of the AI initiatives for the Biden administration. So there will be a new aizar. We don't know who it'll be. There's a really good chance Elon Musk will have a significant say in who that is. One name that kind of popped into my head this morning is Andres Karpathy, who we've talked about, because Andres isn't like. I mean, he's doing his thing with his education initiative initiative, but he's not otherwise, you know, fully employed at a major frontier model company or anything. And he spent five years with Elon Musk heading up AI at Tesla. I don't know. Like, that was just. That was an interesting one. I have no idea if Carpathia would do anything like that, but just a name to keep an eye on. And then I went back to like episode 87 where we presented this AI timeline and I sort of like this road to AGI I sort of laid out. There was two sections I talked about. One was what accelerates progress? Progress. And I'll highlight a couple of those. Clean energy abundance. So wind, solar, nuclear fission. This administration is not a clean energy advocate. Elon Musk is. But overall the administration doesn't love wind and solar energy breakthroughs. Nuclear fusion in particular. I think there's a massive effort to do that that could be part of the Apollo level mission is like let's, let's achieve nuclear fusion. Large scale government funding Apollo type programs is like exactly what I said, like verbatim infrastructure investments, updated electrical grids, more data centers and then more compute chips and fabs, plus a diversity of the chip supply chain. So that all seems to be coming to fruition. That, that does seem to be the direction this is going. But when you accelerate that progress, you also accelerate the likelihood of some of the things I identified that could be triggered that would slow it down. Catastrophic events blamed on AI, social societal revolt against AI due to job losses, politics, perceptions and fears and voluntary or involuntary halt on model advancements due to catastrophic risks. So the government can do all they want to push this forward. But if a year from now Anthropic and others say we have now hit our threshold where we need to pause development because these demodel, these models are getting too advanced and we don't understand all the risks that comes at odds. And if the government has bet hundreds of billions or trillions on this, they're not going to stop. And now all of a sudden private enterprises lose control of the thing they've been trying to build in a controlled way. So it's, it's going to be crazy. I think they're going to do it. I, I think they're going to like open the checkbooks to whatever needs to be put into this and that's going to have good and bad ramifications.
- [64:30]
Mike Caput
Speaking of people opening up their checkbooks, There's a new AI agent startup that just raised a $56 million seed round at a 500 million dollar valuation. Now this startup has a bit of a weird name. The name is literally forward slash dev agents. So we're including the link to the website in the show notes because you cannot type that in easily into your browser.
- [64:54]
Paul Raitzer
I read the headline like five times in decorative. I was like what is this a typo? And I was like oh, that's the name of the company because I even capitalized, it's like so weird.
- [65:03]
Mike Caput
So what they're doing is they're building what they call an operating system for AI agents. So AI agents in their mind being autonomous programs that can handle complex tasks without human supervision. There's a pretty interesting team behind this. David Singleton is the CEO, is a former CTO of Stripe. There's some ex Android and Meta Oculus people involved. One of the people who held a senior role at Figma and Dropbox as well. So basically what they're pitching is just as Android created the foundation for the mobile revolution, they think that AI agents need a similar platform to reach their full potential. Basically they plan to launch their first product in early to mid next year. Their business model might end up mirroring Androids like taking a cut of commerce happening on this platform, maybe charging for subscriptions but there are not that many details if you go to the website which we'll link to, not a ton of information. So when I'm looking at this Paul like it seems like huge numbers, very few details but fair amount of significant pedigree behind this it seems like also Andre Chikarpathi is an investor. Alexander Wang Scale AI also investing. How big a deal is this?
- [66:24]
Paul Raitzer
It's hard to tell. I mean their website is reminiscent of Ilya Sutskevo's and the safe Superintelligence where it's just a page like with basically says nothing and they raised a billion. So yeah, I, I mean one of the ways that you and I evaluate these startups is you know how much funding they're getting and who the investors are. This would certainly check the two boxes of definitely a company worth paying attention to. I assume if agentic AI becomes what all the people in the industry think it's going to be the idea of building and oper trading system for that is a massive market and thereby people are willing to take take some bets here. I I did laugh like this guy's the former cto. So what was his name? David Singleton. Yes. Former CTO of Stripe. Greg Brockman was also the former CTO of Stripe, the co founder of OpenAI. So it's like apparently like being a CTO at Stripe leads to raising a bunch of money and building an AI startup. I don't know. They must have an amazing pitch deck which makes the case for this because it's. It's serious money and those are some serious investors. So we will keep an eye on it. We'll let you know if anything comes out of it. Maybe it's now a race between safe superintelligence and the this company is to like who does something that actually is more than a page on their website or.
- [67:40]
Mike Caput
Right. All right. Some Salesforce news. Salesforce is adding AI agents to Slack so you'll have them as what they claim are digital coworkers right within Slack workspaces. So through Salesforce's Agent Force platform. Basically they're giving agents access to your organization, Slack conversations and enterprise data so they can understand context and take relevant actions to help you work. They're basically positioning these not as passive assistants, but things that can actively suggest and execute actions on behalf of employees within Slack. Paul, there's not a ton of details available here yet because Salesforce says that Agent Force will become available through Slack, but there are no details yet on pricing, availability, et cetera. It's obviously cool to see more agentic stuff getting baked into commonly used products, but I guess like my big question reading this is like how eager are enterprises going to be to open up their conversations and enterprise data to agents using it?
- [68:50]
Paul Raitzer
I don't know. I think this might end up fitting in the category of it's great that you can build these things and release them, but it doesn't mean anyone's going to use them or that has any interest in them. I may be wrong, but I'm also not a Slack guy. Yeah, I've said this before on the show. I just don't get it. I find the user interface for Slack to be wildly overwhelming and I'm just, I don't know, I've tried many times to, to become a Slack user and like see what it is that people love about it and I just personally really struggle with that platform. So I don't know, I'm like, I'll like let you do the research on that one when those agents be. Tell me if they're any good. Yeah, I don't. Are you a Slack? Do you. Do you love. We have Slack?
- [69:36]
Mike Caput
Like honestly the weird thing is obviously not for like work but for our community through Marketing AI Institute, which is not exactly the the same use case as getting work done. I have a Slack with friends of mine so it's a nice like communication platform with like five people. But from what I've heard from everyone that uses it, I feel like a lot of people, at least in a big organization, fall into the trap of I'm spending more time navigating Slack than actually doing work. But that could be anything as well.
- [70:05]
Paul Raitzer
Yeah, I don't know. I just find stuff hard to find in there. Like. Yeah, I mean our community is built on so it's great and we have like 9,000 people in there so I get it that they're. It's valuable and I just as a user, every time I go in there I get anxiety.
- [70:19]
Mike Caput
Like well that's. It gives me anxiety but also like it's fun to be able to follow the community conversations and be tagged in things. I couldn't even imagine how I would find tasks or work that I needed to do in there. I would be lost.
- [70:32]
Paul Raitzer
Personally, I know, again, I know people love Slack and they like swear by it. And that's. I'm not disagreeing that, that it is. I'm just saying personally for me, I struggle with it a lot.
- [70:43]
Mike Caput
Hey, maybe that's why we need an agent to use it for us.
- [70:46]
Paul Raitzer
Yes. I don't ever have to go in there. It just doesn't mean. All right, so speaking of doing things for us.
- [70:52]
Mike Caput
Yeah, yeah, exactly. This is a perfect segue because the next subject we're talking about here are digital clones. Because HubSpot co founder Brian Haugen Paul, who you know well, has been apparently building AI powered digital clones of himself that are trained on his work and that others can interact with. So you can now access one of these clones on a site called Delphi AI. We'll link to that in the show notes. And basically when you go to the link, you can either message, call or even video chat with Halligan's digital clone and you can kind of ask him for advice on anything related to building a business. So I actually tried this out before we recorded. I don't know if I would like immediately use this all the time, but I have to say the video chat was wild. Like, it had obvious visual and vocal flaws. Like, you're not going to be like fooled necessarily, but it's surprisingly good. Like, I had a five minute chat with Brian's digital clone about advice he'd give me on how to find a business ID you. And like, I don't, you know, I don't know him personally. I've seen Speak a bunch. It sounded pretty good. I don't know if it's what he would recommend, but I. It was like, it's wild to kind of try out because you can start to see even if we're not you and I doing this for ourselves. For some reason I could see this being used somewhere. Like, you're like, whoa, this kind of works in certain contexts.
- [72:22]
Paul Raitzer
Yeah. So I, So my initial reaction is like, I'm not really bullish on this. I could see, I could definitely see it working. I could see the influencer space. Like I'm being more like personal influencers.
- [72:34]
Mike Caput
Yeah, I could see that.
- [72:35]
Paul Raitzer
Yeah, yeah, I could see that working. But in terms of like corporate world. So one thing. So I try to put my, like, if I was an investor hat on and someone came and pitched this business idea to me, my first response would be, well, how can. What's going to stop someone from just building an open source version of this without your platform? Like, what are you going to enable that? I couldn't just like go and give 100 podcasts and three books to and every webinar I've presented and like, couldn't I just train a version myself and not pay you 400amonth for it? And, and same being said, like, what's stopping other people from building a clone of Halligan without Delphi? Like, Halligan's got all kinds of stuff online. Like, couldn't you just go and scrape stuff? And like, I'm not, I'm not by the way, like endorsing doing this. But this is how language models are built. Like, right, the language model has everything Halligan's ever publicly said already in it. Like GPT4O has all Halligan's public data in it. Is it that far of a reach to say, like, you could just build something like this? So I don't know. Like, so there's, there's a part of it that's like, thinks about it from the product perspective. There's a part of me that thinks about it as a, like a user. Like, would I pay to interact with hall again? Like, I don't know, because I know there's gonna be hallucinations in there and stuff, but I don't know. Like their positioning says, it was built on the basis of one simple idea. Help the movers and shakers of our world touch more lives. To provide the one on one coaching, tutoring and discussions previously unavailable to an elite few. Modern leaders. YouTube creators, startup CEOs. Domain experts possess potentially life altering knowledge and wisdom. But their time is limited at access is constrained. Our goal is to multiply their impact hundreds, if not thousands of times over in charting the journey they want to achieve. Two feats democratized mentorship. So breaking down the barriers of time and access. And then birth digital immortality. Oh my God. Preserving everyone's unique knowledge forever. Your Delphi clone doesn't merely represent you, it ensures your wisdom transcends time and oh my God, maximizes impact for generations. Oh, just threw up in my mouth. So, okay, so they're pricing. They have like 29amonth starter, a 99amonth advanced, and a 399amonth prodigy package that has like a hundred thousand message credits per. Oh my. This is like a really complicated pricing model. But then there's an immortal package for celebrities, influencers and thought leaders.
- [75:04]
Mike Caput
Oh, my God.
- [75:05]
Paul Raitzer
Unlimited training. Yeah. This is, man, it gets more complicated from there. On the pricing page, we'll put the pricing page and the notes. You can. So I don't know, man. This is. This is a slippery slope for me.
- [75:21]
Mike Caput
Yeah.
- [75:21]
Paul Raitzer
Yeah. I just, honestly, when I first saw Halligan tweet it, I assumed Sequoia, because he's a senior advisor at Sequoia now. I assumed Sequoia was investing in Delphi. So I did a search and I couldn't find a connection between maybe they're investing in their next round or they want to invest in their next round, I don't know. And they're trying to, you know, get in their good graces. But I assume that was the connection, is that, that Halligan was, you know, working with them from an investment perspective. But I don't. I don't see that connection. Maybe he just likes it. So I don't know. I'll. I'll try it. I, I'm, I'm totally happy to be off on this one and that I, I should be bullish on it, but I, I don't, I don't love this space personally at the moment.
- [75:59]
Mike Caput
Yeah, I'm not. Yeah, I'm not sure I would personally use or endorse even the business model. But I think what really just jumped out at me is like, I could see myself talking to one of these things in some context. Would that for one, probably not. Not out of the gate. I mean, it'd be a different country, probably a different use case. Because I don't. To your point, I'm not sure, like, how much I would need a clone.
- [76:22]
Paul Raitzer
Of someone thought of is interesting as an event organizer. So could I get Brian Halligan's digital clone to keynote Macon 2025 without his permission? Could I just, like, say, hey, we're going to experiment. We're going to do a fireside chat with Brian Halligan's digital clone, and I'm going to ask him all about the future of marketing. Does Halligan maintain any licensing rights to that? I've. I'm not asking you the answer to this. Like, sure. I'm starting to now process this business model and what it means to the creator who allows the creation of their digital clone, and what licensing rights are you giving up to your own Persona? And as someone who maybe could take advantage of these, could we then use anyone's Persona that's given permission to create a digital clone to do whatever we want with it? Invited to a podcast guest to have it on? I have no idea. To their terms, but I feel like.
- [77:09]
Mike Caput
We'D be getting a cease and desist from the digital clone of Brown's lawyer. Potentially.
- [77:16]
Paul Raitzer
Maybe. Or maybe. Well, it's interesting, too.
- [77:17]
Mike Caput
Like, if you could. If you talk with one of these. Enough. I only played for it for a few minutes. Like, let's say I recorded a video or a podcast with this clone and somehow was able to get to say stupid stuff or break it or whatever. Right. It's like, then it's kind of like reflecting very poorly.
- [77:33]
Paul Raitzer
Correct. As a brand thing.
- [77:35]
Mike Caput
Yeah.
- [77:36]
Paul Raitzer
And is there. Did you notice, like, I didn't dig into their creator rights, but do they verify the authenticity of your identity to allow you to create the digital clone? Or should, like, could someone go in and spoof and create a digital clone of.
- [77:52]
Mike Caput
That's a good question. I didn't look into the Delphi's particular terms, but I know for a fact you could probably do this using some.
- [78:00]
Paul Raitzer
Other stuff that's 100% going to happen. Like, that is. If that's not already happening. There is going to be a whole world of spoofed digital clones of thought leaders and experts, and people are going to make money selling acts. It's going to be like, you know the crap books on Amazon that people are just pumping out. Oh, my God.
- [78:18]
Mike Caput
You could, like, probably at some stage just max out your Delphi credits and have the official clone train. Your unlicensed clone. Haven't talked to each other. You know, like, horrible.
- [78:29]
Paul Raitzer
Yeah. Like, sometimes I sit down and ponder the future of society as a result of all these things that are getting built, and it's like, I don't really love that part of it.
- [78:38]
Mike Caput
Yeah, maybe this one needs a little pivot. Who knows?
- [78:41]
Paul Raitzer
All right, so more to. I guess we're going to be talking about digital clones next year, too, but that's not the agenda.
- [78:47]
Mike Caput
All right, so next up, we got a pretty cool case study at how large organizations are actually implementing AI. So we have a new report from the Wall Street Journal that looks at how BBVA, which is a very large Spanish bank, deployed ChatGPT Enterprise across a ton of different licenses. So they started out with 3,000 ChatGPT enterprise licenses. They've expanded those by a few hundred more. They have plans to add more next year. What was cool is they created over 2, 900 custom GPTs for specific tasks like translating complex risk terminology or drafting responses to retail banking questions. Now, Interestingly, they say 80% of users said the tools save them more than two hours of work every week. However, it's not all good news. They did say they are hitting some integration barriers. So, you know, ChatGPT Enterprise can handle static documents just fine, but connecting it to BBVA's internal databases and systems. They said they interviewed with the Journal saying that was a really big hurdle and they're trying to attract next year the more kind of tangible returns in terms of savings. So their global head of global AI adoption said that, you know, we think the value of this is much bigger than whatever small savings we could measure today. So they're kind of going to get more signed scientific about measuring that moving forward. Paul, as I was kind of reading this, it definitely feels similar to some of the challenges and opportunities we see and hear about in enterprises. Like what jumped out to you here about this approach?
- [80:29]
Paul Raitzer
Yeah, I think for me the thing that jumped out is the value of custom GPTs. Like this is the thing, you know, we mentioned it earlier, this like have 5 to 10 use cases that are super valuable for people. GPTs have now been around for a year. So November of 2023 is when custom GPTs were introduced by OpenAI. But I think what people need to focus on is how can we build those custom GPTs and eventually Google Gems and anthropic Claude projects, whatever it is, build the things that are going to create value for people right out of the box. So I'm hopeful that OpenAI does a lot more to support and advance custom GPTs in 2025. I'm optimistic we'll see more with Google Gems, but I think that that's the real key here to value unlocking value is build those solutions for people, teach, empower them to build their own, but help them identify the use cases where they're going to get a ton of value in terms of the integrations. Like that's a very real issue that you and I hear all the time. So yeah, I think this is a good example of what we're seeing across enterprises a lot.
- [81:35]
Mike Caput
So kind of related to this and almost like a more macro view of this. In another topic we've got this week, there's been a new interview with OpenAI's Chief Commercial Officer. His name is Giancarlo Lionetti, he goes by the nickname gc. And he basically revealed that the company is targeting a hundred billion in revenue by 2029. They're making an aggressive push into enterprise sales. Specifically to do this, this they actually expanded their sales team to 300 people, currently up from 200 in June. This now makes up 1/5 of OpenAI's total workforce. He said that they have won notable deals with companies like Moderna, Lowe's and a hundred million contract with T Mobile. And very interestingly, he is talking about the evolution of enterprise AI adoption. He says 2023 was about experimentation, 2024 is about solving specific business problems, and 2025 will be about scaling AI across entire organizations. So as he put it, companies are now coming to OpenAI, basically saying help us with our AI strategy from the bottom up. So Paul, like, what do you think of this approach of 2025 being the year of scaling AI? I know we have some OpenAI employees who are podcast listeners. Any insights here?
- [82:59]
Paul Raitzer
They well, I think first thing I noted was they've hired 100 or are hiring 100 salespeople. So that's good news for salespeople. So OpenAI, you know, is still in the process of hiring sales then AGI is not here yet and hasn't replaced the profession. So that's a good sign there. And then, yeah, on the 2025 thing, I think for that to be true, I'd like to think that's true. Then enterprises are going to have to quickly start doing the things that are needed to be part of the scaling AI process. So we have a site dedicated to this. So scalingai.com I teach a free monthly class called Five Essential Steps of Scaling AI. And then we have a on demand course series about this. So it's sort of a bit of a shameless plug, but also educational in nature. The five steps that we teach are launch an AI academy, create an AI Council, develop responsible AI principles and generative AI policies, conduct AI impact assessments on your team, your tech stacks, your partners, your agencies, and build an AI roadmap. So those are the five things. I don't know very many enterprises that are doing those five things well. And so if they think that next year is the year of scaling AI, then we got to really ramp up the education around this and the integration of the technology across those five areas. So again, like, you know, as a starting point, I think our next free classes in January, maybe we'll put the link in the show notes, but again, you can go to scalingai.com and check it out. And that's just an hour class we teach on Zoom every month and it goes through those five steps and then the course series has 10 courses and it goes in depth into each of those five areas.
- [84:37]
Mike Caput
All right, we're going to do one more kind of big topic to cover here, Paul, and then I'm going to Rip through a few final product updates we've got and then we can get this massive week of AI news wrapped up. So first up here, Aiden Gomez, the co founder and CEO of the AI model company Cohere, just gave us a new interview on the popular no Priors podcast. So he covers a ton of range of topics related to enterprise AI. Notably, he sees a fundamental shift happening in how AI is being deployed in the enterprise. Cohere, as a result, is taking a very different approach from the consumer facing AI companies as they work work with enterprises. So he said that Cohere's focus is squarely on enabling organizations to adopt AI in ways that make their workforce more productive and transform their products and services. So rather than competing with something like Chat GPT, Aiden and the team are building a platform that helps enterprises implement AI effectively. He actually said that he thinks many early enterprise AI projects failed in 2023 because custom companies overestimated the model's capabilities. He said they treated them like humans rather than understanding their specific requirements and limitations. So in response, Cohere is actually developing more robust models and creating structured APIs that more rigorously define how to use them. He also believes that we're headed into a flattening of the curve in terms of general capabilities with future gains coming in specialized areas like phys, math and chemistry. The next frontier, he says, is reasoning, giving models the ability to work through problems step by step, etc. And he also says on the broader question of artificial general intelligence, he believes in building generally intelligent machines, but sees this as more of a continuous progression rather than a sudden breakthrough. So Paul, I know you found a lot to like in this interview. It's a really interesting enterprise specific perspective from Aiden who is day in, day out building in that space. What jumped out at you here?
- [86:49]
Paul Raitzer
Yeah, we. So on episode 112, if the name sounds familiar, we mentioned Hayden quite a bit but on episode 112 we talked about his 20 VC podcast and you know he's one of the co authors of the attention is all you need paper that invented the transformer, the Google brain team back in 2017. So, so he's definitely someone we follow closely and I, I like listening to him because he's somewhat contrarian to this, this approach of just keep building bigger and bigger and bigger models and spending tens of billions like he's accepted. I think their role in this is not to go raise 10 or 20 billion and try and build a competing frontier model. It's to be much smarter about the training and how we build these things and the algorithms that go into it it. And so I think he's very well spoken. He's obviously deeply aware of everything that's happening in the industry and I think he very respectfully presents his perspective on how he thinks that gets ahead. But it's also more he's just not approaching it the way the other companies are. And he's, he has very specific buyers and users in mind with their approach. And so I think it's just always good to listen to him because it offers a bit of a balance between, you know, some of the big frontier model companies like an anthropic and OpenAI we talk about a lot with also well funded company. It's not like they haven't raised a bunch of money, but that isn't trying to maybe play that same game. And you know, I think that that's very helpful to give that perspective.
- [88:21]
Mike Caput
All right, Paul, I'm going to take us home here with a few quick product updates that we got related to some of the big players in AI. So first up, Apple is apparently getting ready to launch an AI powered version of Siri. According to Bloomberg, they are developing what employees internally call LLM Siri. It's a major upgrade that hopefully gives Siri the ability to have much more conversational interactions with users thanks to AI powered by large language models. Reportedly, this is going to be built on Apple's own AI models. It's being tested as a separate app on iPhone, iPad and Mac and it's not coming anytime soon. As Apple Intelligence taught us, just because we hear about this doesn't mean we're getting it. So they actually plan to formally announce the updates here of what they're doing as early as 2025, but you won't actually get to use those features until about spring of 2026, which is an eternity from now. Next up, the popular AI music generator Suno released Suno V4, which promises better overall audio quality, more sophisticated song structures, and it can handle more complex musical arrangements and produce sharper, clearer lyrics. They've also introduced a remaster function that can upgrade existing songs you've made to the new V4 quality level. They've got a cover art generator now to create visuals for your music's style that matches it it. And they have a covers feature that can reimagine songs in different styles and Personas, which is a feature that lets users maintain a consistent musical identity across multiple creations. Another update here, the AI voice cloning company eleven Labs just dropped two big updates. Actually a platform for conversational agents and a product that turns written content into podcasts. So first up, this conversational AI agent platform will let developers create agents that can be fine tuned across multiple variables like tone of voice or response length, and they can even integrate their own language models and knowledge bases. So this platform is actually going to feature all these ways to handle real world challenges in conversations, things like customer interruptions, and also collect data during conversations to enrich them. So basically going to result in agents that can eventually have more natural sounding conversations with people.
- [90:56]
Paul Raitzer
So Notebook lm.
- [90:57]
Mike Caput
So. Well, the second feature is Notebook LM because they're coming out with. They also are launching something genuinely called Gen fm, which literally is described as a service that transforms written content into AI hosted podcast discussions, which is the exact same product, I think. So this is actually through their 11 reader app, you can put in PDF article eBooks, turn it into a podcast featuring two co hosts, which again, if you've used NotebookLM, you know what that sounds like.
- [91:26]
Paul Raitzer
How long till OpenAI copies the Notebook LM product? Yeah, it just seems like everybody's gonna copy it now. It's.
- [91:31]
Mike Caput
Yeah, yeah. Now that everyone's seen how popular it is.
- [91:34]
Paul Raitzer
Yeah.
- [91:35]
Mike Caput
And then one other product update here. Runway just unveiled a new AI image model that they claim makes some major advances in controlling style across AI generated imagery. So this is a big challenge right now in AI generation. So this model, which is called Frames, can purportedly maintain consistent visual styles across a bunch of different generations. So you can keep everything consistent over many different images. So they kind of organize the model's capability around this concept of what they call worlds, which are like coherent visual styles you can apply to all the different things you're creating. So this technology is being gradually rolled out through runways, Gen3 Alpha platform and its APIs. Now, last but not least, we are doing a special 25 AI questions for 2025 episode. This is dropping on. I believe we are dropping it on the 19th of December. Sounds right. And we are going to do a special episode where we answer user questions. We've done this before, but to do this we need questions from you. We've already had quite a few come in. But if you can go to bitly bit ly 25 questions episode, we will include this in the show links, but you can drop in whatever question you have about artificial artificial intelligence in business, marketing in the world at large, and we will pick as many of them as possible to answer on that episode. So Paul, I'm looking forward to that one. I always like doing those and I feel like they're really popular.
- [93:15]
Paul Raitzer
Yeah, for sure. I think we said like three of our top 10 episodes all time were our Q and A episodes. So. Yeah, it's always great to see people's questions. And I mean, every time we do like the intro to AI and the scaling AI, we get dozens of questions we don't have time to answer. So, yeah, we love it. It's. It's always helpful to see where everybody's at and, and what they're thinking about and hopefully we can provide some perspective and value heading into next year. And then that will be our final episode of the year. I think that's the plan right now.
- [93:42]
Mike Caput
We're gonna be awesome.
- [93:43]
Paul Raitzer
Yeah, man. After taking one week off for Thanksgiving, I'm like, I don't know if I could do like three weeks.
- [93:48]
Mike Caput
No kidding.
- [93:50]
Paul Raitzer
All right, well, good stuff, Mike. Thanks for running through all those and curating the rapid fire and reminder, everybody. I mean, we're at the end here, but like, like we always put the timestamps in every episode. So, you know, if you, if you ever want to jump around and see different things, you can always go in and, and see where we're at with them. And, and that's. We've always done that for all 100. And was this 125?
- [94:10]
Mike Caput
125, yeah.
- [94:11]
Paul Raitzer
So you could jump around and find what you're looking for. So yeah. Thanks, Mike. Really appreciate everything.
- [94:16]
Mike Caput
Yeah, no problem. Thanks, Paul. Really appreciate it.
- [94:18]
Paul Raitzer
All right, and we will be back with a regular episode next week as well. So we'll talk to you all that. Have a great week. Thanks for listening to the AI show. Visit marketingaiinstitute.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, taken our online AI courses and engaged in the Slack community. Until next time, stay curious and explore AI.