← The Artificial Intelligence Show
The Artificial Intelligence Show
#127: 12 Days of OpenAI Continues, Gemini 2, Hands-On with o1, Andressen Says Gov’t Wanted “Complete Control” Over AI & OpenAI Employee Says AGI Achieved
01:26:24
Loading summary
Transcript79 lines
- [00:00]
Paul Raitzer
If the AI models keep getting better, thinking, reasoning, understanding, imagination, which is a whole separate thing, and we'll talk about at some point if it does those things better than the average human who would otherwise do the job in your business, in your industry, then we got some problems. 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 Kaput, 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 127 of the Artificial Intelligence Show. I'm your host Paul Raetzer along with my co host Mike Kaput. This, Mike, is our final weekly episode of 2024.
- [01:09]
Mike Kaput
Indeed it is.
- [01:10]
Paul Raitzer
All the AI companies saved up all of their big updates for this week. This is. I don't know how I usually start these shows. As I've said before, I don't script what I say for the most part, so I have no idea what I say most of the time. So if I usually say it was a crazy week in AI, that would be an understatement for what Mike and I have to try and get through today on our final episode. And I hope that people don't keep dropping stuff the next four days but we already this morning got stuff from Google, so who knows. We are going to do our best to catch you up on OpenAI's 12 days of ship miss or 12 days of OpenAI. There we are now on day eight. Eight, right, Mike, we're on day eight. Y day eight. We got new models from Gemini. We got a hackathon Mike and I ran on the 01 reasoning model which honestly Mike, when I thought about today's episode that felt like a month ago that we did that and it was just six days ago, I was trying to remember, like didn't we do something with this 01 thing? He's like, oh yeah, we met for two hours on this, like holy cow. So it is just non stop updates. I would have to go back historically and see if there was a week that had more updates. Maybe this time last year. But certainly this was the busiest week of the year that I can recall in terms of updates from basically every frontier model company had stuff. So we got a lot to unpack for this final episode. There is one other episode which I've mentioned a couple of times. Mike and I are doing on December 19th we're going to drop a 25 AI questions for 2025 special episode that we are actually recording Wednesday, right? Is that right? Yeah, the 18th. So it's gonna be fresh. It's gonna be. So today is. We're recording this one on Monday, December 16th at about 2:30 Eastern Time. We delayed it again so we could see what day 8 of OpenAI's 12 days of shipments was. So just a time to get to. All right, so all that being said, this episode is brought to us by the AI Mastery Membership Program. This is the joint membership program offered between SmartRx and Marketing AI Institute. Has tons of marketing content but also has plenty going on for the non marketers and business leaders. So this is a membership that, that we run. It's a 12 month program, gives you access to education, insights, answers that you need to develop a high level of proficiency about AI in your career, in your company. That includes quarterly briefings that Mike and I do, Generative AI Mastery series that Mike runs once a quarter. He goes through a bunch of cool demos. We have a quarterly ask me anything session where you just spend an hour asking, answering whatever you want to talk about. There's unlimited access to our on demand webinars, there's ungated access to blueprints. And I cannot like explain fully what we're planning for next year. But I will tell you that most of my time right now is being spent on the vision and strategy for this membership program going into 2025, working on new course series, new certification ideas, new experiences. And so if you've ever thought about joining the Mastery Program, now would be a great time to do it. There's going to be a ton happening in that next year. So you can go to SmartRx AI and click on Education is a quick way to get there. Just look at AI Mastery. The actual website address, if you want to type it in, is SmartRx AI AI mastery. That link will be in the show notes. And as a special thank you to our podcast audience, you can use Pod150. That is Pod150 and that'll get you $150 off of the membership. And that is good until the end of this year, December 31st. So we'd love to have you join the IMASTER membership program. Like I said, it's going to be a huge focus of ours to really kind of double down on not only the time and investment we make there, but the value we create for people there. So hopefully you can join us leading into next year. All right, Mike, we are on day eight of OpenAI's 12 Days of Shipments. Why don't you give us a rundown of what has happened, which again feels like it's been a month worth of updates. That really has only been the last eight days.
- [05:39]
Mike Kaput
Absolutely. So to kick us off this week for our final, you know, formal episode here, we're covering more of the 12 Days of OpenAI event that's been happening. Like you said, over the last eight days, they're basically releasing new products and updates each weekday for 12 weekdays in a row. So since our last episode dropped, including today, the company has released announcements for days four through eight. So I'm going to run through them really quickly and then Paul, I'd love to get your thoughts on kind of what we're seeing here. So on day four, OpenAI announced the general release of Canvas in ChatGPT. Canvas is a side panel that has responses from ChatGPT on a shared, editable and shareable page so that you can more effectively collaborate with GPT and others on writing encoding tasks. On day five, OpenAI unveiled its long awaited integration with Apple Intelligence. So the team showed off significant improvements using AI to things like Siri's capabilities. The voice assistant is now able to handle complex queries and provide more natural context aware responses. Powered by ChatGPT, users can now seamlessly move between Siri and the ChatGPT app and Siri is able to open different ChatGPT tools like Canvas and Dall E. So this system is also able to work with Apple's visual intelligence, allowing for sophisticated image analysis and processing. On day six, OpenAI finally delivered on its promise of video capabilities for ChatGPT's advanced voice mode. This was a feature that, if you recall, was first previewed during the GPT 4.0 launch back in May. It was a bit delayed after that, but now it is out and it allows users to interact with ChatGPT through their phone's camera so the AI can actually see and respond to what's happening in real time. This feature also includes some screen sharing capabilities which allows ChatGPT to understand and comment on content that is displayed on your device at the same time. They also released a bit of a festive twist. OpenAI introduced a Santa mode for voice mode, so basically you can have it sound and talk just like Santa Claus. You can access this feature by tapping a snowflake icon in the ChatGPT interface. Day 7 we got OpenAI announcing better organization in ChatGPT with a new feature called Projects. This is designed to help users manage their AI conversations more effectively. It works like a sophisticated folder structure. It allows users to group related conversations and resources together in a more intuitive way way. So Projects will appear in the ChatGPT sidebar. Users can create new projects, customize them with different colors, and add instructions to guide how ChatGPT responds. Within this particular project. You can also add and attach files and add existing chat conversations. Now, finally today, just and probably about an hour or two before we're recording here on Monday, December 16, OpenAI announced on day eight improvements to ChatGPT's search functions. So not only is search a lot faster, but they also showed how it is optimized for mobile. So the team actually demonstrated on mobile doing searches about they used an example of finding a restaurant in San Francisco that serves Mexican food and has outdoor patios with heaters. And ChatGPT search on mobile now presents this kind of clean visual list of businesses and the search results for those particular queries. So search is actually now also going to be enabled, which they were displaying it as in advanced voice mode. So the team is able to actually they demoed how you can talk to chat GPT to actually search up to date information on the web. So in their example they actually had Chat GPT look up different holiday events happening in the coming week in different locations around the world and they got up to date info about the locations, the hours of these events at the locations, as well as the weather on certain days. Now, while this is our last formal episode of the year, obviously there are four more days to cover in the event. So we'll link in the show notes to the ongoing post we have summarizing all these updates and we'll be updating that once this event has concluded as wealth. So Paul, that's quite a bit to unpack. That could be honestly half if not the entire podcast episode alone. I'd love to get your thoughts on like what we saw this week. I mean for me, I have to say I'm most impressed so far with advanced voice and video.
- [10:32]
Paul Raitzer
Yeah. So I'll run down a few of my takeaways here. So Canvas is nice. I'm not a power user of canvas within GPT4O, so it wasn't like a huge thing for me. But having it now in custom GPTs is going to be interesting and kind of like as they keep building out that capability. Apple Intelligence plus ChatGPT is very interesting. So this is the first time where I feel like we're starting to see the vision for Apple Intelligence. Just a reminder here, if you have an iPhone, it's got to be, I think, 15 pro or higher, to be able to use Apple Intelligence and then check your iPad and Mac devices to make sure that they're compatible in terms of the OS and the, the, you know, the generation you have of those. But when you connect the account, when you turn on kind of the Apple Intelligence, you actually have to go into Settings and then go to Apple Intelligence. And then you can use ChatGPT free, or you can choose to connect it to an existing Chat GPT account, which is what I did. And then I have a personal account and I have account for the business. And so it actually lets you choose which of those you're going to be sending stuff to. Because the way it works is when you talk to Siri or when you use Visual Intelligence, it's actually sending those inputs, like your requests, your prompts to your existing ChatGPT account. So just some notes on the setup. I text, I tested Siri. And it's interesting because, like, historically, I've been as big a critic as Surrey as anybody. Like, it's just useless, basically. Historically. And it's always been like, why can't it just answer something? Why does it always have to be like, I found it on the web, it's like, I'm driving. I don't, if I wanted to go to the web, have gone to the web, sir, like, tell me the answer to my question. So what it does now is it says, would you like to connect to Chat GPT to get this answer? And then it says, working with Chat GPT. And so I tried one where I said, what happened to Clue and Brown's game yesterday? Figured it was like real time, see what? And it actually did a really good job. It gave me a Chat GPT generated summary in about, I don't know, five seconds. And then it read it to me. And so I thought that was a really good start. Like, I could absolutely see myself using Surrey now way more. I find my guess is I'm gonna get really annoyed that I have to say yes every time to do you want me to use with Chat GPT, it's like, just, I always want a universal yes. Like, if I ask you and you don't know, just go to Chat GPT, which Apple will know that that is a friction point and they will either solve for that or they will eventually build their own version of this. So they don't have to go to chat GPT, which is my guess is what they'll probably eventually do. Yeah, so that's really cool. Visual intelligence is interesting. If you haven't used that yet. The on the 16 Pro, I think is what I have. So there's this new button on the right hand side and so if you just click on that, it actually brings up what looks like the camera function, except you have kind of the rainbow effect going on that you do with the new Surrey. And then you can ask or search are the two options. And what happens is it basically takes a picture of whatever you're looking at and then it can go search for that. So I like, I like Scotch. So I have a bookshelf next to where I'm recording from right now that has a collection of Scotch bottles. And so I just like took a picture of Johnnie Walker Black and, and it just, it went and like found that bottle at different places where you could buy a bottle of Scotch and pricing. And I was like, okay, that's pretty nice. And then you can ask it anything you want. And so it's like, okay, you know, what is this? And it's like, it's a bottle of Johnny Walker Black. And so that was, that was pretty cool. So the Apple Intelligence plus Chat GPT integration is worth testing now. It's actually, you know, very functional. Advanced Voice, actually I'm going to come back to Advanced Voice Projects is interesting. I love that you can now organize your, your chats, your threads that you've got going on. My problem immediately was I got really excited because I'm doing this co CEO webinar tomorrow and I was like, oh, awesome. I can take all of my co CEO threads that I name co CEO so I can find them in my left panel. And I went to put them, add them to the new co CEO project and I realized like, I can't. So I actually went on X and I tweeted at Adam Goldberg who's a go to market lead at OpenAI and then Kevin Whale who's the Chief Product Officer and said like, hey, this is awesome, but it's an annoying limitation that I can't seem to add these custom GPTs to a project folder. And so I'm stuck renaming and labeling these things. I thought projects would solve for that. And the Chief Product officer Kevin actually replied and said it's on the list. So one, I thought it was cool that he took the time to reply on a Sunday and two, it's coming because if you're like me, the vast majority of my use of ChatGPT now is within custom GPT, so I, I can't organize anything without that capability.
- [15:33]
Mike Kaput
Right.
- [15:34]
Paul Raitzer
So projects, I think is nice. And then advanced voice. So as much as we've talked about Project Astra from Google this year and this idea of giving devices vision capability to understand the world around you, it. It's here. Like, if you go into advanced voice mode and you click on the camera, you open up a vision of what you're looking at and you can talk to it and it sees and understands. So, like, if you're watching this on YouTube, you can see my basement. And so I actually went over to the big copy of the Mike, the book from Mike and I, and I said, like, tell me about these authors. And it actually like, told me about me and Mike. It gave the background, it talked about the book. I went like over here and I said, what do you see here? And it's like, oh, that looks like a really fun rocket. And I said, what kind of rocket is like, it looks like the Saturn V LEGO set. I'm like, perfect. It nailed it. I showed it a picture of a puzzle behind me. It's like, oh, that looks like fun. Are you going to frame it? Since it's already done. And so it's like, it's, it's working, it's seen. I've made some mistakes on a couple things, but we now have in ChatGPT the ability to show it something. So if you're traveling, if you're looking at signs and trying to understand it, if you're looking at products in a store, like, you now have vision capability on your phone to do this with. And the other one is screen sharing. You can actually open up and go in and choose to share your screen and then interact with whatever is happening on your screen and talk to Chat GPT about it. So I'm with you. Like, the advanced voice improvements with video. Yeah. And screen share are, are huge and could easily have been lost in a week of craziness since they introduced that on the same day as Santa mode. Which, like, is the first thing I went to was Santa mode. And by the way, if you have younger kids, like pre teens and teens, ask it to do like, you know, towards the night before Christmas or some other, you know, Christmas song or poem in Gen Z slang. It's hilarious. And it'll say, cringe your kids out. Like, my poor daughter was like hiding in the backseat listening to this thing. It was so funny. So, yeah, I mean, there's lots of good stuff. But I would say advanced voice and the Apple intelligence were the things I was the most kind of geeked about.
- [17:46]
Mike Kaput
Yeah. And this might just be my personal opinion, but as a huge advanced voice mode user already, this screen sharing also makes a big case for ChatGPT Pro and getting rid of usage limits. Because if I can suddenly be like pair working with this thing and having it tell me like, hey, how could I be more efficient? How do I do this step? That's really, really fast time to value for me for a subscription like that.
- [18:13]
Paul Raitzer
Yep. Yeah, I agree. And I think that you're starting to see how they could hear their pricing models further in the future now that these things are capable of way more than text in and text out. And that's because again for context here we started the year all of these models were basically text in and text out. They could do some video, you know, image generation but like they couldn't see and talk to you about things you were seeing. They couldn't generate videos the way we're now seeing these. So like we have had a massive shift now in what these models are capable of and we can look out into next year and start to see some of the other capabilities that are coming very closely behind.
- [18:54]
Mike Kaput
So honestly that topic alone would have made for one of the busiest weeks in AI. But there's a lot more that happened because you can clearly tell Google doesn't want to be left in the dust with this big 12 days of OpenAI. And they made some huge announcements this week too. So first they unveiled Google Gemini 2.0, marking what they call their entry into the quote agentic era of artificial intelligence. This is the beginning of rollout of updates to their flagship AI models and it introduces a range of capabilities that allow AI to take more direct actions on behalf of users. So right now at the heart of this initial Release is Gemini 2.0 Flash, which is an experimental model that Google says is not only twice as fast as its predecessor, but can now also generate images and audio alongside text. This model can also directly use tools like Google Search and interact with third party services. And you can currently access Gemini 2.0 Flash in experimental mode using a Gemini Advanced account. Or you can go to Google AI Studio and play around with that. Now at the same time, Google also released a stunning new feature that is attached right now to Gemini 1.5 Pro which is called Deep Research. Deep Research is basically a super powerful research assistant that creates multi step research plans, analyzes information from all across the web and compiles comprehensive reports on complex topics. And it literally does that with dozens or hundreds of different web pages and sources. Third, we also heard that the company unveiled a research prototype called Project Mariner. And this is a Gemini powered agent that can take control of your Chrome browser, move the cursor on your screen, click buttons, fill out forms, and navigate websites. So, Paul, that's a lot to unpack. First, let's maybe get your thoughts on the implications of Gemini 2.0. I mean, even though it's still early, we've just got Flash like this seems like a pretty major release that we were expecting.
- [21:04]
Paul Raitzer
Yeah, there's a ton packed into it and we'll put all the links in so people kind of follow along and, and explore it. The one thing I'll tell You is the 2.0 flash, at the moment at least my last look was only available in my personal Gemini account. Okay. And not in the Workspace account. So if you have a Google Workspace Gemini account, it's still just showing as advanced and I'm pretty sure that's still 1.5, not 2. So I needed to go into there. But like you said, you can also go into the AI studio and there's all kinds of experimental things in there. So yeah, I think it's a really big deal and it's starting to show where their models are going now. Again, you're not going to get the full 2.0. My best guess is April. I just. Because I think that's when their big Next conference is their developer conference. And so I would guess that maybe it's going to be before that. I can't imagine it would be after that. So I would think what you're going to see is this like really full out multimodal model built from the ground up, trained on multimodal, able to output multimodal. And that's where all of these frontier models are going. And so again, it's one of those where you're seeing it, but it's still just the Flash version, it's still previews in a lot of ways. So I don't know, you know, we're going to get the full thing for a little while here. But what they've released is already powerful and so I would definitely play around with it. I haven't, with everything that's been going on, I haven't had a ton of time to like run it through a bunch of use cases over the last few days. But over the holidays it's definitely something that I'm going to Be playing around with a lot now.
- [22:40]
Mike Kaput
I am also blown away by Deep Research and if I'm not mistaken, you are too. Can you kind of maybe give us a rundown of what you found so impressive about it so far?
- [22:51]
Paul Raitzer
Yeah. So Deep Research was one of those wow moments and as we've said on the show before, we don't have too many of those. Like when, when you like pay as close attention to this space as we, we, we have to and we do, you just, you feel like you see everything all the time and it's hard to be like really impressed by something. This is the first time where I used a product since Audio Overviews with Google, Google's Notebook LM product earlier this year where you had that wow moment of like, oh, this like changes things. Like this is, this is different than anything else we had. It's different than what ChatGPT can do. It's different than what Perplexity is doing, in my opinion. So yeah, I was, I was definitely impressed by it. So what Deep Research does is it uses AI to explore complex topics and then it kind of generates them and easy to read reports. And so it's using this Gemini model to do it. And so you put a question in, it creates a multi step research plan. You can either revise or approve that plan. It then goes and executes that and then it takes a few minutes, it goes and browses the web, it does the searching, the finding. It looks for interesting pieces of information and it keeps repeating that process until you get the final product. So I was trying to figure out like, well, what could I do here? And I actually had a research project I was looking at and I was looking into like pricing models and I wanted to look like broadly and then I wanted to go specifically into like five companies that I had in mind. And so this was something literally I was going to try and find time to do over the holidays. And I was like, seems like that's what this thing does. Let me give it a try. And so I put a very simple prompt in and it went and in about three to five minutes completed a research project that would have taken me three to five hours, easily three to five hours. So I put in a single prompt, it created the research plan which again it gives you the choice like edit or approve once it's ready. And the research plan is literally like, I'm going to do these six things like is this what you want me to do? And if you wanted to do something different, it'll do it. So it then visited Analyzed and summarized over 100 websites, and then it turned it into a Google Doc for me that I could review and edit from there. And this is. If you go back a couple episodes, Mike, you'll remember I said this was exactly one of the advantages for Gemini and Google is if I create something in chat GPT, it doesn't live in a productivity doc that I can work with. Right. It lives as this thing I have to have copy and paste over into either Microsoft Word or Google Docs. Well, now I don't have to worry about that. It just automatically creates. It's like nice and streamlined. So my initial reaction was like, this is far beyond anything anybody's doing and it has the benefit of Google search dominance and accuracy and theory. Like, I immediately trust that Google is going to find a way to solve for hallucinations and errors better than any other company because that's what they do in search. And so, you know, you and I had a talk right away, Mike. It's like, shoe man. How can we be using this in content production, course creation, podcast, you know, preparation? So we're already thinking about research, publishing and education, how a product like this that we know is only going to get better, how it can really change the way we do things. So again, I would definitely say give this a test, come up with like a use case or go online or, or I mean, ask, ask Gemini. Hey, I want to test deep research. What would be an example? Use dummy data if you have to. Or yeah, you know, a people use case. It will still make mistakes. Like I'd have to say, like, you're going to use this in your business or for your, you know, your career or for schoolwork or whatever. It's going to make mistakes. It'll still have errors. So do humans. So like, is the error rate going to be lower than a human? Maybe. In some instances. It probably just depends on what you're trying to do with it. But the human still has to remain in the loop. You still have to use your experience and expertise to decide the quality of the output you need to guide the research plan. So this isn't replacing, but this, this changes the way research is done and I, I can definitely see it transforming on us. What about you, Mike? I know you've played around with it too.
- [26:59]
Mike Kaput
Yeah, similar initial reactions, I would say too. It just excites me in the sense of not only doing these complicated and complex research projects so much faster, but also what it enables when you probably start combining it with other AI tools.
- [27:16]
Paul Raitzer
Right.
- [27:16]
Mike Kaput
Like these other Systems and models thrive on giving context and information. So instead of me, you know, I don't know, using a tool like ChatGPT or Claude and prompting it with a couple sentences like, hey, I'm trying to like, write something about, I don't know, the energy sector. I could be like, oh, okay, like we're writing something about the energy sector. By the way, here's seven to ten pages of perfectly researched information based on a hundred plus different websites. Yeah, it really opens up just so many more combinatorial possibilities, I think in my mind that I'm really excited to explore.
- [27:48]
Paul Raitzer
Yeah, I mean, I was trying to think back, like, if we'd have had this when we wrote our book in 2022. Matt.
- [27:54]
Mike Kaput
Yeah, right.
- [27:55]
Paul Raitzer
I mean, so if you combine, I would say, like, so I've written three books and I would say, you know, roughly between 50 and 60,000 words each. My guess is they take roughly four to 500 hours to do, like the full planning and production of the book, like writing of it, and read the research. I don't know how much of that would be shaved off with a tool like this, but it wouldn't be insignificant. Like most of the process of writing the book is, is the research, in my opinion, like, it's developing the initial outline and then doing the research. Not the first party stuff where you're going and conducting interviews and things like. That's obviously not replacing this, but I almost think, like, you'd have way more time to go do those things because so much of the research, like I could, I used to keep Evernotes on this stuff. Like that was the productivity tool I would use and I would have to create all these folders and then I'd go do the research and I would put them in Evernote and then I would have to go through and do summaries of them and I would print them out and I would highlight them. Like, I'm not joking when I. If I were to say if my second book I wrote by myself was 400 hours, 150 to 200 of that was probably researching, organizing the research, extracting insights from the research. Writing the writing part is easy. Like if, if you're a writer, like, that's the easy fun part. It's everything else that goes into it. That is where all the time goes in.
- [29:13]
Mike Kaput
It'll be interesting too, to see this play out in some of the areas that, you know, I know we're interested in getting more into because I look at something like this too, and I think, wow, this just Turned me into a one man public analyst firm.
- [29:28]
Paul Raitzer
Right.
- [29:28]
Mike Kaput
I can, oh, I don't need a research team of 20 people to go do this. I mean it's all public information. Of course there's more to analysis than just that, but it's pretty darn interesting to me.
- [29:38]
Paul Raitzer
Well, yeah, and then last week you and I had on top of that and this isn't in our notes I don't think, but we have CB Insights, right. As a research platform for the company because we do a lot of research for what we write and then we look into a lot of companies. And so we had a call with them last week and they were showing us their new AI capabilities includes like a chat cbi which is connected to their proprietary database. And so like I think your head in my head last week were in the same place of like, oh my gosh, like the future of analyst firms and research is like yeah, before our eyes being reinvented and how you do that. So yeah, when we think about like industries or business models that are going to be impacted very quickly, you could look at tools like this and say wow, you got to, you got to reimagine that business model real fast. Which is exactly what we're doing with SmartR X. That was the whole, whole premise. And again like some of this stuff I think is maybe obvious what we're doing or maybe other people, like they don't stop and think, they don't, they don't really care that much. But like from day one I built SmartRx to re reinvent the analyst firm. Like it's, we want to be a research firm but we want to do it faster, like real time research so it's not stale six month old stuff. Like we wanted to find a way to bring this stuff to market fast and this is the kind of tool that enables that.
- [30:54]
Mike Kaput
So just very quickly to wrap this up. I mean obviously Project Mariner, it's a research prototype but this has like in my mind like huge implications. I immediately just start thinking of everything we do as marketers. Like isn't this, this combined with like deep research, like am I ever visiting website ever again or are my agents going to do it for me?
- [31:16]
Paul Raitzer
Yeah, that's so I mean this is computer use again like we talked about computer use with Anthropic, we've Talked about with OpenAI. We said Google was working on this. We didn't know it was called Project Mariner but like everybody is working on this. We have known that for a while for like seven years. We knew it was happening. It's just the, the breakthroughs are happening now. So yeah, I think the unique thing here is you trust Chrome, you probably trust Google more than you know, you might trust some others. So you could start to see this really impacting, you know, search online behavior. You know, you and I coming from the world of like marketing and the analytics data and organic traffic. Search traffic. Like is it really humans coming to our site anymore or was it just Google Gemini Deep Research that hit our site five times? I don't know. And then, and then even with like the Google Deep Research and, and project manager, do, do we end up clicking on the things that surfacing, right? Like does it or does it just like remove the need? We trust Google so much that we're not going to click on any of the links. Like we don't care. Maybe you'll spot check three or four and be like, yeah, they nailed it. Like we're good. Yeah, I don't know. But yeah, I, you know, I don't know when Mariner is going to come out. I would guess sometime next year again just because everybody's there. But it's also dangerous. Like this computer uses is a really risky endeavor and there's lots that has to be solved by these companies before you can like roll this out at mass market scale and not have those because we've also we're not going to get in today, but there's been a whole lot of stuff in the last few days about the ability to jailbreak all of these models and get them to do terrible things and actually expose their instructions in them and things like that. And so that's a concern here is like the cybersecurity risk that goes into allowing a computer to access your screen for you and things like that.
- [33:03]
Mike Kaput
All right, so our third big topic this week is kind of some hands on experiments we've been doing with the full O1 reasoning model that was announced on the first day of OpenAI's 12 Days of Open AI.
- [33:17]
Paul Raitzer
That feels like three months ago.
- [33:18]
Mike Kaput
I get, I honestly wrote this sentence this morning. I was like this, that can't be right. Wasn't what that was last week.
- [33:26]
Paul Raitzer
Like, yeah, yeah.
- [33:28]
Mike Kaput
So you know, as A quick refresher is Zero1 model. It's notable because it takes time to think through problems. It uses step by step chain of thought reasoning. This makes it kind of operate in a fundamentally different way than previous models and it also means it's very, very good at certain types of complex problem solving tasks as well. Math, coding and Science. Some people even claim that it approaches the level of a PhD at certain tasks. So it's also multimodal, at least in the sense that it can process images and text together. So what we did, Paul, you know, you and I met, we did a bunch of different experiments with the tool and kind of learned, I think, quite a bit in a fairly short amount of time. So I kind of wanted to get your Initial thoughts on L1. What did you use it for? Where did you find it helpful? Where maybe did it fall short? And then I can also share some tests that I ran on the model this past week as well.
- [34:23]
Paul Raitzer
Yeah, so when we sat down do the hackathon, which I think I alluded to on last week's podcast, that we were going to run that and we would share our experiences. So we, yeah, I think it was about like an hour and a half, two hours, kind of met and kind of worked through a couple things. And what we were trying to get at was what are its capabilities? What are the use cases for businesses? Not like, we're not trying to solve complex math problems, we're not trying to, like invent, you know, pharmaceuticals. Like, we're trying to figure out business strategies and, you know, Personas and things like that. And so we wanted to look at use cases and then try and figure out do we want to pay the 200amonth or are we, are we good at this? 20, 30 bucks a month we're paying, Is that sufficient? And how would we use this on an ongoing basis versus, like buy it and forget we even have it and never, never use it. So the one I, again, I'm trying to like, whenever, whenever I'm testing things, I'm trying to use real life situations that I can assess whether or not this would actually make a difference in my life. I don't pick like random, really hard problems to like do an eval against. I don't find them terribly helpful to people. So we try and focus in on like, this is something we do, I do as a CEO or whatever. And if it can help me here, then it can help other CEO. It's kind of how I think about this. So one in particular I was looking at, I mentioned earlier, I'm spending a lot of time on our AI Mastery program and our online academy. And so I'm thinking deeply about like pricing models. And so I actually went in, I gave the same prompt to O1 as I did to GPT4O because I wanted to compare the model we're used to using, which is pretty good to this reasoning model and see what does it do differently in a problem that seems like it would be more complex to solve, more chain of thought required. So I went in and gave it like our AI Mastery membership program pricing model. I gave it a couple of things I'm thinking about doing and then I said like basically analyze this for me. This is my goal. How, how would I best achieve this? You know the outcome I'm looking for. And I told it ask any clarifying questions that you need. So with that simple prompt in mind, it was only with three sentences. One I'll say oh one. Again I give the same prompt to both models. Oh, one asked way more complex and nuanced questions than four. So immediately you could see that it was more deeply understanding and considering what I was asking of it based on the questions that came back to me with the same way you would assess a strategist. Like if I sit down and meet with somebody, the questions they ask often tells me their level of intellect and their ability to do strategy. Well, it gave a 01 gave a much richer explanation up front about like a synopsis of its answer. The scenarios it presented were way more thought out, there was way more reasoning that went into them and then it provided much more context and insights. Overall it was an overall longer output, but there wasn't a waste of characters or words. It was like all good stuff that would have if I was actually had time, I would have continued on working with it and really analyze some of these key areas. That being said, 4.0was formatted nicer like I don't for whatever that's worth. And 4o didn't do a bad job. But side by side, 01 crushed this one over 4.0. So I'll come back to like would we pay for the license? And things like that. Mike, after you kind of walk through your tests.
- [37:35]
Mike Kaput
Yeah, no. That's a great way to set up experiments with this. And I set mine up very similarly in the sense of just going head to head 4.0 with 01. I won't get into every single detail here, but I ran through four again, real world scenarios of problems I'm trying to solve in one context or another. Including we've had a lot of questions like hey, based on our podcast performance, should we be adding another episode doing more limited edition series. So I gave it a bunch of performance data, told it the problem and tried to solve that and got got recommendations on like hey, build a strategic plan for me that answers questions we might have, recommends what we should do did similar things for like building a content strategy. I gave it a bunch of historical performance of our blog and just said, hey, build me like two sentences. Build me a full comprehensive plan to maximize traffic through new content and updates to existing content. We gave it some information on course conversions, people using codes to buy our piloting AI course, and then asked a bunch of got a breakdown of a bunch of questions we should be asking, recommendations. And then last but not least, I just want to dwell on this one really quickly. We obviously run a ton of different workshops through marketing AI institute SmartRx. One of these is like implied AI workshop that I've run with a bunch of different teams. I run it at Macon every year and what you come out of it with is tons and tons of different use cases for AI in your organization. What we do with that is our team of experts sits down, analyzes all that intro and writes, you know, like a brief that's pretty extensive and comprehensive after a workshop. It takes a very long amount of time to fully do, but it's really valuable. And I basically just gave 01 like dummy data from like that kind of just like anonymized and like from one of our workshops that we had ran internally and then said like, here's how the workshop works, here's what's in the final brief. Go make one. And I was blown away. It did an incredible job. It would be like a very good first start. Like we got it from a human strategist. I would be like, yeah, this person did a fantastic.
- [39:46]
Paul Raitzer
Like if you got it from an entry level employee, you'd be like, this person's moving up fast.
- [39:50]
Mike Kaput
Yeah, easily. So I mean overall kind of similar takeaways to you. Oh, one is legit. Like again, these aren't PhD level problems. I'm not even equipped to evaluate a PhD level problem. So I'm sure there's more robust ways to test this. But stacked up against four zero, hands down, more robust, comprehensive and helpful. I hate the fact you can't upload files or spreadsheets yet. I hope that changes the way I got data in there was just hacking it by copying and pasting with a bunch of unstructured data, which actually did a great job with somehow but I would love if we got that. And then of course I see a lot of people talking about this online, so I feel like I'm not the only one here that's struggling. But I still don't get the sense I've pushed it fully to its limits because I Don't know how to structure a PhD level science problem and then evaluate it. But I'm definitely impressed so far, much more so than I think I was during the preview phase when I was like, still struggling to figure out what do I even use this for.
- [40:51]
Paul Raitzer
Yeah, I, I agree. And I think that, you know, that's our, you know, overall takeaway is the, the $200 a month. I think you and I learned that on this mic. It's not worth it unless you plan on like using Sora a ton. Like, because the Sora use is like built into the 200. But there's nothing in the $200 license that's going to get you something on the 01 model that you're probably not going to be good with with the existing license. Is that right?
- [41:18]
Mike Kaput
I think so. I'm interested that right now there are differences in the context windows. But again, it's like if we, if we had like two or three of these workshop type use cases where it's like, hey, we're doing this every month and this saves us however many hours per month on those things, then sure, I could see a use case for it. But yeah, I think your average business, unless you really put in some work to identify some crazy valuable use cases, you probably don't need beyond.01 of what you get in a plus or a team license.
- [41:49]
Paul Raitzer
Okay. Yeah. And I think, you know, one thing that this just highlights for me, you know, Mike and I were talking about like if, if we gave this to an entry level. Personally, I always think back to like my agency life. Like when I was running an agency and at its peak, I think we had like 20 people roughly, and we hired a lot of younger professionals and you would spend years developing strategic capabilities in them. Years. Like it's not something most people maybe like MBA programs and stuff might come out of it that way. But we were hiding out of communication schools, business schools, marketing backgrounds. A lot of times they're not trained to be analysts and you know, strategists in college really, like, you get some peripheral stuff, but like, you need experience to get that. And so when I think about that or you know, even, even now, like with our companies, the question that a lot of industries are going to have to deal with is, is this better in its O1 form, like knowing it's going to get smarter? Is it better than the average human who would otherwise do this? And I think, and this is what, you know, I said when we talked about AGI on earlier episodes, like, I don't even care if we get to AGI if we ever agree on what it is. And we'll talk a little bit more about AGI in a couple minutes. But like my whole premise is it doesn't matter. Like if the AI models keep getting better, thinking, reasoning, understanding, imagination, which is a whole separate thing. And we'll talk about at some point creating this stuff. If it does those things better than the average human who would otherwise do the job in your business, in your industry, then we got some problems in, in terms of the future of work and workforce and the economy. And here's the reality. They're already better than the average human at a lot of things like it, it's. And if you think about your teams, if you think about your organizations, there are very few organizations that are all A players. Like there are some exceptional ones that are dominated by A players, but there's a lot of B and C players in most companies. And my concern is these models are already at B, C player level.
- [43:52]
Mike Kaput
Yep.
- [43:52]
Paul Raitzer
In most use cases in knowledge work. And once you can like string those together and once you have like a full blown plan of how to integrate these models and use them to their fullest capability, who cares about solving biology and math and you know, scientific problems, if they solve business problems. And they're really good at that right now with human oversight. So yeah, this is my, I don't know, concern going into 2025. I just think it's going to become a reality for a lot more people next year. They're going to realize how capable these things already are of doing a lot of knowledge work at or above average human level.
- [44:33]
Mike Kaput
Yeah. I haven't seen a single thing so far to disprove that statement that it's already doing B and C level players easily.
- [44:39]
Paul Raitzer
Mostly doing A. I mean that's the problem. It's actually doing eight more seconds.
- [44:42]
Mike Kaput
I know. Yeah.
- [44:44]
Paul Raitzer
If you knew there was no errors, like if you knew it was above human error rate. Yep. It's likely already at a player level in a lot of what it does.
- [44:54]
Mike Kaput
Yeah, it's going to be a wild 2025 spoiler. There's a bunch of questions around that for the 25 Questions episode that we.
- [45:02]
Paul Raitzer
Do have to laugh. One other note here, we'll drop it in because I don't think we have this in the show notes, Mike. We'll. We'll drop this in as a just like a little kicker for people if you want to check the show notes. So the Klarna CEO related to this. We've talked about this Klarna, like the customer support agent company, whatever. And so he's like doing these interviews where he's saying, we're not hiring any more humans. There's this natural attrition and like 20% a year, our people leave and we just won't backfill. We went from 4,000 employees to 3,500 employees. And you listen to these interviews and everybody's tweeting this thing like this, this sound bite, and it's like, oh, my God, it's happening. Yet you go to their company website and they have 56 job openings. They're which I assume they're not hiring agents, but, like, they're trying to hire 56 humans. So it's just like, I guess it's one of those loans of like, let's all just like, pump the brakes and realize there's going to be a lot of hype still. Jobs are going to be impacted, but it probably isn't going to be as bad as it sounds in a lot of cases. And there's probably more to the story when you hear about this stuff. So don't overreact. Let's, like, step back and analyze the situation. And we're going to do our best to be as objective as possible next year because I think this is going to become a very real thing in some industries, and we want to make sure we don't get caught up in the hype of it all.
- [46:18]
Mike Kaput
All right, let's dive into a bunch of rapid fire for this week to kind of close out the year here. So first up, Perplexity is pitching investors on a vision of rapid growth. They claim they're going to double their annualized revenue next year to 127 million and quintuple it by the end of 2026, which would put them at about $656 million. These projections came from a pitch to investors because Perplexity is currently in talks to raise $500 million at a 9 billion valuation. This was all reported on by the Information. Perplexity's business model seems to primarily hinge on subscriptions it projects growing from 240,000 premium subscribers by the end of this year to 2.9 million by the end of 2026. However, it is also exploring other streams of revenue, like affiliate links within search results, in spite of paying AI providers like OpenAI tens of millions of dollars for their technology, which is what Perplexity uses to do its thing. Perplexity still claims it can achieve gross profit margins of 75% by year's end and are eventually targeting 85% margins. So Paul, I don't know about you, but these sound like some pretty optimistic projections to me. Like, do these numbers bear any resemblance to the reality of Perplexity's business model as you see it operating today?
- [47:42]
Paul Raitzer
I'd love to see the deck. Like, I'd love to learn more. Like, here's my high level take on this. And again, I may be completely wrong. Like anyone listens to the show a lot. Like, we've talked a lot about Perplexity this year. Awesome product. I use it all the time. A lot of times it's better than Chat GPT at certain things. A lot of times it's better than Google at certain things. I think my overall thought here is they're going to get Aqua hired next year. Like, I think if I had to bet money on what happens to this company, yeah, it's really, really hard to sustain what they're doing and hit Matt, mass market like takeoff because they don't have their own models that it's, it's largely, I mean, sure, they've got some amazing algorithms, amazing engineers, they're doing really cool things. But is it anything that Google or Chat GPT couldn't replicate or make better? Like just seeing the deep research project from Google is immediately like, oh yeah, they've got some stuff they haven't shared with the world yet. And you look at that and you think about Google's distribution and OpenAI's distribution and like how many users those other companies have who could likely emulate what Perplexity is already doing. And can Perplexity get to that mass market fast enough to where someone else doesn't just like replicate what they're doing who already has a billion users and then like, who needs Perplexity, right? And that's my, my fear is they're going to realize that their investors are going to realize that. And I could see this company just folding in like Adept or some of these other companies inflection like we'll talk about in a little bit here. I don't know again, they might, they might hit, they might like take off, but 2 million users, nothing like. Right, Kudos like that. It's great for a startup that in the space they're trying to compete in nothing and it's not changing market share one way or the other. Like, and so that's my concern is that they're trying to take on like a massive, massive market with dominant players with way bigger teams and much better technology and probably Much better algorithms. And I think once we see all that stuff coming from those other labs and frontier model companies. Yeah. I just don't know. It sustains. Again, love the product, love the company. Don't love their business practices necessarily. But I also feel the other thing. I know this isn't a main topic, but the other thing that keeps, like, picking at me, like, bothering me a little bit with them is it feel like they're throwing a lot of things out there, like, perplexity for finance, perplexity for sports. Like, it's almost like they're. They're trying to seed a bunch of things to figure out where is the thing that we can take off, what is the market we can blow up in. Right. And I don't think they know yet. And that's the other thing that worries me is, like, that funding can go really fast if you don't hit on something. And I feel like they're just, like, trying a whole bunch of things right now.
- [50:42]
Mike Kaput
Interesting. Yeah, I think that's a smart observation.
- [50:45]
Paul Raitzer
Yeah.
- [50:47]
Mike Kaput
All right, so next up, venture capitalist Marc Andreessen just dropped a bombshell claim about the Biden administration's approach to AI in a recent interview. In a discussion with Barry Weiss on the Honestly Podcast, Andreessen described, quote, absolutely horrifying meetings that he was in with administration officials where they allegedly outlined an intention to, quote, completely control AI technology in the US he told Weiss that during meetings with admin officials, he said, quote, they said, look, AI is a technology basically that the government is going to completely control. This is not going to be a startup thing. They actually said flat out to us, don't do AI startups. Like, don't fund AI startups. They basically said, AI is going to be a game of two or three big companies working closely with the government, and we're going to basically wrap them in a. And he says, I'm paraphrasing, but we're basically going to wrap them in a government cocoon. We're going to protect them from competition, we're going to control them, and we're going to dictate what they do. He says he expressed to the administration skepticism that they could exert such control over AI. And in response, he says, quote, they literally said, well, during the Cold War, we classified entire areas of physics and took them out of the research communities and entire branches of physics basically went dark and didn't proceed. And that if we decide we need to, we're going to do the same thing to the math underneath AI. Andreessen even indicated these discussions motivated his somewhat controversial endorsement of Donald Trump for president. So, Paul, this certainly seems like it gets into conspiratorial territory, but this is pretty wild. If the gist of it is even true. It would indicate, in my mind at least, a much more, like, sinister approach to how the government was looking at influencing AI.
- [52:42]
Paul Raitzer
This is one of the craziest video experts I've ever seen. Like, I haven't listened to enough of Andreessen's interviews to know whether he exaggerates. I, I like, at the moment, I probably need to, like, take him on his word. Now, keep in mind, he wasn't quoting, he was summarizing. He said, I'm summarizing what they said. But directionally, if what he is saying is accurate, one, I would love to hear from someone else who was in the room. Two, I want to know who the government officials were who were telling him this. And three, I want to know when it happened. Because if you recall, Mike, in October of last year, he released the Techno Optimist manifesto that we talked about on this podcast. And I thought it was crazy. Like, I thought that the manifesto was crazy if this actually happened. If. If he sat in a meeting with top government officials and was told that they hid elements of physics from society during the Cold War, right. Or World War II, whatever it was, and that they intended to shut off AI development and investment and pull it all into two or three companies. That is insane. And that is regulatory capture by definition. Like, at its peak. Like, what we talk about, this idea of, like, Sam Altman and a couple is trying to, like, influence regulation and, like, control it all, and then the government would basically invest in only those companies. Like, this is what we pros was under. This idea of regular capture, I couldn't have ever fathomed it would be to the level he's describing here. So if this is true and it can be verified by somebody else, we know who said it and when it was said. A whole lot of what happened in the US election in Silicon Valley makes a hell of a lot more sense now than it did to me seven days ago. So when you saw these people throwing their support behind the Trump administration and, like, not to the joy of the portfolio companies they invest in and their peers, if this is true, I actually understand at a very different level why they did what they did. So I'm anxiously awaiting someone besides Mark to verify this is exactly what happened. But again, I have no reason not to trust what he's saying. I don't. I don't know him and I don't know enough about past interviews with him to judge whether or not he would exaggerate this. But it's nuts.
- [55:17]
Mike Kaput
That's, it's crazy. I mean it's basically on one hand just saying we basically almost nationalized AI or wanted to 100 and also I think it's funny because I think in some pre election episodes you and I may have mused a bit on like, ah, turns out AI wasn't actually that big of an issue politically in this election. But I wonder if behind the scenes it turned out it was the issue. If it was the issue, we'll see. I don't want to overfit that theory, but that would be really interesting to me.
- [55:49]
Paul Raitzer
And I saw him do a second interview with a different journalist about this and he said the same thing. Yeah, so I mean he, he's, he's committed to these talking points. Yeah, I just, like I said, I want, I want to, I want to hear from someone else who heard these same things. If anybody's willing to like step forward and say it, I will be watching very closely for verification from another third party about this. And I really want to know when it happened. Because if it was in September to October of last year, then the Techno Optimist manifesto makes a hell of a lot of sense.
- [56:21]
Mike Kaput
Oh well, it's not the only controversial thing coming out of the world of AI this week. So we actually in our next bit see that a member of OpenAI's technical staff named Vahid Kazemi posted on X that he believes the company has already reached artificial general intelligence. In part of this post he posted, which we'll link to the full one in the show notes, he said, in my opinion, we have already achieved AGI and it's even more clear with 01. We have not achieved, quote, better than any human at any task, but what we have is, quote, better than most humans at most tasks. So he noted that some people criticize LLM saying they can only basically kind of follow a recipe to do what they do. But he did say no one can actually explain what a trillion parameter deep neural net can learn and that if you really look at it, the scientific method itself can be summarized as a recipe. So he basically concludes saying there's nothing that can't be learned with examples, implying there's kind of no real limit here to what AI models will be able to learn and do. So Paul, we've talked episode after episode. There's tons of speculation within different circles of the AI community about Whether or not we've actually achieved AGI or when, if at all, we will achieve it. But should we be taking this prediction more seriously since it's being literally, openly said by someone at OpenAI?
- [57:46]
Paul Raitzer
Yeah. So, I mean, we've talked about this many times. The lines are really blurred because we don't have a uniform definition, we don't have a means of testing it, and they keep moving the goalposts of whether or not we've gotten there. So he used the same explanation I did. Basically this idea that, like, it's already better than most humans at most tasks. So, like, if that's not AGI, like, what. What are we doing here? I think this one, it. It. It does highlight the. The deficiency of current evals. Like, the way they evaluate these models is they run them up against the most complex challenges known to man, like, hardest tests. And, like, it's like, oh, well, if it does this one, then maybe we'll be there. If it does this one, then maybe we'll be there. And it goes back to, in my opinion, what you and I talked about with 01. It's like, forget the evals. Does it do my job better than me? Like, here's the 20 things I do. How many of those 20 things is it better than me at? And I can tell you, I'm willing to admit it's getting better than me every day at something I do. And that's okay. Like, right? I'm okay with that. I got a million other things to go work on. It's okay if it's better than me at things, but a lot of people aren't at that point that they won't even comprehend that these things are getting better than them at something and that they're okay with them getting better at something. Now, I think, like, the other thing it highlights is this uncertainty about what's coming. So I'll highlight. And again, this isn't in the show notes Mike, so I'll drop the link in there. But yesterday, Logan Kilpatrick, who we've talked about before, formerly OpenAI, now he's lead product for Google AI Studio. He tweeted, pre training is only over if you have no imagination. Seems like kind of an innocuous, like, whatever. Like, you know, clever Logan, Vague posting, but here's the track. I actually think he was implying something here. So imagination is something we assume only humans can do. If AI models were deemed to be able to have an imagination, it would mean they could create or synthesize or envision new ideas. New scenarios, new concepts that aren't in their training data. And if he's implying that, then maybe he's implying Google is actually on a path to unlock imagination and metacognition, which means being aware of your own thoughts, in essence. And so like, I, I don't know, like I feel like depending on what you want to call AGI, a lot of AI researchers probably think we're already there and they're already focused on super intelligence. They're already focused on the next next level beyond AGI in their, in their world. But I don't know. We're going to spend a lot of time talking about AGI. We've got a whole new series planned that we'll announce in January. We're going to go deep on this topic, including interviews with experts and stuff. So stay tuned. This is like a huge area of focus for us next year.
- [60:39]
Mike Kaput
All right, so next up, Amazon is setting up a new lab in San Francisco. It's called the Amazon AGI SF Lab and it's actually led by David Luan, who is the co founder of AI startup Adept. This lab aims to develop AI agents that can perform complex tasks across digital and physical environments. That includes handling sophisticated workflows in software tools and executing actions in the real world. This lab is going to focus on AI models improving through human feedback that can self correct their actions and understand user goals. Interestingly, this lab is going to initially be staffed by Adept employees. Amazon says it also wants to hire a few dozen researchers in different fields like quantitative finance, physics and math. This year actually, Adept agreed to license its tech to Amazon and Luan and some of the Adept team joined the company. So Luan is actually working under Rohit Prasad, the former head of Alexa who now leads an AGI team at Amazon. So Adept technology was actually originally designed to help create an AI teammate that can use any software tool. So Amazon seems to be using that to make a big play in this field of agentic AI. So Paul, this struck me as interesting for a few reasons. Like first, it's a pretty open commitment from Amazon that they're pursuing some type of whatever they consider AGI. And I thought it's kind of curious they're planning to hire researchers in all those other areas like finance, physics and math.
- [62:15]
Paul Raitzer
Yep. Yeah. And also interesting, this is the Aqua hire I referenced earlier. So Adept was acqui. Hired by Amazon in. They announced it July. Well, let's see. Yeah, June. July, end of June 2024. So Adept has raised 415 million. They had raised 350 billion Series B in March of 2023, not that long ago, year and a half ago, they raised 350 million. And then they announced June 28 that their mission that they started had been to build general intelligence that enables people, computers to work together. So again, computer use, our plan has been to train progressively larger and smarter multimodal foundation models. They basically admitted in their own post that like, we can't compete on this, like we're not going to keep up on building these frontier models. So that they're going to kind of partner on this. And instead, in addition, the Adept co founders and some of the team are joining Amazon's AGI organization to continue to pursue the mission of building useful general intelligence. Amazon is also licensing Adept's agent technology family of state of the art multimodal models and a few data sets. Actually I never like really thought about that. A few data sets meaning like anything you gave us, we just gave to Amazon, thank you very much. So yeah, that's the backstory there. But more computer use, more agents and Amazon wants a piece of it.
- [63:38]
Mike Kaput
So kind of an interesting corollary to this. So there's an AI agent startup called Sierra that is founded, was founded by the former Salesforce Co CEO named Brett Taylor and he's also the chairman at OpenAI and they're introducing in this article, they came out with this past week, a new way to pay for software in the era of AI agents. So instead of the old seat based model where you would pay a fixed fee for software licenses or even some type of like usage based pricing where the bill would scale based on how much of the software you consume, they've come up with kind of like outcome based pricing that will change only when that will charge rather only when an agent achieves a real measurable result. So for example, if an AI agent resolved a customer issue or completed a valuable task, that's when CRS says it wants to get paid. If it doesn't achieve the outcome, there's no fee. Now the company acknowledges this approach might not be the best for every situation. So they're kind of open to exploring different models, but they basically want to focus on the tangible business impact that agents can achieve and then get rewarded financially for that. So Paul, this is definitely fascinating given the background of Sierra's founder. Seems pretty interesting, maybe innovative, but is this like even remotely feasible? Like not even half joking? Like the invoicing alone for this feels like it would be a nightmare.
- [65:03]
Paul Raitzer
Yeah, I mean I think they gotta find a way to do something like this. And if you. A little bit more on Sierra. We did spend some time on episode 116 talking about when we were talking about AI agents in the Enterprise. We actually covered an interview Brett had done at that time where he kind of went in depth on a lot of this stuff. So if you want more background, Sierra and Brett go check out episode 116. Yeah, I think, I think this, this, this terrifies SaaS companies like the, yeah, the pricing model has been user based seat licenses. So whether you know, like, you know, HubSpot, Asana, Google, OpenAI, like all of them, like, like every, all the tech we use to run our company, we pay a SEAT license for per user. So if, if we're able to train agents to do the job of different people where we don't need them to have a SEAT license, we just train up an agent and it's doing the work of five, what would have been seat licenses? What choice do they have? Like that future is coming there, there will be agents doing the work of humans within every SaaS product. I don't care if it's a financial product, an HR product, a marketing, whatever, they have to find an alternative. And so it can be a usage based thing. You know, you paying for GPU time to do a thing, it could be equivalent of human pay. I think OpenAI has been looking at this, there's been some words and some reports about that that like, well, if you're going to pay a human 150,000 a year to do this and the AI is going to do the equivalent of three FTEs, right. You're, you know, wouldn't you be willing to pay 2,000amonth for that? You know, it's still only 24,000 for the year. You're saving, you know, $350,000. So I think there's going to be a lot of models experimented with. And I go back Mike, like, you know, in the days when I was owning and running PR2020, my agency, all the stuff we went through with HubSpot as their first partner starting in 07, how many iterations of their pricing model did we deal with in like the 12, well, 14 years or whatever that I ran the agency while we were their partner and that was the growing up of the SaaS industry when all everything became SaaS. And so I can imagine there's going to be a whole lot of that experimentation with what is the pricing model where we don't tank our market value like our market cap for publicly traded. In the process of figuring this out. And they're going to have to figure this out fast because I think by 2026, this is a real problem in the SAS industry.
- [67:45]
Mike Kaput
All right, so next up, OpenAI is trying to transition from a nonprofit to for profit company. We've talked about this in the past, but they're facing some new challenges because Meta is now joining Elon Musk's fight against the AI company's restructuring plans. So Meta has asked California's Attorney general to block OpenAI's planned conversion. They argue the company shouldn't be allowed to use assets built as a charity for private gain. This intervention comes as OpenAI actually took the step of formally publishing more details about its history with Elon Musk. They kind of came out guns blazing, revealing that Musk had wanted to convert OpenAI to a for profit structure back in 2017. 17, if you recall, his argument is based on the fact he doesn't think that they should be able to do this. So they're kind of coming out and saying, well, that's BS. So according to OpenAI's account, Musk had demanded majority equity, absolute control, and the CEO position of a for profit Open AI. When they rejected that, he resigned and founded his own AI company, xai. Now Meta, on the other hand, is just appealing directly to the government. They're arguing OpenAI's conduct could have seismic implications for Silicon Valley. It can set a precedent for organizations to launch as nonprofits, collect tax free donations for R and D and then convert to for profit status once their tech becomes commercially viable. So Paul Elon is no longer the only one with knives out for OpenAI. Like, is there any real argument here or is this just kind of like opportunism to like settle scores, slow down progress? Like, are they really worried that everyone's going to be a nonprofit and raise money that way?
- [69:30]
Paul Raitzer
Yeah, so I, I think I mentioned this last time this came up. I, I honestly think Elon's just messing with them and just trying to slow things down. I, I don't think he really thinks that they're going to stop them or like win this, but he's got plenty of money to throw at this just to, you know, entertain himself. So I, I, I, my guess is that's what's going on. But I'll just read two quick Excerpts from the OpenAI letter Because that's the new thing here, is them just kind of laying this out. Now they've talked about this before, but this is like the most direct with all these examples. So the OpenAI letter, which we will put in the show, notes Elon Musk's latest legal filing against OpenAI marks his fourth attempt in less than a year to reframe his claims. However, his own words and actions speak for themselves. In 2017, Elon not only wanted but actually created a for profit as OpenAI's proposed new structure. When he didn't get majority equity and full control, he walked away and told us we would fail. Now that OpenAI is the leading AI research lab and Elon runs a competing AI company, he's asking the court to stop us from effectively pursuing our mission. You can't sue your way to AGI. We have great respect for Elon's accomplishments and gratitude for his early contributions to OpenAI, but he should be competing in the marketplace rather than the courtroom. It is critical for the US to remain the global leader in AI. Our mission is to ensure AGI benefits all humanity and we have been and will remain a mission driven organization. We hope Elon shares that goal and will uphold the values of innovation and free market competition that has driven his own success. And then it lays out everything. All the emails, all the filings. It's somebody on the PR team had a lot of fun putting that post together.
- [71:06]
Mike Kaput
Yeah, I think that Sam Altman might have helped write that one.
- [71:09]
Paul Raitzer
Sam had a lot of fun approving that post.
- [71:13]
Mike Kaput
All right, so Next up, former OpenAI chief scientist Ilya Sutskeverg made a rare public appearance at this past week's AI conference, the prestigious AI conference Neurips this past week. He during that appearance made some striking predictions about AI's future. He said that we're approaching a fundamental shift in how AI systems are developed and trained. He declared that pre training as we know it will unquestionably end. He says we're hitting this concept of peak data so we're running out of new data to train AI models. He says we have to deal with the data that we have. There's only one Internet and this limitation will force the AI field to evolve beyond current training methods. He also kind of confirms some trends that we're seeing. He predicts the next generation of AI systems will be truly agentic and they will develop genuine reasoning capabilities. But he did warn this evolution comes with new challenges. More advanced reasoning means AI systems may become more unpredictable and they may even develop self awareness and desire rights. Though he noted that AIs wanting to coexist with humans and have rights quote is not a bad end result Satska.
- [72:29]
Paul Raitzer
Verde is pursuing, I assume is a bad end Result.
- [72:32]
Mike Kaput
Yeah, yeah, we've got some big fish to fry even if that is the end result I would say. But you know, that's kind of the whole concept behind his company. He's pursuing all of these ideas through his new venture, Safe Superintelligence, which has a billion dollars in funding to I know, figure out AI rights maybe. So we don't often get public comments from Ilya, which is kind of why we're mentioning this. So what did you think of these predictions?
- [73:00]
Paul Raitzer
Yeah, we'll put the link to the full video in the show notes. It's about 26 minutes long I think. Yeah, I, I, I think there's a lot to analyze here probably main topic level analysis that we'll save for the new year. But just keep in mind, I mean he was leading the team building Strawberry which became the O1 model. As we talked about recently it timing wise it seems like he raised red flags internally when they realized that the time test compute this reasoning model approach would scale. While he is alluding to the fact that this like scaling laws as we've previously known them of more compute plus more data seems to be hitting a wall. That doesn't mean that we're going to stop having this exponential growth in the capabilities because that leaves algorithms. We can find more innovative ways to do this. We can introduce imagination. Like there's other things that can be unlocked that seem to continue up and to the right basically for these models capabilities. So yeah, just I'm glad he's back in the public eye and I hope we hear more from him because he has a lot to say and he's back going back to 2012, 2014. He has basically predicted everything in deep learning correctly.
- [74:13]
Mike Kaput
So when Ilya talks, we got to listen.
- [74:15]
Paul Raitzer
Everyone listens, including all the top AI researchers.
- [74:20]
Mike Kaput
All right, just a couple more topics to wrap us up here. But next up, AI expert Ethan Molich just wrote a really great essay on when you should use AI and when you should not use it. Now I would recommend go read Cole essay, we'll link to it. But basically he says look, you can use AI much more effectively when you understand when it can help and when it can hurt. It's most useful for tasks, he says, where quantity matters, like quickly generating many ideas to find a strong one or projects where you're already an expert and can easily spot mistakes. It's also helpful summarizing large amounts of information when accuracy is not absolutely a hundred percent critical. It excels assisting with repetitive and low value tasks so you can focus on more meaningful work, and it's good at producing variations on your writing or offering second opinions that spark fresh thinking. However, he says it cannot substitute for learning deeply or struggling with new ideas. You shouldn't rely on it in situations where perfect accuracy is vital and it fails in unexpected ways. So using it means understanding its limits. And of course, when the effort, the struggle is actually the point of a task, AI shortcuts can deprive you of important insights. So Paul, I thought this was a pretty practical perspective on when you should be thinking about using AI. The guidelines are really helpful. Seems like good required reading. It doesn't take too long to get through. It occurred to me also this would be great to upload to a custom GPT and like determine when should I be using AI. So what did you think when you were reading through all this advice or.
- [75:57]
Paul Raitzer
Throw it into a notebook? For sure. Yeah, I just. I agree. I think I would just go give it a read. For your personal use of AI going into next year, or if you're involved in trying to educate a team or drive adoption within an enterprise, this is a good framework to think about.
- [76:15]
Mike Kaput
All right, so for our last topic, Paul, we've done this a couple episodes in a row here, but we have a ton of really quick like product and funding updates. So I'm just going to go through these quickly like almost mini rapid fire section and if anything jumps out to you or you want to comment on anything, please feel free to jump on in. If that works for you.
- [76:34]
Paul Raitzer
Let's do it. Cool.
- [76:36]
Mike Kaput
So first up here, Databricks is about to potentially make venture capital history with one of the largest private funding rounds ever. The Data analytics and AI company is finalizing a deal that could exceed 9.5 billion and that values the company at over $60 billion. Now, basically, rather than funding operations or expansion, this was apparently going to be used to buy back expiring restricted stock units from early employees. And it kind of mirrors a similar move that Stripe made last year for comparable purposes. So basically, Databricks has positioned itself, it seems, really well in the AI boom. They provide a bunch of tools that help people build and deploy AI applications using their existing data. And they are kind of in direct competition with Snowflake, which currently has a $56 billion market cap. So it seems like they're going after a pretty big rich market. Next up, XAI has announced major upgrades to its Grok AI Assistant. It's now available to all users on the X platform and has some significant performance improvements. Grok 2 runs three times faster than its predecessor, has better accuracy, better at following instructions, and improved multilingual support. It also has web search functionality and citations now, so it can actually get info from both your X posts and the broader Internet. They are trying to also add new visual capabilities through Aurora, their image generation model. And there's a new Grok button which appears on posts across user timelines to give you context and analysis into real time events and trending discussions. Kind of a fun one that I've seen a lot as well. There's like a Draw Me feature where it'll generate what it thinks you look like based on your X profile, so I guess use that with a bit of caution if you're posting a lot of nonsense on X. Another big update, Mark Zuckerberg released an end of year video on Meta's AI plans. In a quick video he posted to Facebook and Instagram, he mentioned that Meta AI has nearly 600 million monthly active users and that Llama has become the most adopted model with more than 650 million downloads. He also noted the final release for the year which we covered last week, which was llama 3.3. This is a text based 70 billion parameter model that performs as well as the company's 405 billion parameter model and runs more efficiently. He mentioned quote the next stop is Rama 4. Lastly, he talked about the company has announced that they're building a 2 plus gigawatt data center in Louisiana that will be used to train Future versions of Llama. Google just today actually unveiled significant upgrades to its AI media generation capabilities. It announced both VO2 for video creation and improved versions of its Imagen 3 image generator. It says both systems have achieved state of the art results against competing models. VO2 represents a big advancement in AI video has enhanced understanding of physics and human movement. It can create videos up to 4k revolution and apparently several minutes in length. On the image front, the upgraded Imagen 3 promises better composition and brighter images and the system. The system is now rolling out globally through Google's Image FX tool in over a hundred countries. Last but not least, they introduced a new experimental tool called Whisk, which combines Imagen 3 with their Gemini AI to allow users to create images by mixing and matching visual elements. Next up, Pika just announced the newest version of their AI generation video generation tool, Pika 2.0. This has a really cool trailer with it and it's kind of clear that both the trailer and the positioning of the update are kind of taking aim at Sora from OpenAI because a digital version of Sam Altman is in the trailer, kind of looking worried about Pika's capabilities. The company touts it as video generation AI that is, quote, not just for the pros, but quote, for actual people, even Europeans. Which alludes to the fact you can't use Sora in the EU at the same time we're all starting to experiment with Sora. OpenAI has ironed out some initial hiccups. There was a really good recap that we'll link to from a 16Z partner, Justine Moore, who said that the tool is really good at photo realistic 5 second videos, but 10 to 15 second ones are hit or miss. It's good at both editing and generating multiple consistent clips in one pass, but as of right now, she says it's not a world model with any type of realistic physics like some people claim that it would be. Microsoft's new Recall feature, which is coming to its AI enhanced Windows 11 PCs, is turning heads from some early testers at the Verge because they found it in a new piece they wrote both unsettling and also helpful. Because if you recall, Recall automatically snapshots everything shown on your screen and gives you this like scrollable timeline that includes emails, chats, personal research. They found that this data logging can be really useful. It helped them locate lost web pages and save a bunch of time on work they were doing. But the Verge kind of concluded that the unnerving nature of this really kind of like ruined the experience a little bit and there's still all these questions about security and data retention, a reviewer said. While I found some early examples of recall helping me out, I still need time to figure out whether I want to keep it enabled. I'm still wary of storing a digital trail like this on my laptop. And then last but not least, to round it out, one more thing from Google NotebookLM, which is Google's popular AI powered research assistant, is rolling out a new interface, an interactive audio feature, and a premium subscription version. The updated design reorganizes the tool into different sections to make it easier to work within the tool. NotebookLM now lets you what they call quote unquote Join the Audio Overview so instead of just listening to the AI generated podcast of your sources, you can actually speak directly to the AI hosts.
- [83:09]
Paul Raitzer
That's crazy.
- [83:10]
Mike Kaput
It's unreal. I really look forward to I haven't had a chance to yeah, me neither. I'd like to really take that for a spin, I think over the holidays. And look, while it's still experimental, it shows that Notebook LM is Kind of really evolving quite quickly. And they've actually introduced Notebook LM plus, which is a subscription option with higher usage limits, more customization, more team collaboration features. So this is going to be offered to businesses and educational institutions with Google Workspace. It's going to be sold separately through Google Cloud, and it's going to be included in the Google One AI premium tier in early 2025. So, Paul, that is a heck of an end to a heck of a year in AI, I would say.
- [84:01]
Paul Raitzer
Yeah. I mean, with a mega episode touching 1,120something here. Yep. Yeah, it's. Yeah, it was a crazy year. And I think it's only, you know, a hint of what is in store for 2025.
- [84:13]
Mike Kaput
Mike, I would agree. I would agree. And we want to make sure we really just tell the audience very quickly how appreciative we are for their support this year. I hope everyone has a happy holiday and, you know, not only get some rest and relaxation, but it's probably going to be a good period to test out some of this cool stuff that just came out.
- [84:32]
Paul Raitzer
Yeah. And to echo that, you know, just grateful for everyone listening and watching. We. I shared on LinkedIn last week, I think we started this in 2021, 10 episodes, 1500 downloads. 2022, 18 episodes, 4, 800 downloads. 2023, 50 episodes, 262,000 downloads. And 2024, 51 episodes, 400,000 downloads. So we really appreciate everybody showing up every week and listening to what Mike and I have to share. It's fun for us to do it. I've said before we'd be doing this if no one was listening, but it's a lot more fun when people are listening and we're getting to hear back and engage with those people. Kind of alluded a little bit to this, but we got some big plans for next year. In addition to the weekly format, which isn't going anywhere, we're going to introduce a collection of new formats and episodes. We're going to bring in some outside perspectives, some experts in different areas of AI and business and society, and talk about AI trends and innovations from some different perspectives. And so we just appreciate being part of your AI journey. And Mike and I both wish you and yours a happy holidays and new year. And so we have one more to go. We'll be back on December 19th for that special 25 AI Questions episode. And then we will talk to you again in the new year with our first weekly of 2025, which is weird to say on Tuesday, January 7th. So thanks again. Everybody. We look forward to being back with you again in 2025. Thanks for listening to the AI Show. Visit marketingai institute.com to continue your AI learning journey and join more than 60,000 professionals and business leaders who have subscribed to the weekly newsletter, downloaded the AI blueprints, attended virtual and in person events, taken our online AI courses, and engaged in the Slack community. Until next time, stay curious and explore AI.