← The Artificial Intelligence Show
The Artificial Intelligence Show
#126: 12 Days of OpenAI: o1, ChatGPT Pro, Sora, New Interviews with Altman, Pichai, and Bezos, Amazon Nova & David Sacks Is US “AI Czar”
01:20:13
Loading summary
Transcript71 lines
- [00:00]
Paul Raitzer
I did see someone over the weekend tweet this idea that, like in a film, the average scene is like three to five seconds. So you know this idea that this could be very disruptive to the ad industry, to the movie industry, when you think about it in that context. Like it can do 20 seconds. But what if it's really, really good at five seconds? 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 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 126 of the Artificial Intelligence Show. I'm your host, Paul Raitzer, on with my co host, Mike Caput. Mike, I was on the road last week and I feel like it's been like three weeks since we did this. I was in Detroit to Philly to Charleston, Monday through Friday, didn't come home. I don't usually do like the road warrior thing where you're just gone for a whole week. I'm usually like one night at a time. Wonderful meetings, great events, but good to be home.
- [01:29]
Mike Caput
I think that's one of your craziest stretches that I can remember. Pretty recently it was.
- [01:34]
Paul Raitzer
And didn't we re recorded the podcast last Monday before I went. Right.
- [01:38]
Mike Caput
Yeah.
- [01:38]
Paul Raitzer
And then I drove to Detroit. Like, the whole thing is a blur. Like I. Yeah, it's like I was like, I don't know what city I'm in, I don't know what day it is. Like the whole week I was just kind of all over the place. But I'm back and I. I think I'm done traveling for a while, so that's exciting. Do you have any talks coming up in December? Are you done with the trip? No, I'm.
- [01:55]
Mike Caput
I'm done for the year, basically, as far as I know. Unless, you know, some. Someone surprises us in the next couple weeks. But yeah, I'm all set. I had a couple things last month and then that's it.
- [02:06]
Paul Raitzer
Okay, nice. And I think you and I might actually end up in the office. Maybe we, like, go out for a coffee or something. Yeah, we see each other like once a month in the office and then we live a mile from each other. Not even. And you know, the Office. We like never can get together the schedules. All right, so this episode is brought to us by our AI Forward CEO webinar. I've mentioned this on the last couple. We're already, I would say get registered soon if you're going to do this. It's December 17th, so we're like 1500 people register. I think we're going to cap this at probably 2000 just for Zoom license purposes. But so December 17th AI forward CEO unlocked the power and intelligence of a co CEO custom GPT. So I kind of told this story before, but we're going to do is December 17th at noon Eastern there will be. We'll make the recording available. So if you can't make it because of time zones or conflicts that day, go ahead and register anyway and then we'll send out the on demand recording and make that available for probably like a week after. So this whole thing came from me just playing around building a GPT for myself to be the co CEO of my companies and help me like function as a strategic advisor. I'm legal and HR and finance and marketing and sales and service and it and like I just wanted somebody to talk to and I figured, well, let me just build a GPT. And it ended up being incredible. So I was just keeping it to myself. And then I mentioned on the podcast a couple times and people kept reaching out and so I finally, I was like, all right, let's just share this thing. So I'm not going to share the one I personally use because it has a bunch of my personal and company data in it, but I'm going to show how we use that one to analyze data, execute tasks, solve problems, build plans and innovate. And I'm going to make a, a general version of it available to everyone. So I'm going to build that actually this week and then I'll put that through a Link on the SmartRx AI website and people will be able to access that and then I'm going to share the prompt. I actually used the full one minus the personal data to build my own. And so, you know, hopefully this is really helpful for people. And we're planning to do some more stuff around like building custom GPTs in the new year. So stay tuned on that. We're actually probably going to announce as of this morning, I think we agreed, right Mike, that we were going to announce something next week. Right. So it gives me like seven days to figure it all out. So what. And my main thinking here is what we see is like custom GPTs are like the fastest way to value for a lot of enterprises. Like build a GPT to help someone do the thing they do every day and they just get it like it becomes immediately understandable to them what the value of AI can be. And so that's what we're going to focus on. So this webinar is going to do what is the co CEO, how does it work, what are some example use cases, who should use it, how to build your own. And it's not just for CEOs it's like for anybody who wants to basically have a CEO mindset or kind of, you know, works with CEOs and wants to be able to communicate better. So SmartRx AI, that is our sister company to Market AI Institute, that's our AI research and consulting firm. Right at the top banner, there is a Register now button for this December 17th webinar. You can grab that and then while you're there, subscribe to the Exact AI newsletter, which is the newsletter I send out every Sunday morning that has editorial kind of and what's coming up for the week ahead. So those are the two things. Go to Smarter X AI and then click Register now and then you can grab the newsletter subscription while you are there. And other than that, it is ship miss, Mike. It is 12 days of ship miss and we are right in the middle of it. You had a chance to look at day three. I have not. So now let's roll right into the OpenAI 12 days of shipmas.
- [05:52]
Mike Caput
All right, Paul, so first up, OpenAI is basically launching this thing they call shipmas. They've also called it 12 days of OpenAI as a Christmas and holiday reference. Basically this is an ambitious 12 day holiday campaign featuring daily, at least on weekdays, product releases, demonstrations and new features from OpenAI. And before you think this is kind of some marketing gimmick, like we've gone through three days so far, including today, and these are pretty significant updates so far. So we're going to go through kind of what's happened on the past three days. Obviously there's going to be one update per day moving forward until we're done. We'll cover those on future episodes, but we're going to walk through these three days. I'm going to get your thoughts on all these, Paul, because there's a lot to talk about. So on day one, this campaign kicked off with two big announcements. So first up, OpenAI said they are doing the full release of their reasoning model 01. And this is coming out for the for everyone in ChatGPT plus and Pro. And if you have never heard of ChatGPT Pro, it's because that was the second announcement. ChatGPT Pro is a new premium subscription tier. So first up, the full O1 model. This represents a big improvement over the preview version. If you didn't know, the O1 we were using before was just a preview and that was available to ChatGPT plus and team users. And according to OpenAI researcher Max Schwarzer, the new version makes 34% fewer major mistakes and processes information 50% faster than the previous model. The model is multimodal, meaning it can process both images and text together, and it's been refined based on user feedback from all this work everyone did with the preview model. So alongside O1 we also got what I referenced just a moment ago, ChatGPT Pro. This is a new subscription tier priced at $200 per month and it offers unlimited access to the new Zero1 model, Zero1 mini GPT4O and advanced voice mode. So if you're hitting any type of usage limit in your ChatGPT plus or team account, ChatGPT Pro might be the license for you. Now on day two, OpenAI announced that it is expanding what it calls its Reinforcement Fine Tuning research program. This enables developers and machine learning engineers to create expert models that are fine tuned to excel at specific sets of complex domain specific tasks. We talked many different times about the importance and the opportunity to create models for specific domains. This is basically OpenAI giving you a better, easier way to do that. Now today, the day we are recording, Monday December 9th, which we delayed recording.
- [08:52]
Paul Raitzer
By the way, for five hours for this we did.
- [08:56]
Mike Caput
Since we are on eastern time for the Most part and OpenAI is on Pacific time, we delayed things so that we could hear today's announcement and bring it to you in this episode. So just a few hours before we recorded OpenAI for day three finally formally released Sora, their video generation model. So Sam Altman and a couple team members demoed Sora, which is getting launched and rolled out in the US and most countries internationally by the time you will listen to this podcast. And they showed off that sora can generate five to 20 second videos from a text prompt or even if you upload an image and it can create multiple variations at once, it can create them in different aspect ratios and resolutions from 4 ADP to 1080p. Now during the demo they showed off some other interesting features. These included things like an Explore feed. So once you log into Sora, you'll see examples of other videos that users have created. It'll also tell you, like, how those videos were created. There's a new tool called Storyboard to direct your videos, basically by describing each scene and placing it in the video's timeline. It's really cool. You can essentially direct second by second how the video works. There's also some features like Remix, which lets you change the video just by describing the changes you want to see, and something called Recut, which allows you to basically add or extend footage that is anywhere in the video. Now, Sora is actually available as its own standalone site@sora s o r a.com now, this site has been a little buggy all day for me, but the way it's supposed to work is if you have a ChatGPT plus or a Pro account, you go to Sora.com, you'll be prompted to log in and then you can use it at no extra charge. If you don't have one of those, I believe you'll have to sign up. But again, both these functions were essentially.
- [10:57]
Paul Raitzer
I can't get in now.
- [10:58]
Mike Caput
You can't get in right now?
- [10:59]
Paul Raitzer
No. Yeah, I did the login part and then it said signups are temporary unavailable. So it's like, yeah, it's getting crushed.
- [11:06]
Mike Caput
So it's. Yeah, it's getting crushed with heavy traffic. So hopefully you can access it when you hear this or sometime this week. ChatGPT plus accounts will right now get 50 generations of videos per month. Now, if you have the Pro account, you'll get what they call unlimited, quote, unquote, slow generations. It'll take more time to do and you'll get 500 faster generations per month. So, Paul, we're going to take these kind of one at a time. First up, full release of O1 certainly sounds like O1 is exhibiting some very powerful reasoning abilities. Can you give us kind of your initial thoughts on this model?
- [11:43]
Paul Raitzer
So I would. I would go Back to episode 113 on September 4th. We talked about Strawberry. So that's. That was the internal code name for the O1 model was Strawberry. And a lot of rumors were swirling at that point that this was sort of imminent that they were going to release this Strawberry model. It then debuted a couple weeks later. So episode 115. We went deep on the 01 mini 01 preview models, Mike, that you mentioned. And any regular listeners to the show or attendees of our Macon event will remember that that 01 came out two hours before the closing keynote at Macon on September 12th. And so Mike and I were scurrying to tell the story of this reasoning model as we were closing our conference. So I think just provide some perspective here because I have not, I've tested it a little bit, but I haven't like pushed it yet because it definitely seems like it's predominantly for harder problems like math, biology, engineering, science related. But in, in the context of why this is relevant for OpenAI and the pursuit of intelligence, if you'll remember, they have these five levels internally. So level one AI is chatbots, which is what we got with the first iteration of Chat GPT back in 2022. Level two is Reasoners, which we now have in the public domain, but they have had since fall of 23. The breakthrough that at least showed them that this was possible is believed to have happened in October 2023. Actually no one Brown who's working on this internally just said that in an interview recently. Level three for OpenAI is agents, which we've obviously talked a lot about agents on recent episodes. Level four is innovators and level five is organizations, like autonomous organizations. So we have moved into level two in OpenAI's world, we are quick, quickly moving into level three and we can talk a little bit about that and Sam Altman's interview on on Dealbook Summit, which we're going to go to in the next topic. And so what what reasoning does is it's like the human cognitive process of drawing conclusions, making inferences, forming judgments based on information, logic, experience. It involves the ability to think logically, analyze situations, evaluate evidence, solve problems. So when we think about this being tied into businesses and when it's kind of baked into large language models like they're doing with L1 now, you can do multi step problem solving, you get more accurate predictions. You talked about Mike, like these things make fewer major mistakes, improved risk assessments, reduced hallucinations and errors, deeper contextual understanding, completion of higher level cognitive tasks like these are all things that get unlocked with a reasoning model. And so in simple terms it makes ChatGPT and the other house that much smarter, that much more generally capable, that much more human like. And now you know, one of the things Mike, I'd flag for you and you and I kind of touched on this beforehand is this comes with downsides. So one of my favorite podcasters is Nathan labenz and he's got the cognitive revolution. So he does amazing interviews and he had sort of an emergency pod with Alexander Menke of Apollo Research and we will put the link to this in. I'm not going to go deep into this episode, but I wanted to read to you an excerpt from the opening of this episode because again, I think it's very important perspective for people to have. So in the opening, Nathan says the O1 model, which is faster, scores higher on reasoning benchmarks and comes with the full complement of multimodal and tool use capabilities. Like many in the AI space. I've spent the last 24 hours testing the model, trying to absorb everything that's been written about it, including OpenAI's 42 page system card which I have not dug into yet. Mike and Apollo's 70 page report entitled Frontier Models Are Capable of in Context Scheming, which will be our main subject today. So again, this is Nathan doing the opening to this episode that we're going to link to. He then goes on to say scheming is when an AI deceives humans to pursue its own hidden or implicit goals. He says, I think we can all agree we don't want that from our AI systems. Some of the examples are properly shocking, for example, trying to overwrite their next version's weights or goals with their own in order to propagate themselves into the future, and then deliberately falsifying data to engineer outcomes that run contrary to user requests. So, he continues, some have tried to downplay these findings, arguing that scenarios that Apollo created are contrived and not representative of real usage. Said while Tess Apollo ran are designed to elicit scheming behavior, they are not conceptually far fetched. The core observation is that when AI's goals conflict with human goals, weird shit starts to happen. This is a legitimately huge problem. He continues, obviously scary to people outside of the field. We in the field shouldn't allow ourselves to get comfortable with it. I'm going to read a little bit more because I think he gives the size of the problem and why this open source movement that we're seeing an acceleration of could create some serious problems. So he says when OpenAI goes live via their API, which it will, the O1 model goes live by the API. Over 1 million active developers building on OpenAI's platform will be able to place goals in system messages, exactly as Apollo did in their research. It seems to me a virtual certainty that O1 will find itself in situations where there is fundamental tension between the standards set for it by OpenAI in the model spec and the goals the developers give it in the system message. Apollo found scheming behavior in roughly 1 to 10% of cases across most of the conditions they tried, even if it's 1000 times rarer in the wild. And even if OpenAI deploys the 92% accurate deception monitoring system they describe in their system card, with hundreds of millions of daily uses across millions of apps, we should expect O1 to be actively scheming with hundreds or thousands of users daily. And then he just goes on to talk about, like, if we're really at this point where, where we have now put a model into the world that we know, schemes that we know tries to overrule the inputs from its human creators and users in ways that are meant to deceive the human. And we know it will happen with high probability. And we think we're only one to three years from much, much more intelligent systems. Shouldn't we be doing more for safety and security? It's kind of like the whole point of this. So I haven't had a chance to listen to this whole episode yet. I will be listening to it this week and I would say anyone who's sort of like, intrigued by this or terrified of this, it's probably a good one to read. And so it does go back, Mike. Like, you know, you'll recall in fall 2023 when Sam Altman was ousted as CEO and we had the whole episode on this and with the talk of, you know, the business world, the I world, certainly for about four weeks, the whole thing was, what did Ilya Sutskova see? Because Ilya was the one that led to the ousting. You know, he had a board seat and he pushed for them to oust Sam. And the timelines, while we still don't have 100 confirmation, it's exactly what happened. No one, Brown said, we realized in October 2023 this was going to scale, that this reasoning approach was going to work. That is when Ilya went to the board and, and alerted them there might be problems. And then Ilya eventually leaves and creates Safe Superintelligence, his own AI company. So it sure does seem that the whole theory all along, that Ilya saw the reasoning model coming to life and became fearful that they were putting something out into the world that the world wasn't ready for yet. It. It definitely aligns that that's a very distinct possibility, at least played a role in what happened back then. So crazy stuff, cool stuff like, it's gonna be fascinating, but, you know, it also comes with its downsides.
- [20:20]
Mike Caput
It sounds like we're getting into a real territory of almost just unanticipated or unintended consequences as these models learn more than just kind of. I think it may be mentioned in here Just more than raw intelligence, it's having these effects that not anyone or everyone can anticipate. And this is all coming from. This doesn't. Isn't any kind of conspiracy theory. This is coming from OpenAI's own efforts that they're releasing to keep this model safe.
- [20:46]
Paul Raitzer
Yeah, yeah, no doubt. And so like, you know, to, I guess, continue on from how you sort of set it up to this Pro license, then it's like, part of me wants to pay the 200amonth just to see what this is like, to see what this thing is fully capable of. I'm not sure, like, what we would use it for. So I had reached out to Mike actually over the weekend. I was like, hey man, we should run a hackathon ourselves and like, push on this model and see what's possible. And not only for the positive uses, but these unintended consequences, like, what are the dangers of this thing? And so Mike and I are actually running internal hackathon on Tuesday. So the day this episode drops. And we'll, you know, summarize for people, because I'm trying to figure out, like, should we be paying the 200amonth? Is there. The value in it for us is, you know, beyond just our research and understanding. But I, you know, I went in to ChatGPT itself and I was like, hey, what. What should we be using the V1 reasoning model for from a business and marketing perspective? And it started presenting some interesting ideas like campaign strategy, audience Persona, performance analysis and insights, content, calendar creation, competitive analysis. I was like, this is kind of interesting. So Mike and I are going to kind of talk about these ideas tomorrow and maybe start building some stuff and report back, like right now, though I would say Most users, your 20 or $30 a month plan is all you need. The $200 a month plan, before you told me that Sora is included in the $200 a month plan, like that. That might change my perspective on the $200 a month plan, actually.
- [22:15]
Mike Caput
All right, so let's talk about day two's announcement. And can you maybe break down for us, like, why is reinforcement fine tuning so such a big deal?
- [22:23]
Paul Raitzer
First of all, this is an announcement for developers. So the average user like you and I, we're not going to be building on this. This is something that's giving developers the ability to take the core model and then do reinforcement learning very quickly by giving it examples it learns from, you know, setting goals and rewards that enable it to kind of learn a domain based on a specific data set. So if you wanted to use this reinforcement. Fine tuning. This is something you're likely going to be teaming up with a developer, the internal IT team, something like that. You're going to need a unique data set that you can use to train this thing in a specific area. But this does absolutely hint to a near term future where every enterprise can custom train their models, maybe even by department and you just have custom versions of it. So imagine like GPTs on steroids. Like now you can actually take the core model and fine tune it and not have to be a developer to do it. To where you and I might could build these fine tuned models the same way we can build a custom GPT. And that's, that's fascinating, like the possibilities there.
- [23:26]
Mike Caput
Okay, so let's talk about Day three and Sora. So you know we've got now. Well, once the website works again, it does sound like looking at the release, the announcement they released along with this, it's included as part of a plus account and it's also included as part of a Pro account. But the usage rates differ dramatically. So what, how are you thinking about this without us having been able yet to test it?
- [23:51]
Paul Raitzer
Yeah, so first I look at the usage limit. So 50 generations a month doesn't seem like a ton. It probably just depends on how good it is, honestly. So like Runway is a company that does video generation, text to video. We've talked about it many times on the show runwayml.com unless they change their URL. And I have a paid account, I think I pay 30 bucks a month for Runway. I haven't been in there in months because every time I go and try and use it, it, the outputs aren't usable. Like they're, they don't maintain consistency. And so like anytime they update their model, I'll go in and play with it. And every time I go, I'm like, God, I have like 900 credits in here. Like, I don't, I don't even know what to use it for. So that's an instance where it didn't seem like a lot. But then once I get in there and use it, I realize there's nothing I can do with this. Then the credits just stack up. I anticipate that Sora is going to be a leap forward in capability and that you could envision using this regularly, especially for, you know, in my world, embedding into videos, you know, things you might create as demonstrations. So if they work well and it creates these quality outputs, then that'll be intriguing. The second thing is the Speed is going to be a huge issue. Like, so Runway can take minutes to do four seconds of video, and it's just not even worth it to me. It's like the effort that sounds so silly. The effort to create a high resolution video taking minutes, but it's minutes. When you get the output, it's like, well, that's not what I needed. And now you just, like, you just keep throwing time at something that's not going to create the output you want. So I would expect that these things are going to take quite a bit of time to generate. I don't think this is gonna be really fast inference time or like, you put it in and three seconds later, five seconds later, you've got your video. Yeah, I would expect this is a slow thing and that's without all the traffic they're gonna have on this site for the next, you know, month or two. So that's interesting. I. I think rate limiting the speed, unless you're paying the 200amonth, like, I could see that being a big thing. It's like, hey, I get faster generation. You get, basically, you get the fast pass, like hit zero point or out of the park. Like, if you're paying the 200amonth, you get the fast pass on your generations. And so I, I guess part of how fast these things generate may be dependent upon the uptake in the $200 a month license realm. So if a bunch of people are like, hey, I'll pay it, then all of a sudden you're, you know, there's a hundred people in the FastPass lane ahead of you, or the TSA lane, I guess, at the. And then the clear lane, like, you just keep adding another way to get that. So I don't know, man. Like, the demos were always super impressive on this, but with video generation, as we've talked about on the show before, it's really hard to maintain character consistency, frame consistency. But I did see someone over the weekend tweet this idea that, like in a, in a film, the average scene is like three to five seconds. So, you know, this idea that this could be very disruptive to the ad industry, to the movie industry, to, you know, from a brand perspective for content creation, videos, things like that. When you think about it in that context, like, it can do 20 seconds, but what if it's really, really good at five seconds?
- [27:09]
Mike Caput
Right?
- [27:09]
Paul Raitzer
And that's enough because then you can just stitch together frame by frame by frame, and you can all of a sudden start building some really incredible things. So I expect adoption to this to be massive if it works really well.
- [27:23]
Mike Caput
All right, so our second big topic this week. On December 4th, the annual DealBook Summit took place. And during this, we got some really interesting in depth interviews with a few of the top people driving the future of AI. So DealBook is the name of a financial news Service founded in 2001 by New York Times columnist Andrew Ross Sorkin. Since then, it's been kind of a core piece of the New York Times reporting in business. And since 2012, the Times has also kind of paired with this the Deal Book Summit. So in this event, they interviewed top newsmakers in business. In the past, they've interviewed people like Elon Musk, Nvidia's Jensen Huang, Vice President Kamala Harris, and Prime Minister of Israel Benjamin Netanyahu. So they get some pretty significant figures at this event. And at this year's event, Sorkin, who kind of emcees the whole thing, interviewed some of the top AI leaders in the world. So there were some other guests that were not related to AI, but the ones we're interested in were the three he talked to who gave us kind of an inside look at where their companies and the industry at large are going as we close out 2024. So in particular, he interviewed OpenAI CEO Sam Altman, Google CEO Sundar Pichai, and Amazon founder Jeff Bezos. Now, Paul, I know you are following these conversations closely. I just want to hit on a few very quick points that jumped out to me and then kind of turn it over to you to kind of reveal to us, like what you took away from these talks. So a couple things with Altman that he said that I thought were kind of notable was he was like, my guess is we will hit AGI sooner than most people think in the world think, and it will matter much less. Interestingly, things will kind of pass through that milestone and kind of go on with our lives in a more abundant future. However, he also did say, I expect the economic disruption to take longer than people think, but be more intense than people think. So he's kind of saying that we might see a lot of changes in the economy. I also thought it was noteworthy, given all the drama around this, that they asked him about his beef with Elon Musk. Will Elon kind of come at him using his newfound influence with the Trump administration? And he said he believes pretty strongly that Elon will do the right thing. It would be profoundly un American to use political power to the degree Elon has it to hurt your competitors and advantage your own businesses. I Don't think people would tolerate that. I don't think Elon would do it. On Sundar Pichai's side, I thought it was kind of interesting. They kind of called him out with some quotes around Microsoft CEO Satya Nadella saying, hey, Google should have been winning this whole generative AI thing. They got caught flat footed. He kind of just said, I'd love to do a side by side comparison of our models with Microsoft any day. And they also said the area we applied AI the most aggressively, if anything, in the company was search. This is essentially what motivated them to be applying transformers way back when. So he basically is saying, look, I'm not worried about our core business. Those search will change profoundly and I think we're going to actually just be able to make it all better and able to handle more complex questions than ever before using AI. Now finally with Bezos, he covered a lot more than just AI in his interview. But what's interesting is he said he's basically, you know, kind of moonlighting back at Amazon, helping specifically with AI. 95% of what he's helping with is AI. And he's talking about the fact they're working on literally a thousand applications internally for AI. We'll talk more in the next topic about their own large language models. They've released the Nova family of models. And basically he said, look, in some ways our models are already smarter than humans because they're multidisciplinary and humans often are not very good at all the things they do in a day. So Paul, I'll turn that over to you, but just some kind of interesting highlights that jumped out at me in this.
- [31:24]
Paul Raitzer
Yeah, they're all worth listening to. I think Bezos was the longest, maybe at 10 minutes or his might have gone 50 minutes, I don't know. But Sam's was like 30 minutes. Sundar was around 40 something. So the, they're all very digestible, especially one and a half times speed on, on YouTube. So I would suggest listening to all of them. I think there's, there's a lot of perspective and context and honestly, like even reading the quotes, which I had read some of the quotes before I listened to them, they take on very different meaning when you hear how they're said. Like even the, the Sundar one, there was an edge in his voice when, because that's how the interview started was he said straight up, like, hey, Satya has been kind of like, you know, taking it to you guys about this. And his comment was like, yeah, where are their models? Like I don't need him talking to me because they don't even have their own. They're using Open AI's models. Was. It was. It was a. It was a tense response. Like, he wasn't super ecstatic about it. And there was quite a bit of emotion because I will say, like, Andrew does an amazing job of just coming straight at and asking hard questions, and then he pushes on the hard questions. He doesn't like, stop. And so he asked point blank, like, at one point with Sam, like, about their thoughts on copyright law and fair use. And Sam's like, well, you guys are kind of suing us. And.
- [32:43]
Mike Caput
And yeah, yeah.
- [32:44]
Paul Raitzer
Sam's like, well, I think it's fair use. And I was like, it's not. And. And he's like, I guess, you know, we'll see each other in court kind of thing. I was like, whoa. Like, that was a weird place without an interview to go. So I would say on Sam's. It's worth listening, especially if you don't like, deeply understand or know the OpenAI origin story, his history with Elon Musk, things. Like, we've talked about the podcast a lot. It was a nice kind of synopsis, and he wove a lot of that into their conversation. You kind of hit on this idea of AGI is like, they're very much now talking about it as almost like this continuous thing and like a mile marker, not like the milestone goal anymore.
- [33:25]
Mike Caput
Yeah.
- [33:25]
Paul Raitzer
And. And I think they might have used this, if not Sundar did this analogy over to like Waymo driving cars, where all sudden the things just drive without people in them. I mean, they're Teller operated sometimes, but whatever. So this idea, though, that we are going to achieve narrow AGI in a way in different fields, and like, life's going to just kind of move on. And so when we talked about this on earlier episodes where we shared Google's levels of AGI, where they look at performance in generality, and so they started talking about this idea that like at like level three, I think it was in their world where it's better than like 75% of the humans that would do a thing. Well, if you think about it, and you start looking at writing and SEO and consulting and eventually accounting and lawyers and doctors, like, we're. We're likely very close to AI that is superhuman in different domains. Meaning it's like virtuoso in. You know, Google's world of virtuoso is like better than 99% of humans at a thing. It. It might be A while before a model like ChatGPT or Gemini is just better than all humans at everything. But we're going to start picking off domains where the AI is better than the human at the thing, better than the best humans at our thing, our discipline. That does not mean the AI takes all the jobs. It. It just means there's like a thing that's, that's probably better at part of your job than you are.
- [35:03]
Mike Caput
Right.
- [35:03]
Paul Raitzer
And, and so that's where, like, they talk. Sam got into, like this idea of super intelligence where, you know, really, it outperforms humans 100 of the time at all cognitive tasks. And he basically is like, yeah, life's gonna kind of just keep going on. He's sort of in this mentality that we'll figure it out and we'll build other models and know whatever he did get into the scaling wall, which we've talked a lot about recently, is there, you know, wall. He, he was very straight, like, no, there's. There's just no wall. And the main reason. And Sundar kind of backed this up and even Bezos did, to a degree. They think of building these models as three main components. The computing power, which is the Nvidia chips. Like, how many chips can we stack into a data center and wire together and get them to do this thing? The data that goes into them, including now synthetic data, and then the algorithms, meaning the ways we find to do this smarter like that these things become more efficient. So we don't, you know, if we have a hundred thousand Nvidia chips, maybe next year we can achieve the same output with 50,000 because we built better algorithms of, like, how to do the learning and things like that. He did confirm some usage data. 300 million active users. That was weekly, wasn't it? Was it weekly or monthly?
- [36:13]
Mike Caput
Weekly active users.
- [36:15]
Paul Raitzer
That was before Sora. So 300 million weekly active users, 1 billion user messages sent on ChatGPT every day, and 1.3 million developers, which is just crazy numbers. Yeah, they got into this whole, like, creators who content has been used to train the models again. You know, he pushed them hard on that. And Sam doesn't really have, I don't know, they just, they just keep standing on this. It was, it was fair use and it's not. But whatever, the courts will decide that. And then he asked all three of them about, like, meaning for humans, like, as AI kept, like, evolving, what does this really mean? And you know, Sam, who has a kid on the way, I don't know when they're going to have Their child, but they, they have a child coming. And so he said, like, you know, you have a child coming, like, how do you think about this? Like, the future for that child. And, you know, he basically said, the economy will grow, jobs will change, but evolution slow and humans adapt was basically Sam's message for everything. Sundar, I thought, like, he asked him about the wall and he said, there's, you know, more breakthroughs needed, but they're, they're coming. The algorithms, especially in planning and reasoning are going to happen. He pushed him on the losing their lead. And the search thing, like you mentioned, impact of hiring coding agents, I thought was an interesting one. He asked him, like, well, are you making changes at Google? You're building these agents, they're more efficient. Have you changed your own hiring plans and budgeting as a result of it? And Sundar kind of sidestepped it. He said, definitely taking into account how to be more productive and efficient. It was like, okay, I don't know what that means. But they talked a lot about regulations. And then he asked him about the economics for creators whose content feed these models. And he said, they're going to be thoughtful, was his quote. And then Bezos you highlighted. I thought it was just interesting that he's back, like spending that much time, kind of like, you know, Larry Page and Sergey Brin, the founders of Google, being back at Google working on AI. So just the fact that this is so significant that these people have moved on from these companies they founded 20 plus years ago, are now back almost full time working on AI. And so it was, it was really cool to hear. And Jeff's like, you can zip. Oh, his was over an hour because it was like at the 56 minute mark or something when they started to 51 minute where they started talking about AI. So there was like all this other stuff within there. So yeah, I thought that he had this one part where he said, you know, these kind of horizontal layers like electricity and compute and now artificial intelligence, they go everywhere. There isn't, I guarantee you, there isn't a single application that you can think of where this won't make it better. And I thought that was interesting because you could start to think about that in your own business. Like, every piece of software you use is going to have AI in it. Every department in your company is going to have AI in it. Every business in your industry is going to have AI in it. And there's just going to be smarter versions of everything. That was where the name Smarter X came from. When I named like Our AI research firm a couple years ago and developed that initially as a consulting practice. It was like Smarter X. Like, whatever it is, just fill in the blank. Like marketing, sales, service industries, everything is just going to get smarter with the underlying intelligence layer. Um, and then there was the last thing I'll say. When he asked Bezos about, like, what will it mean to be human? He had an interesting take. He said, you can always find somebody better than you at something now, and yet that doesn't take the meaning away. So he was saying, like, he's not the best in the world at anything. There is someone smarter than him at every single thing he does or cares about or is passionate about. Yeah. And yet as a human, there's no meaning lost in that. So his point was, if all of a sudden there's these AI models that are just smarter than you at everything, does it really change your life? Like, I mean, yeah, it's on demand now. Like, you can go get it chatgpt instead of having to go find these people. But his whole point was like, we don't derive our meaning from being the smartest in the world at a thing. And like, your meaning comes from relationships and people. And. And I was like, that's a really fascinating perspective and I need to think more about it because I was just watching this, like, know hour before we jumped on this podcast. But I thought that was an interesting take. That, like, there's always people better than you or smarter than you at all these things. And, like, that doesn't change your perspective on the meaning of life.
- [40:33]
Mike Caput
Right.
- [40:33]
Paul Raitzer
I don't know. There might be something to unpack there. I. I have to think about it more now.
- [40:38]
Mike Caput
It was interesting that they, yeah, all three of them did end up getting almost philosophical at some point.
- [40:43]
Paul Raitzer
Yeah, they hit on copyright, fair use for everybody. They hit on competition. They hit on, you know, the future of AI models. And they hit on, like, the meaning of life. Like, that's what I'm saying. Like, to get those three guys on the stage in a single day talking about these things was fascinating.
- [40:57]
Mike Caput
All right, our third big topic this week, Amazon has unveiled Nova, which is a new family of AI models that basically expand their generative AI capabilities. So this was announced at the Re Invent conference that happened just recently. And the Nova suite includes four text generating models, as well as an image generator called Canvas and a video generator called called Real. The four main Nova models, Micro Light Pro and Premiere, offer varying levels of capability and performance. The smallest micro focuses on fast text processing, while the larger Models can handle multiple types of inputs, including text, images and videos. Now, these models appear to feature impressive context. Windows Micro can handle up to 100,000 words and the larger models process up to 225,000 words or 30 minutes of footage. Now, Amazon plans to expand this to over 2 million tokens for some of these models in early 2025. On the media generation front, Nova Canvas creates and edits images with control over color schemes and layouts. Nova Real can generate videos up to six seconds long. They promise these will be two minutes soon enough. And they also have plans for two more models, a speech to speech model in Q1 2025, and what they call quote any to any model in mid-2025. Basically, it'll handle multiple types of input and output. Now, CEO Andy Jassy claims these models are among the fastest and most cost effective in their class. So, Paul, this is kind of interesting pairing with Bezos's interview saying, hey, I'm back in Amazon and by the way, now we've got our own models. We talked last week about how Amazon is both investing billions more in anthropic and trying to reduce its reliance on the company by building its own AI in house. Now this certainly seems like a big win for them, building their own AI in house, doesn't it?
- [43:01]
Paul Raitzer
Yeah, it is kind of like confusing as this whole space is confusing. Right now everybody's building their own models and doing deals with other people who are building models. And yeah, the whole thing is just very complicated. I think in Amazon's case, maybe more than any other company. I mean, Google I could think of is in a similar boat. I would imagine they're building a lot of these models because they see the opportunity to transform Amazon internally. Like it's the same. When they built aws, it was just like they had an internal reason to enable this. And so I think they probably look across their business and say, well, do we want to use anthropic models to optimize our own operations, marketing, sales, service, ops, you know, finance, hr? Or do we want to like devise our own for all of the different things we do from warehousing logistics to running the daily business? So I don't know. I mean, that's my guess is they're largely focusing on internal applications. And then they're, you know, as a byproduct, they're able to also build and open these models up and maybe drive uses of AWS and, I don't know, put it into all their different devices they're building. And it's just one of those, like it was funny because Andrew did ask Bezos about this and he said something about. And Jeff was like, well, you've, you've been busy, Andrew. We did announce our own models yesterday. And Andrew goes, oh, the Nova thing. It was almost like he just blew it off as like not a big deal. You could tell Bezos was a little bit taken aback, like, well, well, we don't, we don't get any credit for launching our own family of models. He's like, they're frontier models, Andrew. Like, these are, these are important models. Yeah. I don't think Andrew bought it. So it will be fascinating to just see again. Like, you know, we've now got, I don't know off the top of my head here. So you have AWS has their own Frontier models. Certainly anthropic. Google, OpenAI XAI, Microsoft is building their own. So. Well, we have what, six main players now. Was that five or six? Yeah. And then you have, oh, Meta. You can't ignore Meta. You got Mistral is sort of playing around cohere. They're doing sort of like the lower level models. So I mean it's, it's getting to be a. Nvidia has their own models now. So I mean you have like seven or so like major 20 of the biggest companies in the world kind of major top 20 building frontier models who have tens of billions of dollars of R and D money every year to build these massive frontier models. And you can officially put Amazon in that category now.
- [45:36]
Mike Caput
You know, I thought, just as a final note, here is pretty interesting, you know, TechCrunch reported there was a comment from the CEO Andy Jassy saying, quote, we've optimized these models to work with proprietary systems and APIs so they can do multiple orchestrated automatic steps, agent behavior much more easily with these models. So sounds like as we've seen with everyone else do not only are they building their own models, but they've got their eye on some type of agent behavior.
- [46:04]
Paul Raitzer
Yeah, I mean the tech companies are going to push the frontiers here. They all know full well what these things are capable of doing and they're going to try and bring those different levels to life. You know, from the reasoners to the agents to the innovators to the organizations. And I think that's going to be where we're going to see the early signs of where corporate America and corporations around the world are going to go is look at what becomes possible. And I keep trying to pay attention to are they still hiring Are these companies still hiring marketing people and salespeople and HR people? Because I feel like the way we're going to know when we're starting to look at the impact of the economy is when the hiring practices of the frontier model companies change. Yeah, because the tech they've built has enabled them to change their way. They hire and promote and retain workers. And if we start seeing a consistent decline from the frontier model companies, that's an indicator that we are now heading toward the job disruption that I expect will start to come next year.
- [47:05]
Mike Caput
All right, let's dive into our Rapid Fire topics this week. So first up, according to some reporting from The Financial Times, OpenAI is potentially going to remove one of its founding principles in its agreements with Microsoft. There is a provision that would shut Microsoft out from accessing their most advanced AI technology once the company reaches artificial general intelligence, or AGI. So under the current terms, if OpenAI creates what it defines as AGI, a system that can outperform humans at economically valuable work, Microsoft's access to this technology would be void. OpenAI's board would be the ones to determine when that milestone is reached. Now, if these reports are correct, OpenAI is potentially considering getting rid of this provision. It seems like they might be considering this move to try to unlock more future investment, especially as they're trying to restructure to become a for profit company. Because Microsoft has invested more than $13 billion in OpenAI, presumably the removal of this clause would enable them to continue investing in and accessing all of OpenAI's valuable technology. So this kind of comes at a bit of a sensitive time for OpenAI. The FTC, the Federal Trade Commission, has actually launched a antitrust investigation into Microsoft with specific focus on the company's deal with OpenAI. So they're kind of looking at whether or not Microsoft's dominance in cloud computing has given it an unfair advantage in AI software sales. So, Paul, this would mark a pretty big change maybe in OpenAI's relationship with Microsoft. Like, what's going on here? And isn't this kind of implying that OpenAI would just be open to commercializing AGI, unlike what it said in the past?
- [48:58]
Paul Raitzer
I think they are. I mean, so if you, if you listen to Sam's interview, they get into this and it's very apparent there's friction with Microsoft. You know, he, he doesn't hide things well, I mean, Sam, Sam I'm sure is an amazing negotiator and, and everything internally, but just watching interviews, he either is insanely manipulative in terms of how he controls his emotions, to present a specific emotion, I. I guess to trigger specific emotions, or he's. He's just very honest and. And open about his thoughts and feelings. And I actually think it's more that. Because you can tell when they started pushing on the Microsoft thing, it's like, yeah, man, it's not all great. Like, it's. And it wasn't even just the usual PR lines of, you know, whatever. You know, we have our differences, but whatever, it's great. He was basically saying, like, no, man, this is hard. Like, it's hard to manage a partnership like this. And we're not the same company we were when we created this. We never intended to be a product company. When we started as a nonprofit, Yuan pulled his money, and we had to go get money from somewhere, and we'd be becoming a product. Like, Sam has gone through stuff as a business leader that there's very few peers to in, like, human history. Like, his last two years has been insane. Like, to grow a company like this with these complexities. And so he seems just, like, really open about it. And. And the Microsoft relationship in particular just seems like a really challenging thing to navigate right now, because even when they entered that partnership, there was a very different vision for where OpenAI was going. And, like, you know, what they thought it was going to be. And so I could see them finding a way to do this. Obviously, I have zero internal knowledge of anything going on here with legal or business sides of this, but from an outside perspective, you could see how the AGI thing is a massive sticking point for both parties. And I think it's part of the reason why Sam has started changing the way he talks about AGI publicly.
- [50:54]
Mike Caput
Yeah.
- [50:55]
Paul Raitzer
And I. You could tell he was softening their stance on that being a definitive moment, and they weren't going to stop building their systems and all of a sudden team up with anthropic because they got the AGI, which is what their charter said they had to do. And so this seems like AGI is becoming this friction point for everybody, and they need to just remove it from that part of the business relationship. So I wouldn't be surprised at all if. If that's what they ended up doing here.
- [51:23]
Mike Caput
All right, so next up, US President elect Donald Trump has appointed venture capitalist David Sachs as the nation's first, what they're calling an AI and crypto czar. So Sachs is a member of the influential PayPal mafia, the kind of team that led PayPal to success. He's more recently very well known as a co host of the very popular all in podcast. And he basically brings an interesting background to this role. He's the former COO of PayPal, he founded Yammer, which Microsoft acquired for $1.2 billion. And he's deeply embedded in tech because now he runs a venture capital firm called Kraft Ventures which has invested in numerous AI enterprises and companies like Elon Musk's SpaceX. Basically this position sounds like it's going to come with some interesting responsibilities. Sachs will guide administration policy on both AI and cryptocurrency. He will also head up the Presidential Council of Advisors for Science and Technology. This role is structured as a special government employee position which allows Saks to serve up to 130 days annually without requiring them to divest assets or make public disclosure of those assets. And it seems like the tech industry has largely welcomed the appointment. Leaders from major AI and crypto companies, including oh Sam Altman, have publicly supported the choice. Sachs will likely work closely with his fellow PayPal alum Elon Musk, who has been tapped to co lead what's been called the Department of Government Efficiency. So Paul, we're both pretty familiar with David Sachs. We've listened in the past heavily to the all in podcast. Like I don't know. I personally feel like I've heard Sachs talk way more about crypto than he has about AI. So for me personally the line from a TechCrunch report on this kind of resonated with me. They said stacks. His views on AI and AI policymaking are less obvious though his than crypto though his policies generally are decidedly right leaning and deregulatory. What did you think of this pick when you heard about it?
- [53:33]
Paul Raitzer
So real quick to continue on the Elon Musk Sam Altman beef. When Sam tweeted to congratulate Sachs and tagged him, Elon replied lll kind of like laughing at Sam that he's like kissing up to Sax now. So yeah certainly right leaning. I mean that, that, that's, that that's not debatable his policies. It does seem there's a good time article I'll link to in the show notes not in favor of regulation. Generally speaking he is going to be very pro open source, very pro acceleration of these AI models with as little regulation as possible. Heavily you know involved in Silicon Valley and the VC world and yeah I just listen to the all in podcast and you'll, you will hear what he has to say. That's the good and the bad I guess of like where we're at. With media today is like, oftentimes the podcast is the media, and, and these people have the platform and you can go learn all about it. So, yeah, I think. I think it fits. You know, we talked about, like, Andreessen Horowitz. I don't. I don't know their relationship. I don't. He wasn't a PayPal mafia guy. But, you know, I think it's going to be a lot of that, like the accelerationist manifesto, like, move fast, drive innovation at all costs, and regulate as little as possible, infuse it into the military. Like, it's. It's going to be accelerated. All costs, I think, is probably the best way to look at what he would bring to the table, which is probably going to fit Elon's mindset and. And, you know, part of the reason why he probably is going to be there. Yeah, like, if I would have stopped last week, I threw out, like, Karpathy is like, just kind of like a. If I would have stopped last week and, like, made a list, I could see that he. He would be on that very short list of people that, you know, probably. But I agree with you. Like, he's very much more crypto. Like, you hear about that. And that is not my world. You're more the crypto guy. Maybe we should come back around to that at some point. But that's not my world. I had a buddy years ago, joked with me, he's like, man, leave crypto alone. You got in early on AI, Leave crypto to the rest of us. And I was like, yeah, I don't want crypto. I don't invest in crypto. I don't really know a heck of a lot about it, so I will plead ignorant on the crypto conversation. I wish I'd invested in some bitcoin not too long ago, but that's about it.
- [55:57]
Mike Caput
Yeah. All right, so next up, we have two leading AI labs have unveiled some really interesting technologies that can generate explorable 3D worlds from single images. So, first up, World Labs, which we've talked about in the past, has introduced tech that can transform a 2D image into a navigatable 3D environment, which allows users to essentially step inside the image and explore it from any angle. So this system maintains consistent physics, lighting, and spatial relationships. So you have features like depth of field effects and interactive elements. Basically, this allows creators to quickly prototype virtual environments and bring static images, including classic paintings, to life in unprecedented ways. In the same time, Google DeepMind has announced Genie 2, which is a more comprehensive, what they call foundation world model that can generate playable 3D environments from prompt images. The system can create interactive worlds lasting up to a minute, complete with physics, character animation and even autonomous non player characters. Genie 2 is also able to respond to keyboard and mouse inputs, effectively turning these environments into playable spaces. So Paul, why are these so called like worlds or world models so significant? This just on the surface is so cool. It sounds like something feels like something out of Harry Potter.
- [57:27]
Paul Raitzer
Yeah, I mean I always immediately think of like gaming and like being able to build games on the fly and imagine worlds and I guess you could eventually do it into like creating your own fictional worlds of like storytelling and the story unfolds in front of you and you know, visually but you know, for the main thing it's, it's trying to understand the physics of the world around us. Like so creating a world model means giving the AI the ability to see and understand the world the way you and I do. That's one of the things I think is still going to come out from OpenAI maybe during ship miss and because they previewed it on 60 minutes last night, this vision model like Project Astra from Google we've talked about before where your device, like your phone or your glasses could see and understand the world around you. So to, to do that like this ability to create and understand and kind of model those worlds matters. So we won't go deep on this now but like episode 115 you had mentioned we explained the World Labs and so I'll just re real quick read through what we talked about there. So they talk about spatial intelligence as the big thing and they said we believe that artificial intelligence will help humans build better worlds. Progress has been rapid, but we've only seen the first chapter of the generative AI revolution. Language has thus far catalyzed this electrifying early moment with text prompted image and video models rising up alongside LLMs. These models have already empowered people to work and create new ways, but they only scratch the surface of what is possible. To advance beyond the capabilities of today's models, we need spatially intelligent AI that can model the world and reason about objects, places and interactions in 3D space and time. Then they went on to say we aim to lift AI models from 2D plane of pixels to full 3D worlds, both virtual and real, endowing them with spatial intelligence as rich as our own. So that's, that's the play. That's why they think this matters. And Fei Fei Li is leading the charge at World Labs who's a world renowned AI researcher. So definitely a company worth paying attention to. Both of these things are in research previews. This is not. You can't go play with either of these. Think of it as almost like when we first saw Sora, and now here we are like 10 months later actually getting to interact with an early form of it. It might be another year or more before we get to like really use these kind of models.
- [59:50]
Mike Caput
All right, so next up, Coca Cola is generating some controversy. Their latest holiday ad campaign has sparked some debate after they revealed that they used exclusively AI to create these iconic Christmas commercials that everyone's kind of expecting from Coke each and every season. So for the first time, their iconic holiday ads were entirely generated by AI. So this campaign has a few videos that are very short, but they reimagine one of their classic campaigns that they ran back in 1995 called Holidays are Coming. This features kind of signature red delivery trucks decorated with Christmas lights driving through snowy towns. Now, three AI studios called Secret Level, Silverside AI and Wildcard collaborated on this project to create essentially AI generated versions, updated versions of these commercials. And they use things like AI models like Leonardo Luma Runway, as well as a newer, a newer model called CLING that was incorporated late in the production to improve the human movements. The results generated a bit of controversy. Like, some critics pointed out that there were all these weird kind of AI generated effects that looked uncanny. They didn't look like real humans. There were technical limitations, like weird proportions, unnatural movements, some subtle errors in background details. And basically, they say the ads rely on extremely quick cuts to mask AI's current limitations, generating consistent, realistic footage. Even in some interviews, you know, one of the studios said that a brief shot of a squirrel alone required hundreds of AI generation attempts to get it right. Coca Cola's head of generative AI defended this approach, saying there are a bunch of benefits in production speed and creative possibilities. They even claim this tech allows them to produce content like this five times faster than traditional methods. Now, unfortunately, some creative professionals see that as an attempt to cut costs at the expense of human artists. So, Paul, this ad seems to have hit a nerve. There's definitely a few angles to unpack here. Like, do you agree with Coca Cola's move here?
- [62:12]
Paul Raitzer
I mean, it's. There's a lot of layers to it. It's all in how you're judging the move. So if the move is push the frontier of creative output while reducing the cost of on site production and talent and all those things, then it's the right move. Like, if your goal is to preserve creativity and human, human artistry, it's the wrong move. And you can understand why there'd be people on both sides. But this is a very clear cut. Like, who wins? The studios, who can use the AI tools and can sit there and run hundreds of variations of a squirrel to create an output, and probably got paid hundreds of thousands of dollars each to do this thing. Who loses? All the people who would have been on site shooting this thing and the production companies and the videographers and, you know, the animators and I. There's going to be winners and losers, but this is the future. Like, there, there's no turning back from this. Like, the question would become, and I would be running this tomorrow if I had, if I was the CMO of Coca Cola or one of these agencies, I would be taking Sora when I have access to it and running the exact same prompts and everything into it and then getting a say, oh, sort of save us 50% of time. Like, awesome. The next one, we're going to save even more time. Like, there is no turning back. There will be brands who choose not to do this for ethical reasons or legal reasons, but there are going to be a whole lot more brands that do choose to do this because either they don't have an alternative and don't have the budgets to do it, or they just look at and say, it's an efficiency thing and we're going to do it no matter what. So, yeah, it's, it's wild. It's just, it's such a high profile brand. It's why it's all of a sudden getting so much attention. But you're going to hear a lot more about this. I, I won't, I won't talk about it, but like, I just heard one last week, another major consumer brand that tried to do this and couldn't pull it off in time, but they were going to run all their holiday ads this way.
- [64:08]
Mike Caput
Oh, wow.
- [64:09]
Paul Raitzer
So, yeah, it's. This is not an isolated incident. I will say that.
- [64:14]
Mike Caput
All right, next up. So back in episode 88 in March of this year, we first talked about the release, or the demo rather, of Devin, which was created by parent company Cognition, and they boldly called it, quote, the first AI software engineer. So at the time, the company was kind of making waves with this demo that showed the tool being able to execute complex tasks using code, all basically kind of on its own without a human programmer. Now, that demo got a ton of buzz. Early testers noted that Devin was a bit unreliable and buggy and the demo is definitely a bit hyped up. We noted at the time that this is a product demo that needed to be taken with a huge grain of salt because we are not yet at the stage of fully autonomous AI coding agents, even though some companies want you to believe, believe that. Now Cognition and their product Devon are back in the news, getting a cover story on Forbes showing basically just how much buzz there is now around AI agents, especially those that can code. So when we first talked about them about nine months ago, they had raised 21 million in funding. So solid. But today they're valued at $2 billion after raising 176 million from investors like Peter Thiel's Founders Fund, which, by the.
- [65:36]
Paul Raitzer
Way, is PayPal mafia. Also, it's Thiel Musk, David Sachs, Reid Hoffman. So, yep, yep. Fingerprints all over it.
- [65:44]
Mike Caput
Yep, yeah. So they are basically back in the news with Forbes at this kind of glowing profile. However, they also are noting that we're kind of seeing some of the same types of uneven results right now from AI agents that can code. So Forbes kind of points out that the company's clients, while they report dramatic productivity gains, some have pointed out that they are facing significant limitations using a tool like this. Independent developers have sometimes found that AI can be slower than human programmers. They can introduce errors. Even in Cognition's own demonstrations of Devin, this system has been inconsistent. But the reason we're kind of like talking about this is that with all the talk of AI agents, it's pretty clear that AI for coding is a market we can't really ignore anymore. So in October, for instance, Google CEO Sundar Pichai said that more than a quarter of new code at the company is written by AI. GitHub says its AI code completion tool accounted for 40% of revenue growth this year. And according to Forbes, PitchBook analyst Brendan Burke told them that AI coding has become the most funded use case in generative AI startups focused on it raised over $1 billion in the first half of 2024 alone. So, Paul, a couple reasons this guy jumps out as important. So, first, AI agents becoming all the rage. I think it's important that we keep reminding the audience where the technology actually stands. This is probably not the last cover story we're going to see about AI agents. But second, we can't ignore the fact that it seems like technologists appear highly incentivized to build AI that can code. Like, what do you think? How are you looking at this space?
- [67:28]
Paul Raitzer
It's definitely A very practical use case in a lot of enterprises right now, especially a lot of tech companies, they are seeing benefits from it. I think there's always something a little bit more behind this. Like Sundar was actually asked about this by Sorkin and he said, well, like the humans write the code and then the AI is assisting in this part and then the humans do the final thing. Like that never makes the headlines, it never makes the pitch decks. Like the human involvement in this automation is, is always sort of like just not really talked about. So I have no doubts that people are seeing these kinds of gains. Like, as a writer, I. I could imagine seeing the same scenario play out in writing. Like if you infuse these things in different ways, like you could absolutely save 50% plus of your time on a different writing project. We had Andy Jesse who earlier reported, he said they saved $260 million and 4.5 thousand developer years, not hours years, by using their internal Amazon Q coding agent. So I think you're going to keep seeing these headlines and I think we'll keep reminding people it's not autonomous yet. And like that's you. You think it is, but it's not. From these headlines it can be misleading. So doesn't mean we shouldn't be paying attention, doesn't mean your company shouldn't be exploring this if they're not already replit. Agent is another one we talk about quite a bit. So yeah, it's going to be huge. We're going to hear a ton more about this stuff going on next year.
- [68:57]
Mike Caput
All right, Paul, So for our final topic this week, I'm just going to run through very quickly kind of a rapid fire within a rapid fire of some product and company updates. And you know, please feel free to interject if you got anything you want to double click on further, but we've got a bunch of others that we're just going to hit really quickly. So first up, Meta has released Llama3.3, a new open source language model that basically matches the performance of much larger AI models while requiring significantly less computing power.
- [69:28]
Paul Raitzer
This is the algorithm thing we talked about earlier, by the way.
- [69:31]
Mike Caput
It's this concept because while it contains just 70 billion parameters, it actually can match, they say, the performance of its 405 billion parameter predecessor. So basically that means it has the same capabilities while requiring only a fraction of the computational resources. So they are claiming the model demonstrates impressive performance across multiple languages. It has a substantial context window of 128,000 tokens. So that matches the capabilities of things like GPT4 and they're making it available through an open source license, though there are some restrictions. While it's free for most users, if you have over 700 million monthly active users, you must obtain a commercial license.
- [70:16]
Paul Raitzer
Also, that update is largely for developers. For people on your technical teams, that is not the average business user is not going to go use that.
- [70:24]
Mike Caput
Next up, HubSpot has announced it's acquiring Frame AI, which is a conversational intelligence platform. Frame AI's technology specializes in taking unstructured data so emails, calls, meetings, conversations and turning those into actionable insights. So basically the interesting thing here is this gives you the potential to combine those conversational insights with HubSpot's existing customer data platforms. So they plan to integrate Frame AI into Breeze, which is HubSpot's recently unveiled an updated AI system so you can do things like do a real time analysis of customer sentiment and behavior. Next up, Humai, which we've talked about multiple times, has unveiled voice control. So this allows developers to fine tune synthetic voices along 10 distinct dimensions, including characteristics like assertiveness, confidence and enthusiasm. And that's without the ethical concerns associated with copying real human voices. They have this kind of slider based approach to voice modification so you don't have to use text prompts or preset voices. You can kind of make continuous adjustments along different vocal dimensions. So this is kind of addressing a big challenge in the AI voice industry. Like companies are struggling to create unique voices that match their brand identity without compromising the quality or running into like ethical issues of voice cloning. So this is currently in beta. The technology is being integrated into Hume's empathetic voice interface so developers can create custom voices through their interface. Next up, Anduril, which is a defense technology company, and OpenAI have announced a strategic partnership. They're going to develop AI solutions for military defense systems. This marks OpenAI's first major entry into the defense sector. It sounds like the initial focus is on improving what are called counter unmanned aircraft systems that protect U.S. and allied forces from drone threats. So as part of this, OpenAI's models will be used with Anduril's existing defense systems and their Lattice software platform. Microsoft has said that Copilot Vision is now in preview for select Copilot Pro subscribers in the US So Copilot Vision basically is an AI browsing assistant. So it integrates directly into the Microsoft Edge browser and basically can look at you, what you're doing and provide real time analysis and insights about web pages as you browse. So it's actively quote unquote seeing and understanding the full context of web pages alongside users. Basically a second pair of eyes and it can help you do different things thanks to this functionality. Now, Microsoft, after some criticism of this feature, is emphasizing privacy and security. The feature is entirely opt in and all the conversation data and context shared with Copilot is deleted at the end of each session. During this initial raw, Vision is only going to interact with a select set of websites and Microsoft has said they're going to be cautious about expanding it over time. They explicitly state Vision does not capture, store or use any publisher data to train its models.
- [73:46]
Paul Raitzer
If anybody sees the prompt to get you to allow this to turn on, please reach out to me on LinkedIn. Like I don't, I'm not, you know, we don't use Windows, so I'm not going to see this myself. I'm very intrigued to see how they try and convince people to use this. I can see the adoption on this product being close to zero. So they're going to have to push real hard. I, I don't know, I'm just, I'll be really intrigued to see how they position it so that you want to let them have this, like what wording they use. So yeah, if anybody gets it, just, I would love to see a screenshot of that.
- [74:23]
Mike Caput
Google Cloud has unveiled two significant additions to its Vertex AI platform. So they're adding Veo, which is their newer video generation model, and Imagen 3, their updated image generation system. So Veo is Google's entry into the image to video generation capabilities ability market. Imagen 3, which will be widely available starting this week, represents Google's most advanced image generation model to date. So it's actually interesting to hear some of the notable companies already implementing these. Mondelez International, which owns brands like Oreo and Cadbury, is apparently using the technology to scale content creation across a hundred plus brands. Wpp, a major marketing agency, has integrated these tools into its AI powered operating system for marketing transformation. Next up, some news about X. X has briefly launched and then apparently removed something called Aurora, which is a new AI image generation feature. It apparently has been removed within like hours because they missed.
- [75:30]
Paul Raitzer
Well, you could create anything. I tested on like Saturday or something and it was like any person, like there was like no guardrails at all.
- [75:40]
Mike Caput
You must have hit the window because it apparently appeared briefly on Saturday and then disappeared for many users. So maybe some people still have access. But basically yeah, it lacked all kind of content restrictions. You could generate photorealistic images or whatever you feel like of anything. So it seems like this raised a huge amount of content moderation issues.
- [76:03]
Paul Raitzer
Yeah, it is gone. I don't have it anymore. Yeah, because when you went in, like, you could, it had all kinds of examples and stuff and you could just create. I was doing like Santa Claus stuff just to like see Santa Claus on a Tesla and just random stuff like that. But you could create anything. Yeah. Yeah.
- [76:18]
Mike Caput
So it's, it's unclear. They haven't really revealed, like, why did this happen, what are we doing with it? But we'll see. And then last but not least, some news about Spotify and Google Notebook lm. So this year, Spotify and Google partnered to create this year's Spotify wrapped experience. So this is when Spotify summarizes all the music you listen to in a year. This time around, they're using Google Notebook LM and their podcast, their audio overview feature, which creates a podcast of material with AI hosts. They're using this to actually have AI hosts analyze and discuss your musical journey throughout 2024, including your favorite tracks, artists, and how your taste evolved over the year. So to access a personalized podcast, users can go to their wrapped feed on their Spotify homepage. There's also a separate URL Spotify.com wrappedai podcast that you can access it at. This is only available in English for both free and premium users. It's in select countries, us, uk, Australia, New Zealand, Canada, Ireland and Sweden. Strangely enough, I don't know how we arrived at that, but it is only available for a limited time. So go check it out. All right, Paul, that is a breathtaking week so far in AI. Just a couple quick final announcements here, then I'll kind of turn it back over to you to lead us out here. We are recording this week and releasing next week a 25 AI questions for 2025 special podcast episode. It is coming out on Thursday, December 17th or 19th rather. Sorry, excuse me. We will be releasing that episode as kind of our last episode of the year. If you could please submit your questions, if you have any. We will try to answer as many as we can on that episode. So go to bitly 25 questions episode. That's bit ly 25 questions episode. It's just a simple Google form. You can submit your questions. And as always, please, if you have not already, check out the Marketing AI Institute newsletter, marketingai institute.com newsletter for a full comprehensive brief of everything going on in AI this week. Paul, thank you again.
- [78:45]
Paul Raitzer
A lot of updates at the end there. Yeah, that and the 25 Questions link will be in the show notes as well. So if you didn't jot that down, just check the show notes and there will be a link there for you. I was, I was trying to log in real quick or before we signed off here I think I was trying to log into Sora with my team account which doesn't look like it's active. You have to either be plus or Pro. So I have to try my personal Pro account. Did you try that yet or did.
- [79:11]
Mike Caput
You try I haven't tried it while we've been talking, but I tried my plus account, my personal plus account and it did not work when I tried it.
- [79:19]
Paul Raitzer
All right, well hopefully you all have better luck getting into Sora than us. Let us know what you think. I'm hoping we'll be in by the time we record next week and we can tell you all about our Soar experience and we'll see what else we get this week from our 12 days of OpenAI ship mess. It should be intriguing. So thanks everyone. We'll be back with you again next week. 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.