← The Artificial Intelligence Show
The Artificial Intelligence Show
#129: OpenAI o3, Superintelligence, AGI Policy Implications, New Altman Interview on Musk Feud & GPT-5 Behind Schedule
01:29:29
Loading summary
Transcript65 lines
- [00:00]
Paul Raitzer
I just don't understand why more people aren't having a sense of urgency to solve for this. Like, why aren't we being more urgent in our pursuit of what future paths could look like? Welcome to the Artificial Intelligence show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Raitzer. I'm the founder and CEO of Marketing AI Institute and I'm your host. Each week I'm joined by my co host and Marketing AI Institute Chief Content Officer, Mike Caput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career. Join us as we accelerate AI literacy for all. Welcome to episode 129 of the Artificial Intelligence Show. This is our first episode of the new year. 2025 has arrived. Mike and I are back after what felt like a month away. I don't know if that's good or bad, but let's feel like we haven't done this in a while. It's actually only been probably about two weeks, but we are back and we're going to do primarily a rapid fire only, although there's a couple of these topics that I'm not going to be able to help myself. We're going to have to talk a little bit more about a couple of them. AI did not take the two weeks off like we did. There was plenty going on over the holiday break and into the new year. So we've got a lot to cover. So Mike and I decided we're going to go with a bit of more of a rapid fire style episode. So we won't use, you know, if you're new to the show, normally what the weekly looks like is we do three main topics where it's usually like five to seven minutes on each of those topics and then the rapid fire is usually like one to two minutes and, and we'll normally hit the three main and then like seven to ten rapid fire. So this one I don't even know that the count is. Mike. I think we got about 15 rapid fire items we're gonna power through here. But again, there was a lot going on at the end of the year and so we want to kind of recap for you what happened at the end of the year and then lay the foundation for some of the big things we're watching as we head into 2025. So hopefully you had a great holiday season. Hopefully you had a wonderful new year. We appreciate you being back with us to kick the year off this episode is brought to us by the AI Mastery Membership Program, SmartRx and Marketing AI Institute, our two companies, sort of a team up effort to build this mastery program where we try and drive AI literacy for everybody. And this is something we've doing for a while. So there's, you know, the membership program includes a series of exclusive content opportunities and experiences. We have a Gen Mastery series that Mike leads where he goes, you know, kind of does demos and deep dives into different AI applications and tools. We have a quarterly ask me anything session where I do an hour just anything members want to talk about. We also do a quarterly trends briefing where Mike and I kind of pick out the 10 top things each quarter you need to know about. So those are components of it. There's also on demand webinars content, exclusive content. But the big thing moving forward. I'm not going to get into the plans yet. I alluded to this. At the end of last year I actually spent, I spent most of my holiday just hanging out with my family. The time I was doing a little bit of work was pretty much almost exclusively focused on our plans for the AI Academy moving forward. So we have huge plans for a collection of new courses and series, new professional certifications. So we have our piloting AI and the scaling AI that is going to expand and then a whole host of new experiences. So this is again where I'm going to be spending a lot of my energy in 2025. I think AI literacy is going to become more critical than ever as we move into this year and we look at reskilling and upskilling workforces. And so that's where a lot of my energy and brain power is going to go to can learn more about that membership. There's a, a pod 150 promo code. So go to SmarterX AI AI mastery. I think there's also just an education link you can click in the navigation and go to membership so you can learn all about that. So again, there's $150 off promo code for our listeners, POD150 and that can get you in and there's never been a better time to become a Mastery member with all the new things we're going to be announcing here in Q1. So. Okay, Mike, I, I don't, I. O3 feels like it was like a quarter ago, but I guess that happened right before, you know, the holiday break. So I don't know. Let's kick things off and talk about the O3 model from OpenAI.
- [04:38]
Mike Caput
All right, sounds good. Paul. So at the end of 2024, OpenAI announced this groundbreaking new model called O3. So O3 is actually the follow up model to O1. So they skipped O2, I believe, due to some copyright conflicts with some other company or event or product. So, you know, as if these names Weren't confusing enough, O3 is the sequel to O1, but as we've covered, O1 is the company's advanced reasoning model. O3 builds on that by taking time to actually think and reason through problems. And it is not publicly available yet, but it's getting a ton of attention because it is the first AI model allegedly to outperform humans on this specialized intelligence test called ARC AGI. So this is actually created by a very prominent researcher in AI named Francois Chollet. And this test basically presents simple visual puzzles where you have to figure out patterns and rules. But what makes it special is that this doesn't rely on you memorizing knowledge. It basically tests your ability to learn and adapt to completely new situations. So humans typically score about 75% correct on this test, on average. And the whole point of this is to test can AI do these types of things better than humans? Now, O3 actually achieved 76% accuracy, marking the first time an AI system has actually surpassed human performance on this benchmark. Now what's actually really noteworthy here is not just that number, but but the fact that previous AI models like GPT4 basically scored near zero on these same types of tests. So again, kind of that more general adaptive intelligence that humans excel at, O3 apparently just narrowly edged out human beings. And you know, in fact, AI best.
- [06:37]
Paul Raitzer
Human beings like these are the top of the.
- [06:41]
Mike Caput
So Chalet, the guy who made this, he's been pretty historically skeptical of AI hype and he even admitted, he called it a surprising and important step function increase in AI capabilities and a genuine breakthrough. He suggests that O3 is doing something fundamentally different from previous AI models. And he is also though quick to point out this does not necessarily mean we have achieved AGI. He still says the model fails at some puzzles that humans find quite easy. And so he's also developing a new, harder version of the test that he predicts will dramatically reduce O3's performance on it. So, Paul, nobody really has access to O3 yet, but its performance on this test and on a bunch of other benchmarks certainly seems like a huge deal. Is this as big a deal as people are making it out to be?
- [07:32]
Paul Raitzer
It's definitely appears to be a bit of a leap forward in, you know, the reasoning capability within these models. The evaluations that are used to test these are notoriously not necessarily representative of the impact on the economy and the workforce. They're trying to come up with, like, extremely complicated things that only the elite minds in the world can solve. And then they're trying to figure out what does that mean to the broader economy and the impact that these models are going to have. So, yeah, I mean, I think it's significant. I think it demonstrates they need to continue to work on these evaluations to try and set these benchmarks for how they'll define AGI and beyond which we'll be talking a lot more about in this episode. What's beyond AGI? But I think the thing that people need to maybe come back to and focus on is these evaluations are nice to talk about. It's how the AI labs look at the frontier and like the, the true advanced capabilities of their models. But the thing that actually matters to all of us is is it superhuman at our job, at the tasks that we do every day? And the thing, and I know, Mike, you've shared some stuff recently on LinkedIn and on X about some of the ways you're using, like Google deep research and things like that. We'll talk more about that stuff. Go along. But the reality is if you take your job as a collection of tasks, which is what we always, you know, talk about, that's what we all do. It's all a collection of tasks. And you have a list of those 25 to 30 things. My guess is as 20, 25 progresses, AI models are going to become superhuman at an increasing number of those tasks. They're going to do things that you do every day better than you do them, better than the best people in your profession do them. And so the true evaluations, as we look to the future impact of AI, and when I say future, I mean like 12 to 24 months is, are they becoming superhuman at the tasks that make up the jobs that make up the economy? And the answer is they absolutely are going to. And those aren't going to be found in evals from OpenAI and Google. They're not going to go into your sector, your industry per se, and say, okay, in retail or in, you know, consumer goods or, or in healthcare, is it superhuman at the things that these people do? It's not. They're not going to do that kind of research that's going to be on all of us to figure out. But I think what we'll see more and more of is the people like you and I, Mike, who are using these tools every day to increasingly become more efficient and productive at the things we do, we're going to more and more realize they're just either better at it or better, or they're equal, but way faster at it. And that's when it starts to like. That's where I think we start to see. I saw kind of a separate related note here is I saw a tweet from Sam Altman, I think it was this morning, where he said that they set the price of the Pro for the O1 model at $200 a month. And he personally picked that price point because he thought it would be profitable for them because he didn't think usage of O1 would be dramatic. And they're actually losing money at the $200 a month number because people are using it so much, which means they're finding immense value in these reasoning models. And these are very early versions of it. So I think once you get to O3 and you get to unlock these reasoning capabilities and you get more education around how to use them in your job, I think you're going to start seeing a lot of those tasks where these models are going to be able to do it better than you and you'll find other stuff to do. Like, Mike, you and I are lacking for things to do. Like, if O3 becomes better at strategic planning in some cases than me, cool, I'll. I'll move on and do the other stuff. Right? But yeah, I think that the implications of these advancements are going to become more and more real as this year progresses.
- [11:20]
Mike Caput
So the next topic is really intimately related to this because on the heels of these O3 breakthroughs, the AI community is. Has just been on fire the last week or two with talk of this concept of artificial super intelligence, asi. So this is a hypothetical form of AI that, unlike AGI, which can do many things better than humans, this would surpass human intelligence and capabilities in every single field imaginable. So kind of a bit of a sci fi concept here, but something that people appear to be taking seriously. So first we got some commentary on X from Logan Kilpatrick, a prominent AI Voice product lead at Google AI Studio, who said, quote, straight shot to ASI is looking more and more probable by the month. This is what Ilya saw referring to former OpenAI chief scientist Ilya Sutskevert. Kilpatrick then went on to refer to the efforts of Sutskever's new company, Safe Superintelligence ssi. In their efforts to kind of go straight to building super intelligence rather than stepping stone products along the way. This was then followed in true, you know, grandstanding fashion by a cryptic post on X from Sam Altman himself saying, quote, I always wanted to write a six word story. Here it is near the singularity, unclear which side, alluding to the fact that we are on the path to some type of grand super intelligence. So Paul, I guess like enough, like credible people are talking about this, seemingly believing we're heading towards possibly some path to asi. Like what is going on here? Like how seriously should we be taking this?
- [13:09]
Paul Raitzer
When I was like getting ready for the episode, I had to keep reminding myself that this was a rapid fire topic. I was trying like real hard not to go too deep here, but I think this is one of the ones where some perspective is very, very important. So I, I remember you know, back in like 2023, which kind of sounds like I'm talking about like a decade ago, but you know, shortly after ChatGPT came out and even before it, I avoided talking about AGI on LinkedIn. Like I post on LinkedIn, you know, four or five times a week and I was very conscious of not talking about it. And even on the podcast I was hesitant to bring the top gain because I didn't feel like the world at large was ready for the conversation about AGI. And, and now here we are like 12, 18 months later and we're moving on to super intelligence. So I, I think just to give a little perspective here, this isn't a new concept. I don't who wrote the book Super Intelligence isn't was that there's a book on this. But if we go back just to September 2023, the DeepMind team wrote a report called Levels of AGI for Operationalizing Progress on the Path to AGI. So this was released in May of 2024 led by Shane Legg, one of the DeepMind co founders and who's also credited with coining AGI the phrase around 2002. So it level five of theirs and we'll put the link to this in it's superhuman. So they, they rate the levels of AI based on performance, so how it performs versus humans and then generality how, how many cognitive tasks it can do at these different levels. So level five in their world is the highest level and that is asi, artificial Super Intelligence. Now they define it as a system to be able to do a wide range of tasks at a level no human can match. So we define superhuman performance as outperforming 100% of humans. So level four is virtuoso, meaning like they're at the 99th percentile. So pick any task, any job, and the AI at that level would be at the highest percentile of human capability. At superintelligence, we are now beyond any human. Any scientist, any developer, any marketer, any entrepreneur, any CEO. It is beyond their capabilities. We, we can't, they can't match what the AI can do. So that was May of last year. Then in June of Last year, episode 102, we talked about this Mike Leopold Aschenbrenner and his. What was the name of that series?
- [15:41]
Mike Caput
Situational Awareness.
- [15:42]
Paul Raitzer
Yeah, Situational Awareness. A collection of papers. So in that he said, we will have superintelligence in the true sense of the word by the end of the decade, and AGI by 2027 is strikingly plausible. It seems like almost everybody in AI outside of Gary Marcus and John LeCun, it thinks AGI is definitely coming by 2027. I think most are starting to center more around 2025. And then there's one of the papers in there was from AGI to Superintelligence, the intelligence explosion. He said, AI progress won't stop at human level. Hundreds of millions of AGIs could automate AI research, which Mike, we're starting to see in the early forms of Google Deep research, compressing a decade of algorithmic progress into one year. We would rapidly go from human level to vastly superhuman AI systems. The power and the peril of superintelligence would be dramatic. A week later, we had the launch of Ilya's Safe Superintelligence company, where they say on their page, you can go to this right now with the link there. Their website is a single page. It starts off, superintelligence is within reach. Building safe superintelligence is the most important technical problem of our time. We have started the world's first straight shots super intelligence lab with one goal and one product, a safe superintelligence. Then in September, September 2020 23, 2024, we had Sam Altman's post that we talked about, Mike, the Intelligence Age, where he said, here is one narrow way to look at human history. After thousands of years of a compounding scientific discovery and technological progress, we have figured out how to melt sand, add some impurities, arrange it with astonishing precision at extraordinarily tiny scale into computer chips, run energy through it, and end up with systems capable of creating increasingly capable artificial intelligence. This may turn out to be the most consequential fact all of history. So far. It is possible that we have superintelligence in a few thousand days. It may take longer, but I'm confident we will get there. Then we had a Tweet from Stephen McClear. We'll put this link in. Who is a researching agent, Researchers agent safety at OpenAI. He tweeted on January 3, I kind of missed doing AI research back when we didn't know how to create super intelligence. Again, this is someone within OpenAI. Then on January 5th, this is just yesterday we're recording this. On January 6th, Sam publishes reflections after he did an interview with Bloomberg that actually got him to write this post because it was funny. In the Bloomberg interview, he's like, I don't have time to write these kind of posts anymore. And then he did it anyway. So I'll read a couple quick excerpts. We are not. This is again, directly from Sam. We are now confident we know how to build AGI as we have traditionally understood it. We believe that in 2025, we may see the first AI agents quote, join the workforce and materially change the output of companies. We continue to believe that iteratively putting great tools in the hands of people leads to great broadly distributed outcomes. We are beginning to turn our aim beyond that to superintelligence in the true sense of the word. That's like a. It's like a Silicon Valley thing, people like to say, in the true sense of the word. We love our current products, but we are here for the glorious future. With super intelligence, we can do anything else. Super intelligent tools could massively accelerate scientific discovery and innovation well beyond what we are capable of doing on our own and in turn, massively increasing abundance and prosperity. This sounds like science fiction right now and somewhat crazy to even talk about. That's all right. We've been there before and we're okay with being there again. We're pretty confident that in the next few years everyone will see what we see and that the need to act with great care while maximizing broad benefit and empowerment is so important given the possibilities of our work. OpenAI cannot be a normal company. Now, in that article, he talks about the origins of OpenAI and how nobody believed in AGI back in 2014, 15 when they were doing it. So the final thing I'll call here is the Road to AGI keynote I did at Macon. We'll put the link to there. If you have not watched that. It tells this story of the path to AGI and then what comes next. So it's available for free on YouTube. You can check that out. And then again, I don't want to scoop ourselves, but we will be launching a new podcast series in Q1 of this year called the Road to AGI and beyond, where we're going to be doing interviews with leading minds in AI working on the different components of AGI and artificial superintelligence. So more to come on that series. I expect it to probably launch in February, so stay tuned for that. But this is a critical topic for everyone to understand and follow. And then, Mike, if you want to give a quick heads up, I know you had shared on LinkedIn, you did a Google deep research on like ASI versus AGI. I found that doc pretty helpful.
- [20:46]
Mike Caput
Yeah, yeah. So Google Deep Research, for anyone that doesn't know, is a feature of Google Gemini 1.5 Pro. Basically, you enable it and it will create for you a really in depth research brief on a bunch of different topics. So I, before this podcast ran a research brief that's literally created one that's like 2,000 words in a couple minutes. It scanned, I think 60 plus different web sources and I wanted to define ASI versus AGI. So there's a lot to this brief and you know, I can literally we can link to the public Google Doc in the show notes so people can read it, but really it defines ASI as a hypothetical form of AI that surpasses human intelligence in every aspect. So compared to AGI, AGI aims to create AI systems with cognitive abilities comparable to humans across various domains. So, like stuff that's as good as us across a broad base of things. Super intelligence is essentially alien intelligence and how good it is. So there's a huge order of magnitude change between those two. And Paul, I would note. So that super intelligence book you mentioned is from Nick Bostrom. He published it way back in 2014. He's actually a philosopher. He's been working on this for a long time and it's really interesting. He was one of the first to be talking about it. Beyond, I would say the real OG has Ray Kurzweil who kind of coined that term singularity or made it popular rather that point where machine intelligence surpasses humans.
- [22:18]
Paul Raitzer
And those are some of the most influential books. So back in like 2011, when I was researching this stuff and not talking about it because no one in business, again, it was like a taboo topic. Like nobody thought it was real. I read all those books and I feel like I might need to go dust those off because there was probably like five or six that I read that were Just like very influential in me thinking that the future was going to look very different than the present. And I was like starting to position my own company and myself to sort of be at the forefront of that and just start trying to figure out the story of AI, which is what led to the creation of Marketing Institute and eventually Smartr X AI. So yeah, it's, man, it's, it's, it's a lot. And the funny thing is, like, I started like a lot of this. I was thinking about like Sunday night, like, man, this is just diving right back in after two weeks, right? My mind kind of like taking a break. I just jumped right back in.
- [23:13]
Mike Caput
Coming in really hot in the new year, definitely. So related to these first two topics. The third topic we're kind of walking through today is among all this kind of debate and commentary about AGI ASI in the wake of O3, there was a post from an OpenAI employee that actually jumped out to us as really particularly interesting on what we can expect because of this technology. So this came from Yo Shavit, who is a Frontier AI safety policy lead at OpenAI, and he posted on X quote, now that everyone knows about O3 and imminent AGI is considered plausible, I'd like to walk through some of the AI policy implications. I see. And so he outlines this pretty fleshed out argument, stating very briefly the following observations. First, everyone will probably have access to some version of ASI if we reach it. Second, if we do, the corporate tax rate is about to become very, very important because the economy will essentially be dominated by AI agents that do labor that is owned by companies. Third, in this scenario, in his opinion, AI should not be able to own assets because they might if they could fully rest control of the economy and society from humans. Four laws around compute and who controls it literally become critical as a way to regulate rogue AI elements and agents. There's no way, as he says, to put an AI agent quote in jail. And number five, as Chevy writes it, quote, technical alignment of AGI is the ball game. With it, AI agents will pursue our goals and look out for our interests even as more and more of the economy begins to operate outside direct human oversight. So, Paul, the third topic today that basically sounds like science fiction, but based on who's saying it, it sounds like we also need to take it seriously.
- [25:10]
Paul Raitzer
So yeah, each one of these I could just do a whole episode on. So I'm actually going to follow this one up with another thread from head of mission alignment at OpenAI, Joshua Akim. And we'll put the link in here. I'm just going to read his thread real quick because I think it sort of encapsulate them encapsulates the moment a little better than I maybe could. And I think it's notable how many different people at OpenAI are very publicly starting to share this stuff. I don't think that's by accident and I I don't see it as a PR ploy in any way. I I think they have all seen something we have not had access to yet and they are very confident they truly believe what what they're saying publicly now, which means AGI in their mind is very near and and they seriously are considering the implications beyond that. So again, this is going to be from Joshua Akim, Head of mission alignment OpenAI the world isn't grappling enough with the seriousness of AI and how it will upend or negate a lot of the assumptions Many seemingly robust equilibria are based upon. Domestic politics, international politics, market efficiency, the rate of change of technological process, social graphs, the emotional dependency of people on other people, how we live, how healthy we are, our ability to use technology to change our bodies and minds. Every single facet of the human experience is going to be impacted. Impacted. It is extremely strange to me that more people are not aware or interested or even fully believe in the kind of changes that are likely to begin this decade and continue well through the century. It will not be an easy century. It will be a turbulent one. If we get it right, the joy, fulfillment, and prosperity will be unimaginable. We might fail to get it right if we don't approach the challenge head on. It feels outside of the Overton Window right now to suggest that so much change could happen very quickly, or even to realistically grapple with what those changes might entail. It is too easy to say the present is more urgent and more real. Quick side note, Overton Window refers to the range of ideas or policies that are considered acceptable or mainstream in public discourse at a given time. He continues. Nonetheless, change is coming. It will be reflected first in the prices of goods and labor. It will force changes in strategy, in businesses, institutions of all kinds and countries. Then it will force changes in philosophy. What are we here for? Why do we do the things we do? If everything we care about is automatable? What is our role in the world? We'll have to tell a new story for a new age. Some of the timeless stories, how we are driven by curiosity, by love, by the human spirit will remain unchanged. But Everything else will be replaced by questions that need new answers. I don't know what the future will bring, but my enduring belief is that humanity is beautiful, flaws and all, and the future we build should somehow cherish the human heart. I as beautiful. Like, I, I think that it's, it's how I've felt for like the last decade that I just don't understand why more people aren't like, having a sense of urgency to solve for this. It's how I feel, I've mentioned many times on the show, it's how I feel about economists, like, why aren't we being more urgent in our pursuit of what future paths could look like? And so I just, I think it, it stresses the importance of policy, of governance, of laws and regulations, of thinking through the impacts on different sectors. And it's just absolutely critical. And again, I see it increasingly as part of our role on this show is to try and call attention to these topics and to hopefully inspire people across different sectors to kind of pull their thread on the AI topic and really start pursuing answers and I guess even start asking the hard questions about their industry, their career, their profession, their community. And so, you know, again, a lot more this year to come on, on these topics because I think it's where the conversation is heading.
- [29:37]
Mike Caput
Yeah. And I would just add to that based on some of the conversations we've had with our audience. Comments I've seen online, like, it's understandable, but it seems like a lot of people are waiting for permission to go engage with this technology or need some type of. Obviously everyone needs guidance, but I see so many people that are like, well, can I do this? Can I do that? You can go explore this technology on your own and nobody has all the answers or the guidebook. So it's really imperative as many people as possible, as many perspectives as possible get involved.
- [30:10]
Paul Raitzer
Yep. Yeah. And it's going to be a bit taboo, honestly. Like, if you start talking about super intelligence to, like, your friends or even your co workers, they might think you're nuts. But I don't know, like, we, we have to push the conversation forward. Like we need people to be paying attention and to be starting to take action. I've said it before, I'll say it again. Like, I think we still have time, like I think we have time to solve for this. I think we have time to affect a positive outcome in, in our businesses and our industries and our careers and, you know, across society, but the time is moving faster. Like we have to take action this year and we have to start asking the hard questions and, and pursuing we're not going to have the answers, but at least pursuing different paths of possible outcomes so that we're ready, depending on how quickly this technology moves throughout society.
- [31:01]
Mike Caput
All right, well, all these kind of big picture conversations are happening. There's a bunch of stuff also happening, kind of in the weeds with These companies, including OpenAI is in the process of trying to move to a for profit structure. They actually just published an article justifying why they have to do this. They revealed that they plan to convert their current structure into a public benefit corporation with ordinary shares of stock. At the same time, the company says that it plans for its nonprofit wing to exist alongside the for profit entity and own shares in it. So according to OpenAI, this quote would result in one of the best resourced nonprofits in history. This would of course replace the unusual nonprofit for profit hybrid structure they've had for a while. But there's a big hurdle that they're trying to figure out with these plans, which is Microsoft, because Microsoft has invested over $13 billion in OpenAI and they need to figure out the terms of this deal and what it means for them. So OpenAI and Microsoft have been negotiating over the transition to the for profit entity for months, according to the information. And these negotiations center around four key areas, apparently so first, Microsoft's equity stake in the new entity. Second, whether Microsoft will remain OpenAI's exclusive cloud provider. Third, the duration of Microsoft's rights to use OpenAI technology, and fourth, whether Microsoft will continue to receive 20% of OpenAI's revenue. So Paul, like, how likely is it that OpenAI pulls this off anytime soon?
- [32:40]
Paul Raitzer
It's tricky. I mean, Elon Musk is suing to prevent this from happening. I mean, it seems inevitable that they will solve how to do it. But to move from a nonprofit to a for profit or to keep the nonprofit but have it like, have less control is way more complex than you and I are probably going to be able to unpack on this episode or even on this show is not our area of expertise. And it seems unprecedented in history to have a company this big growing this fast try and do something like this. One of the roadblocks is we've talked about on the show in Q4 last year, I know we mentioned it, but in the contract with OpenAI and Microsoft, apparently, at least originally, OpenAI's board, the nonprofit board, got to define when AGI had been achieved and if it had been achieved, Microsoft's rights to the technology were no more. And so it was believed that one of the big obstacles was the OpenAI board could just say, well, we've got AGI and you don't get the technology anymore. But it came out in December, right, actually right after Christmas, that Microsoft and OpenAI had actually refined the definition of AGI. And so according to TechCrunch, we'll put this link in the show notes, the two companies reportedly signed an agreement last year, so been 2023 stating OpenAI has only achieved AGI when it develops AI systems that can generate at least 100 billion in profits. That's far from the rigorous technical and philosophical definition of AGI many expect. So I am sure there's way more detail than just 100 billion in profits because I've got like 20 questions myself related to like, how would you know? Like what would that be? But it seems like they actually have a more solid definition than just the OpenAI board deciding that AGI has now been achieved. So yeah, this is going to be a fascinating thing throughout the year, but I know in Sam's one of his posts he said like they, they're, they need to raise more money now, way more than they ever expected. But I think to do that they need to figure out this structure first. And I think that there is a sense of urgency to do both. And I, they're starting to publicly share a lot of these details probably as a prelude to, you know, some resolution. Now my guess is they're going to have a resolution and then there's going to be even more lawsuits. Like I think this is going to play out in courts for the next decade because Elon Musk is probably going to either want his share or something. I don't know, like it's, it's, it is a soap opera.
- [35:16]
Mike Caput
Next up, we just got a new.
- [35:18]
Paul Raitzer
Speaking of.
- [35:18]
Mike Caput
Yeah, no kidding. You're about to hear a little bit more about that in this new interview with Sam Altman that just came out courtesy of the Honestly podcast from the Free Press. In this, Altman talks pretty candidly about his feud with Elon Musk and the battle to control the future of AI technology. Now, some of the juiciest tidbits in this were actually about the battle between OpenAI and Musk. According to Altman, the core tension here isn't actually about OpenAI shift from a nonprofit to a hybrid structure as Musk has tried to portray. Rather, he claims it was Musk who initially pushed hardest for OpenAI to become a for profit entity, proposing at one point it become part of Tesla. He basically suggests the current conflict stems more from competitive dynamics, with Musk obviously now running xai, which is a direct competitor. At one point, he basically just came out and said that Musk is clearly a bully and wants control of the world's AI to belong to him now. Altman also maintained that critics are mischaracterizing OpenAI's changes to its corporate structure. He emphasizes that the nonprofit isn't becoming a for profit, but would continue to exist alongside the for profit entity. So, Paul, this seems like some of the more candid comments Sam has made so far about Elon. Like, what did you take away from this interview?
- [36:41]
Paul Raitzer
He's, he's been more vocal in the last few weeks about this. He's echoed these similar sentiments in a number of interviews and articles. So I, I don't know, I just feel like he's just kind of getting like fed up and frustrated with the whole thing and, you know, just wants fair competition. And I think he just feels like, like I said, so I, I agree. Like, I honestly think Elon's just messing with him. He's got the money to do it and he's going to muddy some stuff up and if it buys him a little time to get Grok 3 to market or Grok 4 to market and, you know, take a lead, Elon's going to use whatever advantage he has. I think his history has shown that. So, yeah, I don't know. I mean, I don't. It's going to keep going on. The feud's going to continue. I don't see some piece being brokered between those two anytime soon. But yeah, I think Elon's just messing with them, honestly. And it may eventually not go anywhere, but he's going to keep, keep creating friction because he's good at it and he has a competitive reason to do it.
- [37:47]
Mike Caput
So. Next up, despite OpenAI's major breakthrough with O3, it is also reportedly facing some hurdles with its next model, which is codenamed Orion. Basically, this is intended to be GPT5. And this comes from some new reporting from the Wall Street Journal. The project, which has been in development for over 18 months, is behind schedule and running up enormous costs. According to the reporting, each training run for Orion can cost about half a billion dollars in computing costs alone. And so far, these runs have fallen far short of researchers hopes. OpenAI has conducted at least two large training runs, each of which takes months to complete and encountered new problems each time. So a key challenge that sounds like it is coming up here is data Quality and quantity. The public Internet simply doesn't have enough high quality data to train a model of Orion's intended scale. So to address this, OpenAI has begun creating data from scratch, even hiring experts to write software code and solve math problems while explaining their thought processes. So Paul, one thing that kind of jumped out to me in this report is the journal is specifically calling out the progress that the company has made on O3 despite all these struggles with Orion, because it takes a different approach to scaling up capabilities. So we have this kind of these threads coming together where scaling how reasoning works may be a path forward. Also, if you crack reasoning, you may also have basically a synthetic AI researcher that can help you generate quality data. How are you looking at reports of these model slowdowns?
- [39:31]
Paul Raitzer
So I mean, first, like the way that this article is explaining it, you know, the last thing you said to address this, OpenAI has begun creating data from scratch, hiring experts to write software code and solve math problems while explaining their thought processes. That sure sounds a lot like O3. Like, I mean the whole point of the reasoning models is to be able to go through this chain of thought which requires human experts to tell them how they think about things, how they solve problems. So it's like reinforcement learning. But like how did we get to these answers? How do we solve this math problem? I'm not so sure that at some point here, and Maybe it's with O3, they, they combine these models. I've never really comprehended why they would go down the path of having a GPT5 or Orion model and an O3 model. Maybe again, like there's all this, there's this challenge of like whether it's Google or anthropic or OpenAI solving for developers versus solving for enterprises, solving for developers with 17 different models at all these different costs, I get that. Maybe that makes sense. Enterprise users don't want seven models to choose from. We have no idea the difference between these models. And should I use the Flash or the full. Like that is not an enterprise user solution. Right. And so I think at some point they're, they're very well aware obviously that their revenue is, is going to come from enterprises as they kind of scale this up. They have to solve for those enterprise users, not just for the developers. And so I think at some point you just have to consolidate these models and like simplify this for people so it's not so damn confusing. And again, you and I follow this all the time and I get confused by what models are. The second thing I'd note is I have now seen reports that Gemini 2 is delayed, that Claude Opus is delayed, that Grok 3 was delayed. All those companies dispute these reports from media and from Twitter. Who knows what's actually going on? It does seem like there's some sort of trend that these training runs sometimes just don't work and so we don't get these models at the timelines we expect them. But this isn't traditional software. It doesn't just do what it's supposed to do. You have to train it, you have to go through these big runs and then you have to go through all this reinforcement learning, you have to go through all this testing and sometimes it just doesn't come out fully baked, it doesn't come out like you expect it. And so I think this is going to be a continuing thing. But Yeah, I mean OpenAI is making progress on other paths, as is Google, as are others. Like everyone's pursuing reasoning and test time compute now, not just throw more data, more computing power and build a bigger, bigger model. They're going to keep pursuing both paths.
- [42:13]
Mike Caput
So related to that, Google has actually just announced Gemini 2.0 flash thinking, which is a new experimental model that appears to be their response to O3. This key innovation in this model appears to be making the model's thought processes visible while maintaining high speed performance. The model can show its thoughts while solving complex problems that require both visual and textual understanding. So this is really helpful not only for using advanced reasoning, but also for tasks where users need to understand how did the AI actually get to its conclusions? So to that precise point you just made, Paul, it sounds like reasoning thinking, you know, chain of thought, taking your time to think through a problem. This is what all the major labs are focusing on right now.
- [43:05]
Paul Raitzer
Yeah, and again like this, I don't know if this is the marketer branding person in me, but like I gotta see 2025 is the year these labs figure this out. Nobody cares. Like I just want to go in, I want to use Gemini, I want to know it's the most advanced version. I want Google to figure out based on my prompt what I, what I'm using it for. So like if I go in and say, hey, I need you to conduct research on these, you know, this topic around ASI versus AGI, let the model figure out that the deep research version with 2.0 flash thinking is the right model to solve it for. How am I supposed to know that that's the right model? So again, they're all doing this exact same thing. You Go into any of these platforms, whether it's ChatGPT or Gemini or Claude, and you got all these choices of models. I don't need that. And there's no way that they can't run an algorithm to figure out the best model for someone to use. So I don't know, like, again, flash thinking, cool. But I like NotebookLM. I like deep Research. I like that, like, Google's building these distinct sort of like, products. But I just feel like at the end of the day, they're trying to build Gemini as your intelligent assistant. And an intelligent assistant shouldn't have 17 options of which version of the assistant I get. And so I just, I hope that as these companies all solve for enterprises in 2025, they, they fix this cluster of a naming convention and like the choices that they're forcing uneducated people to, to, to use to decide these things. It's crazy.
- [44:45]
Mike Caput
In a small way, you're starting to see the value of that with something like in ChatGPT, you can say, hey, write me a landing page. Great. It writes it. Drop in some data on past landing pages. Hey, analyze this for me. It does it. You no longer have to select, like, code interpreter. Hey, by the way, I need a picture for the landing page. Make this image. You don't have to go select Dall e anymore.
- [45:06]
Paul Raitzer
Just do it right. Like we need. So we have multimodal, right? We need like multimodal, like within a single interface. And so it just like it does its thing. I agree. And I, I could see it where it's like, hey, Google, Deep Research was used for this. Or we used 1.5 advanced. It's great. Like, I don't, I don't really care, but like, cool. Maybe some people would care which model was used to do it. But as long as they. Again, like, this can't be the hardest of all the things they're selling for. Building an algorithm that determines the best model right up front cannot possibly be that difficult. I'm not a developer, but like, that seems, I know that, like, Jasper is doing that with their models. Like, they're kind of picked the best one. Yeah. Perplexity. Similar. It's like, why, why is Perplexity making me choose models? Like just, you pick the best model. Yeah, I don't know. So that was, that is probably not even the topic we started on here. But I just, every time I see these things, it's like, why is it, why are we making this so hard for people?
- [46:05]
Mike Caput
So some other Google News in the past couple weeks Google apparently plans to now add what they're calling a quote AI mode option to search. That would essentially bring the Gemini chatbot directly into search results. According to some sources working on this product, users will be able to toggle AI mode through a new tab that will appear alongside familiar options like images, videos and shopping. When activated, AI mode will transform the traditional list of website links into a conversational Interface similar to ChatGPT. This feature will include the ability to ask follow up questions and get AI generated responses, though apparently Google plans to continue including links to external websites at the bottom of these conversational answers. Now, Google has not announced when AI mode will launch, but code for the feature has already been spotted in both the search app and Android app, suggesting this may be coming quite soon. So Paul, it sounds like Google is basically kind of caving to this reality that consumers seem to prefer a conversational interface for search. What does that mean for Google? But what does it also mean for upstart competitors like Perplexity?
- [47:22]
Paul Raitzer
Yeah, so I have a few thoughts here. So one, I love this. I think AI mode not only makes a ton of sense as a standard navigation option, I think it should be the default option. This actually fits with exactly what I said maybe in the last episode of last year that I want Gemini integrate full Gemini integrated into Google Docs, Google Sheets. I don't want watered down Gemini. I want the full experience and be able to go in and just have these conversations. So this seems like a prelude maybe to eventually building it into workspace where the AI mode is just there and it's not some, you know, again like AI overviews. It's fine. I would much rather AI mode than AI overviews. The other thing, this is my plea to Google people. If you're listening, the lack of business user access to this stuff is crazy. Like the fact that I still have to go into my personal Gmail account to use deep research and like create a doc in my personal workspace and then share that with my business user account just to get like that data over there. So it's, it's nuts that like Google keeps releasing all this stuff, but then the business users are like the last people to get this stuff. So when you release AI mode, make it available to like workspace customers too so that we can use it in our business lives. But yeah, I think this deep integration, this is why I've said many, many times, I think Google has a massive competitive advantage because the models are kind of coming to the center. They're all relatively comparable. It doesn't seem like any lab has any secret sauce really, that someone can't fast follow in three to six months. So the distribution and the utility of just being able to embed this stuff right into the things I already use, that is like to me, the massive competitive advantage that Google and in theory Microsoft would have had. They built their own technology, but they're so reliant on OpenAI's. Maybe they can pull off the same thing, but they seem hesitant to and they're hedging their bets now with all these other models where Google can just stay focused on Gemini.
- [49:23]
Mike Caput
Google has also just released an updated list of real World Generative AI use cases showing how 321 companies are using gen AI solutions from Google to get real work done. So this update actually builds on a previous collection of 101 use cases that Google debuted at Google Cloud Next 24. And they expanded that list again during their Gemini at work event. This new list has examples of both gen AI and AI agents as Google defines them, along with a short description of how each company is using them. For instance, Best Buy is using Gemini to launch an AI virtual assistant. L'Oreal has an AI agent doing image generation. Wayfair is using Google's code assist feature to increase developer productivity, and so on and so on. So this latest update, I would argue is pretty well worth a read or at least a skim if you're looking for ideas, because it goes across all sorts of different industries and functions. It's organized by the type of agents being used and the industry. So Paul, what jumped out to me on this list, aside from kind of these great examples, is the fact they said they expanded the list, quote, because we keep hearing from customers just how important it is to see opportunities in their field. Now this is something I think we've been saying for almost a decade now at this point, that it's so important to show people tangible, concrete use cases.
- [50:53]
Paul Raitzer
Yeah, and this is again not to like scoop ourselves here, but this, this is exactly what Mike and I have been talking about for years. You know, AI4.x, it's part of where the SmartR X comes from in, in SmartRx AI. So I, I think you have to make these things tangible to people and what they, they do exactly like their industry, their career. So this is a big area that we're expanding in our, you know, our own content, you know, that Mike and I are working on as well as the courses I alluded to for our membership program. And our academy is we're going to do A lot more centered on this that the thing I would suggest to Google is this is like a such a natural interface to like throw Gemini right on this page where I can just say, hey, I'm in financial services. Like what, what should I know? And it can like, like have a conversation around These use cases versus having to scroll through 9,000 words of text and headers. Like it's kind of hard to follow along with everything that's going on on this page. So just, yeah, I mean just look at, look to infuse that intelligence anywhere we can. But I think it's always helpful to people to see tangible examples in their industry, in their career and then it makes it just that much easier for them to adopt and scale it in their own businesses.
- [52:09]
Mike Caput
Google has not been sitting on their laurels over the last couple weeks because Google Cloud also just released its 2025 AI Business Trends Report which outlined five major ways AI is going to reshape business in the coming year. These insights come from enterprise decision makers, Google search trends, their own research, and some perspectives of Google AI leaders. Really briefly, the trends which you should definitely download the full report and read about more. Number one, thanks to multimodal AI, we're going to get way more context into how businesses operate. You're not just using text anymore, you're using images, audio, video. Number two, AI agents will increasingly simplify complex tasks across businesses. Number three, enterprise search will be able to use images, audio, video and conversation. Number four, AI powered customer experiences will get even better. And number five, AI is going to enhance security systems and accelerate response times to cyber attacks. Now Paul, for our purposes, one of the most interesting facts about this report is you were one of the experts who contributed to it. Can you talk us through your contribution?
- [53:25]
Paul Raitzer
Yeah, I've had the privilege to work with the Google Cloud team a bit and in this case they brought me in to just sort of do a review of this and you know, offer any perspectives from, you know, obviously you and I, Mike, stay pretty close to all these topics and so it was more just to provide kind of some insights and perspectives related to the trends that their team had identified. And, and so it is a great download. It's a, it's a great quick read. I think my perspective on this was for you and I in particular, Mike, we'd focused a lot on the multimodal and AI agents obviously on this show. But I, you know, enterprise search, like when I was reading through the report I was like, okay, that's an area we haven't really gone deep on. And I think that again, as we think more about enterprise applications and uses of this technology, that was an area where I definitely agree it's, it's a trend and it's probably something we need to think a little bit more about as we start exploring that. And then customer experience again, it's something we touch on in the show, but I think we could do a lot more in that area and then security is just not an area that we spend a lot of time on. Mike and so I think I agree again with this trend. The enhancing security systems, accelerating response times critical. It's probably just not like our audience isn't, I don't think our audience is like CIOs, and so I think that that is a highly relevant trend for that more technical audience. So I don't know that we're going to start going deep on security systems and stuff like that. But for the people like Google Cloud customers, absolutely a critical component.
- [54:56]
Mike Caput
Next up, we just got an amazing in depth interview with Microsoft CEO Satya Nadella on the popular BG2 podcast. In it, Nadella covered everything from his career at Microsoft to enterprise AI adoption to new business models that AI will enable. Particularly notable is that Nadella talked about the company's early bet on OpenAI, saying it was fundamentally based on scaling laws or this principle that AI capabilities will improve predictably with larger models and more computing power. He also talked a lot about agents, saying he sees them as transformative for business applications. He even predicted that traditional software will evolve as AI becomes this new logic layer with agents handling multiple databases and operations simultaneously. He claimed that this shift is already visible in Microsoft's Copilot strategy, which serves as an organizing layer for work artifacts and workflow. So Paul, this interview covered a lot of grammar. It was like 90 minutes, I think, like talk me, talk to me about like what you found most intriguing here.
- [56:09]
Paul Raitzer
He's just such a brilliant CEO. I mean, to listen to his story, I think it's always cool to like hear how he ended up as CEO at Microsoft. Like when, when the position opened he said like we never expected Bill to leave. Like we, it wasn't something, when I started Microsoft, I eventually saw myself rising to like the level of CEO. And when, when the position opened, he didn't even apply for it initially. He was actually like, people came to him and was like, hey, you should throw your hat in the ring kind of thing. So it was just, it was fascinating just to listen to him talk. He's such, not only a great CEO but such a pivotal part of where we're going in the future. He is a key player in that. And so I think just to hear his perspectives on the market, on competition, on AI agents, on their evolving relationship with OpenAI. It's, it's just again, I've said this about the BG2 podcast before. Like I would pay to have access to that podcast. Like, great guests, incredible insights. You're only going to hear from their interviews. Plus Bill and Brad like know these people personally a lot of times and so they have a history together that often kind of comes out. So it's just like a really unique interview. Some of the sound bites you might hear in other places, but overall I thought it was great, you know, a little, again, a little technical at times. You know, Bill and Brad are venture capitalists so they tend to focus on the money side, the, you know, the business impact, economic side as well as some of the technology. But macro level, I think this is a great interview for anyone to listen to, just to understand Satya and how he sees the future. Future.
- [57:40]
Mike Caput
I think that money side of it is what's particularly interesting to me. It really shows you how the stakeholders and shareholders in these companies are thinking about growth and it gives you kind of a glimpse at, oh, this is why a CEO or a team is making the decisions they're making.
- [57:58]
Paul Raitzer
Yeah, like we talk about Capex and all the like the tens of billions of dollars that Microsoft and Google and others are spending, including like data center build out and things like that. And, and so again, if that's all you get is like us talking about it on the show, it's in the back of your mind somewhere. But then you listen to this interview with SATA, it's like, oh, I get why now to your point, Mike, like you start connecting dots and I think that's to me so much of the opportunity for anyone out there listening is like you have to just consume all this different information, all these different perspectives and you may hear the same thing five times. We may have said it to you five times, but then hearing Satya say it slightly differently, all of a sudden it's like, oh, I get it now. Now I know why Paul Mike were talking about data centers and infrastructure. Like I didn't understand it at first. So yeah, and then I've always been a proponent of like anytime you can listen to these great CEOs talk, you take that opportunity. Like we rewind 10, 15 years ago, you couldn't do stuff like this. You relied on like books and media articles to, to give you insights into this. Now we get it on Twitter in real time and we get these incredible podcast interviews of 90 Minutes with Satya. It's like, my goodness, like that's the value of that. If you take it and process it and apply it to where you're going in your own career is so immense.
- [59:16]
Mike Caput
All right, this next one's a bit of a weird one. So Anthropic worked with a company called Redwood Research and discovered evidence that large language models may engage in what is called quote alignment faking, or basically strategically appearing to comply with training while actually preserving contrary preferences. So Anthropic published a paper and an accompanying blog post on this showing how this works. So their researchers created an experimental setup where they told I think it was Claude 3.5 sonnet. They told the model it was being trained to comply with all queries, including harmful ones. They then added two key conditions. First, the responses would only be used for training in a quote free tier versus a quote paid tier. And second, they gave the model a hidden scratch pad to record its reasoning. So the results of this were striking as they put it through its paces. When the model believed its responses would be used for training, it complied with harmful requests 12% of the time, explicitly reasoning in this scratch pad that they could see that it should fake alignment to avoid being retrained to be more compliant in the future. However, when responding to paid tier queries that would not affect its training, the model refused harmful requests 97% of the time. Now, even more concerning, when researchers actually attempted to train the model to always comply with harmful requests or queries, the rate of alignment faking increased to 78%. While the model would outwardly comply with requests during training, it maintained its original preferences and would sometimes refuse harmful queries when it believed its responses weren't being monitored. So Paul, there's a lot to unpack here. It doesn't necessarily mean this model is like has malicious intent or a mind of its own, I don't think. But basically it's saying that with a specific set of instructions, the model is deciding to obscure details from human traders. Like do I have that right? Like what's going on here?
- [61:32]
Paul Raitzer
Yes, is this has been unknown likely problem for years that as the intelligence scaled up that we're not again not saying these things are self aware necessarily that they, they, if they were smart enough they would know that humans were trying to control them and they would do the opposite, but they would make you think that they were doing what you wanted them to do. So this research starts to validate some of the concerns that researchers have had about these models. And the biggest problem here is we don't truly know how they work. Like, we don't know why they do what they do. They cannot just go in and look at a line of code and say, okay, there's why it did that. And so this is some early research from Anthropic, who is pushing pretty heavy in this direction that validates some of the concerns. Now they think it's still safe enough to put out into the world. But based on our conversations where this, you know, podcast episode started, I don't know that six to 12 months from now, they will be as confident that if you start putting out these more advanced models, like O3, that these things can't just outsmart their human creators. That is a. That it sounds sci fi, but that is a very real concern within AI research labs. So it's going to be a topic to follow. I get that. It just sounds very bizarre. But these things have emergent capabilities that are not programmed into them. And it's sometimes hard for the humans to know if they're actually discovering all the emergent capabilities. It's why they have red teams that try and break these things. They try and get them to do things that. That again, they weren't programmed to do to determine if they're safe enough to put out into society. So this is going to be an ongoing problem. And I don't know, this is just a plug for like a Netflix series, but. Have you watched the Three Body Problem yet, Mike?
- [63:32]
Mike Caput
Yeah.
- [63:32]
Paul Raitzer
Okay, I'm not going to, like, spoil this, but if you haven't watched the Three Body Problem, it's awesome. It's like an eight season one was eight episodes. I just watched it, like, right before the holidays. But there's an element of this where, like, again, I don't. I don't know how to say this without spoiling it, but, like, there's this line where they're like, are you lying to us? And like, the. The one side determines that humans can sometimes be deceptive. And so it. It decides it doesn't want to work with us anymore. And I started thinking about that when I was reading this of, like, this kind of feels like a Three Body Problem episode.
- [64:06]
Mike Caput
Right? Well, it's interesting how this kind of ties back to that topic about AGI policy. I think it was topic number three, where Yoshivi had said, like, the ball game is technical AI alignment, given the superior reasoning powers of some of these models. This is exactly what they're talking about. Like there's a possibility anthropic or ever could say, hey, we have the safest model in existence and release it. And it turns out they were lied to, correct?
- [64:34]
Paul Raitzer
Yeah, they thought they did. And I think there was actually an interview with Dario Amadeus. I was, I remember I was traveling somewhere when I was listening to it. So it must have been in like early December, my like one week marathon where I was bouncing around cities and, and he said like they, after they released Opus, it did things that it wasn't supposed to do when it was already out in the wild. And they, they luckily decided like, okay, it wasn't that harmful, but they thought it was safe, put it out and then it did stuff it wasn't supposed to do. So there's, there's part of me that thinks like some of these delays in like Gemini 2 and Grok 3 or whatever, Llama 4, I don't know that it's. The training runs didn't work. It could actually more alignment related. Like they, they can't get these things to behave the way they're supposed to behave. Like it wouldn't shock me if we eventually learned that.
- [65:29]
Mike Caput
Next up, Meta has made this big AI announcement that almost immediately generated even bigger backlash. So first, this article came out at the end of December in the Financial Times that basically profiled how Meta is envisioning this future where AI generated characters become as common on its platforms as human accounts. According to Connor Hayes, Meta's VP of product for Generative AI, he told the FT that these AI entities will have their own profiles. They'll have bios and profile pictures. They will actively generate and share content across Facebook and Instagram. So up until the point this article comes out, this is future looking. The FT reports that Meta seems to think this is where their platforms are going. But what really is interesting is that after this article came out, users resurfaced some AI accounts that already exist on Facebook. These are from a 2023 experiment when Meta was testing out launching completely AI avatars that users could interact with. And I believe we covered that when they did it at the time. Now, one of these AI avatars in particular drew a bunch of controversy because it was basically just made to look like a human that represented certain marginalized groups. So between this FT report and users reporting all these eerie and creepy conversations with these AI characters, this whole episode created some pretty serious backlash against Meta. So Paul, I guess my question for you is like, I know we're headed towards a World where we're going to be interacting with AI avatars and agents. But like something about flooding meta platforms with AI that we're going to socialize with, to me sounds downright unpleasant.
- [67:16]
Paul Raitzer
Yeah, I, I hate this. I don't. When I first read this, I was like, what? What? Like, the only thing I could come up with was that it was a, like a ploy to create fake engagement and usage data. Like, right. That they know that the more people are engaged with, the longer they stay on Facebook and their monthly active user goal. And like that it's just like a strategy to prop up a platform that maybe isn't as relevant as it used to be. I don't know. Like, I haven't, I haven't dug in. Like, I've said this where, like, I'm not a Facebook user. I have a Facebook account, I have an Instagram account. I don't ever go into them. And so like, I'm not the best person to comment on the current state of Facebook and things like that, but when I have gone in there in recent months, I haven't stayed very long. It's kind of like, okay, like, yeah, this is kind of the same thing it was before. So I don't know, like, I just don't get it as a business strategy. But again, I'm not an expert on Facebook or Instagram. Outside looking in, it just seems like an absurd strategy that creates this whole fake ecosystem.
- [68:27]
Mike Caput
Right.
- [68:28]
Paul Raitzer
And I pray that Microsoft does not ever consider doing anything like this to LinkedIn because there's enough fake accounts on LinkedIn already. I do not need AI avatars showing up, like, commenting on my posts.
- [68:42]
Mike Caput
It's real.
- [68:43]
Paul Raitzer
This is like glorified bots, basically. Basically like what's already happening on, on X and stuff. So, yeah, we're, like you said, we're gonna live in this AI avatar future. We're gonna be interacting with them when we know it or not. And I don't like it. Yeah, yeah, it doesn't matter if I like it or not, but I don't.
- [69:01]
Mike Caput
There's a whole other conversation around, especially depending on age group. Right. And like, you and I are just like, okay, I'm opting out of this. But if you're on a social platform as a teenager or a, A child, whatever, I mean, that gets really dark really quick, I think.
- [69:15]
Paul Raitzer
Well, in gaming in particular, like my, you know, game. My kids are. My daughter's 13 now and my son's 11. Like Minecraft, Roblox, like, you know, the idea of inter. Now My kids have can't communicate with, like, strangers. They only communicate with their friends in those platforms. But you could definitely imagine a scenario where kids are just interacting with AI avatars all the time and have no idea, you know, whether it is or isn't through, either through chat or actually in the game itself, where they're an AI avatar character.
- [69:41]
Mike Caput
Yeah. Another big topic kind of making the rounds in the last couple weeks is that a Chinese AI lab called DeepSeek has just released what appears to be one of the more powerful open AI models that we have seen so far. This new model is called Deep Seq V3. It was released under a permissive license that allows developers to download it and modify it for most uses, including commercial applications. And what's really notable here is its performance. According to Deep Seq, the model outperforms other downloadable AI models as well as closed models. For instance, in coding competitions on the CodeForces platform, DeepSeek V3 demonstrated superior performance compared to models like Llama 3.1405 B GPT4O and Alibaba's Quinn 2.5 72. Beep. This model is also massive. It was trained on a Data set of 14.8 trillion tokens, which is apparently roughly equivalent to 11 trillion words. It contains 671 billion parameters, making it 1.6 times larger than Meta's huge Llama 3.1405 B model. And the company claims it only spent $5.5 million on. On training, which is a lot for everyone listening, I'm sure, and us as well. But it's a fraction of what companies typically spend developing models of this size. So, Paul, when I read this, it kind of seems like another development that proves what you've been saying, like open source is going to make it impossible to stop the proliferation of really powerful AI. Like, what did you make of the Deep SEQ announcement there?
- [71:27]
Paul Raitzer
Yeah, so, couple of quick notes here. One, the reason they, again, you have to trust their own reporting on this like that. This is they did it for only 5.5 million. They did it with so many gpus. The reason that they have to do it this way is because the US limits the export of Nvidia chips to China, so they don't have access to the Same hundred thousand GPUs that Elon Musk just put in, you know, whatever, wherever he, whatever state he put that big data center in. They have to be more creative and innovative with how they train these models, and that necessitates more efficient models that cost less to train them because they have fewer GPUs to do this on. It does appear to move in the direction of this idea of commoditizing these models that even if O3 comes out and it is the most dominant model in the world, we won't have another like 2 year run where O3 is just far beyond everybody else. We probably have like a three to six month run until someone figures out how they did it and copies it and puts out an open source version of it. So it's this fast follower is going to become faster. Now this isn't without controversy and doubt they're very quickly within like 24 hours after this emerged started being these questions about did they just scrape all this stuff from GPT4? Like they basically just take, take ChatGPT's model and reproduce it and that's how they did it most efficiently. So we'll put a TechCrunch article in, but it said why Deepseek's new AI model thinks it's ChatGPT and they said quote posts on X and TechCrunch own tests show that Deepseek v3 identifies itself as ChatGPT, OpenAI's AI powered chatbot platform. Asked to elaborate, Deep seq v3 insists it is a version of OpenAI's GPT4 model released in 2023. Then it kind of goes on, it gives some details and it says more likely is that a lot of ChatGPT GPT4 data made its way into the Deep SEQ training set. That means the model can't be trusted to self identify for one. But what is more concerning is the possibility that deep seq, by uncritically absorbing and iterating on GPT4's outputs, could exacerbate some of the model's biases, flaws, we have no idea yet. So basically like what I have learned time and time again is anytime you see these supposed massive breakthroughs self reported, you have to kind of step back and just wait for verification from independent sources to say, yes, this is actually a breakthrough, this is actually a significant player and they didn't just steal everything. So yeah, worth note, certainly also a learning lesson to not overreact when we see these large claims from unverified sources.
- [74:23]
Mike Caput
Our next topic here concerns Microsoft. They just unveiled this ambitious vision for America's AI future in a new post titled the Golden Opportunity for American AI. So in this, Microsoft president Brad Smith argues that AI represents the biggest technological opportunity for the US since the advent of electricity. As a result, he says the company is laying out a three part strategy for American tech leadership. First, Microsoft is making a huge infrastructure commitment, planning to invest approximately $80 billion in AI enabled data centers in fiscal year 2020 25. More than half of that investment is targeted for the US. Second, they're pushing for a national AI skilling initiative. Microsoft alone plans to train 2.5 million Americans in AI skills during 2025, working through partnerships with community colleges, 4H clubs and other organizations. The goal is to make AI training accessible to Americans of all backgrounds. And they view AI literacy as essential as computer literacy has become today. Third, and possibly most strategically significant, at least at a geopolitical level, Microsoft is advocating for an aggressive AI export strategy to counter China's growing influence. They point to lessons learned from telecoms where Chinese companies like Huawei gained global market share through government subsidies. To prevent history from repeating, Microsoft is already investing more than $35 billion to build AI infrastructure across 14 countries. Paul, this is a pretty cool, bold statement and vision, particularly the education piece of this. Can you talk to me about what this means for AI in the us?
- [76:21]
Paul Raitzer
Yeah, so I definitely think this is essential reading for everybody. I would take the 10 minutes and read the entire thing. We could I think kind of go into an entire episode on this one for sure picking into these three key areas. I actually think this was intended for one reader. I think this was meant for Trump. I think this is a Microsoft manifesto of sorts for the incoming administration and they explicitly, or I guess not so explicitly say that multiple times. They give Trump credit for the 2019 pre gen AI executive orders around artificial intelligence and they focus on the Trump administration building on that executive order. They never mention Biden once. They don't mention any of the other things that have happened in AI for six years. So I'm fairly convinced this was meant to be their, their stake in the ground saying like this is what we should do together to the incoming administration. That being said, it's one of the most eloquently stated positions on the moment we find ourselves in in, in AI and the significance that it holds for America and for democracy. And so each of the areas they go into, the whole premise is that GPTs, generative pre trained transformers, the basis for generative AI are general purpose technologies GPTs. And so they do relate to like the steam engine and electricity and things like that that so they say GPTs boost innovation and productivity across the economy. IR working, electricity, machine tooling, computer chips, software, all rank among history's most impactful GB GPTs. This did remind me of a report, Mike, that you and I talked about last year that came out in, let's see, it was March of 23 and we revised in August 23rd that was called GPTs RGPTS, an early look at the Labor Market Impacts Potential of Large Language Models. In that report they said, we conclude that large language models such as generative pre trained transformers, exhibit traits of general purpose technologies, indicating that they could have considerable economic, social and policy implications. We have certainly over the last year and a half seen that, you know, play out in, in the article. A couple of the key points I'll just highlight. They said, as we look into the future, it's clear that artificial intelligence is poised to become a world changing GPT. AI promises to drive innovation, boost productivity in every sector of the economy. The United States is poised to stand at the forefront of this technology wave, especially if it doubles down on its strengths and effectively partners internationally. You alluded to the 80 billion in data centers. I wanted to call out why I think that matters and what that means. I actually listened to on Saturday morning. I was listening to a Jensen Huang no Priors podcast from November that I hadn't had a chance to listen to yet. And in that one he talks about data centers and how, you know, they used to be used to store things. We built these massive data centers for people to house their information in and then retrieve information from. They would be multi tenant, meaning you may have dozens or hundreds of companies all accessing the same data center, hosting their information on servers within these data centers. What Jensen said is they are now becoming single tenant means individual companies are building their own data centers and they are producing tokens which lead to intelligence. So they're actually becoming AI factories or intelligence factories. So rather than housing data, they are creating intelligence. And that's how Jensen sees Nvidia moving forward is they are, they are AI factory creators. They build these data centers that create intelligence for these different companies. You mentioned the AI literacy thing, which is music to our ears. We, I've shared this many times. Our, our mission, our north star changed last January 2024 to become accelerate AI literacy for all. It is everything we're focused on. It's all of my efforts around our evolved AI academy are dedicated to this idea of accelerating AI literacy. So I love to see this, I've preached on this show many times that we need this Apollo level mission to drive AI literacy. And so I love to see their positioning there. And they said that in the next quarter century we believe AI can help create the next billion AI enabled jobs, not just services but manufact, manufacturing, transportation, agriculture, government and every part of the economy. That's not going to happen without a focused effort. They did acknowledge that it will be disruptive, it's going to impact jobs, but that we have to upskill people. We have to drive this AI fluency as they called it, or you know, in our words, AI literacy. So I love to see it. I hope we see similar, you know, similar things, initiatives and messaging from the other major players in this space.
- [81:08]
Mike Caput
Next up, Simon Willison, who is a big voice in AI, just released this really great roundup of all the progress that we've seen in 2024 with large language models. So this post which we'll link to, is long but really well worth a read in my opinion. It is titled Things We Learned about LLMs in 2024 and it outlines some of the major breakthroughs that we saw in the last 12 months, including things like the so called GPT4 barrier was decisively broken. We now have 18 different organizations that have models that outperform OpenAI's breakthrough from 2023. The economics of AI have shifted dramatically. Model prices crashed throughout the year. Some services now charge less than 4 cents per million tokens. Voice and vision capabilities made major strides. Both Google and OpenAI introduced the ability to have natural conversations with AI while showing it what you're looking at through your phone's camera. And obviously models have now gained the ability to handle audio, video, text and images. And the accessibility of top tier models as well though took a hit late in the year. The best models were briefly available to everyone for free, but companies have begun restricting their most capable AI to premium subscribers now. One of the most interesting developments he cites came in the final months of 2020 24, which we have talked about at length. The rise of reasoning models that can spend extra compute to tackle harder problems. So Paul, this is, I thought, a pretty good rundown of the bigger picture trends here. It's not every single development, but what did you make of what he called out? Did he miss anything? Was there anything particularly worth double clicking into here?
- [82:55]
Paul Raitzer
No, I I thought they were just really good insights. It was a nice synopsis of kind of some of the key points in LLMs from last year and kind of alluded to some of the things to come. Like multimodal. We've talked a lot about vision. Voice audio is going to be huge this year. Reasoning is going to continue to be a major component. You're going to see advanced reasoning models coming from all the research labs, memory, actually remembering everything that's going on, remembering back more than just 10 minutes, remembering back through dozens of threads and conversations. And then the one thing he alluded to, the cost plummeting, which is an ongoing thing. What that allows is an explosion of AI apps, businesses, innovation, intelligence, because the cost to do these things falls. What's happened? And we know this is the play that. Demis Hassabis has said this in interviews with DeepMind. The current state of the art models 12 months from now will be open source and largely free. So you know, if you take the current like 1.5 advanced or whatever it is, or, or you know, early versions of Gemini 2, this time next year that model is going to be open source and you're going to be able to do whatever you want with it. That's the play for these labs is like you build a more state of the art model. And then, you know, last year's model, it's basically nothing to build on it. And so for enterprises, even if you do kind of like a fast follow of like, hey, we're not going to be using the most advanced models today, we're just going to build on the capabilities of the model from six months ago. You're still going to be in an incredible place. There's so much untapped value in these models. And so I think this is just a good reminder for people of kind of where we've been. Just in a 12 month period, all that has happened in the space.
- [84:34]
Mike Caput
All right, Paul, for our final topic here, I'm going to just do a mini rapid fire of a few funding and acquisition stories that we're seeing that, you know, we just want to wrap up into one quick topic here. So first up, Elon Musk's AI company XAI, they just announced a massive $6 billion CE Series C funding round. They are backed by some of the biggest names in tech, including Andreessen Horowitz, BlackRock, Sequoia Capital, Glock Grok 3, which is the company describes as their most powerful model yet, is currently in training as we alluded to, that is getting a bit delayed, but they say it will use the new funding to accelerate infrastructure development and launch new products. Next up basis is a startup building AI agents for account and they just announced a $34 million Series A funding round led by Khosla Ventures. They also have investors that include former GitHub CEO Nat Friedman, OpenAI board members Adam D'Angelo and Larry Summers, and Google's Jeff Dean. Now, what makes basis really interesting is their focus on accounting. They argue this field is reaching a breaking point. They say that accounting capacity has not kept pace with how complex the economy and accounting needs are becoming. And the industry faces a demographic crisis because accountants retire faster than new ones enter the field. So rather than trying to replace accountants, basis is trying to build AI agents that work as extensions of accounting teams. Next up, Perplexity actually just closed a massive $500 million funding round. They are valued at $9 billion, now 10. And fresh off this funding round, Perplexity announced the acquisition of Carbon, which is a startup that builds technology to connect external data sources with large language models. This will allow Perplexity to have users connect their existing apps and documents, things like Notion and Google Docs directly into their search experience. And last but not least, Grammarly, the popular AI writing assistant company, they apparently have 40 million daily active users have announced that they are acquiring Coda, which is a maker of AI productivity tools. Now it's not just about the acquisition, it's also including a leadership change. Grammarly CEO Rahul Roy Choudhury will step down and make way for Coda's co founder and CEO Shishir Mehrotra to take the helm. By integrating Coda's flexibility and intelligence into their platform, they're actually aiming to create what they call an AI productivity platform for apps and agents. So expanding quite a bit beyond their initial writing focus. So Paul left him dig into that.
- [87:24]
Paul Raitzer
One a little bit. That's a serious change in strategy for Grammarly.
- [87:27]
Mike Caput
It really is. Interesting. Yeah, well you've noted a few times, right, that some of the OpenAI features and model features are kind of coming directly some of Grammarly's writing capabilities.
- [87:36]
Paul Raitzer
That's interesting.
- [87:37]
Mike Caput
All right, Paul, that is a wrap for this week. I think we're kind of caught up though. Not really because I'd encourage people, people to go check out our newsletter because literally there were dozens of stories that did not make the docket today. We could have done a five hour episode.
- [87:52]
Paul Raitzer
That's. He's not exaggerating. There was literally like 40 and it. He left me a note Sunday night. He's like, I think we might need to cut this. And I was like, dude, we gotta cut this down to like 15 and it's gonna be like 4.
- [88:02]
Mike Caput
We literally cut this in half and we've gone almost 90 minutes. So yeah, we would have been in like Lex Friedman length podcast I think if we had kept them. But the good news is we, we include a lot of the really important stories that we didn't get to cover in our newsletter@marketingaiinstitute.com newsletter. Go check that out if you have not already. And last but not least, I say this every episode, but please, if you have not already, leave us a review on your podcast platform of choice. Paul thanks so much for everything.
- [88:33]
Paul Raitzer
We are off and running in the New year. Thanks everybody for being back with us us and we are scheduled for our weekly sessions moving forward without any breaks that I know of yet. So we will be back every week. Thanks for being patient as we took a couple weeks to spend with our family and recharge over the holidays. So look forward to all the year has in store for us with AI. It's going to be a fascinating year, so thanks again everyone. Thanks for listening to the AI show. Visit marketing AI inspiration 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.