← The Artificial Intelligence Show
The Artificial Intelligence Show
#120: OpenAI-Microsoft Drama, Major Study on AI Job Impact, Sequoia’s New GenAI Market Analysis, NotebookLM Updates & Adobe Max 2024
01:19:24
Loading summary
Transcript66 lines
- [00:00]
Paul Raitzer
There is massive disruption coming. No one seems to be talking about it. Like, economists don't seem to be talking about it. Industry leaders don't seem to be talking about it. Government leaders don't seem to be talking about it because they don't seem to really realize what's about to happen. And we don't have answers to what happens to jobs in all these different industries.
- [00:21]
Mike Kaput
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.
- [00:57]
Paul Raitzer
Welcome to episode 120 of the Artificial Intelligence Show. I'm your host Paul Raetzer, along with my co host Mike Kaput. We are coming to you on well, it's Monday, October 21, 10:00am in Cleveland. A beautiful fall day in Cleveland. I must say I've got my pumpkin coffee with me. I am like, I am all falled out this morning. I love it. I was thinking about we're, we're just under a year away from Macon. 2025, Micah, like, because next year's event is October, I'm gonna get this wrong because I'm saying it off top. I had like 14th to the 16th. Oh my gosh. So it's literally like one year away from Macon. But it's fun because you see like the beautiful fall colors and it's just such an amazing time of year in Cleveland. So I'm excited to have everybody back here in Cleveland next year at this time. In the meantime, I'm just going to enjoy my pumpkin coffee. I have like one month out of the year where I like, I have to have a pumpkin coffee every day. All right, we've got some big items to talk about. Last week wasn't like another crazy week of launches and new models and stuff, but there's a few articles and reports that came out that Mike and I are going to really drill into today that I think provide some really great like macro level perspective about what's going on. So I'm excited to talk about those first and say this episode is brought to us by Rasa IO talk about a common challenge we all face, making our email newsletters truly engaging. Well, Rasa is changing the Email Newsletter Landscape Imagine each of your subscribers receiving a newsletter tailored just for them. Sounds impossible, but Rasa IO makes it possible with their AI powered platform that makes personalization easy. We've known the team at Rasa for about six years now. They've been a longtime supporter of our Marketing AI Institute. And if you are running or planning to launch a newsletter for your business, you should definitely check out Rasa IO. Head over to Rasa IO. MaiI and use the code 5 number 5 M A I I and you'll get a 5% discount. Give it a try. Your subscribers and your engagement rates will thank you and the episode is also brought to us by our second annual AI for Agency Summit. This is a virtual conference that's taking place noon to 5pm Eastern time on Wednesday, November 20th. So that's the live version and then you can also purchase on demand. So if you can't make it on November 20th from 12 to 5pm Eastern, you can always watch the replay so the AI for agency summit is designed for marketing agency practitioners and leaders who are ready to reinvent what's possible in their business and embrace smarter technologies to accelerate transformation and value creation. During the event you'll join hundreds of other forward thinking agency professionals to consider ways to recruit AI savvy talent and upskill your team. Explore AI tools can boost creativity, productivity and operations. Hear insider stories I think we have six different agency leaders presenting on their case study within their agency. So it's going to be all about the agencies that are actually doing this and what they're learning and being able to share that with you. Understand how Ampics impacts your pricing models and service offerings. That's a topic we're going to talk a little bit about in today's episode. A lot of things happening in that space and then connect with like minded agency professionals and leaders who are in the midst of their own AI transformation journeys. It's all presented by a group of world class agency leaders and experts. You can get your tickets@ai4agencies.com and click on Register now. So when you do use the code pod100 that'll get $100 off of your ticket. So that's AI for agencies. That's f o r4agencies.com and code pod100 for $100 off. All right Mike, there's some brewing drama. We've thought that there was some stuff going on with OpenAI and Microsoft. There was some some rumors but the New York Times sort of blew some stuff out with a bunch of inside sources and yeah There's a lot going on in that Microsoft OpenAI relationship. The former best. What was it? Best relate bro. Bromance in tech was what it was called. Maybe not so much anymore. So let's, let's get into that.
- [05:13]
Demis Hassabis
Yeah. So first up, as you mentioned, OpenAI CEO Sam Altman used to call the partnership between OpenAI and Microsoft, quote, the best bromance in tech. But according to some new reporting by the New York Times, that partnership has started to sour a bit. So it's. If you recall, this began with Microsoft investing 13 billion into OpenAI, basically giving them a bunch of essential funding and importantly computing power. However, after OpenAI's board briefly ousted Altman last November, apparently Microsoft started consider reconsidering its approach to further investments. According to the times. Now, OpenAI has a lot of financial challenges. They have expected losses of 5 billion this year, and they've been seeking additional funds and computing resources from Microsoft specifically. However, apparently they have been hesitant to increase that commitment, which has led OpenAI to explore other options to fund and support their growth. In response, OpenAI has also been attempting to renegotiate its deal with Microsoft, aiming to secure more computing power and reduce expenses. They also recently, which we reported on, broadened their investor base, closing a $6.6 billion funding round recently. Now Microsoft has also started to hedge its bets. It sounds like in March the company invested at least 650 million in what you might call an aqua hire to get most of the staff from Inflection, which was an open AI competitor. Mustafa Suleiman, Inflection's former CEO, now oversees a new Microsoft group working on AI that could potentially replace what OpenAI offers. And that's also been causing some friction, it sounds like, based on this reporting, OpenAI executives and employees expressed frustration over Suleiman being at Microsoft and concerns about sharing of technology between the two companies. So, Paul, there's never a dull moment at OpenAI. Like, can you maybe break down some of the complexities for us here? Like, Microsoft has this huge vested interest in OpenAI success, at least it seems like. But they also appear to be creating obstacles to the company's success, kind of helping them grow, but not too crazy. OpenAI increasingly seems to be in a competitive, not a collaborative position here. Like, what's going on?
- [07:48]
Paul Raitzer
Yeah, it's wild. We've talked about this relationship a number of times on episodes over the last year, year and a half, and I recall in the last couple months just the first time that Microsoft identified OpenAI as a competitor officially. I think within their earnings reports. And we've obviously when inflection was brought kind of through that aqua hire, that seemed odd. You certainly could understand how that could ruffle some feathers on OpenAI side by putting Mustafa in the position he was in, especially as sounds like made the main point of contact for Microsoft and OpenAI. So I think this article again I would highly recommend people read it. It's by Cade Matz, who's the author of Genius Makers and heavily sourced within all these companies. Mike Isaac and Aaron Griffith at New York Times. There's a, there's quite a few details in here. I had not heard yet that usually every time Kade is involved in doing an article on AI, there's something comes out that you previously weren't aware of. So I'll zip through a few of the things that I did a double take on or kind of made specific notes on as I was going through it. So I had not previously heard that OpenAI was trying to raise billions from Microsoft prior to November 2023 when Sam was temporarily ousted. If that had been reported, I had not seen it anywhere. So the article said Nadella was initially willing to keep the cash spigot flowing. But after OpenAI's board of directors briefly ousted Mr. Altman last November, Nadella and Microsoft reconsidered, according to four people familiar with the talks. And then you know, as you mentioned, that he was shocked and concerned by the firing. They OpenAI has been trying to renegotiate the deal. This makes sense in the context of we know they're trying to shift to this for profit entity, but not only do they have the nonprofit current structure that is restrictive, their agreement with Microsoft is highly restrictive. And for, for the for profit change to occur and for them to eventually ipo, I would imagine there needs to be a reimagination of the OpenAI Microsoft relationship. And Microsoft may be in no hurry to do that because if I'm not mistaken, they have 49% ownership of the for profit entity. I think we reported on a recent episode there was like a certain percentage of their revenue. Like I think it was 20% of all revenue or something goes back to Microsoft. And I believe there was a profit cap too where like the first hundred million or a hundred billion or some crazy number was Microsoft's money too. Like it's, I mean Microsoft's got a sweet deal here plus they get inside access to the technology and it's just like it's not ideal for where OpenAI wants to go. It was like fundamental to get them where they are, but it makes sense that they would be trying to negotiate this. So the article said over the past year, the AI company repeatedly tried to renegotiate a lower cost on compute and allow it to buy compute from other companies because they didn't feel like Microsoft was giving them enough. It did say Microsoft agreed to an exception in the contract recently that allowed OpenAI to sign a roughly $10 billion computing deal with Oracle. I had not seen that. Reporting that one, I assume has been out because that's a pretty big deal for Oracle. So I would guess that information is out there somewhere. I said, in recent weeks, OpenAI, Microsoft renegotiated change to a future contract that reduces how much Microsoft will charge the company for the computing power. You mentioned that there's these concerns internally about, you know, Mustafa's management style and maybe what Microsoft is doing. Said and some have complained that if another company beats this is an interesting one that leads to an issue later on. Some have complained that if another company beat it to the creation of AI that matches human the human brain. So basically AGI Microsoft will be to blame because it hasn't given OpenAI the computing power it needs. So that's kind of interesting. The, you know, some of the issues around Microsoft, like, it almost seems like what Microsoft is doing, and this may not be correct, but it seems like OpenAI thinks they're slow playing them on the compute access they're giving them so they can build their own internal competing tools and not be reliant on OpenAI anymore. So you bring in Mustafa. Mustafa now owns the relationship with OpenAI, but he's also in charge of building OpenAI or Microsoft's own capabilities with Kevin Scott and the team there. And so it's like, let's just slow play this. Over the next 12 months, we'll build up an equivalent model to what OpenAI has. We have all the inside information on what they're building next. So, like we can get there where they get there and then we're not relying on them anymore, which I asked. Like, there was two episodes I go where I said, sure. Seems like Microsoft just relies on OpenAI for all this innovation. I'm wrong. Someone from Microsoft, like, call us like. But it does appear that this is basically validating what I was saying, that Microsoft's aware of the two relying on OpenAI, so they're trying to build around them. So imagine Microsoft sitting there a year from now, maybe they have a competing model, they're not relying on OpenAI's tech anymore, and they still own 49% of the company. So Microsoft seems to have all this leverage. But then the most fascinating part of the article brings back a unique nuance of the relationship that some people maybe either forgot or didn't know, which is the contract is void if OpenAI achieves AGI. So in the original deal with Microsoft, with all the, you know, the access to the technology, Everything else, if OpenAI achieved AGI, Microsoft doesn't get access to that technology. And that was a clause they put in because they wanted to make sure and supposedly that it would be used responsibly and they didn't want a third party having access to, like this AGI technology. And so that's what the article said. The clause was meant to ensure that a company like Microsoft did not misuse this machine of the Future. But today, OpenAI executives see it as a path to a better contract. According to a personal familiar with negotiations, under the terms of the contract, the OpenAI board would decide when AGI has arrived. That is the most fascinating part of all of this. So imagine Microsoft has all this leverage, like, we're not going to give you more compute, we're going to charge you exorbitant fees for the computer, we're going to build our own tech. And OpenAI seems kind of helpless in this scenario, relying on Microsoft and their contract, but they have the trump card, which is, oh, yeah, by the way, GPT5 is AGI, our board has decided, so you don't get access to it. So both sides are going to try and play leverage. And if the contract truly is that, you know, cut and dry, that if AGI is achieved, which is decided by the board of OpenAI, which is what we've heard over and over again, then all the board of OpenAI has to do is say GPT5 is AGI and then we're done. Microsoft doesn't get it. And so that's the game we're playing. Which makes me go back to last year, Sam Altman did Lex Friedman podcast, or maybe it was early this year, and I remember Lex asking him about GPT4 and did he think it was AGI? And Sam's like, I don't know, do you think it is? And knowing how Sam approaches things, there's this part of me that now looks back and thinks that was a totally intentional thing to like a shot over at Microsoft saying, maybe it is. Maybe we're, maybe we're already there in this, this contract situation we're having is, you know, we're willing to play that Leverage if we need to. So that led me to like, oh, maybe we're going to get to AGI in society just because OpenAI wants out of their contract with Microsoft.
- [15:34]
Demis Hassabis
Yeah, that struck me too. Is like, we know and we talk about all the time that the leaders at these companies do appear to genuinely believe they are on the path to building AGI, but now there's a direct incentive for them to also contractually say what they have is fascinating.
- [15:54]
Paul Raitzer
Yeah, so I definitely read the article. There's a lot more that I, you know, I'm not touching on here, but there's a lot of really fascinating elements to this. And they got into like, Apple's relationship with them a little bit. And like, the one I thought was fascinating is. So back in 2022, as OpenAI was developing the technologies that would drive ChatGPT, Altman and Kevin Scott, the CTO of Microsoft, met with executives at Apple to explore those three companies working together. This is before ChatGPT. I had no idea they were talking with Apple back then and that Microsoft would be involved in doing a deal with that. I mean, it's just wild. Like, I would love to like click into that story angle because they just kind of left that paragraph hanging there. It's like, whoa, hold on a second. This feels like a whole nother chapter of the book.
- [16:39]
Demis Hassabis
I'm selfishly hoping that this is a prelude to Cade's next book.
- [16:44]
Paul Raitzer
Me too.
- [16:44]
Demis Hassabis
All this, I've joked online with them.
- [16:47]
Paul Raitzer
A couple times, like, when is Genius Makers 2 coming out? Yeah, right.
- [16:51]
Demis Hassabis
So to kind of wrap this up, like, their OpenAI is obviously facing this issue with Microsoft, but also, like, in the past weeks, like we've talked about Ilya starting his own venture, OpenAI tries to restrict its own investors from funding that company, along with a few others. This week we got a report that former CTO Mira Muradi is now raising purportedly a ton of money for her new startup. As $100 million quoted, the startup aims to build AI products based on proprietary models. According to Reuters, like we've talked about, there's only a handful of companies out here that can actually, like, build these frontier models. But like, how worried does OpenAI also need to be about, like, its ex employees forming companies to compete with them?
- [17:35]
Paul Raitzer
It seems like probably very, I don't know, I mean, like, their secrets are just going out the door every week with all these leaders. I mean, there was a visual I saw in one article where it was like, I don't know, like 14, 15 founding members and actual like co founders of OpenAI. And the only people left are Greg and Sam and I think there's one other person on that list, but all, all of them are gone and seemingly doing their own things. So yeah, it's, it's, it's a fascinating story. I mean, I don't know. That's why like every week we end up talking about OpenAI because it's just always new angles to the story and it is so instrumental to where this all goes.
- [18:18]
Demis Hassabis
All right, so our next big topic this week we just got a new report from the Brookings Institution, which is a well known nonprofit. And in this report they take a data driven approach to analyzing the potential impact of generative AI on the American workforce. And some of these findings are worth paying a little bit of attention to because the study highlights that generative AI could significantly disrupt a wide range of jobs. With over 30% of workers potentially seeing at least half of their tasks affected by AI, 85% of workers could see at least 10% of their work tasks impacted. And to assess this, Brookings quote utilized estimates shared by OpenAI relating the predicted chat GPT4 exposure level to of thousands of the tasks that make up the hundreds of occupations defined by the Department of Labor's O NET database. Now, unlike previous waves of automation that impacted routine blue collar work, generative AI, they say is poised to affect cognitive and non routine tasks, which means it is getting into middle and higher paid professions. They highlight fields such as stem, business finance, law, office administration. All of these are the most exposed to potential generative AI disruption. And they also make the point that while there are opportunities and risks for workers with generative AI, society is currently underprepared to address these challenges. Now Paul, we don't, I don't think we directly know anyone at Brookings Institution, but this is sounding very familiar. So a note on the methodology here. Brookings used estimates shared by OpenAI relating the predicted exposure level of thousands of the tasks that make up occupations using the O NET database. That sounds a lot like the methodology that you use to create and part jobs GPT and campaigns. GPT, like how seriously are you taking this study? Can you talk a little bit about this methodology? Why it might make sense?
- [20:28]
Paul Raitzer
Yeah, I think it's wonderful. I mean this is what is needed. So this is what I've been, you know, kind of pleading for is more high level involvement from institutions that can move the needle, have the ears of key government leaders and industry leaders. So I am very happy to see this effort. I like the questions they're asking, so you know, the report starts off with how do we ensure workers can proactively shape generative AI's design and deployment? What will it take to make sure worker benefits meaningly from the gains? And what guardrails are needed for workers to avoid harms as much as possible. I like that they're taking a multi year approach. They said this is just a start of the initiative. Brookings Metro has embarked on a new multi year effort focused on raising awareness and shaping societal responses. Wonderful. They said they're drawing on insights from a workshop where they convened 30 experts from policy, business, innovation, investment, labor, academic, think tank, research, civil society and philanthropy. Brilliant. More of this, like this is all phenomenal. What they did, from best I can tell is they actually took the GPTs. Our GPTs paper from August 2023, which is where the exposure methodology originated from. So this is an OpenAI paper where Mike and I have talked about this on the show before. We talked about it in the Jobs GPT episode. I think it's 110, which is when GPT4 came out. So GPT4 comes out in March 2023. OpenAI, I think in partnership with Microsoft, or it might have been mostly OpenAI created a paper, a research paper called GPTs Are GPTs generative Pre trained Transformers? The foundation of language models are general purpose technologies was the point. That these AIs are capable of doing many cognitive tasks is kind of the premise of a general purpose technology. And so it seems like what Brookings has done is zoomed into that research. So this doesn't seem like new research on their part that's looking at like, you know, 30% of all workers could see at least 50% of their occupation task. I think they just further analyzed OpenAI's data, which means the data is a year and a half old that they're referencing. So it's probably not even that relevant per se. But if that's what they did, which is what it seems to be, then I have the same, I don't know, concerns isn't even the right word. Like the same reason I built Jobs GPT and the exposure key that I developed, which goes beyond current, and it looks at future capabilities of advanced reasoning and persuasion and digital action and computer vision, you know, having agents in the world. So when we built Jobs GPT, the idea was to project future impact on jobs, to look at the exposure of the models we know that are coming. And so I think that there's a lot to build on with what Brookings is doing if we can assist them in any way, like, please, someone reach out, we're happy to provide any data or help wherever we can, because this is the kind of initiative I think we need more of, is just these higher profile institutions really aggressively pursuing answers to very difficult questions. And I think at the end of the day, like if you go Back to episode 105, Mike, you and I were talking about this, the Carl Schulman interview, one of the co founders of OpenAI, the early people there, he did that 80,000 hours interview and in that one he talked about the economic impact. And so we kind of went pretty deep into there. And I went back and I pulled the excerpt where I said, like, what Carl was talking about in that episode and what I think Brookings is hopefully trying to get at here is there is massive disruption coming. No one seems to be talking about it. Like, economists don't seem to be talking about it. Industry leaders don't seem to be talking about it. Government leaders don't seem to be talking about it because they don't seem to really realize what's about to happen. And that even if we don't get to AGI and we just get smarter models that follow these scaling laws, that's sufficiently disruptive and we don't have answers to what happens to jobs in all these different industries. So, yeah, I mean, at the end of the day, I'm just really happy to see this. And I hope they like really aggressively pursue this plan that they have. They talk about defining and supporting good employer practices, enhancing worker voice and power, and developing public, public policy solutions as sort of like their three main priority areas going forward. But they address that. There's just so many unknowns and so many questions that need to be explored. And that's kind of where I'm at, is like, I don't have the answers, but it sure looks, when you look out a year or two from now, like we're going to see some pretty significant disruption across a lot of different industries that no one is preparing for.
- [25:24]
Demis Hassabis
And kind of like you alluded to in the Schulman episode, you expressed, I would say, correct me if I'm wrong, like, bafflement that more people aren't talking about this. And it made you kind of feel like when you were kind of very early to AI and you were like, I can see where the impact is going to be. I'm surprised more people aren't talking about it because a thing that they mention in here is that they said this data does not even attempt to project future capability enhancements from next generation AI models likely to be released. So they're saying like, these are the numbers we're citing to of exposure of just today.
- [25:59]
Paul Raitzer
Right?
- [25:59]
Demis Hassabis
And we've talked about we're on the cusp of GPT5 already.
- [26:03]
Paul Raitzer
Right. And that's. I like, I get why they wouldn't do that yet, but I also don't see why you couldn't. Like, I mean that's. I built the Jobs GPT exposure key by just studying the space and looking at what all these research labs are telling us. It's pretty obvious what they're all working on. Like, you don't have to search really that hard, right, to to reasonably project what the next versions of these models are going to be able to do. And so why can't we model that to specific industries? And that's what needs to happen is take the exposure key I've created, like, go to SmartRx AI and go to the Jobs GPT page and like, it's the exposure keys right there. Think about your industry. If you're an accountant, customer service, lawyer, doctor, I don't care what you are, entrepreneur, take that exposure key and think about your own career paths and industries. Like, this isn't that hard to do. But for whatever reason, we haven't had economists and institutions like Brookings do it yet at scale. And that's, that's what has to happen is like, we can only think kind of horizontally and we think obviously about marketers through our Marketing Institute brand. But like, I spent Saturday doing a talk on AI and entrepreneurship in K through 12 education. And so for that, that afternoon my brain is locked into like, what does this mean to educators? How do we prepare students at. You know, my daughter's in seventh grade, so I was actually talking at like a, this morning I was advising the 7th grade entrepreneurship initiative at the school. And so like, sometimes I'll zero into a specific vertical, but most of the time we're just zoomed out and you and I, Mike, are talking like macro and then hoping people will take that inspiration and go run into their domain. And that's where I think this needs to go is in Brookings isn't going to answer all that either. So if you're a listener to this podcast, like, take this knowledge and you go figure out what it means to your industry, your business over the next one to two years because nobody else is doing it right now.
- [27:58]
Demis Hassabis
So just very briefly, it definitely touches on, you know, in our last episode when we talked about one of the main topics was Dario Amade's AI manifesto, how this would AI would impact work in society. And you kind of ended with this really interesting call to action. I thought we were like, look, we don't have the funding to do this massive whatever research project or deep study, but someone like the government or Google or Microsoft, somebody needs to make this their thing and do deep studies about the future of the economy. You had said, and we noted that we kind of have listeners of some of these companies. I just want to push like a little further as we wrap this up. Like if you did have the funding or partners with that kind, the kind of funding it might take to actually take this seriously. Like, would you both go broad and deep like you just described? Like, what would be kind of that next big step that needs to happen?
- [28:49]
Paul Raitzer
Yeah, I, I think so. I mean, you and I talk so much about, you know, even when I'm thinking about our own business model and how much, you know, we try and tackle. Yeah, we always come back to empowering other people. Like, I think that more than anything what I have focused our energy on is distribution of knowledge and information and tools that empower other people to go figure it out in their own domain. Because this stuff's going to move too fast for one council or one government initiative to achieve everything. And so I think by distributing the knowledge into people who can then take it and apply it to their domain. That's critical. But I do believe that there needs to be like an Apollo level mission on literacy and reskilling and upskilling workforce, not just on the building of the technology. And it seems to me like most of the focus at the big frontier model companies and at the government level is all about like how do we build the smarter technology and maintain a competitive advantage over other countries. Not what does this actually mean to our society and our workforce and our educational systems and how do we, in a very quick time period prepare for that? And I think, I don't know if I'm gonna use that analogy on the last episode, but for better or for worse, the pandemic is the closest thing we have to how quickly we can mobilize change in business and society and educational systems. And I'm not saying we need like a 30 day plan, but it sure wouldn't hurt to have like a next 12 month plan from somewhere on high that's pushing for dramatic action. You know, again, I don't, I don't know exactly what the change looks like, but this isn't a 5 to 10 year play like if Dario and Demis and Sam and all these other people are right, we don't have five to 10 years. Like we got two to five maybe before. Like just completely disruption and transformation. So yeah, I just, if you're in a position of authority or if you're in a position of power at one of these bigger entities, you know, non profit government, frontier model company, we gotta do something way faster on the societal, educational, workforce side of things.
- [31:06]
Demis Hassabis
So in our third big topic this week, the famous VC firm Sequoia Capital just released an updated market analysis of the generative AI landscape. Now they titled this generative AI's Act 01 in reference to OpenAI's new 01 model. And in it Sequoia partners Sonia Huang and Pat Grady break down how the generative AI ecosystem is evolving. They write, quote, two years into the generative AI revolution, research is progressing the field from quote thinking fast rapid fire pre trained responses to quote thinking slow reasoning at inference time. This evolution is unlocking a new cohort of agentic applications. So they find things like the foundation layer of generative AI is actually stabilizing around some of these major players like Microsoft/OpenAI and Google/DeepMind. Focus they they believe is now shifting to the development of a reasoning layer. The latest most significant advancement being of course OpenAI's O1 model, which enables this more deliberate system two thinking we've referenced on multiple episodes as opposed to say a quick pattern matching that earlier models have done. And they also think we've gotten to a new scaling law that's emerging. And they claim that the more inference time compute that's given to a model, the better it can reason. So we're shifting a focus from massive pre training to scalable inference clouds. So basically the authors argue this is going to have some major, major effects on business as usual. They even anticipate what they call a future AlphaGo moment in generative AI where these systems demonstrate truly novel and superhuman capabilities. So Paul, I know you found a lot to like in this analysis and obviously Sequoia's like investing on the bleeding edge of AI. So they're absolutely worth paying attention to. Like what was really worth a look within this report.
- [33:10]
Paul Raitzer
I think for our regular listeners it's a really nice concise summarization of what we've been talking about for the last like you know, eight to 12 episodes. So as we knew that Strawberry was coming, so we were talking about strawberry, you know, 10 episodes ago and, and then we eventually got the O1 model. We've talked since January about the System 1, System 2 thinking that actually came. Andre Karpathy had that intro to LLMs video in January of this year on YouTube that we featured and talked about. That was one of his big things is giving the machine time to think. We talked about with the Noam Brown episode and his efforts in poker play and diplomacy that the more time you give the computer to think, the more, the smarter it seems to get or the better it seems to be at solving things. So there wasn't anything like groundbreaking and new that came out of this. If you're a regular listener. If you're not, it's a really good five to seven minute read where you can kind of catch up on some of the key things that are going on that you highlighted in your overview there. Mike There's a few things that I thought were interesting just from Sequoia's perspective that I wanted to kind of click into for a minute. One is the foundation models. You know, you mentioned this idea that they thought a couple years ago that that there would end up being like a single model company, which I assume they assumed was OpenAI at the time or maybe Google that basically just made everybody else irrelevant and everyone else would kind of play for the, you know, the scraps. That doesn't appear to be true. When we talked about this, I don't know, maybe five, six, six episodes ago, where I said the key to me is, is GPT5 a breakthrough or is it a continuation? And just like more compute, more training data went into it, more reinforcement learning. Because it seems like everyone has sort of caught up to OpenAI for the most part, like with Metta and Google X AI is, you know, on their heels to a degree where it's hard to know how they're going to truly like leap ahead and spend another year and a half with a better model than everybody else. So it, Sequoia was sort of saying this that it just seems like we're now like every three to six months the other model companies catch up to whoever had the most recent, most powerful model. The inference time compute thing came out of the 01 paper from OpenAI and again Noam Brown, who was at Meta and is now at OpenAI. This is like his main thing. It's why he went to OpenAI. It's to work on this reasoning thing, which is the basic premise is the System one, System two thinking. So System one is what's the capital of Ohio? And it's like, okay, it's Columbus. Like real quick, it is Columbus Right. Like I'm thinking of myself. Okay. So I paused as well.
- [35:59]
Demis Hassabis
Yeah, is it?
- [36:00]
Paul Raitzer
That's system one. Like it's just a fact based thing. There's no real thought process other than memorization that goes into this retrieval and memorization. That's it. That's system one thinking. System two thinking is, you know, why did they decide to make Columbus the capital of Ohio? Well now I gotta stop and actually kind of like think about this and I gotta do some research and I gotta go through some steps like that's a more System 2 style thinking. And so that's the basic premise is what they're saying is when we give the machine time to think, it seems to be able to do much more complex things in like math and biology and chemistry and all these things. Business strategy, you know, when you start to bring this back to our world. And so that's what they're looking at. The, the one thing that they didn't really say like explicitly within this, that I sort of like was thinking about is the more and more I think about this, the more I believe that what ends up happening is Google Meta OpenAI Microsoft probably gets back in the game with their own frontier models. Nvidia. They're all going to spend their anthropic, they're going to spend their billions of dollars building these massively powerful approaching AGI, eventually getting to AGI models that are generally capable of almost all cognitive tasks. But those models are always going to cost the most money to run and to use. And the reality is many of the use cases in business, like helping us write our emails or brainstorming ideas, building a marketing strategy, those don't require a $10 billion frontier model to do. We could do that with like GPT4 level stuff like that might be good enough.
- [37:44]
Demis Hassabis
Right.
- [37:44]
Paul Raitzer
So I'm envisioning this world where there's four or five dominant frontier models that are all approaching or at AGI a year or two or three from now. But we don't use those models daily in our lives. What's happening is that model almost functions as like the project manager, like the overseer of all the other models and agents that live within it. And so when we go into chat GPT instead of having to pick from one of the four models, which makes no sense from a user experience, how am I supposed to know which model to pick? I just put in my prompt and then the most powerful model figures out which model is actually best to solve that calls on that model to do its thing and maybe there's like a symphony of dozens of these things. And unknown to us, we're just putting our prompt in. But behind the scenes you start to get this symphony of agents and models working together to do the thing for us. And I'm like, I'm almost like 99% sure that's what ends up happening. Because us picking models for ourselves makes no sense from user experience. Only for developers does that make sense to the average user. A marketer, an entrepreneur, a lawyer. Like, how do they know which model to pick? 01 preview 01 mini 4.040 advanced Gemini 1.5 like, who knows, right? So I think that that happens. And then the thing they called out is like, and you and I have talked about this, this idea of like these AI software companies that are just wrappers for the model, meaning if I work directly with OpenAI, use ChatGPT, I'm using the model directly through OpenAI. If I go use, I don't know, like Khanmigo, I'll say, because I was using that example. If I'm using Khanmigo, that's a wrapper for OpenAI's models specifically tuned for education. They don't build their own models at Khan Academy, they're building on top of it. So they're a rapper, software rapper. Those rappers kind of got, no pun intended, a bad rap. Like a year ago, we all kind of assumed that they would just be eaten up. What did, what did Sam say? They would steamroll them or something, I think, like if they were useless. So we assumed these wrappers that in the first, you know, six months after ChatGPT were raising 10 million, 20 million, whatever, that they were just going to become worthless. What Sequoia is saying is. No, that these wrappers are actually critical because you re it requires domain expertise to build like a legal assistant or a customer service assistant or, you know, a marketing agency assistant, and that that's actually where the knowledge or the value will accrue in the venture capital world is at the wrapper layer for people that build these domain specific things. So I don't know, like, again, nothing in this article to me was like, oh my God, I didn't, I didn't know that because we're living it every day. Yeah. But it definitely summarized it in a really nice way and it allowed me to sort of step back and, and have some more macro level thoughts that I maybe haven't like given my brain time to think about for a few weeks or a few months. So again, a really good read. It's a Pretty approachable read. It's not highly technical. If you're struggling with it, drop it into NotebookLM and like, you know, ask it to explain it to you at a seventh grade level. Like, and I'm not joking about that, like, that it's a truly great function of NotebookLM to do it that way. So, yeah, again, all three of the things we've covered, these main topics, all really good reads and really good kind of macro level stuff for people to understand.
- [41:08]
Demis Hassabis
Yeah, that's why I like the big picture, like zooming out a bit. Because even if you are a listener, it can be a regular listener and you're following this stuff. It can be really hard sometimes to frame. Like, you know, this is a very different landscape than it was six months ago or a year ago. Like, that doesn't seem like that long of a time. But really we're starting to see how this is evolving and just you can't use the old mental model of like, oh, well, it's like just a prediction machine or it's just like making predictions. Or it's like, ah, the prompts are okay, like, no, go read this article. You'll see quickly how things are moving.
- [41:41]
Paul Raitzer
Yeah. And it's helpful just to like, I mean, it's a fire hose. Like, what we're all trying to, like, process right now is just a fire hose of information and names and companies and model numbers and it's a lot. And so whenever you can get these articles that just do a nice job of giving us, like, hey, here's the three to five themes to be thinking about right now. Like, that's what we try and do with our podcast. But we love it as just, you know, people seeking knowledge too. Like, I always love when I can just step back. Okay, what are the three to five things they're saying? And does it jive with what we're thinking, or is this actually like a different perspective that we should be considering when we're talking about on our show? So, yeah, good. You know, appreciate Sequoia and Brookings and New York Times for putting out some good stuff.
- [42:23]
Demis Hassabis
All right, let's dive into this week's rapid fires. First up, we're going to talk some more about Google's Notebook lm. This is the company's AI powered research assistant. This thing is just quickly becoming a darling of the AI community. Google actually says over 80,000 organizations are already using it and it's getting even more powerful updates. We've had a slew of recent updates. And first off, NotebookLM is actually getting rid of the experimental label, it is now ready for primetime. Millions of users appear to already be using the AI powered notebook to engage with intricate topics. The star of this that kind of put Notebook LM really got a bunch of its current buzz and put it on the map was this audio overview feature where you had your own personal AI podcast hosts that discussed the content you had uploaded to the notebook. They just made an update to enhance this feature where you can now actually guide the virtual hosts so you can customize their focus and their expertise level. It's basically they equate it to like slipping the hosts a note just before they go live, like shaping how they present the material that you want to consume. So if you want them to focus more on specific topics, you can do that. You can have them adjust their expertise level to suit their audience. And you can also listen to them now while you're actually working within NotebookLM, which is a new feature so you can be querying your sources, getting citations, all while listening to this audio overview. Interestingly, Google is now also rolling out NotebookLM business, which is going to be available soon through Google Workspace. This is tailored for organizations, universities and businesses. It is also focusing on data privacy and security, so you can rest assured that you're keeping your information safe and secure as you're using NotebookLM. So Paul, like Notebook LM just feels like it's kind of caught lightning in a bottle at the moment. I mean, it's been around for longer than we've been talking about it, but right now it's just having this incredible moment now. For instance, we've been using it like, inspired by some experiments from our podcast listeners that they posted about stuff they were doing with it. I actually just put all of our 20, 24 podcast episodes into a single notebook, which I can then query and converse with. We were literally using it right before recording. Quickly find some like, obscure quotes that we'd been talking about and Mike was.
- [45:00]
Paul Raitzer
Like, oh my God, it actually works. Exactly. We were looking for. With the time.
- [45:04]
Demis Hassabis
Yeah, I typed in some very vague search of something we were trying to remember if Paul had said. And it narrowed it down quite quickly. It was really impressive. Have you. How have you been using Notebook?
- [45:14]
Paul Raitzer
Interestingly enough, last night. So when I was prepping for the podcast today, like the way I, like I've said before, but if you're a new listener, like this is kind of how we do it. Every article we talk about, podcast we talk about, video we talk about, I listen to, watch or read every single One of them, I don't use AI to summarize these things for me and then just regurgitate bullet points for people because my feeling is like, I don't grasp the topic, then like, I don't deeply understand it if I don't personally consume it. So the way I do things is like, let's say I'm listening to a podcast, or like, go back to the Sequoia example, I will read it and I will be copying and pasting excerpts from it into a kind of a sandbox. And then I'll boldface the things that I want to specifically call out. So what I tested last night for each of the main topics is I created a new notebook and notebook om for each of them, gave them the source note. New York Times doesn't work because it's a paid subscription and you can't get it into there, even though I have a paid subscription. So maybe future integrations with paid subscriptions would be cool. But then I created a briefing doc and a table of contents as the starting point for them. And what I then did is I went through and looked at the themes I had pulled out personally and compared them to the themes and summary that NotebookLM created to see if I missed anything or if it called out something that maybe was more interesting than what I was selecting as a theme. And so I actually, last night was my first time to sort of pilot using it as like a research assistant almost in prepping for the podcast. And, like, it was cool. Like, I haven't found. I haven't perfect the workflow at all, but it definitely helped help me again, not replacing doing the work, but definitely assisting me in doing it.
- [46:54]
Demis Hassabis
Yeah, for lack of a better phrase, I feel like it's extremely helpful in this, like, connecting the dots research where it's like, not just dropping in one complicated source to understand it. It is very valuable for that. But like, for instance, I dropped in three of the top papers or manifestos on AGI last week. Dario Amade, Sam Altman's and then Leopold Aschenbrenner's situational awareness paper. And then you can start saying, like, what commonalities did they hit on? What did they disagree on? Things like that. It's really insightful if you want to kind of dial in on a topic.
- [47:28]
Paul Raitzer
I mean, you and I could probably sit here and brainstorm like 50 ways to use it. Because as you're saying that I immediately is like, okay, so if we're profiling Demis Asavas, let's go grab the transcripts from his last five interviews and let's like summarize, you know, those things. Yeah, it could. Yeah. Or we could like if we were talking about like AI in the workforce, which is the thing Mike was trying to look up before we started is like the last time I talked about jobs in the workforce, we could go back and say let's just grab every episode where we talk specifically about jobs. Let's create a notebook LM dedicated to the transcripts of those. And now we have everything that we've talked about previously about jobs. So like yeah, it's again, like pick a single tool with chat, GPT notebook and like go deep on. Let's find three to five use cases that are just going to be really valuable to us and let's, let's lock those in and like adjust our workflow and then you can always experiment, keep finding more. But like just drill in and nail those and you create a ton of value for yourself.
- [48:23]
Demis Hassabis
All right, next up, Adobe just wrapped up its annual Max event and during this they announced a bunch of new and interesting AI updates. So the star of the show was Adobe's first Generative AI video model. This has been teased for a while. It's called their Firefly video model. It is now launching across a handful of new Adobe tools, including right inside Premiere Pro. So there's a tool now in Premiere Pro in beta called Generative Extend, which can be used to extend the end or beginning of footage that's slightly too short or make adjustments mid shot, like correct shifting eyelines or unexpected movement. Adobe also announced a bunch of AI powered features across its Creative Cloud apps. One, for instance, is in Photoshop. It is called Distraction Removal and it can automatically identify and remove common distractions in images like people or wires with a single click. Adobe is also pushing the boundaries with experimental tools. They have one called Project Turntable, which allows designers to rotate 2D vector images as if they were 3D objects. This would typically require you to completely redraw the image. Another interesting development is something called Project Know how, which can help combat misinformation by tracking image ownership across various platforms. Adobe is also teasing some future developments, including something called Project Concept, a planning app that allows real time collaboration on mood boards with AI powered image remixing capabilities. And interestingly, Adobe during this event signaled a shift in its approach to AI. So Scott Belsky, the Chief Product Officer, announced that the company is moving away from the prompt era of AI, which he suggests cheapened and undermined the craft of creative professionals. And instead Adobe is entering what they call the control era, focusing on integrating AI in more specific ways to enhance creative workflows without replacing the human touch. Now Paul, I confess that 12 to 18 months ago with everything we were seeing all this stunning stuff coming out in image and video generation, I was like, I worry Adobe's in a lot of danger and they are not going to move fast enough to deal with this given their established business. But I don't think I had to be worried because they have like been on a tear with embracing generative AI. They've got their own video model ahead of Sora. They beat out Sora to actually get to market. Like how bullish are you on what Adobe is up to?
- [51:04]
Paul Raitzer
So interestingly enough, if you go back to like 2019, 20, 20, 2021, when I was doing keynotes about artificial intelligence, I actually often featured Adobe as one of the forward thinking companies. They were doing a ton in, you know, what we'll call like the machine learning era before generative AI really took off in 2022. Like I specifically remember this, like I think it was their CEO and they had a slide. This is back in like 2019 and like over 100 and some AI features within the Adobe platform. So it's not, wasn't thinking about AI and doing AI, but there was definitely that window when generative AI emerged in 2022 where it's like, what is Adobe doing? Like they just seem to get caught flat footed by the innovation and image and video generation and editing. And then when they came out with like their first version of Firefly, it was kind of unimpressive. And so yeah, they do seem to be catching their stride now. And I was following along online with some people who were at their event and it seemed like people were responding very positively. I know they focused on, I think it was their Firefly video model that they were positioning as like a responsible model that's actually trained only on licensed data and their own internal stuff. So yeah, I mean it definitely a company to watch. I think that while we thought there was a lot of probability of disruption in the early days, almost going back to kind of how Sequoia was talking about this with these, you know, kind of wrapper companies that were showing up and they're going to threaten Adobe, it does sure seem like we've shifted more toward the incumbents who figure out how to apply AI seem to still have an advantage. They have the data, they have the money, they have access to the compute to build models. And yeah, so it's, you know, I don't know where Adobe stock has been, like how Wall Street's been responding to their moves, but it does seem like they're heading in a good direction.
- [53:00]
Demis Hassabis
All right, we've got a couple other design and imagery focused updates here. So Midjourney, which is the company behind one of the most popular AI image generation tools out there, has announced plans to release an upgraded web tool that allows users to edit any uploaded images from the web using Midjourney. So this new feature is apparently set to launch sometime this week. According to their CEO, this will enable you to edit existing images and also retexture objects within them. So users will be able to essentially repaint colors and details of objects based on text captions, which opens up all sorts of creative possibilities. However, there are understandably some concerns here. The ability to easily edit and manipulate existing images is raising questions about copyright infringement and potentially spreading deep fakes. Midjourney to address this, is planning initially to restrict the release to a subset of its current community and implement increased human moderation along what they call new, more advanced AI moderators. So Paul, given just like how popular and powerful Midjourney already is, this seems like kind of a big deal and certainly something I imagine Adobe is looking at pretty seriously.
- [54:21]
Paul Raitzer
Yeah, and it's that, you know, it sounds like Adobe is really steering toward their traditional customer base of the design community, where Midjourney is likely going to open up to the non design community and more and more cater to people like me who have zero design capabilities but still wants to mess around with logo concepts and tweak designs and improve images that otherwise I would have no right doing. So yeah, it's kind of interesting, the responsible rollout. Good luck. Like anything Mid Journey can do, someone can do with open source. Like it's just not a solution. So I just, when I see messaging like that, it's like, yeah, okay, like it's just like playing the PR game of trying to, you know, sound good. But let's all be realistic that this tech is going to be readily accessible to anybody and they're going to be able to do whatever they want with it.
- [55:10]
Demis Hassabis
So Playground, which is another popular AI graphic design tool, just released Playground V3, which is their latest text to image model that achieves state of the art performance across a bunch of testing benchmarks. And they basically said that our new model's focus was to be the best at prompt understanding and control. There's that control term again. Going beyond aesthetics, which has saturated as a benchmark, it outperforms all the most popular image foundation models in its class. Now, the company says that they actually evaluated this model across popular graphic design categories. So users consistently chose plate round B3's designs over human made ones in categories that may sound familiar. Things like logos, social media, post designs, cards and invites, and even memes. It can also handle prompts with more detail and longer token links than any other image model. According to the company, it excels at generating accurate text within context in the image, which is something that historically these models have struggled with. And they say that Playground V3 shines in all these areas thanks to its LLM integrated structure. It understands and follows detailed composition, layout and style directions while also grasping cultural references like holidays, memes, celebrities, sports teams, and more. Paul, we've talked a bit about Playground in the past. Certainly not as much as the usual suspects like Adobe and Midjourney, but it's really interesting. They're pretty clearly pivoting or positioning themselves as AI for graphic design. Should we expect more of this in the graphic design space?
- [56:55]
Paul Raitzer
Yeah, I haven't tested Playground in a while. It's probably been four or five months since I've been in there. But I used to enjoy it because you could choose the different models and you could kind of play around with the creativity and the temperature, I guess, for lack of a better way of saying it. So yeah, I'll have to. I'll dive back in and play around with it a little bit. I don't. I mean it sounds like they're saying they're building their own models. I find that hard to believe. But maybe they are versus tuning on top of someone else's models. But yeah, I mean it's. They've been around for a while. They've been kind of. I forget they rebranded at some point. It wasn't called Playground.
- [57:28]
Demis Hassabis
Yeah, they did. I forget what the initial name was. Yeah, yeah.
- [57:31]
Paul Raitzer
But yeah, it's again like it's worth using as a, like as a non designer I was able to kind of get in there and do some stuff. But like Imagine from Google Gemini is getting really good. Dolly's, you know, increasingly getting better. I think both of those will probably have these kind of capabilities native within it. So I think it's a hard play. Like if I was getting pitched to invest in a company like this, I would probably struggle to understand how they're going to differentiate 12 months from now when Dall? E and Imagine have all these capabilities baked in and Adobe's got them all. But maybe there's a market there for it. I don't know.
- [58:09]
Demis Hassabis
So next up, Elias Touris, who is the former VP of engineering at HubSpot and a co founder of Drift, which sold for 1.2 billion, has launched a new AI startup called Agency. Now Agency just emerged from stealth Mode. It secured $12 million in seed funding led by Sequoia and HubSpot Ventures. And the company's mission, they say, is to automate many of the tasks traditionally handled by customer service managers, CSMs, things like onboarding, training and upselling new features to users of things like complex B2B software. So Taurus actually conceived the idea for Agency while consulting for OpenAI in early 2023. He was basically helping work on AI solutions for OpenAI's enterprise customers and realized that everyone could benefit from AI powered CSM work. It's basically designed to understand each customer deeply by analyzing data from different sources like email, CRM chat and PH conversations. This allows agencies AI to anticipate customer needs, effectively automate routine tasks like scheduling follow ups, customer onboarding and meeting prep. The product is currently in an invite only beta phase, but it is being tested by companies like heygen. In a post that describes the company Taurus wrote about the name quote, the company is called Agency because that's the vision. Just like hiring an agency, our product will handle the work for you and without the meetings, contracts and back and forths. Paul, we're obviously very familiar with both HubSpot and Drift. Elias's background alone makes this worth paying attention to. Like what do you make of Agency given your experience with his background and these companies and this kind of problem set. I'd also kind of love to get your thoughts on this name because this is total speculation on my part, but this is customer success focused right now. Now but that last quote I read sure sounds to me like this is meant to expand to other areas reserved for agencies. Am I wrong in that?
- [60:14]
Paul Raitzer
Well, I mean interest so Sequoia the investment. Brian Halligan is at Sequoia now. Brian Halligan, the chairman of HubSpot Co founder, former CEO I think Brian is involved obviously in in this deal. Brian is the guy back in 2000. So my agency again longtime listeners know this. So I created PR2020 back in 2005. We were HubSpot's first partner back in 2007. We were the origin of their partner program that at one point accounted for 45% of their revenue. So yes, I have intimate knowledge and experience with HubSpot. I built an agency on the backbone of their agency ecosystem. Man, I probably have a lot of thoughts about this one, but this is just a rapid fire, so I'll be concise here. As soon as I saw the name and as soon as I saw the description, I thought, well, that seems like a pretty direct service. As a software play, which Halligan has previously tweeted about and Sequoia touched on in their paper that we didn't really get into, but this idea that the AI provides the services and it sure sounds like Torres is directly saying that like we're just gonna, you don't need the agency, like we'll build the agency for you and you'll automate it. It might be a collection of models and agents like I was explaining earlier, the symphony of agents and, and it does the work for you. So yeah, interestingly enough, I mean this is so my keynote for agency summit on November 20th is like AI agents and the future of the agency. And I don't, I don't even know what I'm going to say yet, honestly. And it's like a month from now. But this is the exact thing I was trying to prepare agency leaders for. So I don't obviously own an agency anymore. I don't really have a stake in the game. But when I look from the outside in, knowing what goes into running an agency for 16 years, I, I would be very seriously exploring what the future of the agency world is when people are literally creating AI companies called agency and saying they're going to do your job for you, but the client doesn't have to deal with all the BS that goes with managing an agency relationship. So I don't know, maybe, maybe they're going to sell two agencies as like a future. Maybe that's a distribution channel for them, I don't know. But yeah, from the outside looking in and without doing a bunch of additional research or talking directly to them, this sure seems like a direct shot at saying let's just go take on the agency world and you know, the multi billion dollar industry that it is. Let's, let's go get a piece of that. And I honestly, it's there to be had. Like I. Yep. You know, I think it's a threat to agencies. I think it's a very smart market for Sequoia and, and it's interesting that HubSpot Ventures is involved because I mean HubSpot was built on the back of agencies and they have over the recent years, let's just say that partner program has evolved in its focus, so. Fascinating.
- [63:09]
Demis Hassabis
So next up, another kind of related Bain and company, which is one of the giants in the world of consulting has announced a significant expansion of its partnership with OpenAI. So the two companies have been collaborating since 2022. They had a global services alliance announced in 2023, but now they're expanding in a couple different ways. So Bain is establishing a dedicated what they call center of Excellence CoE staffed by a team with extensive experience in OpenAI technologies. Bain and OpenAI will co design and deliver initial solutions for they specifically call out retail and healthcare life sciences. With plans to expand to other sectors. This center of Excellence will be equipped with the technical resources to use OpenAI frontier technology to deliver client solutions. And so far, Bain has already deployed OpenAI platforms, including ChatGPT Enterprise to its employees worldwide. They say that the partnership has already delivered concrete results for clients like Coca Cola. And in addition to this partnership, Bain is going to continue to offer AI transformation consulting services that includes stuff like strategy development, process change and organizational development. So Paul, we have covered plenty of partnerships between these consulting firms like Accenture McKinsey with OpenAI and other AI companies. This is kind of, you know, the latest expansion of this type of team up. I found the focus here on retail and healthcare interesting. Is that kind of a signal that those two industries are what people have their eyes on when it comes to AI transformation?
- [64:51]
Paul Raitzer
They're just big market value industries, obvious use cases. But yeah, I mean, what's going on here like can go back to HubSpot. In 2007, what HubSpot eventually decided was that the way to push and distribute their software into the market was to go through trusted relationships with agencies. So you know, you build an agency partner program with people like, you know, my agency and you introduce HubSpot software into that client base. And so that's what's happening. And this has been going on for a couple of years. I've been involved in some of these conversations with some of these bigger firms where, you know, if you're open AI or you're Google, Gemini or Anthropic or whomever it is, whatever the deals we've been hearing about with Accenture, McKinsey and Deloitte, instead of building your own salesforce and scaling it up immediately and trying to go in and sell to these enterprises that you don't have relationships with, it's way faster to train up an existing client base or client relationship team at Bain or Accenture or McKinsey and and then have them introduce your technology through their solutions. So you're just leveraging the trusty relationships of Bain and Accenture McKinsey and and let them go do the work and the onboarding and sell services on top of it. And I think back in the day with HubSpot, it was like for every dollar of software people spend, I'm going to get the number wrong. But I want to say they spent four or five dollars on services. And so it's the same premise here. If you go spend 2 million a year on OpenAI enterprise licenses for your entire organization, you're probably going to spend 10 million on the services to onboard and implement and function the change management, build the solutions around it that it's. This is a playbook that's been going on in tech for 50 years. So it's a very natural thing to see happen.
- [66:41]
Demis Hassabis
All right.
- [66:42]
Paul Raitzer
Until agency AI puts them out of business. I'm just kidding. That's not going to happen.
- [66:47]
Demis Hassabis
Oh boy. We might have to roll that into another main topic next week. I feel like there's more to unpack there. All right, so a couple final topics here this week. The UK government appears to be planning to consult on a controversial, what they're calling opt out model for AI content scraping. So the Financial Times is reporting that under this proposed opt out model, basically AI companies would be allowed to scrape online content from publishers and artists unless those parties specifically opt out. The UK government plans to unveil their consultation on this opt out model in the coming weeks. And this is described as the government's quote, preferred outcome by sources that are close to the matter. Publishers and creatives are not exactly happy about this. They say this kind of model of regulation or legislation is unfair and impractical. They claim it would create a huge administrative burden for smaller companies. The creative industry, you know, prefer, prefers this opt in model which allows for licensing agreements and fair compensation. Interestingly, the European Union has a similar opt out model in the AI Act. Paul. Obviously this is exclusive to the UK at the moment. Kind of interesting though to see that this approach which was heavily lobbied for AI companies, is the preferred outcome. Like given that the major cash and influence that AI companies wield, not just in the uk, should we expect to see this approach kind of become more normal?
- [68:23]
Paul Raitzer
I have no idea. I mean, I feel like, I think Japan has their laws. Basically there is no, like copyright doesn't matter. Yeah. So I'm kind of surprised to see in the uk, but maybe I, I misunderstand how they generally have approached AI. I mean the EU certainly has taken what seems to be quite a conservative approach. Yeah. And this does seem to almost be the opposite. But I don't know, like it's, I'd be interested to see, like, we should dig back into the, like the U.S. copyright Office and see if there's been any updates. To my knowledge there hasn't been. From the listening sessions they were doing like 2023. I don't know if there's been any movement, anything. But yeah, we'll have to circle back around and see if there's any updates on what's going on in the US on this topic. I mean, even back in May, kind of September, there wasn't anything new that I'm aware of and we had that whole panel on copyright.
- [69:18]
Demis Hassabis
I don't, I don't think I've seen anything come out that's definitive. All right, so next up, Microsoft has announced that there are new autonomous agent capabilities coming for copilot. So first, you'll be able to access a public preview of the ability to create autonomous agents within Copilot Studio starting next month. And second, Microsoft is introducing 10 new autonomous agents in Dynamics 365 for things like sales, service, finance and supply chain. Here's a quote from Microsoft in this announcement on kind of what they're doing here. They say, quote, think of agents as the new apps for an AI powered world. Every organization will have a constellation of agents ranging from simple prompt in response to fully autonomous. They will work on behalf of an individual team or function to execute and orchestrate business processes. Copilot is how you'll interact with these agents and they'll do everything from accelerating Legion and processing sales orders to automating your supply chain. So this feature within Copilot has been private. Previously, it's been used by a few select customers, people like McKinsey and Thompson and Reuters. Now it will be in a public preview, which means more people have access. Microsoft also provided some examples of like what these Dynamics 365 agents look like. One is a sales qualification agent that basically prioritizes sales opportunities. Another is a customer intent and knowledge management agent that improves customer service by learning to resolve issues. Now, obviously Microsoft is heavily interested in promoting this service, but says that some of their early results from using agents include a sales team achieving 9.4% higher revenue per seller and 20% more closed deals, and an HR team having 42% greater accuracy answering employee questions with an agent. You can apparently start building agents in Copilot Studio today, but this autonomous agent capability is rolling out next month. So Paul, kind of a couple things just jump out here. Like one, we're obviously like all in on agents. Microsoft is not the first one to be doing this. We've talked about Google Salesforce and others 2. This kind of struck me as like a really broad definition of agent. Like they seem to be considering them as anything from simple AI assistance to fully autonomous. Like, is this going to get really confusing for buyers and users if you're not. Not following this closely?
- [71:55]
Paul Raitzer
Yeah. I'll tell you the thing that's confusing me is why is Satya Nadella tweeting this at 6:30am Eastern Time on a Monday morning? Like, I'm seriously, what. What else is happening this week that they felt the need to get this out? Events that I'm aware of. I just did perplexity search. So what is Microsoft doing for AI this week? And I can't see anything. The only time you see news like this is when like OpenAI is about to announce something and they're getting out ahead of it. Or maybe anthropic or something. But like very weird. Like what announced? Like it specifically says today we're announcing new. It's like, okay, why today? Like, this is stuff you normally announce at an event or not at 6:30 in the morning. So which again, Eastern time. Like they're not even West Coast.
- [72:38]
Demis Hassabis
Right, Right.
- [72:39]
Paul Raitzer
So. And it's a CEO tweeting. Like usually when Sundar or Satya tweets something, there's something more to the story. So, yeah, I don't, I don't know, it just seems like a continuation of like, we know they're building agents, we know it's under copilot. Like, I'm not sure what exactly is actually new here that wasn't already out per se, but I think I'm more intrigued by now what happens the rest of the week.
- [73:02]
Demis Hassabis
Because, right.
- [73:03]
Paul Raitzer
If they did this just randomly on a Monday morning, then it just isn't following the patterns that AI leaders have been following over the last two years, which is to preempt each other on other news.
- [73:15]
Demis Hassabis
Interesting. We're in a bookmark this.
- [73:17]
Paul Raitzer
I haven't looked at Twitter on this. I wonder if anybody's announced anything while we're sitting on this.
- [73:23]
Demis Hassabis
All right, so our last topic this week is that Demis Hassabis, the head of Google DeepMind, recently sat down with the Times, which is a British publication, at their recent tech summit to discuss building AGI with safety in mind. And Paul, I know you were paying close attention to some of the things Demis was saying here. Do you want to kind of talk us through, like what he was talking about? That that is worth paying attention to here.
- [73:51]
Paul Raitzer
It kind of goes back to that Sequoia article where, you know, if you're a regular listener paying close attention. Nothing groundbreaking per se, but when I listened to the podcast there were definitely a few things that jumped out at me, so I'll just call it a few pieces. So when asked specifically about timeline for AGI, he said probably within 10 years. He then did give his definition of AGI, which, you know, we've quoted numerous times and it changes a little bit, but I thought it was worth noting how he defined it here. So he said the goal of DeepMind is to get to AGI, which means a general system that's capable out of the box of doing any cognitive tasks that humans can do. So fully general, capable of computing anything that's computable. Now interestingly, he does not distinguish their between the performance level, which is in the levels of AGI that Shane Leg and the DeepMind team published earlier this year. It's so it says doing any cognitive task at what level? At a 50% tile level of like 50 of all humans. At a virtuoso level, like 99th percentile of humans, like PhD level. So again, like a definition, but with some vagueness to it. He does say that multimodality is a key to AGI, which we know that the models from the ground up, like Gemini, are being built with image and video and audio capabilities and coding capabilities and reasoning right within the model. It's not just a text in, text out model, which is what GPT3, GPT4 were. He said he thinks there's two to three big innovations needed from here until we get to AGI. And then he kind of hints at what areas those might occur within. So he said Project Astra, which we've talked about, which is kind of like the vision capabilities of your phone or of glasses that can see and understand the world. He specifically says memory personalization are coming in next gen universal assistance. So memory and personalization are two key things to think about. He talked about current chatbots are passive question and answer systems. They want agent based systems. So again agents coming in. He said they need to be able to do planning like chain of thought reasoning. Take actions have basically nearly infinite memory. Remember everything about your interactions with them and be personalized so it remembers your preferences. And so those are the kind of the keys, like somewhere within those are the two to three breakthroughs. He's talking about that if we can get some breakthroughs that allow us to achieve infinite memory, true personalization, advanced reasoning and planning that we can then get to that point. The one that like kind of stuck with me was he was asked about the, the Doomers versus the people like the techno optimists who think that every all acceleration of technology is great. And they said like, why are you sort of more of a cautious optimist? Like why, why are you concerned? And so he said specifically, so if you go Back to like AlphaZero, which is a system that they designed that could learn like gameplay basically from scratch. He said, I've seen this in the microcosm of games, teams, something I understand well, like playing chess where you start with a system Alpha 0 that's random in the morning, by morning coffee break, it can beat me. And he is a like world class chess player. By lunchtime it's better than the world champion. And then by afternoon, within eight hours, it's the best chess playing entity the world has ever seen. I've watched that process over an eight hour period. So he's basically saying all these techno optimists were like, we'll figure it out, we'll give it goals. It'll only do what humans want it to do. He's saying, no, I've seen them from zero go to like virtuoso, world class superhuman at a thing in eight hours. So anyone who thinks we can't have a fast takeoff doesn't understand how quickly these things can learn when developed this way. And so I think it was just kind of like a call of caution, but a very practical way of like I've been there, I've built the systems that do it like it, it can take off. So I always, I mean it's only like a 25 minute interview. We'll put the link in there. I always just love listening to Demis talk. I'm going to learn something every time.
- [78:00]
Demis Hassabis
Yeah. And I certainly wouldn't characterize him as someone who is overhyping things often. So if he says something like that, I would pay pretty close attention.
- [78:08]
Paul Raitzer
Yeah.
- [78:10]
Demis Hassabis
All right, Paul, that's all we got this week. Just a couple quick housekeeping notes. If you have not checked out the Marketing AI Institute newsletter, it is marketingaistitute.com newsletter. It is called this Week in AI and we'll give you an in depth breakdown of everything we just talked about, plus all the other news we didn't get to in this episode. Last but not least, if you have not left us a review and you have the ability to do so on your podcasting platform of choice, we would very, very much appreciate it. It helps us get better and reach more people with the show. Paul, thanks for demystifying AI for us this week.
- [78:47]
Paul Raitzer
Good stuff as always. Thanks Mike, and we'll talk with everyone again next week. Appreciate you listening.
- [78:53]
Mike Kaput
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.