DeepSeek’s New Model, Potential TikTok US CEO, Microsoft CEO News, AI Agenda Live NYC | Oct 1, 2025

0:13Welcome, everyone, to the Informations TI TV. My name is Akash Basricha. It is Wednesday, October 1st. We have got a great show planned for you today. We are recapping our AI Agenda Live conference from yesterday with all of the highlights for you. We have also got the CEO of ClickUp coming on to get his view on the corporate data wars, and we are taking stock of the current deal-making environment with the Futurum group. We also have a ton of news to get to. We've got a big TikTok story that we are going to break down, But I want to start with the news from Microsoft today. Microsoft announced this morning that CEO Satya Nadella will offload some of his duties to the company's chief commercial officer, Judson Althoff.

0:57Althoff will become the CEO of Microsoft's commercial business. And Nadella will start focusing his time more on Microsoft's core technology. It is a big announcement, and so I want to bring on our Microsoft reporter, Aaron Holmes, to tell us more about what he thinks of all this. Aaron, welcome back to the show. It's great to have you. Happy to be here. So we got big news today. Gosh, I wasn't expecting this one. Why do you think Microsoft is making this move? You know, it's really interesting. I think that what Satya Nadella said in his memo to staff this morning is that he wants to be laser focused on leading, you know, some of the engineering tasks that Microsoft has in front of it, like, you know, developing cutting edge AI, as he put it, or, you know, working on systems architecture and data centers, and is somewhat, you know, delegating more of the sales and go to market role to Judson Althoff.

1:53At the same time, you know, it's hard to judge exactly how big of a change this is just because Satya is still, you know, above Judson in the org chart and is effectively still the CEO overseeing everything. So I think we're going to have to wait and see whether this actually, you know, meaningfully changes his role in that regard. And we should say, I mean, Microsoft has a ton of CEOs inside the org, right? Yeah, I mean, this is a company that is kind of known for, you know, for role inflation. It has probably like dozens of executive vice presidents and over 100 corporate vice presidents, not to mention several CEOs, including, you know, Mustafa Suleiman, who's the CEO of AI.

2:32So it's not uncommon for people to have these, you know, really lofty titles at the company while still being somewhat lower on the org chart. Right. Well, look, I want to talk about this role that Seth and Adela will now play. You know, I think to understand what sorts of elements of the business he's now going to oversee a little more closely, I kind of want to take stock of where Microsoft's core technology is and how it stacks up against other big cloud players like Google, like AWS, Meta is getting in the game. I mean, talk to us about where their models sit right now, because I think we think of it as synonymous with OpenAI, and yet they're obviously developing their own stuff too.

3:11Yeah, I mean, so we've seen Microsoft attempt to develop its own AI models over the past two plus years. And so far, you know, the progress on that has been somewhat halting. They finally just debuted their AI models for the first time last month, but they still haven't done a full release of those models. And, you know, they early on stacked up a bit lower than the models from OpenAI and Google. So I think, you know, the company is still somewhat playing catch up there. And at the same time, you know, their efforts to develop their own chips, as we reported, you know, a couple months ago had to be scaled back because they were behind schedule.

3:49So I think there are a few realms where, you know, Nadella might really want to put his head down and work on catching up and getting Microsoft back into, you know, the top five or first place on those fields. So we've got chips, we've got models, any sort of signal from them as to how much is going to be focused on data centers and quantum as well? Yeah, I mean, the company has put out a lot of research about quantum computing in recent years and signaled that it's important, but it's still not clear when that will become something that's actually usable as part of its systems. At the same time, it is spending as much as many of the other hyperscalers on data centers.

4:31The company spent more than$80 billion last year and is on track to spend even more than that in the coming year on CapEx. And Satya said in his memo today that he essentially wants to focus on the highest ambition technical work like those data centers as part of his role. So I think we'll see him get even more hands-on in that realm. Right. Great. Well, Aaron, I take it this is just the start of what we have to find out about the new ways that Microsoft plans to operate. And so I look forward to having you on again. That was Aaron Holmes, our Microsoft reporter here at The Information. Okay.

5:07Well, yesterday was our AI Agenda Live conference in New York. We had a number of great speakers, including Sarah Guo from Conviction and other speakers from Reflection AI, Weka, and OneX. I want to bring on our AI reporter, Stephanie Palazzolo, who hosted the event to share more about her big takeaways from the discussions yesterday. Stephanie, welcome back to the show. It's great to have you. Thanks. Thanks for having me. So look, I want to talk all about what your biggest key takeaways were. But before we do it, I do want to play a clip. We've got a clip from Sarah Guo at Conviction. Let's take a look.

5:43I think most tools take a cycle of adoption that is a lot longer than a few years. We spent a bunch of time recently with really large private equity firms that are figuring out how to like transform the companies they own. And they're like, okay, we have a five-year plan that involves like education, adoption of tools, training, you know, re-skilling, performance management. And I think that is more of the timeline that people should expect of just, you know, real human beings who are used to doing something for a long time are going to need to figure out how to use these tools. And that is a lot of education.

6:21So this idea here of a five-year timeline, I mean, that's the, I thought it was a good point that Sarah made. You know, it's also a long time to, you know, sort of wait for a return on your investment in some cases. What were your big key takeaways from that discussion? And then we'll get into later parts of the conference in a second. Yeah, totally. I mean, so Sarah was responding to a question that I'd asked her around kind of this disconnect that we've written a lot about in the world of AI, where on one hand, you have companies that are obviously investing, you know, tens, if not hundreds of billions of dollars into this infrastructure data centers.

6:56And then we have people like Sam Altman who are saying, you know, AI will soon be able to cure cancer and fix our education system. But then on the other hand, we've also written a lot about businesses that are just having a lot of trouble getting AI up and running in a reliable way or finding useful use cases for it with products like Agent Force from Salesforce or Copilot from Microsoft. And so I think Sarah's point here is that these things are never as easy as what people think. And the timeline that we should expect to kind of see some of these products start working in a reliable way is more on the, you know, more in the scope of, you know, five plus years versus a couple months.

7:38Right. And I should say, you know, two of the other points that I thought were really interesting is we also had speakers from SAP and from AMD on the stage yesterday. And some of the points that they raised, which resonated with me, are this idea that, hey, you know, a lot of companies are spending on pilot projects right now, and they might not be seeing the ROI initially from these pilot projects. But you just kind of got to stick with it. And, you know, just because a pilot doesn't go well doesn't mean you stop investing. It's actually more reason to keep finding the thing that can make you a lot of money.

8:08And the other point that they raised was that it's really the boring use cases of AI that are being the most practical uses right now for the technology. It's not the stuff in the promo videos. It's like reading through documents quicker and PDFs and stuff like that. Yeah. No, I mean, I think that's definitely very true with the customers that we talk to. And I think maybe what the issue is, is just how big that disconnect is between, again, the very high expectations and the promises of AI and what they're able to do in real life, which a lot of the times is just searching through PDF or taking an image and taking a text to put into an expense report.

8:53So I think the fact that it's taking a while is not a surprise, but just you know, reality versus expectations, that gap, I think, is just quite large. What about the panel you had at the end of the day? You had some folks on from Databricks and from Anthropic. You were talking about sort of the core technology, right? I mean, reinforcement learning, AGI. What were some of the reflections you had from that? Because they didn't necessarily agree on some of the points that they were both talking about. Yeah, I mean, I think that was a fascinating panel. So I think part of what I wanted to get from both of them is a bit of a kind of vibe check of sorts on the, you know, AI research world.

9:35Because again, you know, last year, there have been a lot of stories, you know, kicked off by one that we wrote in November around, you know, the issues the labs are running into whenever they just simply tried to train models on more compute and more data um they weren't getting the improvements that they had expected and so uh interestingly i i think yesterday was a bit of a shift from that with uh people like sholto douglas from anthropic um showing some kind of cautious optimism there he kind of made the comment where he said you know i really think we can get to AGI or like AI that's on the level of human experts in certain domains just by using techniques that we already have today.

10:20And we understand largely how to use like reinforcement learning. And so that's obviously a pretty striking point just because AGI is kind of, you know, it's like the pie in the sky and the gold. Right. It's like, what is AGI? You know, this is the thing that Microsoft and OpenAI are debating right now. It's like, we're never going to know when we get there, really. Yeah, yeah. I mean, it's very hard to define. And it kind of feels like the goalposts keep on moving a little bit. But I mean, I think it was surprising and promising that he was so optimistic on our chances of reaching AGI, however you want to define it, without needing new types of models or brand new kind of scientific discoveries.

10:58Right. Well, Stephanie, it was a great event yesterday. There was so much more, including topics about robotics and also some of the more scientific applications of AI that I'm excited to bring to our audience later on throughout the week, and we'll be sure to have you on more to talk more about it. That is Stephanie Palazzolo, our AI reporter and host of yesterday's AI Agenda Live conference here in New York City. Okay, well, speaking of Anthropic, Anthropic wasn't the only company to launch a new model this week. DeepSeek also released a new model, which it is calling experimental right now.

11:32But given the way that DeepSeek has a track record of moving markets around the world, everyone in AI plays close attention to every update that they put out. I want to bring on Nick Patience, VP and practice lead of AI at the Futurum Group, to talk to us about how he is thinking about this new model release. Nick, welcome to TI TV. It's great to have you. Thanks, Akash. Thanks for having me. So talk to us about what exactly this new model is, because I get kind of confused. You've got the decimals in there. You've got the experimental nature of it. What exactly did they release this week? So they released what they call DeepSeq version 3.2 experimental.

12:10So given, as you kind of hinted, the word experimental is in there and a 3.2. This is based on the version 3 of the model they released earlier this year, but it's not version 4. So this is very much, and they made no bones about it. This is an experiment. um and so yeah the key uh innovation that they talked about was what they call deep seek sparse attention um so it kind of makes the model more efficient um while you know and more cost effective to use uh without a noticeable drop in in performance and so you know this is obviously a big deal because as i was just talking about you know the costs of running um you know large ai models can be you know cost prohibitive and so yeah anything that can do uh it can be done to make models more efficient would make them more useful.

12:56And so when we hear sparse attention, basically, we are just to think that, hey, we use less compute and it's cheaper. Is that the idea? Yeah, and that's exactly the idea. And as a kind of whether there was a correlation causation here thing, they also announced at the same time they cut their API prices in half. Oh. So, you know, they're implying. So they're going for the price leader in every respect here. Yeah, exactly. And I guess they were implying that because they've got a more efficient model, they can charge less for it and make it up on volume of API calls, I guess. I mean, it's always hard to know exactly, but that's the implication.

13:32And remind us, we've heard that several models are good for several specific tools and applications, I guess. This model, is there any sort of consensus on, is it good for coding, for text generation? Where does it sort of specialize? Is it a broad model? It's fairly broad. I mean, the one thing, they didn't give a lot away, but the one thing they did talk about was, you know, long context windows. So you're talking there about document processing of large volumes of information, of text, essentially. That's what it's probably good at. They haven't really indicated that it's, you know, particularly good at code generation or anything like that.

14:08It's more the, you know, it can input a large volume of information and then process that, do analysis on that, that kind of application. Right. And from the people you talk to, I mean, how good is it? um well it's hard to get much details yet um on how good it is i mean because it's extremely early but yeah early occasions is it's it's pretty good um you know it's it's you know there's difficult to you know i'm slightly skeptical of the kind of benchmark race um because i'm not sure some of these things have real world use cases um so yeah it is it's not in you know it's not obviously being used in a real world use case at the moment so it's it's it's pretty hard to determine right I mean, you follow so many of these different models that are being released and the different companies that are behind it.

14:50I kind of want to take a little bit of your pulse on the models coming out of China right now, because a lot was made of DeepSeq when it first came out and sort of shocked the world. I mean, since then, we've seen a lot of other companies in China, not the least of which are ByteDance and Alibaba also making progress with their own models. Give us a little bit of the lay of the land in terms of how DeepSeek's technology now stacks up against some of those other big companies in China that are releasing their own tools. I think it does stack up pretty well. I mean, with Alibaba and the others, I think China is appearing to take a fairly pragmatic attitude towards this.

15:27This is not a kind of march towards AGI, quest for AGI. This does seem to be more focused on practical applications in business and in academic research and things like that, rather than big flashy announcements and kind of great statements that we're just one model improvement away from AGI. So I think they have a kind of fundamental different attitude to Western Europe and US, certainly model makers. And also embracing open source and releasing a lot of not just DeepSeat, but other Chinese vendors releasing a lot of things open source, which is something a lot of the, you have some elements of it with Mistral and Cohere.

16:12And then obviously more recently, OpenAI realized they had to go and do that as well. Are those other companies, the companies I mentioned in China, for example, are they catching up to DeepSeq in terms of compute efficiency, not being able to have to use as much? Where are we there? I'd be slightly, I'm slightly skeptical about the claims that DeepSeek made in the first place back in January. Because obviously they were basically implying, and they said so in a paper then and a later paper, that they actually done most of the work in advance, like 90 % of the work, and then basically spent a few hundred thousand dollars on the last 5 % of training.

16:51So there's multiple training runs that went on before that. So I don't quite buy the notion that they are so super efficient and everything just costs six figures versus millions. I just don't really think that's the case. So I think they've done an incredibly good job of pitching that. But there's obviously some elements of it. They are. They are as good as anybody else at efficient model making. It depends. But there's a whole load of other things around. Models are not applications. Models are not platforms. They're things you can use and then get a development environment and bring them in and things like that.

17:30So it's kind of, you know, I think they stack up pretty well against their Chinese counterparts, which is not bad considering where they came from spun out of a hedge fund. But there's a lot more going on there, I think, under the hood than they sometimes make out. And I haven't really released a lot of detail on this one. And so, you know, taken together with all this into consideration, then, you know, the question I have for you is sort of what you see the long term future of DeepSeek specifically being, you know, you talked about now how the other companies are coming into the picture. You talked about how there have been sort of some clarifications about, you know, how the model was actually developed.

18:08But is DeepSeek sort of a flash in the pan in the long run of AI? Or do you see it as sort of an enduring business and an enduring technology that actually gets adopted very widely around the world? I think they're getting adopted pretty widely in China, and even especially in the large EV market there. There's obviously, there's a whole other discussion about how many EV vendors you need in one market, but they're very widely adopted there. I'm slightly skeptical about a long-term future, I'll be honest, of them. Because as I say, it does all seem to be focused on the model itself. And I think the open AI has realized early on and the others are realizing early on, there is more to life than models.

18:48And, you know, for instance, if you're a CIO of a large bank at the moment, are you focused on deep-seek's experimental model or are you focused more on the challenges of getting your AI applications into production? I suspect it's the latter. And so I think the companies that will succeed in this market are the ones that build out more of a stack. Maybe that makes them a good acquisition candidate in that case. Yes, possibly. Yeah, exactly. They're obviously good at what they do. The model-centric nature of what they do, they're good at it. There's no two ways about it. And they're obviously attracting good researchers, and that's another battle that goes on in the AI world.

19:24And so, yeah, you're right. I think they probably are a good target, if that's what the owners want to do, of course. Right. Well, Nick, thank you so much for coming on the show. I know that these models are coming fast and furious. We didn't even get to talk about Anthropics' new model this week, but we'll have to have you back on next time we get a new big release. That was Nick Patience from the Futurum Group. Okay. Well, earlier this week, we had an executive from Snowflake on the show talking about their answer to the corporate data wars that are shaking out in AI. And Snowflake is not the only software company that has been outspoken on that issue.

19:59I want to bring on Zeb Evans, the CEO of ClickUp, to give us some of his thoughts. ClickUp is a project management and product development software company that was last valued at$4 billion. Zeb, welcome to the show. It's great to have you. Thanks so much for having me. Happy to be here. It's a great shirt you're wearing, man. Holy moly. I always have crazy shirts. You know, unfortunately, I branded myself as this. So if I don't wear something crazy, then I'm like, what's wrong, Zeb? Well, it's funny. The instruction I get from our team is they say, wear the quietest shirts as possible. So maybe I'll have to tell them now.

20:35I say, you know, look at what Zeb was wearing. I mean, you know, he looks, you know, he's having fun out there. Anyway, let's talk about the fun of corporate data wars, because it's something that we've written about a lot of the information. Very quickly, just in 20 seconds, tell us about what your business does and the software that you sell and sort of your relationship to this issue at hand. Yeah. So ClickUp builds what we call converged software, which is single application platform primitives where you can build pretty much any software on top of our platform. And so we see the future as you purchasing software and AI from one vendor.

21:14And we see this as like super software providers that occur in the future. And that is what we are building towards today. Got it. And so tell me, how big of an issue has the corporate data wars in AI been for you? I mean, have you had trouble accessing data, you know, as you try to build out AI tools and stuff like that? We have. So with two examples would be Slack and with Figma. Both of those have started locking down data access, depending on which scope you have with their APIs. And, you know, the big thing here is that I think most customers and most people don't understand how important that context is yet.

21:56because the B2B value has yet to be realized in AI in the vast majority of companies and the vast majority of use cases, outside of engineering, let's say, and copywriting. And those things will become so much more valuable over time. And that's, again, why we see the world as converged software. You have 100 % context, and you'll need 100 % context for AI to actually do its best job. Have you gone about quantifying this issue at all in terms of like, hey, this has costed us this much revenue. We could have grown our user base, I don't know, 10 percentage points faster if we had this. Have you put any numbers around it?

22:38The value and the loss of value is more for the customer. It's really, in certain companies, for sure, when their business model is based off this, they're hurt. Like Glean, for example, right? That's something that we predicted would happen is that these other companies, first of all, enterprise search would become a feature on everybody else's platform. I think that's what will happen with nearly all software that is built in the next year or two years. I mean, if you believe that engineering efficiency will gain in productivity, you have to believe that every company will build more software, not less software.

23:13So it's really the customer that gets screwed in these situations. Not the company, unless that company is one of those providers of that actual enterprise search tool. Right, and I guess that's what I was getting at. Presumably the companies that sell enterprise search tools or your company that sells, you guys sell these types of AI tools as well? We do, we do. But in this case, we've actually just seen a surge in customers adopting our chat product because of this. Our chat product is free and it is the same feature parity of something like Slack or Microsoft Teams. So in many of these cases where their data access was cut off and that customer was already realizing value from AI having access to their team communication, a big portion of them would just switch to ClickUp.

24:09And then that gives them 100 % context. And there is no worry about are they going to cut off data again? How much access are they going to allow? So we've actually seen it as a benefit for our platform. So I'm sure you followed the news of Snowflake and the consortium that they put together earlier this week with a bunch of companies signing on to say, hey, you know, we need open data access. You know, that is the future of this sector. Was that sort of the step that you were looking to take? When we had Snowflake on the show, you know, he sort of made clear it's a bit of a handshake agreement.

24:43But, you know, the idea is everyone abides by it. Was that the solution that you were looking for? Are you just looking for everyone to join that pledge? Like, what is the solution here? I believe that the only thing that is realistic that will happen when you think about business perspectives and interests is that over time, you will have super software providers. You will have several software providers that provide all software and all AI for companies. So you only go to one. So you're saying we're going to get even, it's going to be consolidation even more. Yep, 100%, 100%. And so I think everybody can sit there and say those things.

25:23So are you saying that, I just want to make sure I understand. So you're looking at a company like Salesforce, the owner of Slack, right? We've written about, and as you said, they've put up some of these walls for data access. You're saying that a company like Salesforce is going to get even bigger and sort of engulf all of the tools in one? they're going to put up more walls and and over time yes and and they will attempt to provide all of your software okay for for your business and i think that's where the that's the future of of software and b2b okay so i'm just so then so what is the future of the startups then uh yeah like i have a 40 page memo on this i'm happy to share um but i i think it's very difficult for startups.

26:09I don't think you can really... Startups like yourself. I think early stage startups, it is going to be impossible to create a public company and B2B software in the future. But I think that there are some platforms that have built large enough already and have the scale to be able to go public and to be able to compete against the giants and that can provide all of that software with last mover advantage, not just bundling every single software product under the planet and it not actually talking well together and not actually connecting together and having totally different user experiences.