On the Ground at UBS’ Private AI, Software and Internet Conference | Menlo Park

0:07Welcome everyone to a special edition of the Informations TI TV. My name is Akash Pasricha. It is Wednesday, July 15th and we are here on the ground in Menlo Park at UBS's Private AI Software and Internet Conference. But we're going to be talking to a number of people here about the current state of the software sector. We're going to talk about the current state of AI research. We're going to talk about the future of robotics and what that has to hold. We're going to have some great conversations for you. Before we get started, I want to turn it over to Rocket Drew, who is in our San Francisco Bureau.

0:38He has a quick news update for you. Rocket, over to you.

0:45Thanks, Akash. The Information's Asia Bureau published exclusive reporting that strong revenue growth at DeepSeek is empowering the company to raise its second funding round of about$7.4 billion and plot a potential public offering in Shanghai next year. Their annualized revenue reached between$400 million and$500 million recently. We're going to have more from our Asia Bureau on that story tomorrow. Another piece of news, Stripe and PE firm Advent International have submitted an offer to buy PayPal for more than$53 billion, according to reporting from Reuters. The outlet says PayPal has not responded to the offer and that Stripe and Advent want to advance discussions in the coming weeks.

1:32We'll keep our eye on that story. Meanwhile, Cursor, the coding startup set to be acquired by SpaceX for$60 billion later this year has been mapping out a major transformation. Let's bring in our colleague Grace Kay for more on this story. Welcome back on the show, Grace. Hi, Rocket. Great to be here. So you heard about a company-wide meeting earlier this year. What was it that Cursor CEO Michael Truel told his staff at that meeting? Yeah, so Michael outlined kind of a new strategy for the company um you know what has famously been a coding startup is now pivoting towards general ai models he put some deadlines out there they want to be one of the top ai companies by the end of the year they want to have the most compute you know by 2027 and be pushing the boundaries of ai by then um he also addressed some concerns from staff around the spacex acquisition so you know it was a pretty big meeting for the company.

2:31Does this shed some light on why SpaceX AI was interested in acquiring Cursor in the first place? I mean, there's always been a little bit of a mystery around that to me. Yeah, I think like initially going into it, my thought was, you know, obviously they want the coding data just because they've been struggling so much, you know, with their Grok coding model. But one thing that Michael noted in one of the meetings shortly after you know this partnership was announced he talked about how spacex was really excited because of cursor's brand and you know some of the enterprise relationships they have a lot of fortune 500 companies are you know their customers they also have a really large go-to-market you know sales marketing team much larger than xai's it's funny like xai didn't have a great brand going into this it's hard to imagine they would look at any company and think that brand has a worse brand than ours but i guess cursor does have like a great brand so it makes sense how are staff feeling about the acquisition right now uh at cursor yeah i think feelings are mixed i think there there's some people who are maybe excited to partner with spacex you know they have the ipo you know it's a it's a big company right now but there are also a lot of people who are concerned about the acquisition i talked to some people who you know they looked at elon musk's acquisition of Twitter, you know, when he cut three quarters of the company.

3:57And, you know, there's obviously concerns about cuts, I think, with any acquisition. But with Elon Musk, he's kind of known for that. So that's definitely a concern. Yeah, he's kind of hard to work for at best. And then when he acquires your company, it can be a little ruthless, I guess. How is Cursor changing ahead of the acquisition right now? Yeah, so Cursor is, you know, trying to pivot to become more of a generalized AI company. They're working on this Claude Cowork competitor, which is internally being called Sand, and that they've rolled out and are testing. They're also looking at some chatbots.

4:33They're looking at different ways they can push beyond coding. Michael internally has kind of characterized this as something that customers have been pushing for, but it is interesting that some of the goals he outlined in these meetings with staff are very similar to goals that Elon Musk has outlined at XAI. So it's kind of similar how like right now they're in parallel, but like they're aligned on a lot of things. Sand is kind of a funny codename for your upcoming really exciting new product. Sort of like naming it like dirt, but I guess it's kind of cool. It's like people joke about AI chips being just like sand that we melted down and taught how to think.

5:07So maybe there's something kind of appealing about that. Yeah, I think the codenames are always interesting because like, you know, there was garlic at one of the companies. There was avocado at Meta. Yeah, it's like a technique. Yeah. So what do antitrust rules require for a company like this that's pending an acquisition? They're in this awkward spot where they want to do all this work together. They're probably making plans, but the deal hasn't totally gone through yet. Yeah, it's really interesting. So ordinarily, you know, they wouldn't really be able to work together until the deal went through.

5:42But because they have this separate partnership that includes working on models and other AI products together, it's very complicated. I think right now they're trying to keep separate, but there's also a lot of ways that they're working together very closely because of that other partnership. Is Elon taking a big role in the acquisition himself? So initially I'd heard at XAI, Elon Musk was meeting with Michael. He was meeting with Amon, you know, the co-founders of Cursor. And, you know, Michael has been seen at XAI's office. But on the Cursor side, you know, they're not seeing Elon Musk as much.

6:18He hasn't really met with rank and file or anything like that. He has met with some higher ups at Cursor. So it seems like right now, like Elon isn't as involved, you know, and it probably won't be until the deal goes through. Okay. Do we know yet how Cursor hopes to fit into the bigger picture at XAI? I mean, And if it's making all of the models itself, it's making these products, does that leave any room for Grok? Yeah, I think that's something we'll kind of like wait to be seen. Right now, it seems like Cursor is really building up their sales team and trying to sell themselves on being the enterprise.

6:55They've added 60 % to their GTM team since April, so that seems to be something they're really bulking up on. I also think, you know, they're going to be a huge part of the coding push, obviously, with their data that they're coming in with and their expertise that way. For sure. Well, it makes sense that Cursor would maintain a strong focus on coding, even as its purview expands to include other kinds of models and products. So thanks for coming on, Grace, and breaking that down for us. Thanks. The information published exclusive reporting that more startups are turning to newer cloud providers to secure the NVIDIA GPUs they need during the ongoing AI compute crunch.

7:34Our Amazon reporter, Catherine Perloff, wrote about why startups are choosing these newer providers over AWS. Catherine joins me now with the details. Welcome on the show, Catherine. Hi, Rocket. So you opened up your piece talking about an open source AI developer, RC, which had committed$8 million to AWS, but couldn't get the NVIDIA chips they needed. How common is this issue for AI startups right now? I think, you know, it is a, it's not an uncommon problem throughout the industry. And it's, you know, not just with AWS. AWS. The, you know, NVIDIA chips are hard to come by and there are capacity issues everywhere, even at some of the NeoClouds.

8:22But, you know, the startups I talked to found that it was just really hard to get the NVIDIA chips they needed at AWS. Sometimes that they were just too expensive or they were only available in sort of bigger chunks or bigger commitments than they needed and they wanted a kind of more flexible arrangement. And there are some other reasons that companies might find a NeoCloud or another type of cloud startup more suitable for their needs. But yeah, in the hunt for capacity, AWS doesn't always have it. I see. So it's not always that these startups are getting turned away at the door and being told there's literally no capacity.

9:08Sometimes we say, well, we have capacity, but if you want it, you're going to have to accept these terms that maybe aren't what you're looking for. Yeah, I think it's like, you know, it's like sometimes it's like, OK, do we is there capacity? And it's like, no, but you look tomorrow, maybe there's something, but it's too expensive. Or it's like if you want exactly what you need, you have to buy it for like a year and you don't have the money for that. So it's kind of a combination of things. You know, AWS says like they don't have a minimum. And if you want to buy one chip, you can buy one chip or, you know, rent, rent, buy compute for one chip.

9:41You can. So I think, you know, like I think the options exist, but, you know, in practice, they're not always what a startup might be looking for. Yeah. You wrote that Nebius, one of these neoclouds, is getting 75 % of its business from startups who have already maxed out at the big cloud firms. Is that right? So what are the pros and cons of switching to a NeoCloud from a big provider like AWS? I think, you know, yeah, it's interesting. So in that kind of like context, so 75%, you know, they've already tried to use Microsoft, AWS, Google, and they can't get any more capacity, so they go to NetBS.

10:23I think, you know, the pro of a NeoCloud or sort of like these infrastructure providers that I don't know if you'd quite call them neoclads, like these together and RunPod. They're also sort of in the conversation. They sort of offer more like inference as a service, but people are still running workloads on those companies instead of AWS. The advantage is, you know, sometimes it's easier to find capacity there. And sometimes they can be a lot more flexible in what they can offer, or they're more likely to sort of do a deal with the startup. on the fly. Another advantage I've heard is that on that flexibility, I talked to one startup that kind of specializes in post-training.

11:06And they said, you know, a lot of the current cloud architecture doesn't really work for us, but we were able to kind of like work with run potted together to kind of create an environment that was better for us. So I think it's just like a bit more customizability and flexibility. You know, having said that, But the NeoClouds are not a panacea and the NVIDIA chips are hard to come by. So sometimes the NeoClouds can be more expensive. They're not always cheaper. Sometimes they have capacity issues.

11:40But I think it's more just sort of flexibility and at times a cheaper price, which can also just come from being able to rent less upfront. Yeah, still that flexibility is probably really appealing to a lot of startups, even if there wasn't this dramatic compute crunch where they were getting turned away from larger providers. But you asked AWS, of course, what do they think about this? And they gave you this statement that like, frankly, to me, Red is a little bit defensive, but they said that, you know, a few anecdotes does not make a trend. So they pushed back on the idea that there was a trend happening here.

12:12What did you think of that? Is it a trend in your opinion? You know, I think that, But it's undeniable that AWS is doing well. They accelerated their growth to 28%, four percentage points from the prior quarter. And sometimes some of the rise of some of these upstarts might benefit AWS because some of the infrastructure as a service provider, some other cloud startups like Vercel and Render run on AWS. So it's a complicated story. And a lot of top startups do use AWS, obviously, OpenAI, Amphoropic, if you want to call them startups. You know, they do. So I think it's a complicated picture. But I also think it's undeniable that a lot of these neoclouds and, you know, infrastructure providers are growing.

13:06We report on their, you know, funding rounds and revenue in, you know, some of my colleagues do. I feel like AI infrastructure might be one of the hottest areas of venture investing right now. So, you know, if these companies are growing, if there's a thesis behind them, you know, something has to be propelling that growth. So, and, you know, I talk to a lot of startups and that, you know, they're seeing these as, you know, viable alternatives for at least, you know, some of their workloads. Right. Well, it seems like the compute crunch is here at least for a while longer. So I think we're going to continue to have to learn about these dynamics.

13:42So thanks for explaining the trade-offs to us about sticking with AWS or going to a neocloud. It's very helpful. Thanks again, Catherine. Thank you. All right, Akash, back over to you.

14:02I'm here with Daniele Magadzani, Chief AI Officer at UBS. Daniele, thank you so much for having us. Really appreciate it. Thanks for having me. So before you were at UBS, you were a professor of AI. And so I want to get into some of the technical topics of AI with you. My first question is, what are the technical challenges that AI researchers have yet to solve right now? Yeah, it's a good question. There are a number. I would say what is critical for the actual implementation of this AI in critical processes is the ability to trust this model. Therefore, being able to assess the reliability and accuracy of this model is very important.

14:43And from a technical point of view, there is a lot of research going on on evals, which you can think about them as test cases that given different questions, assess the answers. However, given that this AI, as we know, is probabilistic, it's a type of statistical assessment. It's a hard problem, but it's doable, and there is a lot of research going on. The other thing is that you can actually mathematically prove that the agents are doing what they are supposed to do. It's not easy, it's hard, but it's doable. And I believe that that research in that direction will be very, very valuable and very important.

15:20So let's get into both of these challenges then one by one. So evals, I mean, this is the the effectiveness of the models themselves. We have these benchmarks. There's a question around which benchmark is the best, how do we even benchmark the models against each other? What are the challenges with that? I mean, here at the information we've written about, for example, when the models know that they're being evaluated, they can pretend to be better than they are, I guess. I've never done an eval before. What are the other flavor of challenges with respect to evals and models coming up? Look, I would say one challenge is that When it comes to business, use of AI, the knowledge to evaluate the models is actually in the human experience, human knowledge.

16:02Subjective. Correct. And based on, again, the expertise and the experience that humans doing the job, you know, had during the year. So the challenge is how do you translate the human knowledge and human expertise into a machine-readable, you know, format so that they can actually use to assess the model. That's definitely one challenge. And as you said, you can fool the model in a different way than you fool humans. So it's hard, but it's very important. And for this, I guess you really need AI researchers working very closely with the domain expert in the business. And then the effectiveness of the agents, the mathematical proof that they're doing what you hope for them to do.

16:44I mean, is that not just another flavor of eval? I mean, what are the questions that are coming up with respect to agent effectiveness? I guess one difference is that in regards you assess one LLM into a specific task, whereby in the mathematical proof, you really want to assess the behavior of an agent across multiple tasks that this agent is tasked to. Again, it's a hard problem. It's not easy, but it's doable. So more and more research. And just so I understand this, I mean, is this the idea that chance alone wouldn't have been able to produce that result? Is that what you mean by mathematical proof?

17:23Correct. So it's the difference between a statistical assessment versus a proof of correctness. And by the way, verifying the correctness of AI systems is an old topic in academia and in research before the LLMs. Therefore, I guess one opportunity is to leverage what has been done before and see how, if and how, can be adapted to deal with this new generative AI. I want to ask you about safety. Is safety keeping pace with the innovation and AI models right now? I guess the focus on safety remains extremely high. Also for UBS, for example, trust is paramount, is the key element of how we work. So also when it comes to AI, we want to make sure that we can trust the AI we use and therefore our clients can trust the way we use AI.

18:15Of course, as you mentioned, the pace is very rapid evolving, so we need to catch up. But also So that's why we want to go faster, we want to accelerate, but without compromising in safety. So that's where there is always a gap between what you have outside versus what we can deploy in production in a bank. Right. That's because we care about safety. And so here at UBS, I mean, you guys are a bank, financial services, financial institutions. This is one of the most regulated industries that there is. Are you using the frontier models? Are you using open source models? Do you have to wait a year to work with these models?

18:52And what's your process? So first of all, we really want to diversify. So we use frontier models as well as small language models, which are good for a growing number of tasks. My view is that you do not need a frontier model for a non-frontier problem. Therefore, one clear strategy we have at UBS is to make sure that our colleagues use the right model for the right questions or for the right tasks they want to solve. And is this generating recommendations for clients? Is this coding your own tools internally? What do you end up using these models for? Well, a couple of use cases I'm happy to share is that when for sure we have a number of AI generated insights to better serve our clients, our focus is always how we can use also AI to better serve our clients.

19:40It's still a business around trust, but AI can help people to provide better insights. Of course, we use AI when it comes to developer productivity. We are using it a lot. And when it comes to research, the way we consume information, we produce information is really enhanced by the AI. The way we see AI is really how to empower people to work in an even smarter way to better serve our clients. That's our focus. And you mentioned small language models. Is that the dominant way for how you optimize for costs of AI as well? Yeah, look, it's important from a cost perspective as well as for a sustainability point of view.

20:21You really don't want to use the latest model for very, very simple questions. So you want to leverage, you know, the appropriate model for the question you have. And to do so, the strategy is to have, you know, a proper model garden where you host a number of models. and then to have an AI that can help root the right question to the right model. We don't expect everyone to select the models, but we want to do them for them in a very optimized way. I wonder, as you've led the process of getting UBS internally to adopt AI more and more, I mean, what's been a cheat code for you to adoption? We hear a lot about the cultural changes that have to happen in terms of using AI, teaching people how to use them, building trust.

21:06How do you actually do that? I mean, is this running webinars? It's a good question. Is this people holding your hand? Do you have forward deployed engineers on the ground sitting next to associates saying, hey, maybe you should do it this way, that way? What do you do? First of all, it's a good question. And I always say it's people first challenge and opportunity. As important and powerful the technology is, it's really the willingness of our colleagues, ourselves, to willing to learn how to use this technology. I believe a few things I noticed in terms of what is that slowdown adoption is, first of all, maybe people tried models a year ago, weren't so good, but now models got better over time, so they really need to go back.

21:51The other thing, people need to, we all need to invest time to learn how to use this AI. A pattern I've seen constantly is that as soon as anyone finds a very productive way of using AI, they cannot stop using it. So one initiative we started this week actually is what we call AI Power Hour, whereby we really want every UBS employee to be able to invest one hour a day a week to learn about AI by doing. And we have a number of resources, including agents that can teach you how to use AI. So the idea of investing time to learn about AI is something which is very good for the company as well as for each individual in the company because they learn.

22:31But it's really about the willingness of people to want to learn. It's non-negotiable anymore, I guess. Right. Let me ask you one last question. Again, back to your background as a professor. Yeah. Recursive self-improvement. Does it scare you? Does it excite you? Where do you land on it? Look, it's both, I would say. Exciting and scaring. I do believe that given the great focus around guardrails, controls, validation, the exciting part wins over the scaring part. But I totally agree that remaining careful and cautious, also aware of the limitation and the risk is absolutely critical. Great. Well, Daniele, I want to thank you for joining us.

23:17Thank you very much. That is Daniele Magazzini, the Chief AI Officer at UBS here on TI-TV.

23:34I'm here with Carl Kirstead, head of AI and software research at UBS. Carl, thank you so much for having us. I really appreciate it. Thank you for coming. So, Carl, I want to start with the software landscape, broadly speaking. I was looking at a basket of about 80 software companies that I track regularly this morning. That index is down about 26 % in the last year. It's up 26 % in the last three months, though. Yeah. And so with that in mind, give us the overview here. Where is the software sector at right now? Are we still amidst the SaaS-pocalypse? What's the story? We are. I would say investor sentiment on at least the SaaS or application software stocks remains very depressed.

24:20The stocks have had a bit of a rebound, I'd say 20-ish percent off the bottoms in late April. But I think that's primarily a function of a fade in the semis trade and a broader portfolio rotation into defensive cheaper stocks. I don't think it's because investors are picking up evidence of a fundamental improvement in the application software space. So at least the view of our team is that given that the environment for application software still remains tough. Procurement officers are trying to limit their spend. I worry a little bit that the upcoming results across the SaaS space might be a little bit soft and that rally could fade.

25:00Why is that? Well, I think it's partly a function of AI, to be honest with you, where even in the IBM results that were pre-announced this morning, we're seeing pretty strong evidence that rising AI spend inside Fortune 500 enterprises is beginning to crowd out spend. It's crowding out traditional software spend. It's crowding out IT services spend. You can see that in the results of Accenture recently in the Indian firms. And as per the IBM pre-announcement this morning, it's beginning to crowd out non-AI hardware spend. So this is a fairly dominant theme. So I think that's probably the main one.

25:38And then I think the secondary one is that a lot of the traditional SaaS firms have fairly mature end markets now. Now, there are pockets of the software space that are doing relatively well. So as you know, Snowflake, Datadog, those things. Anything backend seems, you know, we've been talking about this on the show, this idea that if you can build the backend yourself and create your own application layer using AI, that seems to be the formula that everyone is going for, which is good news. Or put another way, the bottom of the tech stack. Infrastructure, data, cybersecurity is doing fairly well.

26:15The layer on top, the application layer has struggled. And fundamentally, that's been the call of our team for the last year to summarize long infra data security, cautious apps. So what are investors looking for then? Is it just growth rates going up? Is that the number one priority? Yeah, at least when I talk to large, long-only investors, hedge funds, and I ask them the question you just asked me. What do you look like? The consistent answer is we need to see a bending of the growth curve. So that's what we're all waiting for. I want to ask you about a couple specific names. Microsoft, the stock is down about 20 % in the past year.