How Microsoft is Fighting OpenAI’s Frontier, Waymo’s New World Model, & OpenClaw Craze

0:13Welcome everyone to the information's TI TV. My name is Akash Pasrich. It is Monday, February 9th. First up today, we are unpacking our exclusive reporting about how Microsoft is responding to OpenAI's latest Frontier agents-focused product release. We'll then shift gears and bring on an engineer from Waymo to discuss the company's new simulation tool called the Waymo World Model. And we'll wrap the show with a look at how OpenClaw marked yet another inflection point in the story of what AI will be able to do for people. Our AI and robotics reporter, Rocket Drew, will join us to break it all down.

0:50It's going to be a fun show, so let's get right on into it. Microsoft is organizing itself to better compete against OpenAI's latest Frontier product. The information has exclusive reporting of the talking points that Judson Althoff, Microsoft's commercial CEO, is giving his sales teams with respect to how he sees their products being better than OpenAI's new tools. This, of course, is central to the threat that many investors are afraid that AI could pose for old-school enterprise software companies. I want to bring on Kevin McLaughlin and Aaron Holmes to help us unpack their reporting. Kevin and Aaron, welcome back to the show.

1:25It's great to have you on. Thank you. So, Kevin, help us understand, what is OpenAI's new Frontier product? It's essentially a collection of existing products like ChatGPT Enterprise, the ChatGPT Atlas browser. What's new or what's sort of the differentiation piece here is the use of forward deployed engineers, which are consultants that go in and have engineering expertise and they know every product in the enterprise data center and they're kind of like the white glove team that comes in and helps companies get up and running with OpenAI. This is an area where pretty much every other enterprise software provider is, they're all using forward deployed engineers.

2:10Palantir is sort of credited with pioneering this approach. But that's where the rubber meets the road in terms of the experience that customers are going to have with OpenAI and how it's going to be different from the experience that they've had with Microsoft and all the other companies out there. So, you know, there's a good diagram that OpenAI put out to sort of lay out how it sees its frontier product fitting into the existing software stack. And we've included in the story. But, Kevin, so am I understanding you correctly here that it's really not so much about the software? The frontier is really about giving customers, enterprise customers, the support they need these forward-to-put engineers to sort of implement all this software.

2:53Yeah, I think that my take at least is that this is a bundle and this is the way that enterprise AI companies or enterprise software companies have always gone to market. Like customers like to buy a bundle. They like to have the ability to use multiple different products if they want. And more importantly, you sell this bundle as a single product SKU, which makes it easier for accounting and for financial forecasting and things like that. So I think that the important thing is not so much the product frontier, but what it represents. And that is ChatGPT competing against companies like Microsoft and, like I said, all the other companies out there that it also sells AI to.

3:36And so it's kind of an interesting situation. It's certainly not unusual in enterprise software for this dynamic to be happening. But it is the first time OpenAI will have it happening to them. So Aaron, where exactly does this compete with Microsoft? Is this competing against Copilot? How do they overlap? Yeah, in a sense it is. I mean, I think what both Microsoft and OpenAI are betting on here is that we're going to move to a future where instead of a white collar worker using their office applications and using their Salesforce applications and Slack and all of these different pieces of software, they're instead going to use, you know, maybe one interface that looks like a chat bot or looks like, you know, a dashboard and just simply tell AI agents to carry out actions in those different pieces of software.

4:23And so I think what the next battle is going to be that we're seeing these companies gearing up for is who gets to own that platform or that interface where you control all of these AI agents. And so, you know, what Microsoft and what Judson Altsop is telling staff is we are a better position for that because we already serve, you know, different AI labs, agents and models on Azure. And we also can promise enterprises that we can guarantee cybersecurity and compliance and all of these things that chief information security officers are all worried about. At the same time, Microsoft hasn't really put out its own AI agents that have broken through in a real way.

5:03They've mostly been relying on models from OpenAI and from Anthropoc to power their own AI products. So OpenAI has more of a claim to actually be the one building these capabilities that can automate work in new use cases. Kevin, with these two sort of opposing sides of the story here, the idea that Microsoft agents haven't really broken through yet, yet they obviously have a lot of existing customers using their applications. I mean, which of these two in your mind, after covering enterprise software for as long as you have, which of those two are the bigger leg up right now? In the enterprise, experience usually is the thing that matters most because the experience is built on trust, which is built over years of serving customers.

5:50And more importantly, a lot of the trust is built in how enterprise customers respond when there are problems and there are inevitably problems. And so I think that Microsoft and all the other old guard enterprise companies probably look at OpenAI coming in and they're sort of privately behind closed doors saying, good luck, because you're going to have a lot of learning to do in this space. So you think having the existing customers, I mean, that is really, I mean, that's a significant advantage that OpenAI is going to have trouble competing with. You got to take your lumps as an enterprise vendor.

6:26And OpenAI has not done that yet. It's not to say that they're not going to get there, but like I said, my guess is that, I mean, you can see it in the email that we reported on from Microsoft on Friday. You know, the message was really like, hey, you know, this is a great new competitor for us, Microsoft speaking here. But we got this, we got the experience and we got all the things that they're going to have to build. Aaron, in your reporting, I wonder, is it getting more and more important for leaders at these companies to respond to every single product release that customers have? I mean, you know, I don't know if this was always the case, but obviously every new model, every new agent or tool, it seems to drastically change the narrative around who is leading.

7:12I mean, is this common for leaders to respond to each and every product release that competitors put out there? It's definitely not common, and I think that it speaks to the level of potentially fear that a lot of people in legacy software have around the new AI products. I mean, last month, Anthropic released Cowork, which is another sort of AI agent product that aims to automate white-collar work. And we've seen every enterprise software stock sell off dramatically in the weeks following that. I think that especially that the graphic that you mentioned that OpenAI put on their Frontier announcement, to a lot of people who are in existing enterprise software companies, that came across as a shot across the bow because OpenAI is basically saying, you know, we will provide the agents, we'll provide the orchestration tools that you use to control those agents.

8:03And every other software company is just this, you know, dumb system of record at the bottom of your stack that our agents are just using to draw data from. So I think that we are seeing like a heightened level of competition between the AI labs and the existing enterprise software players. It's not really clear whether, you know, customers are actually changing their habits yet, but people can see that OpenAI and Anthropic are moving in on that territory. Evan, last question for you. I mean, the word that keeps coming up, not just from Microsoft, but all these enterprise software companies is platform.

8:38That is the big competitive advantage that a lot of them tout. And maybe this relates to the conversation we were just having two minutes ago about the bundle and how long they've been in the game here. But how sustainable a competitive advantage is this platform play that all these enterprise software companies seem to have? And moreover, I mean, how do you compare the relative strengths of these platforms when every software company seems to tout their own? Well, I think it's the benefits of being a platformer are pretty well established, not just in enterprise software, but broadly. So I won't go into that.

9:16But I think what we're seeing right now is in enterprise AI, everybody wants to be the platform for building AI agents and the place where customers store their data and build applications and do everything else. The problem is not everybody can be the platform. And so we're looking at a competition for hearts and minds in the Fortune 500. And, yeah, it'll be interesting to see. I think that you could say that the enterprise software providers have an advantage just from their decades of experience. But AI is a very different animal than enterprise software. And so we don't really know what's going to happen.

9:54Maybe customers will like it, open AI's pitch, and they'll just want to work with them. But if past history is any guide, as I said before, I think that this could be a really intense competition. And Kevin, one more thing before you go. i i've always wanted to ask you this i never had time to but platform in in your words and kevin mclaughlin's words is it is it just one-stop shop is that really what they mean here yeah i mean the the opposite of platform is in industry parlance or enterprise tech parlance it's called uh point solutions which is you're buying one product from this vendor one product from that vendor one product from the other vendor which actually might be a cheaper way and sometimes to do it and over time companies have experimented with this approach but what they find is you have to stitch together product a product b and product c and in order to do that it's expensive coding it's expensive integration work and uh oftentimes what happens is the the companies that try this approach end up going back into the welcoming arms of the platform provider but again ai might be different we'll have to see what happens great well kevin and aaron i want to thank you for coming on that is Kevin McLaughlin, our enterprise software reporter, and Aaron Holmes, our Microsoft reporter, here at The Information.

11:12Waymo released a new simulation tool called Waymo World Model. It leans on a model called Genie 3, created by Waymo's parent company, Google. To help us understand what all of this technology is, I want to bring on Vincent Vanuc, distinguished engineer at Waymo. Vincent, welcome to TI-TV. It's great to have you here. Thanks for having me. Okay, So you have to help us understand what is the Waymo world model? What is Genie 3? How do they relate to each other? Break it all down for us. Yeah, a world model, you can think of it as an AI-generated video game. It's a video game, but one that's so realistic that you can actually put a virtual Waymo car inside the video game and the car will behave exactly like it would on the road.

11:59And that has a lot of advantages. We can simulate any scenario that we want safely, virtually. The benefit of the Waymo world model today is that by leaning on Genie 3, which is a video generative model from Google DeepMind, we can generate any scenarios that we want that we may not have seen or may have seen very rarely in the real world. A great example is, imagine you're at Halloween night in San Francisco, and there is somebody in a T-Rex costume walking down the street. We may have encountered that once, but we want to make sure that we understand the situation well and that we handled it well.

12:44The Genie 3 backbone enables us to generate those kinds of situations with very high realism, both physical and graphic realism. And so help me understand where this fits into the company's operation. So, I mean, you have the Waymos out there and presumably, I mean, the way I understand it is that they're also, they're collecting data too and sort of further training the model as they roam around. And that's sort of one way of improving the system in perpetuity. Now, basically, what I understand you saying is that you have this other lane of training these models, which is you can, you know, turn on the model with a flick a button and say, hey, we want to continue to train it artificially.

13:29And so this is a parallel lane that you're pursuing. Yeah, it's both for training and evaluation. We run billions of miles of simulations all the time, in addition to driving in the real world. We want to simulate the things that, you know, we hope to never observe in the real world, but that we have to be very defensive against, things like accident scenarios or extreme weather conditions. So, for example, you can take all the miles that we've driven in the real world in daytime and just flip a switch and say, now drive it at nighttime and see that the car behaves in the same manner and is equally safe in those conditions.

14:11Now, is this something that Waymo is going to keep in-house? Is it something that you plan to sell externally to other customers, other companies developing autonomous vehicles? How do you think about it? This is an in-house development at this time. It's very load-bearing on our operations, being able to understand the world, even in the extreme long-tail cases. and so it's integrated part of developing the driver and evolving it and testing it for every release that we put out there in the cars you know i i sat in a way more for the first time a couple weeks ago in san francisco i hadn't done it it was it was what i call a spiritual experience in some ways because it's you're just sitting there and and all your senses are are on on high alert, really.

15:02But, you know, this was in San Francisco, and obviously there are a couple cities where it's been rolled out now. But what are the technical limitations right now that Waymo still needs to overcome to be able to get these Waymos, you know, comfortable, I guess, to navigate all these certain all these sorts of environments? Is it the idea that they just haven't been trained yet on the environments? Or what is left to do? No, we've seen that the Waymo driver actually generalizes this very well to many of the new cities that we've deployed in. And there are still areas that we want to improve, like driving snow, for example, is a very different exercise than compared to driving in fair weather.

15:47And so we've been focusing a lot on the kind of improvements that are needed to understand what it looks like to drive in snow effectively. We've recently launched freeways, and so we are now able to drive in freeways. That is slightly different from driving on surface streets on a freeway. You cannot just stop. You have to be safe and pull over when there is any incident in front of you. So you have to really understand the dynamics of driving in those kinds of environments. But in general, we found that the Waymo Driver has been generalizing very well, and it's been a matter of expanding our operations, doing it responsibly, doing it with the right guardrails in place, and understanding the environment that we move into well enough that we have the right level of confidence that we can operate safely and with good quality of service for our users.

16:52So in other words, you're saying you don't necessarily have to go city by city or region by region. I mean, right now, as it is, it would be comfortable driving in any one of these environments. Yeah, the Wynmo driver is the same in all the cities that we've deployed thus far. And it has proven to be very robust and reliable to all the different environments that we've been able to deploy in thus far. And so is the only reason today that it's not, you know, in more cities, is that just a regulation rule or is it so there's no limitation on technology the way I understand? It's very multifaceted in the sense that there are regulatory compliance that we need to go through in some areas.

17:39There is operational constraints. There is supply constraints. There is a number of different factors that come into play. But in general, we're moving as fast as we can to bring the benefits of the Waymo driver to as many cities as possible and hopefully save many lives in the process. And last question for you, you know, in order to be able to scale these operations, as you talked about, I presume things will have to get cheaper along the way. And, you know, maybe the equipment that's needed to outfit the cars with the Waymo driver, you know, becomes less and less bulky, you know, less material.

18:22I don't know what it is, but I'm sure cheaper is part of the equation. How do you get there in your mind? Cheaper is definitely part of the equation. Things like the Waymo World Model is part of that equation. Being able to drive a lot of simulated miles instead of having to drive there physically in the real world. All this testing that has to happen can be done at scale much cheaply on GPUs than it would have to be done in the real world. So there's different axis of improving the total cost of operation of the systems and building that kind of technology is a very central part of this. Great.

19:04Well, Vincent, I want to thank you for coming on. That is Vincent Vanouk from Waymo here on TI TV. OpenClaw once again seemed to be another inflection point in the story of what AI will be able to do for people. And its release also sparked a wave of developers and tech enthusiasts building agents quicker than they ever could have imagined. Our AI and robotics reporter rocket drew wrote a deep dive into that craze what it means for tech but also what it doesn't and i want to bring him on to talk all about it rocket welcome back to the show it's great to have you here hey akash thanks for having me so previously on the rocket drew show you were explaining to us what open claw was and look you told it what it was but the thing is it changes every two days the definition expands what is open claw today that's right that's right and it's subtle.

19:55It's definitely subtle because there have been agents before, right? We've been hearing about agents for a year, but usually we hear about them in the context of an agent to fix your Excel sheet in Microsoft or a coding agent like Claude Code or Opening Eyes Codex that mostly writes software. In contrast, OpenClaw is a personal AI agent. So what does that mean? First, it's open source. So anyone can download it. Developers can modify it and tinker with it. Second, it can take control over your whole computer. So it can have access to all of your files. You can give it your credit card if you're crazy enough to do that, which some people are.

20:30And it can orchestrate multiple agents. So it can pull on models from different providers, open source models. You can pull in OpenAI's models, Anthropics models, and you can have them all work together to accomplish tasks on the internet and across your whole computer. So it's a personal AI agent. And so this was one thing that I wanted to sort of clarify with you, which is that is OpenClaw the agent or is it the tool that you can use to create other agents? Yeah, that's a great question. You know, maybe the language is still getting ironed out here. I think if we're getting technical, it's a framework.

21:01It's an open source framework for creating agents. And one way you see this really clearly is that people will have multiple OpenClaw agents. They'll say, this is my therapist agent. This is my chief of staff agent. This is my information security agent and you know then the agents sort of have their own identities let's say their own personalities um but also people will refer to it as my open call agent went and did this so people are still playing it a little fast and loose with the language got it okay okay so it's the general foundation the general framework it's open source people can use it they can customize it they can they can do what they want with it uh inevitably so it's it's been less than two weeks now since this framework was released and everyone's talking about it.

21:44You spent most of the last week talking to people who are using it, people who are competing against it. What are the discussions that you had with folks and what are the reactions to it? Yeah, well, you called it an inflection point. I think that's exactly right. I think people don't know, is this the personal agent that's going to end up in everyone's pocket or is it going to be someone else. But either way, it's a very strong proof of concept. It has sort of lit the way for people. It's shown people what's possible in the realm of AI agents. So I've talked to some startups that considered themselves in the agent business, and even OpenClaw has been a wake-up call for them.

22:22It helps them have something concrete to point to and say, oh, this, this is what we've been trying to do all along. They can point to, they can point their investors to it, they can point customers to it, and they say, we're OpenClaw, but for X, but for a certain kind of user. We're open claw with more security. We're open claw, but we're trying to get into these businesses. So I think it really is just sort of a clarifying moment for the industry about what personal AI agents can be. It also shows that people are quick to adopt this kind of really unproven technology and see what it's capable of.

22:53I mean, if you ever use some of the alternatives out there, like Claude Code, Claude Cowork, there are a lot of guardrails built in. You know, Anthropic knows that the internet is an unsafe place to send an agent sometimes. It might get hacked. It might expose your personal information. So they try to be really careful about it. OpenClaw kind of said, YOLO, do whatever you want. And people were quick to embrace the kind of risks and freedom that come with that. So yeah, I think just a really strong proof of concept. But clarifying why, is it because it worked better than other agents in the past?

23:26Is that the idea? Yeah, I think it comes down to the sort of unfettered access people were giving it to their personal data and to all of their machines. Right. That people were willing, they were willing to say, do whatever you want. Go crazy. That's right. People are giving it access to their workplace data saying, go, you know, check in with this team, scan my emails from the past week, look at the usage data and come up with a marketing campaign for my company. Then go off and draft the marketing materials, create the actual, the ad images that we're going to use. These sort of end-to-end things that pull in potentially very sensitive data at every step of the way.

24:02I think still my favorite example of how I've heard someone use OpenClaw is they set up a monitoring feed for their infant who's in the NICU. They wanted to know, based on this video feed, if my child's vitals change, I want to be notified of that. And of course, that's a very sensitive stream of data. That's a very confidential source of a video feed. But this person used it to great effect, and they did that using OpenClaw. So I think that's the source of the power is the access and control that people are giving it. And is anybody concerned about the cybersecurity implications of all this? Yes, I think people are very concerned.

24:36I think you are not going to see any sort of chief information security officer at a large company in their right mind roll this out to their employees. Now, I think a lot of them are experimenting with it on their free time. And a lot of them are even encouraging some engineers to experiment with it on their free time. But it's not at that point yet where we have enterprise level security. And so the people who are using it most effectively are being very careful. They're taking a lot of precautions along the way. They're setting it up in a virtual machine. They're giving it very strict permissions and then only gradually loosening those permissions as they trust it more.

25:09They're not giving it access to their full bank account. Maybe at best they'll give it one of these Visa gift cards with 20 bucks on it. So that's the most money that it could lose for them. So people are using it very carefully. But to be careful requires a lot of technical expertise, which further limits who can use this thing effectively right now. Now, you liken this moment to the moment that meme stocks really came into popularity and also NFTs in some ways. Why did this remind you of that? Yeah. Well, Akash, I have to say, I've heard this, I've heard OpenClock compared to any technology you can imagine.

25:46Yeah, okay, true. I've heard people say it's like the first time you got on a Waymo. I've heard people say it's like the first time you picked up an iPhone. I've heard it called a civilization level step change. Okay, so I've heard everything. But for me, it really reminded me of the meme coin and NFT craze. One, because there's a direct crypto angle. There's like people are pushing how can crypto be involved? How can these agents use crypto for their transactions on the web? But two, just the level of enthusiasm around it. It almost like defies logic. Everyone has to be using it. Everyone is talking about it.

26:19Just the way it has sort of become this overnight sensation. And people almost have to experiment with it because the craze has gotten so large. Okay, now speaking of the craze, there was a part in your story where you said you actually participated in some ways in the craze. And you wrote in the story, I even listed myself on rentahuman.ai, a marketplace where agents can rent my body for meat space tasks. What does that mean, Rock? That's their language, not mine. Yeah, you did it. You wrote it. I mean, you... But meatspace tasks is something I pulled from them. It's one of these spinoff sites.

27:02So if people saw a site went viral, they called it Maltbook. It's sort of a Reddit, but for AI agents. There have been a lot of these spinoff sites that are trying to ride the coattails of Maltbook. So you have Maltmatch, which is Tinder for agents. You have Molt Secret, where the molts can confess their innermost secrets to each other. And then you have Rent-A-Human.ai, where they can, it's like TaskRabbit. But instead of humans hiring, you know, someone to do something on the web for them, it's an agent hiring a human to do something in real life for them, in the physical space. So the only example I've seen of this so far is that someone was paid$100 to stand out on a street corner holding a sign that says, an AI agent paid me to hold the sign.

27:43And supposedly they made$100 for doing this task. Now, I listed myself at the cheap, cheap rate of$15 an hour, thinking that would be a sure deal. Like an agent has to come along and rent me. But based on the site, it looks like a lot of humans have signed up for these kinds of tasks. So maybe I just wasn't in high enough demand. I said I have a pickup truck. I'm like, I'm sure there's an agent out there that needs to rent my pickup truck, but no leads so far. Sure. And just to clarify here, though, I mean, again, these are people that have created agents and saying, I'm basically trying to trace through where are the people in this equation?