Loading summary
Transcript55 lines
- [00:00]
A
Foreign. Welcome back to AI to roi, the Big Story edition. I'm Ray Reich, founder and CEO of Benchmarket. And joining me as always, is my co host, Peter Buchanan.
- [00:20]
B
I'm Peter Buchanan, I'm the managing partner of New Plan. And Ray, this week we are talking about something that went from 0 to 1.5 million agents in action in two months and has every major AI company scrambling to respond.
- [00:39]
A
Well, sounds like you're talking about OpenClaw. It's interesting. It was a lonely soul Australian developer who built it in about, I think it took more than an hour, but he built it in a few days back in November and it sparked this entire new category of personal AI agents, or as the techies call them, clause. And by the way, for the audience out there, one of my favorite podcasters besides Peter Buchanan and Ray Reich is Lex Friedman and he had an amazing interview with Peter, the founder of openclaw back in February. But be prepared, it's a three hour long podcast, but well worth it. And by the way, just what is openclaw? It's an always on autonomous AI agent that, that's running on your computer and it's executing these multi step tasks without waiting for you. So Peter, you know there's a little bit of a spoiler alert here and that is the open call actually proved something. It proved that there's millions of people out there who want an always on AI assistant right there beside them on their computer. But the harder question for me is whether companies, large enterprises, will actually allow employees to bring their agents to work. It's a little bit like, I still remember my wife worked at JP Morgan and when you first had smartphones come out, they would not let you bring your smartphone to work. Why? Because the security risk are real and in open claw, those security risks are even heightened. So, so let's start with openclaw and actually you know what it is exactly and why Exploded.
- [02:22]
B
All right, so in November 2025, this Austrian fellow, Peter Steinberger, a superstar guy, really well known, he built the first version of openclaw incredibly quickly. He called it claudebot, but then Anthropic called and he changed the name to Molt Bot.
- [02:42]
A
And then it was Molt Bot initially.
- [02:45]
B
Right, right. And then it became Open Clause. So it had three names in about two weeks, but within two months he, you know, he'd renamed it twice and he had one and a half million active agents chunking away. It was a little bit more expensive than he expected when he started building it because to maintain it, it was costing him $10,000 a month. So Sam Altman called him a genius and he also opened his checkbook because he hired Steinberger in mid February to join OpenAI to drive the next generation of personal agents. And they put openclaw into basically one of those open source foundations where anybody can work on it. And then the next closed version from OpenAI, whenever it shows up, it's being led by Mr. Steinberg. So Ray.
- [03:36]
A
What Peter, I can ask you a question because that channel noise said that Peter. And Openclaw was worth about a billion dollars when OpenAI bought him. Has that been confirmed?
- [03:49]
B
I've never seen a number. I mean I was actually quite curious about that because it's only him and he had a couple of helpers so he could go to sleep. That helped keep, keep the instances up. But I've never seen that number and God love him if he got a big number.
- [04:09]
A
Okay, but a lot of our audience maybe doesn't know exactly what Open Clause is at a more detailed level. So let me tell you why it's different. So Open Claw is a self hosted, so it's on your computer, it's not in the cloud like the large language models from OpenAI and Anthropic. It's a personal agent specific to you. It can connect to any major AI model. So it's kind of user option, what model do you want to use? And it takes its commands through messaging apps. I think it started with WhatsApp and they added Telegram, iMessage, Slack and now 20 others. So you actually speak to your agent through the message app you're already using. It's continuously browsing the web, reading and writing files and scheduling and executing tasks. And it actually works when you're not at your computer, so you don't need to keep going back and forth and prompting the LM. It's running 24 by 7 even when you're not there, right?
- [05:13]
B
So you better give it good instructions. And the uptake was fantastic. And it wasn't just in the U.S. so the first thing that you started reading is that smart tech savvy people were saying, well I don't want to put this thing on my computer with all my other stuff, so I'm going to buy another computer. And there was this little boomlet of sales of Mac Minis that were only running openclaws on them. So a few weeks ago in our newsletter we actually had a chart of the Mac Mini boomlet from all tracked by all the retailers through. So we were cautious in the US In China not so much. So Baidu, the big retail and search company, they planned to embed OpenClaw into their main smartphone app. And Tencent actually hosted setup sessions in Shenzhen that drew retirees, students, tech workers. People were really, really excited.
- [06:14]
A
Yeah. And even Peter actually told, and this is Steinberger, not, not. Buchanan told Bloomberg that the dynamic between the US and China was stark. Initially, in the US he felt that some companies, you get fired if you brought Open Claw to work. So I guess you can't call Open Claw your child. And in China there are many companies. Thank you for the little laugh there.
- [06:36]
B
Yeah.
- [06:37]
A
And in China there are so many companies where if you don't bring open cloud work, you may get fired. And that captures attention in the enterprise environment between us and China. And we're going to talk later. Governments have a different perspective than companies, Peter.
- [06:54]
B
They do. And very soon after launch, it was clear that Open Claw wasn't built with enterprise security in mind. So of course this came out, it was a big hit. So naturally security researchers had to rain on the parade, justifiably, as it turns out. And so they sounded the alarm quite quickly. So the Cisco AI security team found that a third party skill could easily exfiltrate data and execute a prompt injection without user awareness. A Northeastern University study found that agents could be manipulated into disabling themselves. So basically hari Kari for agents. And one of openclaw's own maintainers, people who helped Peter Steinberger out, said that it's far too dangerous for non technical users. But Ray, nevertheless, by the end of March 2026, even the Chinese government was restricting Openclaw. So how serious were these risks and are these risks?
- [08:00]
A
Well, from the ccp it was very serious and they actually restricted all their state agencies from running OpenClaw on office computers due to security risk. Now, I don't believe this actually expanded to the private companies, even though we know there's a very cozy relationship between the government and private companies in China. But right now it seems to be restricted. The use seems to be restricted only in governmental and state agencies.
- [08:31]
B
Right. So here's the Central Tension is OpenClaw is an autonomous agent and it's got unrestricted access to the machine and that's simultaneously its best feature. And it's where the security risk comes from.
- [08:46]
A
Yeah, well, when you give an AI agent the ability to execute these shell script or commands in, browse the web, read and write all your files and schedule tasks, all without any human approval, even the let's go do that, and then do I approve that you take the next step, you're creating enormous value. I love having my agent working 24 by 7, but I also don't appreciate the fact that there's still so much risk. So it's really a use at your own caution or peril right now.
- [09:20]
B
Right. But here's the thing that really came out of this. This basically is like putting miracle grow on a market. It's literally the dawn of a market. So we have now we have a personal AI agent market, idc, Gartner, all these companies will track it. They'll be tracking clause actually everybody's calling them claws. And so you know, we, we've talked about how it works, but what are, what are people actually using these agents for? You know, I read, oh, this has made me so much more productive. My life is so much better. But how are people making their work and personal lives better?
- [10:05]
A
Yeah, and I must admit I have not used openclaw yet myself. It's much more technical than cloud cowork and we're going to compare that in a minute. But one of the top use cases are developers are remotely delegating coding tasks, things like push commits and resolve pull requests from their phones. So it makes sense. Highly software technical resources doing it knowledge workers are triaging their email, summarizing documents and managing calendar logistics. By the way, some of the stuff you can do with Claude cowork consumers are using open Claude to order groceries, book their travel, manage their entire household. One was actually one user had his agent managing his nanny's hours and booking dates via text, via imessage and another one's running 50 actually pre scheduled marketing analysis tasks daily at 6am so that they're provided kind of a summary when they start the day. So a lot of different use cases but still most users are highly technical.
- [11:12]
B
Right. And so the hype machine though has started. So Jensen Huang had his GTC last week with his three hour keynote and he said that every enterprise and software company needs an open claw strategy. And Harrison Chase, the founder of LangChain, he was even more direct. He said, I guarantee that every enterprise developer out there wants to put a safe version of openclaw on their computer.
- [11:44]
A
Once again, every enterprise developer, because I would love to have it also as a non developer because every day I've got maybe 20 different tabs open my computer and I'm always doing context switching. Right. I have some things that need to be done every day, some things like my social media posts, et cetera. I could automate all that kind of knowledge Worker Repetition. And these agents could communicate across every channel in a company. Stack, Slack, WhatsApp, et cetera. So I don't think there's a real question of if personal AI agents will be really successful and popular. I think there's two questions. One question is how soon will us mere mortals be able to use them on a regular basis? I think the second question, which is even bigger from an adoption and economic benefit, is whether enterprises will actually allow and be able to deploy these agents in a secure environment.
- [12:46]
B
Right. So let's dive into the vendor landscape because sort of clause without being called claws were already under development. Openclaw showed up. It was really exciting. It had this big red lobster logo. But it's not the only option. Every major AI player either has something that's in the same ballpark or they're going to release a personal AI agent soon. So let's go through the landscape. We'll start with Nvidia's Nemo Claw. So we talked about Nemo Claw in the news newsletter, the big story on Nvidia week before last. And it's basically a security wrapper around openclaw. And it also installs in a single command. So there's three layers. It has the open shell runtime, which basically creates an environment so that you can have commands and give it commands. It has a privacy router which basically protects your data and allows you to communicate with the crowd. And then it connects to Nvidia's open source Nemotron models. And you know, they, that, that's, that's basically. And also there's a governance layer. So. So it feels like Nvidia did this awfully quickly, but I think they just wanted to be ahead of the game. They had GTC coming. Jensen couldn't resist. Nevertheless, it makes progress.
- [14:15]
A
You know, I never knew that Jensen and Nvidia had a little bit of a sense of humor because in the movie Finding Nemo, we almost remember that Nemo had one little fin that was hurt.
- [14:27]
B
Right.
- [14:28]
A
Very well. So now Nvidia is being nice enough to give Nemo a claw instead. So we hate.
- [14:35]
B
Right.
- [14:36]
A
So now that I've gotten my little child movie humor out of the way,
- [14:40]
B
you know, Nvidia, the first one, there might be more, there might be, there might be more.
- [14:45]
A
Nvidia did not go to loan. Now we know that they're partnering across the entire AO ecosystem, AI ecosystem. So with Nemoclaw they have, I think it's almost 20 partners, including Box and Cisco as launch partners. And Box we're showing how they're using it to enable agent workflows which within enterprise class file systems with human matching permission controls. So they tried to address the security issues with that human matching and then Cisco demonstrated a claw that autonomously responded to a zero day vulnerability advisory in under an hour, producing a full audit trail, once again trying to get some Nemo Claw and AI explainability built into the process. So here's my question. I know we covered this on a newsletter edition last week where Nvidia now is the AI maestro, including a full stack platform software platform. So do you think Nvidia can really succeed here with the personal AI agent? I think
- [15:53]
B
it's possible, but they're going to have a lot of competition and they're in many different markets at the same time, not just the software. In fact, they're just becoming a software company now. So they have. The good news for Nvidia is they have technically elite products from the bottom of the stack at the chip layer to the chip software. They've got a great ecosystem of people that support them, that want to use their, particularly their open source models on top of Nvidia infrastructure and software. And so I think that's possible, but I also think they're blockers.
- [16:33]
A
Yeah, well there, there is. And not to get beyond my technical capability, but I know one kind of governance expert told Shio magazine that Nemo Claw, even though it's supposed to provide all the security mechanisms to make a personal agentic AI possible, it still lacks some of the basic safety features that it teams expect. So I don't think it's going to be taking off quickly, not until you can prove who's doing it. How do you go backwards and undo mistakes and have that audit trail of who did what and why and according to what policy or governance? So I think at this point in time Nvidia is not going to be a big winner in the short term. But let's look at another player. That's Perplexity, because I believe they built a personal AI agent also.
- [17:26]
B
Yeah, they actually have a suite of products that are interesting. So you'll remember six months ago Perplexity was the AI search company and it was voted at big conferences as the high growth AI company that was most likely to fail. But no one is voting for that anymore because they have created three products that work together to perform this personal AI function. So the first is a cloud based product called Perplexity Computer and it's basically multimodal orchestration. So if you're building an application, you're A developer building an application and you're working with models and each model has a different purpose inside your application. Well, you can connect to 19 AI models simultaneously and, and do work with them through Perplexity and have them managed and orchestrated to run your application. And price is pretty reasonable. Then there's a product called Personal computer that provides 24. 7. It runs on a Mac and it provides the 24. 7 local file access, compute, et cetera, et cetera. And then there's something called Comet Enterprise, which is an AI native browser that you use for researching, summarizing, completing tasks. And it connects to both of these things and connected together it's pretty powerful both for knowledge workers and developers.
- [18:54]
A
Hey Peter, this sounds like it's almost More geared towards B2B in the enterprise market than the consumer market, is that correct?
- [19:03]
B
I think that's absolutely right. For one thing, it'd be awfully expensive by the time you Comet doesn't cost anything, but you add the other things up and you bought the capacity you needed for demanding workloads, you'd be at least a couple hundred dollars per user. So I think they're on that enterprise track for sure.
- [19:22]
A
Well, their CEO Arvid Srinivas actually summed up it on a philosophic basic. He said a traditional operating system takes instructions, but an AI operating systems take, take subjective. So it's focused on the goals, not the input. And I thought that was interesting.
- [19:41]
B
That is very interesting. So. And Perplexity has some validation. So the Samsung Galaxy phones are the number two family of phones behind iPhones globally. And Samsung has given Perplexity OS level access to power their Bixby agent, which is their equivalent of Siri or Alexa or whatever inside of the Samsung phones. And it's on the new flagship S26 phone. So Ray, why would Perplexity win?
- [20:15]
A
Well, they have made a pretty hard pivot in the last 12 plus months to focus on B2B with this model agnostic architecture. So they're not tied to their LLM or SLM and they have a architecture that was also built for real time search. So I think Perplexity could be a winner, but I'm not predicting that. I still think I'm on the God I love anthropic bandwagon. But I think the fact that they can go and orchestrate between different LLM that others can't match, that could be a differentiator.
- [20:53]
B
Right. So I think the biggest blocker is actually these, these big even better funded companies. Right now Perplexity is valued at $21 billion, which seems like a lot of money until you look at what everybody else is valued at and how much access to capital they have and how much brand trust they have with enterprises. Microsoft, now anthropic, increasingly OpenAI, Google, Amazon, they can walk in at the CIO, CTO, Chief AI officer level. I think that's going to be harder for Perplexity.
- [21:30]
A
Yeah, I agree, but I still think, and I mentioned this just a couple minutes ago, I think Anthropic is really well positioned with their approach, but they're taking a different path. So can you share a little bit more about the path that Anthropic is taking?
- [21:44]
B
Yeah, I can. They don't appear to be building a personal AI project. What they're doing is they're embedding this personal capability into existing products. So Claude code already connects to messaging apps, including Telegram and iMessage. And they released an extension called Dispatch that enhances that. Just this week, Cowork gives Claude direct control of the Mac. So that's clicking, scrolling, navigating operating apps, just as a human would. And Cowork also makes users ask for permission before editing files and running a virtual machine. So it has this basically caution yellow light layer built into it, which is very on brand for Anthropic.
- [22:37]
A
I'm interested. This is the second time we've talked about Mac that people seem to be designing Mac first, Windows second. Just a little bit of aside, but, you know, one of the things about Anthropic is they have been the AI model company primarily focused on B2B since the early days. And I also believe that they really launched the first killer enterprise product out there that's a native. And that was Claude Code I read the other day. I think it scaled to over $2.5 billion very quickly. I remember we were talking about lovable getting to 100 million, 200 million, 2.5 billion, Peter.
- [23:18]
B
In nine months.
- [23:20]
A
Yeah, in nine months. And they also, you know, not too long ago, maybe this was six weeks ago, introduced Claude codework to automate file and task management. And that's kind of a different way to have a personal agent. So I actually love that. In fact, the information reported that Claude is gaining ground in that personal AI agent category because it offers those comparable agentic features to OpenClaw, but with much lower setup friction, better security predictions, and not needing to be a developer.
- [23:54]
B
Right. Plus they have this partnership with Microsoft to integrate cowork into Microsoft 365 copilot. So the new copilot offering, called Wave 3, the official name of the offering is Copilot Coworks, so somebody was bad at marketing there. But anyway, Claude Cowork is the multi step reasoning engine powering the new Cowork feature in Copilot and I seen it's officially available as of yesterday. The hot take reviews are off the charts in terms of positive capability because Copilot Cowork combines models from OpenAI to Anthropic and chooses the one that does the task the best. It's expensive on paper, but it's probably going to be really useful to Microsoft centric customers. If I looked at low friction, high performance, high trust relationship, mid to large size enterprises, the traction that Anthropic's had over the last four to six months in the enterprise. I saw a chart recently that interviewed enterprise customers and they buy tools from all sorts of companies but they said what tool do you buy first to do technical work? And about 66% of them, that was another chart from the information, 66% of them said they bought Claude first. And ChatGPT was in the 30s somewhere. So.
- [25:48]
A
Right, that's a big difference. When we talk about this stuff, our time flies so fast. So I'm going to pivot and try to kind of get to a wrap up here in the next two to three minutes. Peter so here's a question I have have will personal AI agents really land and survive in the enterprise? I'm just going to ask you to tell me what you think.
- [26:13]
B
The honest answer is yes, they are going to do that, but it's sort of going to be like everything else that lands in the enterprise. It's going to be observed, managed, walled off. The data they're going to, you know, companies are going to make sure that the data that goes into these personal AI managers or it's data for the company, not data for the employee. So I mean you've had, I'm sure times in your career where you have data on your laptop and you left it there and you wanted to work in a laptop and someone wanted it and they said why isn't it in SharePoint? Well, the same thing's going to really happen with these AI things. I think that the process of making these personal AI agents productive means that they're going to be put into a stack and also that not every knowledge worker is going to get it because to run some of these personal AI agents you need a lot better hardware and a lot more storage and dedicated infrastructure. So choosing who's going to have access to these personal AI agents In the enterprise, it's going to be a difficult question.
- [27:38]
A
Unlike the SaaS industry where one of the reasons SaaS grew so fast was because it was a decentralized cell central. It didn't have any saying on it because it really didn't touch your infrastructure. This is going to be a slow roll in enterprise because it's going to be an IT first and CISO first approval and that's going to take some time. So kind of to wrap up today. You know, the key takeaways, you know, pioneers get the arrows and settlers get the land. And I've been in visionary categories for so long and I really believe that OpenClaw is a pioneering product and, you know, a great kind of a developer sold both a personal and a market need. Right. He's made and OpenClaw made the promise of personal AI agents real. But the Enterprise version of personal AI agents I still think were, if not months, maybe one to two years away.
- [28:35]
B
Yeah, I'd say that. But I also think that the model that Anthropic and Microsoft and probably Google will pursue is going to be more evolutionary, where the functionality won't matter whether really it's on somebody's device specifically. It'll be more like, I created a place for you to have a sandbox connected into the work environment of the company and you can have your personal space and it'll be more productive and you'll get more work done, but it's all going to be part of the bigger plan. And frankly, the bigger companies understand that plan better than the upstart in this case.
- [29:15]
A
Yeah, I think if you're in a big company, if you're a department executive, if you have a chief AI officer, I think you should start playing with this at home or in that secured development environment area, not touching any production systems and just look at the potential. And, and I think once people see the potential that they're going to want to try to promote it and get the CEO and board maybe to push it and security to move a little bit quicker. So that's my perspective, Peter.
- [29:44]
B
Yeah, I think you're right on target. I think we're really aligned. Just think of your personal AI agents as something that sits next to you, like your email and your calendar and your Slack, and that there are rules around it. And as exciting as it's going to be, well, companies will do their absolute best to take the excitement out of open claws and just turn them into quiet productivity.
- [30:13]
A
Well, you know what I'm going to try to do? I'm going to try to use it here at home to manage my home administration. Hell, maybe I'll even see if I can get it to manage my wife. No, I'm just kidding on that one.
- [30:23]
B
I met your wife. I think it's going to work the other way around. I think she already has an open claw and you don't even know.
- [30:28]
A
Exactly. Well, hey, Peter, thanks so much for joining me on this week's AI to Ri the Big story. And for the listening audience, right. If you want to read even more detail to a lot of the latest trends happenings in our analysis, go and subscribe to the AI to ROI newsletter. All you need to do is go AI to ROI. That's ai2roi.substack.com and you can subscribe. We're not charging. We put out five different editions every week. And instead of having to read 100 to 200 articles that Peter, I and our AI agents read, you can read the top five to 10 from us. Thanks, Peter.
- [31:07]
B
All right, see you next week.
- [31:08]
A
Bye, everyone.