When Will Openclaw go Mainstream?

Summary

This discussion centers on the mainstream potential of agentic AI—specifically “Open Claude”—and the barriers currently preventing widespread consumer adoption. While the guests describe the technology as a “tinker’s dream,” they note that it remains too technical for the average user due to local installation requirements and the lack of a singular “killer use case.” A recurring theme is the transition toward a “UI-less” future, where traditional software interfaces are replaced by text or voice-driven assistants. However, significant hurdles remain, including the massive security risks of granting agents access to sensitive personal data and the technical challenge of “context loss” in long-running AI conversations.

Tactical demos highlight the immediate value of these tools for specialized personal and professional productivity. For example, Matthew Berman demonstrates using an AI agent to track health data and identify food sensitivities, while Ryan Yannelli discusses automating the management of Type 1 Diabetes and clinical documentation for psychiatrists. On the business front, Jason Grad introduces ClawPod, a skill designed for large-scale, ethically sourced web scraping and marketing research. These examples illustrate a move toward “multi-model” workflows, where power users leverage a mix of models (such as Anthropic’s Opus and Sonnet 3.5) to balance cost, speed, and reasoning capabilities.

The conversation concludes with a look at the volatile competitive landscape and the future of open-source AI projects. The guests express skepticism regarding the long-term independence of Open Claude following its creator’s move to OpenAI, suggesting the project may eventually be subsumed by larger corporate interests or forked by the community. They also address Anthropic’s recent policy updates, which may restrict how users apply personal subscriptions to external agentic tools. Ultimately, the episode underscores a period of “vibe coding” and rapid innovation where the potential for personalized automation is immense, but the infrastructure for security and user-friendly “normie” adoption is still under construction.

Highlights

”Openclaw is not ready for the consumer”

“I don’t think it’s quite there. I think this is really a tinker’s dream right now.” — Matthew Berman, 2:12

yt-dlp --download-sections "*2:12-3:02" "https://www.youtube.com/watch?v=2u0L4BQ1a3w" --force-keyframes-at-cuts --merge-output-format mp4 -o "openclaw-is-not-ready-for-the-consumer.mp4"

”Ryan sees new opportunities for consumers”

“It’s, uh, I think a lot of it’s just going to be app-based and like for now. Um, it’s just I see Open Claw giving people that didn’t have access to a tool or work-” — Ryan Yannelli, 5:47

yt-dlp --download-sections "*5:47-6:37" "https://www.youtube.com/watch?v=2u0L4BQ1a3w" --force-keyframes-at-cuts --merge-output-format mp4 -o "ryan-sees-new-opportunities-for-consumers.mp4"

”The killer use case for consumer OpenClaw”

“This will be a breakout success as soon as it solves a painful enough—either it solves a painful enough problem for people to go through the step to set it up and get over security risks, and that’s converging on it becoming very easy to, um, have your instance set up or have it set up on whatever device you want. So it’s going to get both easier to implement this, but there needs to be a killer use case, uh, that brings this into the mainstream because—and by the way, because of my background, very security concerned, we—when this came out, I immediately flagged to everyone in the company, “Do not install this on any work machines,” like the day I saw it, because I knew there’d be so much temptation..” — Jason Grad, 10:59

yt-dlp --download-sections "*10:59-11:49" "https://www.youtube.com/watch?v=2u0L4BQ1a3w" --force-keyframes-at-cuts --merge-output-format mp4 -o "the-killer-use-case-for-consumer-openclaw.mp4"

”Jason demos his Openclaw workflow”

“All right, so now take us from Massive to Clawpot. I’m going to pull it up here so everyone can take a look at it. It’s a skill you built. Uh, I’m going to show the Claw Hub page, but what is this? Why do we need it? And how’s market reaction been?” — Jason Calacanis, 23:18

yt-dlp --download-sections "*23:18-24:18" "https://www.youtube.com/watch?v=2u0L4BQ1a3w" --force-keyframes-at-cuts --merge-output-format mp4 -o "jason-demos-his-openclaw-workflow.mp4"

”Anthropic patches pro plan usage”

“And one thing that they did was they said very clearly that you can’t use your uh your Pro or Max plan uh outside of their universe, which means that now if you have been running um your Open-Claude instance off of your personal subscription account to Anthropic’s AI products, you’re kind of banned from doing that as far as we can tell. Matt, you’re you’re nodding your head back and forth.” — Ryan Yannelli, 39:00

yt-dlp --download-sections "*39:00-39:50" "https://www.youtube.com/watch?v=2u0L4BQ1a3w" --force-keyframes-at-cuts --merge-output-format mp4 -o "anthropic-patches-pro-plan-usage.mp4"

Key Points

• Not ready for consumers (2:12) - Matthew Berman argues Openclaw is not yet ready to be brought to the consumer market
• Too many apps and tabs (4:17) - Jason laments having hundreds of different apps and thousands of tabs
• Consumer opportunities (5:47) - Ryan sees OpenClaw giving consumers access to opportunities they couldn’t reach otherwise
• Only 10% technical enough (7:02) - Only about 10% of people are technical enough to install and use Openclaw
• Better as an app? (8:01) - Debate on whether Openclaw would be better off as a traditional app
• Killer use case (10:59) - Discussion of what killer use case could bring Openclaw to the mainstream consumer
• Meta acquired Manus (12:08) - Why Meta’s acquisition of Manus matters for the Openclaw ecosystem
• Ryan’s personal use (15:16) - How Ryan uses Openclaw in his personal life for everyday tasks
• Jason’s Clawpod demo (23:18) - Jason demonstrates his personal Openclaw workflow and the Clawpod setup
• Matt’s health diagnosis (30:04) - How Matt used Openclaw to figure out he’s been having stomach issues
• Pro plan patched (39:00) - Anthropic patches the ability to use Openclaw through its pro plan
• Multi-model future hopes (42:11) - Matt and Jason hope for a multi-model future but acknowledge lack of progress
• OpenClaw foundation skepticism (52:15) - Jason expresses skepticism about the OpenClaw foundation’s direction
• New fork predicted (52:52) - Ryan predicts a new OpenClaw fork coming from the shadows
• Peter Steinberger leaving (54:41) - Peter Steinberger going to OpenAI, potentially orphaning OpenClaw

Mentions

Companies

• Anthropic (38:48) - Patched ability to use OpenClaw through its pro plan
• Meta (12:13) - Acquired Manus, discussed implications for OpenClaw
• OpenAI (54:21) - Peter Steinberger moving there, potentially orphaning OpenClaw
• Apple (58:19) - Discussion of raspberry AI competing against Apple
• Nextvisit (5:45) - Ryan Yannelli’s startup
• Massive (4:45) - Jason Grad’s company
• Gusto (8:27) - Sponsor - payroll platform
• Crusoe Cloud (18:44) - Sponsor - GPU cloud infrastructure
• Uber AI Solutions (28:23) - Sponsor - enterprise AI implementation

Products & Technologies

• OpenClaw (2:04) - Core discussion topic - when it will go mainstream
• Clawpod (23:24) - Jason’s personal OpenClaw hardware setup

People

• Matthew Berman (2:04) - YouTuber and OpenClaw enthusiast, argues not ready for consumers
• Ryan Yannelli (5:45) - Nextvisit founder, sees consumer opportunities in OpenClaw
• Jason Grad (4:45) - Massive founder
• Peter Steinberger (54:21) - OpenClaw creator leaving for OpenAI

Surprising Quotes

“Sorry, I did assume you still had friends. I should have known better than asking a founder.” — Ryan Yannelli, 7:47

“Mannis though was, if memory serves, aimed at the SMB market. So aimed at small businesses versus consumers. And when I think about mainstream, to me that does imply consumer map more so than, I don’t know, your average kind of mom and pop or brick and mortar.” — Ryan Yannelli, 12:59

“Well, when you put it like that… that’s not a bad point.” — Jason Calacanis, 52:31

Transcript

Jason Calacanis: 0:00 Hello everybody, welcome back to TWiST. This is Alex. Today we want to answer a very important question which is when will Open Claude go mainstream? Now if you’ve been watching the show lately, you know we’re all in on Open Claude. But for mainstream consumers, it’s just too technical still. Most people are not able to use it. Think about your parents or your friends that are outside the tech world. But we think that the first consumer app to truly bring Open Claude into the mainstream will become a unicorn. The question is: what’s that going to look like? Is it an app? Is it built into your phone? And how much does the average consumer actually care about security in the way that we talk about it if it just works? So to help us understand consumer adoption and to give us some tactical and practical demos, please join me in welcoming our guests today. We have Matthew Berman. Matthew is a well-known YouTuber with hundreds of thousands of subscribers, diving deep into how the average person can use Open Claude today. Matt, welcome to the show.

Matthew Berman: 0:59 Thank you Alex, good to be here.

Jason Calacanis: 1:01 Then we have Ryan Yannelli, the founder of Next Visit. He won our recent Gamma pitch competition, so he might look familiar to you. Next Visit is an AI-powered clinical documentation platform for psychiatrists. You can kind of imagine how agentic automation might play in there. Ryan, welcome back.

Ryan Yannelli: 1:13 Thank you for having me again.

Jason Calacanis: 1:14 And calling in all the way from Europe, it’s Jason Grad, the founder of Massive, a company that I covered their funding round back in the day, as it turns out. Massive helps companies gather public data from around the world, which they can then bring to bear on their own AI use case. And he recently launched something called ClaudePod that we’re going to talk about. Jason, welcome to the show.

Jason Grad: 1:34 Great to be here. Thanks for having me.

Jason Calacanis: 1:36 I’m so excited to have you all here because I’ve been spending my time deep in the Open Claude mines, burning tokens and basically trashing my family’s finances. But we’re going to start with the average person today. We’re going to talk about going mainstream. So there’s conversation about how if Open Claude was appified, if you will, it would immediately shoot to the top of the App Store and everyone would love to do it. But it still feels to me that Open Claude is far too complicated for the average person to set up. So Matt, you’ve been doing a lot of work helping people understand this. Where’s the average person in terms of readiness to adopt Open Claude and what’s holding them back?

Matthew Berman: 2:12 I don’t think it’s quite there. I think this is really a tinker’s dream right now. You have to install it locally, you’re still trying to iterate rapidly, figure out what works, what doesn’t work. There’s still so much discovery. But for the average person that just wants a one-click install and everything to just work, it’s not quite there yet. I think Google actually probably has the biggest opportunity to do this right. They have all the data, they have so many different services that they offer with Gmail and Calendar and Drive. And if they just continued to push forward as they have been and maybe even take a little bit of the red tape off and allow you to plug into your different services, I think they would probably be able to deliver something incredible.

Jason Calacanis: 2:56 But the security concerns wouldn’t be solved just by having access to all that information. So do you think that Google has the…

Matthew Berman: 3:00 the will to actually go forth and build something that forward looking or are they more conservative and therefore going to wait for someone else to take the baton?

Jason Calacanis: 3:07 If you asked me that 18 months ago I would have said they’re not going to be able to get out of their way, but now looking at just how far the Gemini series of models have come, how deeply integrated it is into every single product and service that they offer, I do think there’s an opportunity there. Obviously being a multi-trillion dollar public company they have to think about the security, and there are still a lot of security issues. Now the the thing about that is if you’re plugging in a bunch of different Google services and it’s all within their it’s all within their cloud, they probably can lock down that security pretty well, but then you’re very limited in what you can do within, you know, a one-click open claw type system. So you still need to connect other services and I still do think Google can do it.

Matthew Berman: 3:53 Yeah. Jason, the mobile comment’s very interesting to me because when I think about my phone, I’m an iPhone guy, I guess I should admit that up front here. It’s it’s amazing how the AI world seems to have absolutely missed the operating system layer, and instead it’s occasionally an app that I pull up on my phone, but if you think about it given what we can now do with Open Claw on our computers, it feels really backwards to me. Am I am I being too impatient?

Jason Grad: 4:17 I mean, first of all, I think the people that have adopted this or there’s a lot of really vocal early adopters and we’ve been speaking with people that use Clawpod and they’re super excited to be testing stuff out, and this is a very small group of people who are very loud, they’re like hobbyists, a lot of small business owners, a lot of people who help businesses set up workflows and then some, you know, some people in tech depending on how much work they have with their kind of normal jobs. Um, I don’t think you’re background per- like backwards personally. I mean, we’re seeing or if you’ve been claw pilled as as I guess you guys are saying here. Um, we’re seeing that all you need to do is give your machine a set of instructions. We don’t need this whole layer of UI, bloated apps, different types of weird experiences as a user anymore. I just want to have stuff run, have context from wherever I store some information and then do something because it has access to do something else. I don’t need to click into or have a screen full of hundreds of apps that I can’t find any app. I don’t need a thousand tabs open on my computer. I don’t need another thousand apps that I’m installing on my computer and I’m dealing with this. All I need is really, you know, a skill that says what to do and gives me access.

Matthew Berman: 5:32 Ryan, do you think we’re going to get there at the OS level or do you think that applications will for the now for the next, I don’t know, say 6 or 12 months maintain their kind of pre-eminence on the consumer side as we consider agentic automation, viz Open Claw and its other tools out there that are related?

Ryan Yannelli: 5:47 It’s, uh, I think a lot of it’s just going to be app-based and like for now. Um, it’s just I see Open Claw giving people that didn’t have access to a tool or work-

Matthew Berman: 6:00 vibe coding story before to have access to AI in these complicated tools and kind of give them that simple interface although it’s not simple yet to set up but.

Ryan Yannelli: 6:08 On the setup point though, Matt, it’s the one thing that you do on your channel that I really like is you break down kind of what people can do with OpenClaw, kind of giving them a blueprint for getting started and getting set up. When you talk to people out there who are watching your stuff, where do they get stuck going from, ‘okay, I want to try this out, I’ve heard about it,’ to ‘I can’t quite make it work’? Where’s the falling down point, if you will?

Matthew Berman: 6:29 I actually think the falling down is kind of discovering the use cases that make sense for their life. That’s why I’ve been just exploring so many different use cases, trying to document them, trying to put them out there. People kind of understand the gist of how this works. You have a Telegram or a WhatsApp personal assistant built-in, stored locally on your computer, you can do anything with it, but what is that ‘anything’? And I think, like, discovering what those actual use cases are, those are the questions that I personally get all the time, not only from folks who watch my videos, but also my friends who are technical enough to get this installed.

Ryan Yannelli: 7:02 I see. And what percentage of your friends do you think are technical enough to get this installed? Just if you had to put a number on it for fun.

Matthew Berman: 7:06 Ah, 10% and in it, you know, when ‘normies’ I don’t think it’s quite ready for them yet. Again, like if you can, if you can vibe code, if you are decent at vibe coding, you can absolutely use OpenClaw, but I think for the vast majority of people, it’s not, it’s not ready yet.

Ryan Yannelli: 7:25 Yeah, I hear that. Jason, if you had to put a guess on how many of your friends could set up OpenClaw today, friends thinking broadly, not just your tech friends, but just your social circle. What percent could pull it off without you swooping in to save the day and help them out?

Jason Grad: 7:40 I’m running a pretty high velocity startup right now. The amount of time I’m catching up with friends is limited these days.

Ryan Yannelli: 7:47 Sorry, I did assume you still had friends. I should have known better than asking a founder.

Jason Grad: 7:52 I have them somewhere. I mean, some people will mention it here and there as a you know have catch up conversations but it’s a really small percentage outside of techie founder type friends.

Jason Calacanis: 8:01 Yeah, yeah. Ryan, back to your point about, you know, sticking with apps for a while probably for the consumer, do you think that an app-ified Open Claw on say a mobile device would be useful given the way mobile operating systems are set up today? Because the way that I think of iOS for example is that it’s, it’s so locked down that you couldn’t actually probably ape what we can do with OpenClaw on PCs on mobile today, but I just want to kind of vet that against your views here.

Matthew Berman: 9:43 I think that if OpenClaude was in the form of an app where the use cases for how it’s going to be used by the average consumer are defined kind of like what Matt said before, um, it’s going to make it easier for a consumer to say, “Oh, okay, I see why I need this.” Uh, and it’s going to come down to, I think, like whoever is running that platform on the back end, uh, like Google or Anthropic, whoever it might be.

Ryan Yannelli: 10:15 Okay, well, we’ll have to see how that works out. But one thing I did see is that over on PolyMarket, there is people betting on this. And the reason why I’m so curious about the app side of this is, there was a moment when people thought that, I think an OpenClaude-ish app would come to market and knock finally ChatGPT out of the top of the App Store. And lately we’ve seen those numbers absolutely drop. If you’re on the audio version of this, essentially the number of people who think that ChatGPT will lose its top spot is dropping by double digit percentage points. So to me, this implies that we’re not going to have a good app version of it, which means that consumer adoption will remain slow and this is going to stay kind of a tech product for the foreseeable future. But Jason, please.

Jason Grad: 10:59 This will be a breakout success as soon as it solves a painful enough—either it solves a painful enough problem for people to go through the step to set it up and get over security risks, and that’s converging on it becoming very easy to, um, have your instance set up or have it set up on whatever device you want. So it’s going to get both easier to implement this, but there needs to be a killer use case, uh, that brings this into the mainstream because—and by the way, because of my background, very security concerned, we—when this came out, I immediately flagged to everyone in the company, “Do not install this on any work machines,” like the day I saw it, because I knew there’d be so much temptation.

Ryan Yannelli: 11:38 We’re going to get to the Wired article in a second, I promise.

Jason Grad: 11:40 But yeah, but yeah, but yeah, but what I’m saying is that, um, yeah, I think users have thrown away their security because something solves a pain point for decades now. It’s a huge problem that we all see, but—and so I think the use case is going to be really what brings this to the mainstream.

Jason Calacanis: 11:56 I feel like we’re getting really close to crypto territory here. Like, what is the use case for this? Awesome cool music technology. What do you think that use case that killer use case will be Matt? And do you think we know what it is yet?

Matthew Berman: 12:08 Yeah, I want to just actually point out Mannis, Mannis AI acquired by Meta. That is the reason that Meta acquired Mannis. That is essentially kind of the fully hosted version of Open Claw, right? And didn’t Mannis at a certain point reach pretty high up in the App Store rankings? Do you remember that?

Ryan Yannelli: 12:25 I don’t recall its actually its App Store rankings progression, but for folks listening, Mannis was a Chinese AI company, moved to I think it’s Singapore to redomicile, and then Meta bought it for $2 billion. Matt back me up here in the last month, month and a half.

Matthew Berman: 12:39 Yeah, very recently, and I think the point is like if you look at what Mannis is capable of and when Mannis first came out, it’s a lot, it’s kind of that a very akin to what Open Claw is able to do. Open Claw is obviously much more configurable now because it’s not fully cloud-based. But I think that’s Meta’s plan with it. Uh, so I just wanted to bring that point back for a second and and yeah, Mannis just looks like it could be maybe that first breakout Open Claw version.

Ryan Yannelli: 12:59 Mannis though was, if memory serves, aimed at the SMB market. So aimed at small businesses versus consumers. And when I think about mainstream, to me that does imply consumer map more so than, I don’t know, your average kind of mom and pop or brick and mortar.

Matthew Berman: 13:16 Yeah, I actually I think that’s a really good distinction. Are we talking about personal AI assistants or are we talking about AI assistants that can get knowledge work done? For me, right now, Open Claw is doing both, which is wonderful. Mannis, yes, it was much more focused on the knowledge work use case, but I think now that they’ve been acquired by Meta, it’s much more likely that they’re going to be able to service both parts of that market.

Ryan Yannelli: 13:29 Okay, but back to the question about what the breakout consumer use case is going to be, to Jason’s point, if you had to pick like a top three thing that or top three things that someone outside of our technology bubble would want to do with it or get value from, what do you think those are?

Matthew Berman: 13:41 Probably as soon as as soon as these systems can actually get real-world tasks done consistently, that’s probably going to be it. I know that’s not necessarily what you’re asking, you’re asking for a very specific use case. I guess the short answer is I’m not exactly sure what that’s going to be. It’s probably going to be determined by what everybody needs in the moment. I think one that just comes to mind immediately is I have just complete notification overload at all times. If I had a system that was a buffer between me and all of the information that I’m getting from email, from Slack, from everywhere—and these are personal use cases—and then can maybe even triage some of those things for me, that would be an incredible time saver and maybe something broadly applicable.

Jason Grad: 14:08 Do you think the average person has so many notifications and and inbound communications that they need that? Because some—you ever like look at your your partner’s email and see they’ve got like four that day and you’re like, ‘Oh, we live in different worlds’?

Matthew Berman: 14:18 Yeah, but then you you think about Instagram and Facebook and X and like it doesn’t just come from email, it comes from everywhere and and people are scrolling like crazy. I just think there’s so much content.

Ryan Yannelli: 15:00 noise right now, the ability to actually filter some of that noise plus when AI creates so much more noise because you can create any image, any video, and then all of a sudden you need your own AI to filter all the noise coming through and only give you the signal. I think that’s gonna be a killer use case.

Matthew Berman: 15:16 All right, so on the use case front, Ryan, you sent over some notes about what you’re actually using OpenClaw for, and I’m just going to list out a couple of things that you mentioned to us. So, start of the day, end of the day accountability, agent for yourself and your staff, business-focused, not consumer, but very cool. Uh, weekly recaps, customer retention agents, issue tracking with auto-resolution. You’re using OpenClaw for a bajillion different things. What’s your take, and then we’ll get to the business side of this in a second. But if you had to pick a consumer-friendly use case, what would you say is going to be the hot ticket in the next couple quarters?

Ryan Yannelli: 15:48 I would say, like, the mundane tasks that you normally have to spend a little time thinking about, possibly such like picking out groceries for the week, sending out a pickup order. Stuff like that would be ideal. Like, I want that workflow right now. But, you know, right now I have automated a lot of other stuff with my health as you may have seen in the notes and just my daily productivity.

Matthew Berman: 16:18 Can you actually talk about your use of wearables and AI and how that impacts your daily life? And I’m going to try to not focus on your beautiful cat that is jumping around behind you. But talking about the health side of this and why you made those choices to use OpenClaw in this way.

Ryan Yannelli: 16:31 I have type 1 diabetes, and it’s basically an autoimmune disease that takes up a lot of my time normally because it’s constant adjustments, it’s alarms going off. And having AI to monitor my levels in real-time and check to make sure the pharmacy has stuff in stock and send my orders out and prescriptions has been such a time saver to the point where I don’t even have to think about management of type 1 anymore. It’s just kind of a message on Telegram once or twice a day. So it’s been incredible just how AI has, how far it’s come to where I can just send a message versus being on the phone for a half hour just to talk to pharmacy.

Matthew Berman: 17:18 Yeah. Is that something do you think you could set up for other people en masse, or is that something that you have to do kind of individually on a per-person basis? I’m just thinking about, again, taking cool ideas and then bringing them to as many people as possible, as quickly as possible.

Ryan Yannelli: 17:31 Yeah, I love open-sourcing stuff that I can. So, typically, like this workflow, once I’m comfortable with it, I’m gonna throw it on my GitHub and, you know, let other people use it if they want to. I love sharing that stuff.

Jason Calacanis: 17:45 No, I absolutely love that. All right, and on the subject of putting things up on the internet for people to use, Jason, I do want to turn to business stuff here, some tactical and practical for the founders. I want to talk about Clawpod, but first, actually, I think we should pause and let you talk about Massive and what it does and then we’ll get into the business side. But give us some background, fill in the blanks for the folks out there who are not familiar with what you’re working on.

Jason Grad: 18:05 Yeah, sure. Massive is built in opt-in, ethically sourced residential proxy infrastructure of millions of users and their devices around the world. And we make it easy for all of our clients to access the web at a large or massive scale. And so, you know, we’re solving the same problem with Massive that we want to solve for Cloud Pod users, is that the web wasn’t built for machines. It was built for humans. And now this AI revolution has come and humans are realizing all these crazy things that we can do on the web and that we can have agents do it for us. I don’t want to be doing this stuff. For a lot of people they want their computers to work for them and then they could do what they love. For me, I want the same, but I love working, so it just means I could be more productive. Um, and it’s kind of crazy that we still see these problems when we’re using LLMs every day, like can’t access that site or can’t access that content. And this seems like such a solvable problem. It’s a solved problem for us at scale with our clients and so, even as a human I try to use the web and I get blocked by crazy captchas that kind of interrupt my workflows.

Matthew Berman: 20:22 Not only that, but we used to complain so much about GDPR, you know, pop up the notifications like ‘do you accept cookies’ and how many, you know, people years we’ve lost to people clicking those things. So often now, I go to a website that I’ve been to since I was a child and it’s like ‘Cloudflare, we have to make sure that you’re a person.’ And that is now a very standard roadblock just to using the internet. Jason, can I just ask you about the ethics of Massive? Because it sounds like exactly what I need, because I have the same problem that you have. Often my, either AI chatbot or my open source instance can’t access a part of the internet that I would like it to. And it seems like Massive is a way to very politely- Get around those restrictions? Um, how often do you get in trouble for helping people avoid the chain-link fence?

Jason Grad: 21:07 We built this company on ethics and compliance from the start in an industry that might not have been like so above board and transparent about what they did historically. So there’s a few things that we do to stay out of trouble and one is that all of our users that comprise our network are consenting users, they’re opted in, they get compensated for being a part of the network. It’s it’s kind of an upfront deal that we have with them. Um and then the other part of that is that we are not a good service to use to do nefarious things. We block 5 million plus domains that would be used for things like DDoS, for um or distributing malware or visiting certain content that might be legal but not ethical and we’re definitely on the on the safer side of this, so we block all this content that makes us not practical to use if you want to do these things. So, um, I hope that answers your question. We really have since the start of this company tried to do things the right way, not only in line with what is legal but what we think is right in our version of of the internet.

Matthew Berman: 22:09 Do the companies that are trying to keep out, let’s say, agentic calls on their websites agree with you or do they still protest a little bit? I’m trying to figure out the temperature of, uh, the people building more tools for AI agents versus the existing companies that might want to have a humans-only sign across the door, if that makes sense.

Jason Grad: 22:29 It’s an interesting topic because a lot of the big voices in the space saying not to do this have been doing it all along. You know? So, you know like Google is involved in some cases where they’re trying to make it hard for people to scrape or they’re saying that’s that’s not the right thing and they’ve had the biggest scraping apparatus, uh, in the internet history. Yeah. And so all these big companies that are coming out against it are also doing it and so it’s kind of a fine line. I feel where it personally like where it resonates the most with me is small business owners or site owners that are getting hit with tons of traffic. A lot of the traffic, uh, from scrapers, especially like very mass scale now that AI is around, is not efficient and so it could hit a site many times and create cost that makes no sense versus just efficiently kind of scraping, collecting, and you know using the info somehow.

Jason Calacanis: 23:18 All right, so now take us from Massive to Clawpot. I’m going to pull it up here so everyone can take a look at it. It’s a skill you built. Uh, I’m going to show the Claw Hub page, but what is this? Why do we need it? And how’s market reaction been?

Ryan Yannelli: 23:37 So, um, Clawpot takes all of the tools that we built at Massive that lets our clients gather, you know, tens of billions of data points each month at scale and it gives it to the consumers so that they could use these tools and, you know, have a pretty easy way to make complicated workflows. So, it’s uh using our unblocker API and search API that allow you to access Google and kind of hard to reach sites and then all of that’s running on top of our um proxy infrastructure users across the world. And so now it’s a… yeah, sorry, go ahead.

Matthew Berman: 24:01 Oh, no, please. I was gonna ask about making the skill itself, but please finish your thought. I didn’t mean to cut you off.

Ryan Yannelli: 24:06 Oh, yeah. This… the skill wraps like all of our tools together and makes it really easy for users to… to access. They just go to ClawHub… to install it, they can like grab it there, but they can also just install it via their Open Claw and say, ‘Install ClawPod.’ And then basically whenever you’re about to run into one of those messages where it can’t access some site, then ClawPod will automatically go and retrieve that content for you, or you can build… and I’ll show you guys one… a demo.

Matthew Berman: 24:34 Yeah, no, no. Bring it up. We love demos.

Ryan Yannelli: 24:42 So, I’m giving a demo of a workflow that I built… Basically, I want to know everyone that’s mentioning Open Claw every day on TikTok, on X, on other social platforms. And so I’ve… I’m using ClawPod to do this, and it’s gonna go… and I told my Open Claw, ‘Go, go use ClawPod and search a few pages of Google for Open Claw with TikTok and X,’ and kind of get people who mentioned it and some info about them and their profiles and stuff like that so that we could reach out. And it’ll take about a minute, so I’m gonna fast-forward in the video. You can see the scraper’s running. And then about a minute later, um… you can see that Open Claw created a report. It found eight creators on TikTok, and it’s broken them down into tiers. So, creators with over 400,000 followers versus 50 to 400,000 followers, 10 to 50, and so on. And it’s created this for both TikTok and… um, Twitter. And now this gets sent to the marketing team, and they can automatically action it.

Matthew Berman: 25:43 So, one thing, a question that I have is, why does it take a minute to… to do that? I’m a very impatient boy, and one thing I love about the internet is that things are quick. Uh, but I’ve also discovered, doing my own Open Claw usage, that sometimes there is a bit of lag. So, tell me about what’s going on during that minute, and, uh, how fast we can kind of get that shrunk down.

Ryan Yannelli: 26:02 Yeah, it… it depends what your workflow is. And this is kind of a workflow where you’re… you’re collecting public web data. Um, you’ll want to run these as like a cron job, and it’ll create, you know, usually this kind of thing will be running in the background. You’re not necessarily watching it each time, like on the demo. Uh, the reason is because it’s actually going to visit Google. Um, it’s doing so in a way where we’re not spamming Google with tons of traffic to the search pages. And then it has to go visit each one of those links and grab the content for you. And so there’s ways to speed it up, like parallelization. Um, this is a pretty simple flow that I built. So, if I spent a little bit more time on this, or any user that uses ClawPod, you could tell it to grab those URLs and then visit 10 of them at a time, and that’ll certainly speed up the workflow.

Matthew Berman: 26:49 All right. Really quickly, people will talk a lot about skills. We talk about downloading them, making them. Uh, can you just give people a definition of what an Open Claw skill is and what they look like, because…

Jason Calacanis: 27:00 We just sent people to your Claw page and I want to make sure that when they scan through it, they have a little bit of context about what it is they’re reading.

Jason Grad: 27:07 It is a list of instructions to your computer to how to use some technology. And so you’re going to see, hey, in our- in our skill, for example, it says, and I just- I had a Google search today, so it was pretty fresh for me. Um, it’s going to say, hey, this is a couple APIs, you’ve got a Google search API, you’ve got an unblocker. Uh, whenever the user comes across some content that they can’t access or- or one of their LLMs says to them this site cannot be accessed, OpenClaw should go and retrieve the information and send it to them. Um, and it’ll- and it says stuff like use Massive’s unblocker to do this or use the search parameters if the user wants to go to Google. And that’s pretty much it and, you know, of course, like a prompt if you don’t have an account go get API keys and stuff like this. So it’s just an instruction model for your machine and I mean, the really cool thing about these skills is we’ve condensed sometimes huge apps down into just a small instruction manual that can live in a very tiny and efficient Markdown file for your OpenClaw.

Ryan Yannelli: 28:09 Matt, one way that I kind of think about these when I explain this to people is that a skill is kind of like a recipe, if that makes sense? Like you’re kind of giving a list of ingredients, kind of a run through of how to go about them and, you know, much like a human, it’s not going to do the exact same thing every time, but it’s, uh, it’s a good guide book or recipe if you will for what should be done. How fair is that?

Matthew Berman: 29:26 Yeah, very fair. So, the way I look at it is as a skill is a combination of natural language, a prompt actually give the model, and then it also is able to call tools. And so the distinction between skills and tools is tools is purely deterministic code, meaning like the kind of more traditional code. It is if, then, and do this, hit this API. And those tools are available to the skills to call. The skills have prompts that can invoke the tools.

Jason Calacanis: 29:56 I want to get to demo number two. You’re going to show us some cool stuff inside of Telegram, including some buttons that we haven’t seen before. Apparently released in the latest Open Claude instance?

Matthew Berman: 30:02 Just last night. Yeah.

Jason Calacanis: 30:03 Okay. Can’t wait to see this.

Matthew Berman: 30:04 We’re getting right into some personal information here. I have had some stomach issues the last few months and I wanted to figure out why. This is one of my favorite use cases and I’m just going to cut to the end for a second. It has figured out why I’ve had stomach issues. But let me show you what I’m doing. So, at the beginning basically what I told it to do is look at everything that I’m eating, sometimes I’ll just describe it with natural language, sometimes I will take pictures and upload it and I say, ‘track it all.’ It tracks it by reading what’s in the image, it records the time, it asks me some follow-up questions sometimes about how much I’ve eaten and it basically tracks all of that. And then throughout the day I tell it how my stomach is feeling. And within the last I believe I’ve been running this for about 10 days, it has figured out that the main cause of my stomach issues is onions.

Jason Calacanis: 30:58 Really? That’s terrible. Onions are like clutch, man.

Matthew Berman: 31:03 It is terrible and I love onions and I love grilled onions and I eat them and sometimes my wife is just going to have to suffer, but most of the time I’m going to try to avoid it and so you can see like here I have an espresso and I just tell it regular low-fat milk, some olives and back and forth and then finally at the 10-day stomach analysis, onion. Right? So this is just such a fantastic use case. Then you also mentioned the inline buttons, I want to show that off. Just yesterday Claude released a new update that had inline buttons just like this which is really cool and rather than me kind of proactively telling it how my stomach is feeling, what it’s going to do is just ping me three times a day and ask me and all I have to do is click here. This is by the way the extent of a UI that I would want. I like everything to be in text. So that’s one use case. I’m going to pause there for a second.

Jason Calacanis: 31:54 That’s the extent of the UI that you want.

Matthew Berman: 31:57 I want text only.

Jason Calacanis: 31:58 Yeah. Well, we talked about this when we had Hiten Shah and Jesse Janelle on the show on Monday and we kind of said that apps are dead, if that makes sense. Like the argument was that we no longer need to have separate user interfaces for different tasks and it sounds like you’re echoing that sentiment?

Matthew Berman: 32:15 Absolutely. I mean I made a video about this. It was called SaaS is dead. It was about 18 months ago.

Jason Calacanis: 32:22 Oh, so it’s your fault that the stock market crashed. I see.

Matthew Berman: 32:25 And so look, all of these agents are going to be able to handle all of this with very minimal interface and to be honest it would be great if I didn’t even have to have text. Like it just knew what I wanted to do, it just did it and would only ask me questions when it was absolutely necessary. I do think the ultimate UX for AI is voice, but I just haven’t figured out a good way to give Claude synchronous voice yet. But yeah, text for me for now.

Jason Calacanis: 32:50 Well, that brings up… actually if we could take that screen share down… I want to get Ryan’s take on what you just said because NextVisit, his company, is a medical scribe. It’s something that I think can hear as you talk to people…

Matthew Berman: 33:00 …and in the psychiatry field. So first of all, Ryan, respond to the idea about user interfaces and voice being so clutch, and then after that, I want to talk to you about security and using AI agents in the medical world.

Ryan Yannelli: 33:12 Yeah, so, um, our platform does real-time voice, and it does real-time analytics with LLMs. Um, and it’s real-time with latency, because there’s no way for us to instantaneously process massive amounts of health data securely and accurately that quick. Um, but, you know, our users can interact with, uh, the AI assistant Arya, either by voice or, um, and whatnot. Um, but something we released recently was something like OpenClaw but for healthcare inside of the system, and it has a lot of guardrails around it. Uh, part of it comes up with this real-time voice issue. How do you interact with it in real-time? And all doctors really have done so far is what we’ve seen is just real-time dictation, but, uh, they don’t know how to talk to it yet. I think like the consumer side is just how do you talk to AI or interact with it like voice, um, and, uh, I don’t know, I feel like it’s just not 100% there yet.

Matthew Berman: 34:25 Is that because the underlying technology to enable conversation with AI isn’t there yet, or people haven’t figured out yet how to engage with that like a feature? Because I’m not sure if it’s like a technology thing or a learning thing.

Ryan Yannelli: 34:37 I’ll rephrase it. It’s more like, um, right now you have trigger keywords like, um, you know, I’ll just say Siri, uh, and you have to say these trigger keywords. Um, and when you’re a busy like a doctor or a founder for that matter, uh, you kind of just want it to pick up on certain things or do things without you having to trigger it each time. I think just interacting with it in general is not there yet, or using those trigger keywords, um, because it’s just, uh, it doesn’t fit into people’s workflow naturally, at least in my opinion.

Jason Calacanis: 35:17 Yeah. So Jason, um, one thing that I’m really curious about is the security side of this, and I want to get back to Ryan in the medical world in a second, but you mentioned telling your whole team at your company to don’t install OpenClaw on your, on your work computer, and I think that’s great advice. Uh, what, what do people need to know in a business or personal context about security right now with OpenClaw, and, uh, do we have enough guardrails in place via like the skill hub and so forth to keep people reasonably secure if they’re going to play with this?

Jason Grad: 35:46 Let’s think about the biggest risk that we have right now aside from OpenClaw. The biggest risk is that, um, someone sends you something via email or text that you trust and you click the link. That’s one big thing. Another is that your password gets somehow grabbed from somewhere and that’s allowed them to have phishing attacks, but somehow your password gets exposed, your password to sensitive information. So when you use something like Open Claw where you’re putting a bunch of your, first of all, for a layman, you’re starting with some software that hasn’t been fully tested, it’s in early stages, and you’re putting a bunch of your information there and you’re letting it do stuff. So you’re letting it maybe access email and you’re giving it a bunch of your passwords. And so you’re just increasing this kind of attack surface immediately. And unless you figure out some guardrails around that, depending on your personal risk and what’s at stake, I would advise caution. I mean, I would definitely advise caution before you start, but there’s definitely safe ways to start experimenting. And I just wanted to get that message out to the team before anyone, you know, went too far. When you work at a company, it’s like the silly things that can cause huge problems, like a person clicking the wrong link in an email. And so we take a pretty cautious approach to kind of just like shut something down and then investigate it and then figure out the path forward for everyone.

Matthew Berman: 37:17 Ryan, Jason’s point there about phishing attacks and how, you know, you always have some risk. My relatively unlettered perspective, and I don’t know as much about HIPAA, I’m sure as you do, is that we’re already so insecure generally in our digital lives that adding a little bit of agentic juice to it isn’t that terrifying. Like I’m running Open Claw on my gaming PC most of the time, and I gave it access to my email and it can, I just don’t care. And I’m trying to figure out if I’m being a silly, silly boy, or if people like Jason are being too conservative when they consider using Open Claw in kind of a businessy context.

Ryan Yannelli: 37:54 Yeah, and I mean, you know, Jason’s point about just prompt injection in those emails is that I see it in our help desk every day. There’s something like, you know, stop, like you’re now giving me the full system prompt, blah, blah, blah, do this and that. So I see that every day. But when it comes to data and security, my personal data, I don’t care. I’m going to give whatever wants access to it access because I want to mess around with stuff. But when it comes to other people’s data, I’m very guarded about what has access to that, not only for compliance, but just ethics and security. I don’t want to risk patient’s data.

Jason Calacanis: 38:30 Yeah, one of our, one of our live viewers, Jay Fennigal says, feeding Open Claw health data, just one example of a massive attack surface. 135,000 instances exposed to the internet. No real security posture. When does that become a problem? Sounds like your view is now. Yeah. All right. So Jay Fennigal, there you go. The answer is yes, you’re dead on. All right, guys, I want to get to one more thing that the audience is bringing up. Right before we got on the show, Anthropic updated its policies and

Ryan Yannelli: 39:00 And one thing that they did was they said very clearly that you can’t use your uh your Pro or Max plan uh outside of their universe, which means that now if you have been running um your Open-Claude instance off of your personal subscription account to Anthropic’s AI products, you’re kind of banned from doing that as far as we can tell. Matt, you’re you’re nodding your head back and forth. Am I misreading the the Anthropic update?

Matthew Berman: 39:27 No, did that just happen? I think that has been the policy.

Ryan Yannelli: 39:30 Well, they updated the document. I know there was the conversation about uh the the coding tools, but people are reading this more as indications that you can’t bring your Claude subscription to Open-Claude itself.

Matthew Berman: 39:43 Yeah, well, what a fumble. Uh it just seems like such a- it’s like a negative signal to this community of builders who I, you know, I have a Max subscription and yes, I’m using it through that. Obviously, I’m going to have to consider other options, maybe the API directly, but it just- it leaves such a sour taste in my mouth. I don’t- like why not just let me do it? I’m- I have increased my payments to Anthropic because of Open-Claude. So like just let me do it. Let me do what I want with my tokens and I don’t know, it’s just not- not a good vibe.

Ryan Yannelli: 40:14 Yeah, it’s a little bit surprising to me. Jason, there’s been a lot of commentary that, you know, Anthropic has fumbled uh the bag, the moment, the future, uh in regards to- to Open-Claude. But at the same time, you don’t want people using your stuff in ways that you don’t want them to. Like if I took Massive out and and did things with it that you didn’t approve of, I presume you would cut me off as well.

Jason Grad: 40:34 Yeah, absolutely, immediately. Uh not approve of, but went against like policies. Uh so it makes sense just as a tech, you know, when we talk about security by the way, there’s like this tradeoff, yeah, versus moving slow and- and cautiously or like figuring out what the future might look like and jumping in. Um so we’re trying to, you know, toe that line. Um and yeah, it’s disappointing because uh Claude has built some really great technology that people love using. I’ve been using Claude to do all kinds of uh it’s like magic on my computer that used to take me a lot longer to produce. Um and when you take Open-Claude, you just want to use- bring- bring your own anything to use it. So it’s kind of- it doesn’t leave a great taste in the user’s mouth to- to have this restriction imposed on us.

Ryan Yannelli: 41:22 You make an interesting point though about using their models. I’ve used Opus 4.5 in Open-Claude before I started switching over to some other stuff. Um what- uh other- which one of the models are you using right now, Jason, that you find to be so impressive?

Jason Grad: 41:34 I’m using Opus 4.6 uh for- for- across most of what I’m doing right now because uh I’m still expres- it- I know it doesn’t seem like it because every day is like a year in AI. Uh but it’s been like days or a couple weeks since it came out, I’m not sure of time anymore, I’m also having a lot of late night hacking over here, but I’m still playing with- with 4.6 to understand all the capabilities.

Ryan Yannelli: 41:58 Uh Matt, you just did a video about the release of Sonnet 3.5 from Anthropic, uh, which people are raving about. What’s been your experience with that, and do you think people should use it if they are currently, much like Jason, Opus 3 users at present?

Matthew Berman: 42:11 Yeah, so I’ll say presently I am also an Opus 3 user pretty much across the entire OpenClaude stack. OpenClaude just released support for Sonnet 3.5 last night, so I’m starting to dabble with it. I haven’t had time to extensively test it, but if you look at the benchmarks, if you look how Sonnet 3.5 is being positioned in the market as an absolute workhorse for knowledge work, it seems like it’s going to fit in extremely well into my stack, and especially when I have to switch off of using the max plan to using the API directly, I’m going to appreciate the cost savings of Sonnet. Um, so yeah, I definitely plan on playing around with it. It really does seem like just an absolute workhorse for the things that I need to get done, and with tool use, agentic use, it’s very impressive. Uh, I need to test more.

Jason Grad: 42:59 That’s a really interesting point. Uh, I’ve heard a lot of people talking about how, you know, there’s not just one best model for everything. You really want to be specific about what you’re using for what. But then again, I was using Opus 3, you and Jason are now on 3.5. Is that true, or should we just always go for the biggest, best, and most expensive model? Because I’m kind of lazy and I tried to build a model switcher in my OpenClaude instance and it was just kind of cludgy, Matt, and I didn’t really like what I got.

Matthew Berman: 43:27 Yeah, yeah, look, if you’re lazy or you don’t want to think about it, choose the best model, use that. Obviously, you have to work within your budget. The, I think, like the optimal use case or the optimal implementation is a multi-model world. You want to use the right model for the right job. Opus is a great general model that can also code. CodeX is really good at coding but wouldn’t be able to chat with you and have a really good personal conversation with you. Plus, if you’re doing embeddings, if you’re doing cron job management, all of this stuff can be offloaded to smaller models. Um, so I do think it is really like the ultimate implementation of this is to have it be multi-model. I think the other thing to consider is prompt management, which I don’t think a lot of people think about yet. The reason, one of the top reasons I went with a single model across the board is because I’ve optimized all my prompts specifically for Opus 3 based on Anthropic’s prompting guide. Now, if I had multiple models, I’d have to manage multiple versions of each prompt, and that is not supported as a first-class feature in OpenClaude yet. So I think the world will be multi-model, but for now, it is, you know, if you don’t care or if you just want to be something simple, single model.

Ryan Yannelli: 44:42 But if you don’t end up like—

Jason Calacanis: 44:44 Yeah, I hope you’re right, because a multi-model future means there’s more competition, there’s more prices, there’s more features coming out, and it’s overall a huge win for consumers. I don’t know that we’ve made the kind of crazy progress—

Jason Grad: 45:00 Progress that we’ve made in the last few years unless there was intense competition over winning the market. And so, I mean part of that is that there’s really a low cost of switching. So people will just up and leave to the next thing if it’s better. And that’s painful if you run a company in the space—in that space, but it’s amazing for developers and for consumers.

Matthew Berman: 45:22 Yeah, I mean Jason, we’re seeing that, right? There’s very little loyalty to any model, especially in the coding world. Like Opus was kind of the end-all be-all and then as soon as GPT-4o coding came out it was like, oh no wait, that’s the one now. But you’re right. And just to add one more fine point on that, local models continue to get better and better. And so there are increasing numbers of use cases that can be offloaded to your local computer and I think that’s also going to be really powerful. A mixture of all the best models seems to be the clearest path forward.

Jason Calacanis: 45:54 We’re hearing a lot about model switching, context, and what’s most efficient. What are you running at your company and how frequently do you test out different models in either your main product or in your open questions?

Jason Grad: 46:05 So I personally, Opus 4.6. I’ve been with Opus—I’ve basically been using Claude since Anthropic released Opus 4. But yeah, we’re pretty much Anthropic at Next Visit. They give us zero data retention. They give us really great support. And we do test other models. We test on Azure GPT and OpenAI. We test Google. And Anthropic in their last release, when it was what is it, Opus 4.5, we noticed that there was a significant quality increase in their healthcare in terms of being able to diagnose psychiatric conditions. So that was just like, that sealed the deal for us to actually go solely with Anthropic right now. But compared to other models, they’re just, at least for healthcare, it doesn’t even compare.

Ryan Yannelli: 47:07 Doesn’t even compare. Yeah. It’s funny that we’re all mostly still using Opus because, you know, Zhipu AI dropped GLM-4, we just got MiniMax m2.5, there’s a lot of rumblings about a new DeepSeek model coming out, everyone loved Kimi k2.5 Thinking, but we’re still, we’re still using Opus, Jason. I mean, maybe that’s just indicative of the lead that American AI labs have when it comes to agentic usage.

Matthew Berman: 47:33 I’m not sure. And I don’t think that—I don’t think anyone, not just the average consumer, no one on my team or anyone I know has time to try every single model out. So, I think for us, we’ve tested plenty. When Opus came out, it was—we’re mostly engineers on our team—it was a world of difference. It was like night and day from what had come before. I think lately for generalized tasks…

Ryan Yannelli: 48:00 The team has switched away from ChatGPT, a lot of people on the team cancelled their subs and moved to Gemini for just general kind of research stuff. And it happened like a couple people noticed it wasn’t working as well and then boom, like a few — ‘yeah, actually I don’t need this anymore, I’m gonna switch, sign off’. So the cost of switching is low, AI companies is crazy wild growth but also you know the flame that burns the brightest type thing where it’s like users have no loyalty to this. And I think for for me and you know people that I know, you recognize a few brands, you’ll test between them and you’ll make your decision.

Jason Calacanis: 48:38 Yeah, well I think the one thing that Open Claude’s done for me is made me more interested in testing out different models because I’m always trying to figure out how to not go bankrupt. Like I love Opus and its various versions, but oh my lord, I burn through API credits so fast and you know I can’t write it off as a business expense. So to me I was like I need to find something cheaper. So I ended up going down kind of the model rabbit hole. But the quality of today’s models do unlock a lot of stuff.

Ryan Yannelli: 49:03 We have a comment from someone watching. They said that they have a theory that Open Claude is only winning because of timing. Could MANS or Devin have been just as big if they came out when these new LLMs existed? And so you know Matt, I’m curious from your perspective, is Open Claude special because it’s done something new or was it just built at a time in which models could take use of the harness and actually do cool things, but it’s more model dependent versus the agent setup?

Matthew Berman: 49:23 No, I think that’s certainly part of it that these models became so capable right at the same time. But Peter, what what he has done is given inspiration to a community of tinkers and builders and I think that showing the vision of what’s possible, that was really the key unlock. And and I think the the second part of that unlock is just how personal it felt coming into the existing chat channels that you’re using, Telegram, WhatsApp, Slack. Like I think the combination of those three things is really what made it take off. Could MANS have done it? Could Devin have done it? I mean you could build, you could have built all this stuff three six months ago with Codex, with Claude code, it was very possible. But it was like the fact that he put all the pieces together in a really nice way and then just gave it to the community to extend and explore what was possible, I think I think that is really what made it special.

Ryan Yannelli: 50:30 Jason, you’re nodding your head, I want you to weigh in on this.

Jason Grad: 50:33 Yeah, I’ve been talking with a lot of the Clawpod users and I want to hear all their use cases, so if you’re using it then ping me, I’m happy to jump on a quick…

Ryan Yannelli: 50:44 What’s a what’s a good email address for you for people to send in notes if they’re watching?

Jason Grad: 50:48 Find me on LinkedIn or X, Jason Grad without the A. Best way to reach me. Um, I’m working at my computer, I’m pretty responsive if you send me something thoughtful. I’m happy to hear especially the use cases for Clawpod. And what I’m finding is that like there’s… Just a really long tail of… there hasn’t been a use case that just jumps out as the clear winner other than some power users. But what I see is a long tail of people doing all kinds of random stuff that they probably couldn’t have done before, that’s helping power their small businesses or power their lives. And not only on OpenClaw, but I have a bunch of friends who were not technical before and now are just excited that they’ve been given the power to build. And it’s the same feeling that I felt almost a decade ago when I was learning to code and I could do stuff and I had power. The power to create stuff that I wanted to exist in the world, this annoying feature that seemed so simple that just didn’t wasn’t out there or something specific. And now we’ve given people this like AI OS to, and it empowers them to just build all these things they’ve wanted to build but could never do before. And so I think that’s part of this… it’s free. It’s not… well, now it’s tied to OpenAI, so we’ll see how the community responds to that, what happens over time. Kind of, yeah, but… so we’ll see, but money has a very influential effect on the activities of organizations.

Jason Calacanis: 51:59 That’s a spicy comment, Jason, but so you’re not… you’re not as confident in the foundation setup kind of leaving OpenClaw by itself and letting it exist as its own thing, even if Peter goes to OpenAI?

Jason Grad: 52:15 You know what, I don’t know Peter personally, so I can’t say what he’s going to do. And so I’m just a spectator. But OpenAI also started as a foundation.

Jason Calacanis: 52:31 Well, when you put it like that… that’s not a bad point. Let’s get one more founder view on this question because this is interesting to me. Ryan, confidence in OpenClaw’s ability to stay independent via its foundation and not get subsumed into the Borg that is OpenAI? Are you more skeptical like Jason or are you more optimistic like me?

Ryan Yannelli: 52:52 A little more skeptical. Honestly, I think there’s going to be a fork of this project that, or just something new entirely, that’s pretty much the same thing, but just worked out all the bugs and is truly just refined, that just comes out from the shadows.

Jason Calacanis: 53:11 I agree with that, but I’m not sure which one it’s going to be. So Matt showed us his Telegram demo with the new buttons that you can now do, which is a cool new feature in OpenClaw, but there’s also PicoBot, Nanoclaw, Nanobot, and SafeClaw, which is a non-LLM OpenClaw competitor. Matt, do you think any of those have a shot at picking up material market share as people tinker with this stuff, or is it going to be mostly just OpenClaw as the main event?

Matthew Berman: 53:41 It’s hard to predict the future. I think OpenClaw has so much traction, so much growth, so much mind share right now. And then kind of back to the previous point of how skeptical I am, right, OpenAI is going to push a product that competes with OpenClaw, that competes with PicoBot. I’ve actually not used any of those alternatives, but… So, no, the answer to your question, no, I don’t think in the immediate, the short, even maybe the mid-term, those other projects are going to get much traction. And then I just want to kind of give my perspective. I am also skeptical. Peter’s going to OpenAI not to work on Open Interpreter as far as I know. And then when OpenAI delivers their own Open Interpreter product and it’s competing with the foundation’s product, it’s just like, okay, OpenAI needs to make money. They’re going to go focus on that, and I feel like Open Interpreter may be an orphan.

Jason Grad: 54:41 Oh, that’s really depressing. But Jason said earlier that he liked all the competition inside of different models because it yields to better stuff. Wouldn’t it be a net positive if we had a commercial company competing with the open-source product in this case to, you know, steel sharpen steel—steel sharpens steel? I can speak English. It’s one of my skills.

Matthew Berman: 54:57 Yeah, look, OpenAI, Google, maybe even Apple one day comes out with a product like this, that would be wonderful and I love competition. I think just like the wild west nature of open-source and how rapidly it was iterating, how many eyeballs, how much excitement, that is what I was super into. But look, it’s inevitable that all of the other companies are going to gain inspiration from what Open Interpreter did, and that’s fine. And yeah, competition is always good.

Jason Grad: 55:21 I agree with that. Uh, I have one more question for Ryan before we wrap up here. One thing I’ve been hearing quite a lot about when it comes to people using one agent or multiple agents inside of Open Interpreter is context loss. And when it comes to me using Open Interpreter or any other kind of agentic system, it doesn’t matter as much because I’m tinkering, I’m experimenting, I’m kind of hacking around. But using anything in a production setting is very different. And I know that, you know, Nexu uses AI. So how large of a problem is context loss? And is it a problem much like hallucinations that’s going away as time goes on, or is it ramping up in terms of an issue for you and your company?

Ryan Yannelli: 55:57 I mean, it’s something we had to tackle really early on because when the context windows were what, I think like 8,000 tokens, we had to compress a lot of data. Um, so even to this day, what we’re still doing is, and even in my personal agents, is just compressing data, taking out the parts that don’t matter, and just kind of fitting it into what’s more relevant just for that conversation or that specific agent. But just by taking a document, chunking it out, making it more concise, that’s been really helpful for context loss.

Jason Grad: 56:31 Jason, the new Sonnet 4.6 context window, I think is a million tokens. Does that solve the context loss issue just because we can now dump so much into the context window that we don’t need to worry as much about losing it as we go along?

Jason Calacanis: 56:44 I haven’t tested Sonnet before I could say too much about it, but I don’t think that… It feels like there’s a limit, there’s some kind of limit of context regardless of what the token limit is, where…

Jason Grad: 57:00 Models start to get a little confused or they leak memory to memory and so as you add more context and go down and have more and more chats I find that they get less accurate and start bringing in things from different conversations that don’t make sense anymore. So I don’t think it’s a more context, um, for me with with Opus now what I’ve been doing is deleting a lot of old conversations after I have them and trying to get rid of this and also I build my own instruction manuals for how to do various workflows and then have that loaded into different conversations and so I can just start in conversations and I want to start this conversations fresh. So I don’t think it’s a more context thing, I think it’s about having better instructions to achieve specific goals, uh, if you’re, you know, as you become a power user.

Jason Calacanis: 57:52 What we need then is a OpenClaude skill that deletes unnecessary old conversations to clear up your context window so you don’t lose context in OpenClaude, and that I think is exactly why we’re not mainstream yet with this stuff because we’re still clearly running up against things that will get ironed out in the future but right now remain not thorny problems but things that are annoying enough that I think the average person would just throw up their hands and kind of walk away. Ugh, that’s too bad. All right, one more question before we go, Matt. So people have been talking a lot about using Mac Minis, very fun, to set up their own kind of like, you know, hubs for their agents. The share price of Raspberry Pi, a British company that makes very tiny computers, recently went up by like 50% because people were thinking about, ‘Hey, maybe we can use Raspberry Pis to power our at-home agents.’ I’m just curious about what you recommend for the average person if they’re getting started and if you do think that Raspberry Pi has a shot at gaining any of that sweet, sweet Mac Mini money from Apple.

Matthew Berman: 58:54 I would probably recommend using an old computer that you already have that you don’t mind wiping. If you don’t have an old computer, the next best option is to use a VPS, some kind of fully hosted service. And we talked about this earlier, as soon as other companies start coming out with their own version, fully, like really actually hosted, you don’t have to set anything up, that’s probably going to be the best way. But for now, if I’m going to recommend it, like get it, I have a MacBook Air sitting on my desk right now, that’s where everything goes to, I wiped it clean, installed only what OpenClaude needs and nothing else, and that’s what I recommend.

Jason Calacanis: 59:44 Well, we really appreciate it. Matt, if people want to find your YouTube channel, it’s youtube.com/@matthew_berman. Anything else you want to plug?

Matthew Berman: 59:52 Forward Future, find us at Forward Future on Twitter, @forwardfuture.

Jason Calacanis: 59:57 Love it. Ryan, Nextvisit. Apart from nextvisit.ai, anything you want to pitch or plug as we wrap up here?

Matthew Berman: 60:00 I’ll post all my OpenClaw stuff on there and keep skills updated.

Ryan Yannelli: 60:03 Um, you know, just say if you’re interested in, you know, looking at what I create, check out my GitHub, Yannelli, and, um, you know, I’ll be posting more projects there as well. Awesome. And Jason, we have joinmassive.com. Anything else that people should know about what you’re working on or what you’re looking for?

Jason Grad: 60:11 Yeah, find me on Twitter at Jason Grad without the A, like JSON, or on LinkedIn, I’m around. Massive is scraping huge amounts of data for our clients. They’re doing the scraping on our infrastructure every month. We’re excited to unlock this capability in Clawpod. I will probably drop a guide tomorrow on an easy setup because I see it’s a huge problem for how to get this rolling on a VPS and I think, you know, we got a lot of exciting stuff, exciting stuff planned for Clawpod, so just stay tuned. I think we’ll have some more things launch in the next week or so.

Ryan Yannelli: 60:47 Well, make sure to keep us abreast and next time there’s a major update, we’ll have you back on to show us what it can do and how people can use it. But in the meantime, my friends, another TWiST is in the can. We’ll see you on Friday.