1 00:00:00,120 --> 00:00:04,800 This episode is about Open Claw, the open source AI agent tool 2 00:00:04,800 --> 00:00:09,000 that went from a side project by just a single Austrian developer 3 00:00:09,000 --> 00:00:12,440 to what the CEO of NVIDIA, Jensen Huang, a couple of weeks 4 00:00:12,440 --> 00:00:16,200 ago described as probably the single most important release of 5 00:00:16,200 --> 00:00:19,480 software probably ever. He compared it to Windows, 6 00:00:19,520 --> 00:00:22,840 Linux, HTML, and announced that NVIDIA were building their 7 00:00:22,840 --> 00:00:25,520 entire enterprise agent strategy on top of it. 8 00:00:25,760 --> 00:00:29,920 In today's episode, I'm going to explain how this project became 9 00:00:30,320 --> 00:00:34,520 such a global sensation. I'm going to explain exactly why 10 00:00:34,560 --> 00:00:38,520 it's one of the most important software updates ever, why it's 11 00:00:38,520 --> 00:00:41,600 potentially more important than the web browser, than Linux, 12 00:00:41,880 --> 00:00:43,640 than even the iPhone operating system. 13 00:00:43,640 --> 00:00:47,520 And I'm going to also explain what it actually is, what's 14 00:00:47,520 --> 00:00:51,320 innovative about it, what's new, and what it means for all of us. 15 00:00:51,320 --> 00:00:53,720 This is In the Loop with Jack Harton. 16 00:00:54,400 --> 00:01:07,290 Hope you enjoy the show. Let's start with what open core 17 00:01:07,290 --> 00:01:09,650 actually is, because the first thing to understand about open 18 00:01:09,650 --> 00:01:11,490 Core is it's it's not just a normal agent. 19 00:01:11,490 --> 00:01:15,010 You know, yes, it is a chat bot still because, you know, all 20 00:01:15,010 --> 00:01:17,650 agents are mostly chat bots and it feels mostly the same, but 21 00:01:17,650 --> 00:01:19,370 underneath it's something very, very different. 22 00:01:20,090 --> 00:01:22,970 Essentially what open core is, is a long running no dot JS 23 00:01:22,970 --> 00:01:25,210 background service, often called a demon. 24 00:01:25,410 --> 00:01:30,730 So in software terms, that is essentially a computer program 25 00:01:30,730 --> 00:01:32,560 that runs as a background process. 26 00:01:32,560 --> 00:01:36,840 So it doesn't rely on a user to initiate the process. 27 00:01:37,200 --> 00:01:40,520 So it might handle typically on a, you know, a device, maybe 28 00:01:40,520 --> 00:01:42,800 something like automatically connecting to your Wi-Fi. 29 00:01:42,960 --> 00:01:47,200 Now open claw connects to your file system, your browser, any 30 00:01:47,200 --> 00:01:49,800 APIs, your calendar, your e-mail. 31 00:01:49,800 --> 00:01:52,200 It's on your laptop. It can run on a schedule. 32 00:01:52,200 --> 00:01:55,280 It monitors all those applications for things that are 33 00:01:55,280 --> 00:01:58,160 happening called events. It can change a series of 34 00:01:58,160 --> 00:02:01,240 actions together and skills together, and it can message 35 00:02:01,240 --> 00:02:05,240 you, you know, on text or Messenger or WhatsApp or 36 00:02:05,240 --> 00:02:08,039 whatever it might be when it needs your input or attention. 37 00:02:08,240 --> 00:02:12,200 Now, if you compare that to ChatGPT, which is a conversation 38 00:02:12,200 --> 00:02:15,960 that you must start, whereas as long as your PC is open or you 39 00:02:15,960 --> 00:02:20,160 got something to host it and keep it running, it's literally 40 00:02:20,160 --> 00:02:22,120 always on. Then there's the interface, or 41 00:02:23,000 --> 00:02:25,520 in this particular case, the lack of a web interface. 42 00:02:25,520 --> 00:02:28,840 So there's no open claw, you know, web app that you make an 43 00:02:28,840 --> 00:02:32,560 account with and log into. Instead, it works inside over 15 44 00:02:32,560 --> 00:02:34,920 different messaging platforms that, you know, WhatsApp, 45 00:02:35,440 --> 00:02:39,000 Telegram, Slack, Discord, Signal, iMessage, Microsoft 46 00:02:39,000 --> 00:02:41,680 Teams, Google Chart. And you can just very simply 47 00:02:41,680 --> 00:02:44,640 text it the same way that you'd text someone that is in your 48 00:02:44,640 --> 00:02:48,360 team, you know, reschedule my Thursday meeting or, you know, 49 00:02:48,640 --> 00:02:52,080 go to my emails, summarize them all, triage and prioritize them 50 00:02:52,080 --> 00:02:55,320 all, write e-mail responses, and then I'll approve them. 51 00:02:55,480 --> 00:02:58,960 Summarize the PDFs that have been sent to me across all these 52 00:02:58,960 --> 00:03:01,200 different, you know, Microsoft Teams are Slack chats. 53 00:03:01,480 --> 00:03:03,600 And I have to say, these types of things, when you try them, 54 00:03:03,600 --> 00:03:06,320 they are quite magical. And that zero friction approach 55 00:03:06,320 --> 00:03:10,720 to an agent doing something for you is why the adoption curve 56 00:03:10,720 --> 00:03:14,120 has been so fast. I mean, staggeringly fast. 57 00:03:14,400 --> 00:03:19,400 It's faster than ChatGPT, roughly 20 times faster by a 58 00:03:19,400 --> 00:03:22,680 many visitor growth numbers. You know, every AI company spent 59 00:03:22,680 --> 00:03:25,800 billions trying to get people to come to their own interface, 60 00:03:25,800 --> 00:03:27,840 whereas Open Cloud simply went to them. 61 00:03:28,000 --> 00:03:30,920 And you know, Anthropic and all the big platforms have of course 62 00:03:30,920 --> 00:03:34,360 copied almost all of it. They copied the messaging first 63 00:03:34,360 --> 00:03:37,680 approach with Cloud Code or Cloud Code channel, shall I call 64 00:03:37,680 --> 00:03:39,920 it. They've got Bishop, Telegram and 65 00:03:39,920 --> 00:03:43,240 Discord integrations within a few more weeks that brought out 66 00:03:43,240 --> 00:03:45,880 dispatch, that brought out schedules. 67 00:03:45,880 --> 00:03:48,640 So many of the different features that Open Claw built, 68 00:03:48,920 --> 00:03:53,400 they've provided to users as part of the $20 or whatever 69 00:03:53,400 --> 00:03:57,240 membership T you're on, but in a much more secure and Safeway. 70 00:03:57,360 --> 00:03:59,760 And I think the speed of their response tells you a lot. 71 00:03:59,920 --> 00:04:02,080 An open eye even went and bought Open Claw. 72 00:04:02,240 --> 00:04:05,600 Now buying Open Claw is more buying the talent and the 73 00:04:05,600 --> 00:04:08,640 people, which was basically just the main Austrian developer who 74 00:04:08,640 --> 00:04:12,120 built it all because it was all open source and everybody could 75 00:04:12,120 --> 00:04:14,280 access the technology and see how it was built anyway. 76 00:04:29,760 --> 00:04:33,280 Now the skills and MCP ecosystem is where it gets really 77 00:04:33,280 --> 00:04:37,920 interesting because a skill in Open Core is just a markdown 78 00:04:37,920 --> 00:04:41,080 file, so a document type, so just a document type containing 79 00:04:41,320 --> 00:04:44,560 human readable, so normal text instructions for a specific 80 00:04:44,560 --> 00:04:46,720 task. Now if you've listened to last 81 00:04:46,720 --> 00:04:49,360 week's episode, you'll understand exactly what a skill 82 00:04:49,360 --> 00:04:51,320 is. I really recommend going listen 83 00:04:51,320 --> 00:04:53,560 to that episode if you haven't. And it's one of the most 84 00:04:53,560 --> 00:04:58,160 underutilized new capabilities that all AI systems globally are 85 00:04:58,160 --> 00:05:00,440 now adopting. And any of those skills are able 86 00:05:00,440 --> 00:05:02,920 to call into and connect to MCP servers. 87 00:05:03,480 --> 00:05:08,040 So model context protocol, which is basically a wrap around of 88 00:05:08,040 --> 00:05:11,000 integrations APIs. So it makes it really easy for 89 00:05:11,000 --> 00:05:13,800 an agent to take actions in different systems. 90 00:05:13,800 --> 00:05:17,320 So suddenly this agent has access to a huge ecosystem of 91 00:05:17,320 --> 00:05:21,920 skills and a huge ecosystem of integrations and control of your 92 00:05:21,920 --> 00:05:24,760 entire PC, and it can run around the clock for you. 93 00:05:25,760 --> 00:05:29,520 You can see why it's an incredibly innovative approach 94 00:05:29,520 --> 00:05:31,240 to how agents can do work for people. 95 00:05:31,280 --> 00:05:32,880 And that community really blew up. 96 00:05:32,880 --> 00:05:37,400 So Claw Hub hosts more than 13,000 community contributed 97 00:05:37,400 --> 00:05:40,960 skills and basically each skill is like a mini app for an agent. 98 00:05:41,120 --> 00:05:42,960 And then there's also memory open. 99 00:05:42,960 --> 00:05:46,720 Claw stores everything in just markdown on your local machine, 100 00:05:46,720 --> 00:05:49,000 such human readable. You can look at it, you can edit 101 00:05:49,000 --> 00:05:52,520 it, and it essentially builds a three layer memory system. 102 00:05:53,040 --> 00:05:55,240 And then what's really interesting about open claw is 103 00:05:55,240 --> 00:05:59,200 the token economics. So obviously tokens are when 104 00:05:59,320 --> 00:06:02,520 word start getting generated by an LLMA tickle. 105 00:06:02,960 --> 00:06:07,280 A typical ChatGPT prompt uses maybe a few 1000 tokens and a 106 00:06:07,280 --> 00:06:11,120 genetic open claw task where the agent plans, executes, checks 107 00:06:11,120 --> 00:06:16,600 your PC does work, re pawns, re executes, all whilst you know 108 00:06:16,640 --> 00:06:20,840 you're having a shower. It can consume 1000 times more 109 00:06:21,160 --> 00:06:23,360 tokens and chat chi BT. So you can just imagine a 110 00:06:23,360 --> 00:06:26,360 continuously running background agent, the kind that Jensen 111 00:06:26,920 --> 00:06:30,080 Huang is envisioning everyone having, is going to burn 112 00:06:30,120 --> 00:06:34,640 millions times more tokens per day than a single chat chi BT 113 00:06:34,640 --> 00:06:37,040 conversation. Now obviously the CEO of the 114 00:06:37,040 --> 00:06:40,880 company that sells everybody computer chips wants this to be 115 00:06:41,160 --> 00:06:44,240 how everyone consumes credits, because obviously that means 116 00:06:44,240 --> 00:06:46,480 everyone needs more computer chips from them. 117 00:07:02,120 --> 00:07:04,480 So yeah, I think this dovetails really nicely into our Jensen 118 00:07:04,480 --> 00:07:07,920 firing is so excited because he called Open Claw the operating 119 00:07:07,920 --> 00:07:11,760 system of agentic computers. It's huge big conference just a 120 00:07:11,760 --> 00:07:14,560 couple of weeks ago. So he said, and I quote Mac and 121 00:07:14,560 --> 00:07:16,440 Windows are the operating systems for the personal 122 00:07:16,440 --> 00:07:20,440 computer and Open Core is the operating system for personal 123 00:07:20,440 --> 00:07:22,480 AI. You know, Windows and Apple made 124 00:07:22,480 --> 00:07:25,440 personal computers accessible just to everybody, and Open 125 00:07:25,440 --> 00:07:29,360 Claw, in his framing, does the exact same for personal agents. 126 00:07:30,000 --> 00:07:33,640 And the deeper analogy is Linux. You know, Linux has become 127 00:07:33,640 --> 00:07:35,840 basically the invisible infrastructure of the Internet. 128 00:07:36,120 --> 00:07:39,240 You know, every cloud server, every Android phone, every smart 129 00:07:39,240 --> 00:07:43,440 devices runs on a Linux, and they built all operating systems 130 00:07:43,440 --> 00:07:45,880 on top of Linux. And Jensen is betting that Open 131 00:07:45,880 --> 00:07:49,760 Claw becomes that invisible infrastructure of the agent era. 132 00:07:49,760 --> 00:07:53,600 Linux took 30 years to reach ubiquitous global adoption. 133 00:07:53,600 --> 00:07:56,760 Open Claw, on the other hand, surpassed that in three weeks. 134 00:07:56,920 --> 00:08:00,200 Yes, you heard that right. Linux is one of the most 135 00:08:00,360 --> 00:08:03,520 important open source technologies ever invented, is 136 00:08:03,520 --> 00:08:07,680 literally the basis of most technology, and Open Claw has 137 00:08:07,680 --> 00:08:11,920 become more widely adopted than Linux in just a few weeks. 138 00:08:12,040 --> 00:08:15,880 So yeah, NVIDIA have obviously announced massive plans to make 139 00:08:15,880 --> 00:08:20,200 Open Claw the fundamental pillar of its strategy so it can then 140 00:08:20,200 --> 00:08:24,160 increase token generation globally from say 2 million 141 00:08:24,160 --> 00:08:27,120 tokens per second, which is where we're at today, to about 142 00:08:27,120 --> 00:08:32,640 700 million tokens a second, which is about 350% increase. 143 00:08:32,799 --> 00:08:35,960 The scale of what Jensen Huang is describing is obviously just 144 00:08:35,960 --> 00:08:38,200 ridiculous. You know, NVIDIA currently has 145 00:08:38,640 --> 00:08:44,800 43,000 employees and he projects 75,000 human employees alongside 146 00:08:44,800 --> 00:08:48,840 7 and a half million open call like AI agents over the next 10 147 00:08:48,840 --> 00:08:50,760 years. For NVIDIA alone, that's about 148 00:08:50,760 --> 00:08:54,880 100 digital workers per person and the things that you can get 149 00:08:54,880 --> 00:08:58,440 open call to do, and very soon all the main platforms are going 150 00:08:58,440 --> 00:09:01,440 to enable you to do the same on their systems. 151 00:09:01,600 --> 00:09:05,560 The scale of production is quite outrageous and the number of 152 00:09:05,560 --> 00:09:09,240 tokens that are going to be consumed is ridiculous, which 153 00:09:09,240 --> 00:09:11,960 means that obviously Nvidia's chips are going to be very, very 154 00:09:11,960 --> 00:09:13,920 important. So for those that you know, want 155 00:09:13,920 --> 00:09:18,040 to jump into testing this, what I'd say is that you can a wait 156 00:09:18,480 --> 00:09:22,040 for this to be fully released on all main systems that you feel 157 00:09:22,040 --> 00:09:23,560 really comfortable with and trusted. 158 00:09:24,120 --> 00:09:27,400 When that starts to happen, you know, core code and core Co 159 00:09:27,400 --> 00:09:29,360 works, let's say, has already started doing that. 160 00:09:29,360 --> 00:09:33,280 I'm going to do an episode as a full breakdown, but if you 161 00:09:33,280 --> 00:09:35,680 really want to get into it and I would recommend people having a 162 00:09:35,680 --> 00:09:38,760 go, I'll give you a few tips here and also how to do it 163 00:09:38,760 --> 00:09:54,320 safely. So if you want to set it up, 164 00:09:54,840 --> 00:09:58,160 ideally you have clock code maybe running in an IDE like 165 00:09:58,160 --> 00:10:00,520 Visual Studio Codes. And then it makes it really, 166 00:10:00,520 --> 00:10:03,200 really simple because it will just help you get through it 167 00:10:03,200 --> 00:10:04,880 all. So if you want to do it, you can 168 00:10:04,880 --> 00:10:10,440 go and look for Open Claw on GitHub and obviously just pull 169 00:10:10,440 --> 00:10:12,240 it into whatever ID you're using. 170 00:10:12,480 --> 00:10:17,200 You can go and search for Open Claw, clone the repo, go and 171 00:10:17,200 --> 00:10:21,160 find the web URL, pull it into Claw code and you'll be able to 172 00:10:21,160 --> 00:10:23,480 get guided through it. Quite frankly, I pick a use 173 00:10:23,480 --> 00:10:25,120 case. It's quite simple, you know, 174 00:10:25,360 --> 00:10:29,040 daily updates through my emails, for example, all that triaging 175 00:10:29,040 --> 00:10:30,440 of emails. That's quite a nice one to get 176 00:10:30,440 --> 00:10:32,320 into. If you've not got claw code in 177 00:10:32,320 --> 00:10:36,160 like an ID or whatever, maybe go and check out Rep or cursor. 178 00:10:36,160 --> 00:10:38,200 It's really, really simple. It honestly takes just a few 179 00:10:38,200 --> 00:10:40,560 minutes if you haven't. The experience is a bit 180 00:10:40,560 --> 00:10:43,680 friendlier if you're not used to being in, you know, Visual 181 00:10:43,680 --> 00:10:47,600 Studio Code. And again, just find the repo, 182 00:10:47,600 --> 00:10:49,920 get it in there, and you'll get guided through the process. 183 00:10:50,040 --> 00:10:54,480 The one tip I'd really give for everyone is do not run this on 184 00:10:54,480 --> 00:10:57,720 your main machine. So do not use this on your main 185 00:10:57,720 --> 00:10:59,600 laptop. It's why people have been 186 00:10:59,600 --> 00:11:02,440 rushing out to buy Mac minis. I think in the future, Mac minis 187 00:11:02,440 --> 00:11:08,320 are going to be very, very, very popular because as we need 24/7 188 00:11:08,320 --> 00:11:11,440 agents, you need your local machine to be turned on for an 189 00:11:11,440 --> 00:11:14,800 agent to be doing work. And so I think just Mac Mini's 190 00:11:14,800 --> 00:11:17,120 running in the background all the time and going to just have 191 00:11:17,120 --> 00:11:19,000 agents. And it means I think we're going 192 00:11:19,000 --> 00:11:21,840 to have Mac mini's always on in the background just to enable 193 00:11:21,840 --> 00:11:25,160 our agents to do work for us. But anyway, don't set all of 194 00:11:25,160 --> 00:11:27,320 this up on your local machine because there's a massive 195 00:11:27,320 --> 00:11:29,840 security risk. You know, I think it was like a 196 00:11:29,840 --> 00:11:35,240 third of all public skills made had some form of malicious 197 00:11:35,240 --> 00:11:39,280 software or malicious prompt injecting hiddens inside it. 198 00:11:39,280 --> 00:11:41,040 So do just be super careful there. 199 00:11:41,160 --> 00:11:44,680 Once you give this access to your PC, it can literally go and 200 00:11:44,680 --> 00:11:47,200 access your personal e-mail, your banking details, 201 00:11:47,200 --> 00:11:48,880 everything. So I'm not just saying this to 202 00:11:48,880 --> 00:11:51,640 be, you know, fear mongering because it's very easy to be 203 00:11:51,640 --> 00:11:55,040 safe, but just make sure that you don't use your local main 204 00:11:55,040 --> 00:11:57,080 PC. So where does this leave us? 205 00:11:58,120 --> 00:12:01,680 Well, Jensen Huang called Open Claw probably the single most 206 00:12:01,680 --> 00:12:04,120 important release of software probably ever. 207 00:12:04,720 --> 00:12:08,000 And the claim is obviously incredibly self-serving The it's 208 00:12:08,000 --> 00:12:12,560 a trillion dollar thesis on how can we sell loads of computer 209 00:12:12,560 --> 00:12:16,080 chips because the world is consuming loads of tokens from 210 00:12:16,080 --> 00:12:19,400 large language models. I think it's probably correct or 211 00:12:19,400 --> 00:12:22,440 at least closer to correct than it may be first sounds. 212 00:12:22,520 --> 00:12:25,600 This is the greatest leap towards the dream of agents 213 00:12:25,600 --> 00:12:28,760 that, you know, people have talked about online for so long 214 00:12:29,400 --> 00:12:32,640 and every model provider is racing to add open clause 215 00:12:32,640 --> 00:12:36,320 features to their products. So I think very, very soon 216 00:12:36,320 --> 00:12:40,200 you're going to see these type of capabilities democratized and 217 00:12:40,200 --> 00:12:43,000 accessible to everyone. Anyway, thanks for listening and 218 00:12:43,000 --> 00:12:44,680 I'll see you next week.