1 00:00:00,120 --> 00:00:02,840 One of the only steps that we're seeing in the direction of 2 00:00:02,840 --> 00:00:05,080 bringing that power to the average person and letting them 3 00:00:05,080 --> 00:00:07,280 actually reap the benefits of this new technology. 4 00:00:07,280 --> 00:00:09,560 This one is a safe, improving, very personal. 5 00:00:09,600 --> 00:00:12,880 I think it's a new paradigm for how to interact with AI to have 6 00:00:12,880 --> 00:00:14,680 your own age. And that's a major advantage. 7 00:00:14,680 --> 00:00:16,800 And it's like people aren't even thinking on that. 8 00:00:16,800 --> 00:00:18,960 But it's, yeah, it's beyond easier than ever. 9 00:00:18,960 --> 00:00:21,880 Without an open source alternative, it's just all of 10 00:00:21,880 --> 00:00:26,520 the power in the hands of them very small people for this 11 00:00:26,560 --> 00:00:28,400 civilization changing technology. 12 00:00:28,640 --> 00:00:32,040 If there is some asymmetrical intelligence advantage where two 13 00:00:32,040 --> 00:00:35,320 or three labs have access to most of the intelligence and 14 00:00:35,320 --> 00:00:38,200 everyone else is basically on like the Kitty version, you're 15 00:00:38,240 --> 00:00:40,400 in a super asymmetric power advantage. 16 00:00:40,760 --> 00:00:48,200 That's a terrifying thing to me. There is a race happening right 17 00:00:48,200 --> 00:00:51,200 now for what the future of your AI agent should be. 18 00:00:51,200 --> 00:00:54,800 One vision is that held by some of the largest AI labs in the 19 00:00:54,800 --> 00:00:57,000 world. Another was demonstrated through 20 00:00:57,000 --> 00:01:00,920 one of the fastest growing projects introducing the masses 21 00:01:01,040 --> 00:01:04,480 to what AI assistance could be. And the third is put together by 22 00:01:04,480 --> 00:01:08,400 a team whose goal is not even the tool itself, but what comes 23 00:01:08,400 --> 00:01:11,160 next. I spend nearly all of my time in 24 00:01:11,160 --> 00:01:13,560 this space, and I feel like I can barely keep up. 25 00:01:14,120 --> 00:01:17,680 It moves so fast, so much changes, that the only way you 26 00:01:17,680 --> 00:01:20,600 can get a true idea of the state of things today and the 27 00:01:20,600 --> 00:01:24,320 trajectory we're on for tomorrow is by having conversations with 28 00:01:24,320 --> 00:01:26,720 people who are actively building in the space. 29 00:01:27,080 --> 00:01:30,240 So today I'm talking with five people ranging from Co founder 30 00:01:30,240 --> 00:01:34,040 of an AI research lab to non-technical builders to AI 31 00:01:34,040 --> 00:01:36,760 engineering veterans so that I can get the full picture. 32 00:01:36,920 --> 00:01:39,520 What are these agents actually capable of today? 33 00:01:39,520 --> 00:01:42,520 And what's the gap between the hype and the reality? 34 00:01:42,720 --> 00:01:45,520 And does it matter which agent you actually choose? 35 00:01:45,600 --> 00:01:49,400 This episode is me trying to fill those gaps for all of us to 36 00:01:49,400 --> 00:01:54,240 learn from. So just to make sure everyone's 37 00:01:54,240 --> 00:01:58,640 on the same page, it's mid 2026. Ambient agents have replaced 38 00:01:58,640 --> 00:02:01,440 chat bots. If you have to initiate every 39 00:02:01,440 --> 00:02:04,560 interaction with your AI, you are falling way behind. 40 00:02:04,600 --> 00:02:07,480 Every query that you send that doesn't contribute to your 41 00:02:07,480 --> 00:02:11,000 system is lost potential, and every recurring task that you 42 00:02:11,000 --> 00:02:13,400 have to manually initiate is a waste of time. 43 00:02:13,680 --> 00:02:16,520 I like to think of it like this. Chat bots are your gateway to 44 00:02:16,520 --> 00:02:19,520 intelligence. Ambient agents are your 45 00:02:19,520 --> 00:02:22,520 intelligence flywheel. One cool thing with these agents 46 00:02:22,520 --> 00:02:24,960 is just the depth of information you get. 47 00:02:24,960 --> 00:02:27,680 So instead of just telling you what the weather is, it can 48 00:02:27,680 --> 00:02:30,320 inform you how the weather will impact your plans and how to 49 00:02:30,320 --> 00:02:32,440 adjust accordingly. Instead of just telling you your 50 00:02:32,440 --> 00:02:34,640 schedule for the day or your calendar, it can help you 51 00:02:34,640 --> 00:02:36,680 prepare for every commitment you have that day. 52 00:02:36,720 --> 00:02:39,760 And it can help you optimize your scheduling in the future. 53 00:02:39,800 --> 00:02:41,440 And you get more depth to research. 54 00:02:41,440 --> 00:02:44,680 Instead of what are the top restaurants in the area, you can 55 00:02:44,680 --> 00:02:47,000 actually have it planned out a full special occasion. 56 00:02:47,000 --> 00:02:49,440 And instead of just getting and in answer to a question, you can 57 00:02:49,440 --> 00:02:52,000 have it cited sources. You can understand the why 58 00:02:52,000 --> 00:02:55,160 behind it, the how behind it instead of just being fed an 59 00:02:55,160 --> 00:02:57,680 answer. Also for this video, ambient 60 00:02:57,680 --> 00:03:00,320 agents, harnesses, AI assistants, they all mean the 61 00:03:00,320 --> 00:03:03,280 same thing. It's a system that leverages AI 62 00:03:03,400 --> 00:03:04,920 in order to interact with the world. 63 00:03:04,960 --> 00:03:07,240 And these systems are supposed to be designed so that they 64 00:03:07,240 --> 00:03:09,640 align with the user. They can interact with the user 65 00:03:09,640 --> 00:03:12,080 wherever they're at and they can do meaningful work. 66 00:03:12,240 --> 00:03:16,120 So with these ambient agents, this isn't new technology. 67 00:03:16,360 --> 00:03:19,560 Nothing in these harnesses was was invented in the last six 68 00:03:19,560 --> 00:03:21,080 months. I mean, look at Open 69 00:03:21,080 --> 00:03:24,640 Interpreter, it existed nearly three years ago and AI has been 70 00:03:24,640 --> 00:03:27,200 available over API endpoints much longer than that. 71 00:03:27,640 --> 00:03:30,360 So what is it now that has caused this explosion and 72 00:03:30,360 --> 00:03:32,600 harnesses to start percolating everywhere? 73 00:03:32,680 --> 00:03:35,400 The models weren't really there yet, You know what I mean. 74 00:03:36,640 --> 00:03:41,160 With the advent of 4.5 opus, a 4.6 opus and then GLM 5 minimax 75 00:03:41,160 --> 00:03:45,000 2.7 like today's models a 2026 time right? 76 00:03:45,920 --> 00:03:48,560 The models are there. They're there for you to do 77 00:03:48,560 --> 00:03:50,840 useful work, economically valuable work. 78 00:03:51,280 --> 00:03:53,600 And that's the thing. The models are finally at the 79 00:03:53,600 --> 00:03:56,160 point where they can do economically valuable work 80 00:03:56,160 --> 00:03:59,080 without being cost prohibitive. Yes, they can still be 81 00:03:59,080 --> 00:04:02,120 expensive, and yes, they can still blow simple tasks, but 82 00:04:02,120 --> 00:04:04,480 adoption doesn't require perfection. 83 00:04:04,560 --> 00:04:07,160 I've been working on AI tooling for four years, and I've seen 84 00:04:07,160 --> 00:04:10,320 many systems of scaffolding built to support much worse 85 00:04:10,320 --> 00:04:12,920 models. And now you don't need to jump 86 00:04:12,920 --> 00:04:16,640 through hoops to drive value. Countless smart people have been 87 00:04:16,640 --> 00:04:20,040 improving every single aspect of the stack to get us to this 88 00:04:20,040 --> 00:04:22,000 point. And if you feel that things are 89 00:04:22,000 --> 00:04:23,880 moving faster than ever, you're correct. 90 00:04:24,080 --> 00:04:27,520 There are so many compounding technologies, and the flywheel 91 00:04:27,520 --> 00:04:29,440 effect is spinning faster than ever. 92 00:04:29,680 --> 00:04:31,800 Robert says it best. I can't even believe what's 93 00:04:31,800 --> 00:04:34,200 happened in the last two months. It's mind boggling. 94 00:04:34,400 --> 00:04:36,600 Now there are so many harnesses these days. 95 00:04:36,920 --> 00:04:40,280 Clawed code started as strictly A coding tool, but people 96 00:04:40,280 --> 00:04:42,360 started using it in non coding workflows. 97 00:04:42,760 --> 00:04:46,160 Openclaw went mainstream and introduced many people to what's 98 00:04:46,160 --> 00:04:49,920 possible with AI agents, and a plethora of variants started 99 00:04:49,920 --> 00:04:52,720 showing up. Open claw, Nemo claw, Iron claw, 100 00:04:52,880 --> 00:04:56,280 Nano claw, 0 claw. It was getting to be a bit much, 101 00:04:56,400 --> 00:05:00,400 but one day a lab released something fully open, MIT 102 00:05:00,400 --> 00:05:02,600 licensed, with a strong community around it. 103 00:05:03,000 --> 00:05:06,320 A thoughtful approach with taste driven by emission. 104 00:05:12,200 --> 00:05:16,120 When Hermes Agent came out, I switched to it two weeks, 3 105 00:05:16,160 --> 00:05:21,200 weeks ago because of that continuous instability of open 106 00:05:21,200 --> 00:05:23,280 claw. So I decided to just give it a 107 00:05:23,280 --> 00:05:26,560 try. And Hermes Agent has been so 108 00:05:26,560 --> 00:05:29,080 flawless out-of-the-box. That was so amazing. 109 00:05:29,080 --> 00:05:31,840 I think we have ended up the age of personalized software. 110 00:05:31,840 --> 00:05:33,560 You have your agent. You want a feature. 111 00:05:33,560 --> 00:05:36,720 You tell your agent it has access to the soft code, it has 112 00:05:36,720 --> 00:05:39,840 access to the docs, and it will just implement whatever you need 113 00:05:40,240 --> 00:05:42,320 and make the changes. Yeah, I can ask the agent 114 00:05:42,320 --> 00:05:43,840 anything. It looks at the source code and 115 00:05:43,840 --> 00:05:47,080 it changes it and it learns. I do think it's like one of the 116 00:05:47,360 --> 00:05:50,680 better agents that out there and one of the ones that's more 117 00:05:50,680 --> 00:05:53,320 available to people really and more understandable. 118 00:05:53,320 --> 00:05:55,400 It's less of a black box. I think which is the really 119 00:05:55,400 --> 00:05:58,400 important thing that people are afraid of with AI right now is 120 00:05:58,400 --> 00:06:01,000 that there's it's completely opaque and you don't know 121 00:06:01,000 --> 00:06:02,480 anything. About it, Hermes was very 122 00:06:02,480 --> 00:06:07,880 capable out-of-the-box, so I'm using a lot of the built in 123 00:06:07,880 --> 00:06:13,400 browser searching functionality, created my own set of core 124 00:06:13,440 --> 00:06:17,480 skills that Hermes uses, and then Hermes also has the ability 125 00:06:17,480 --> 00:06:21,160 to iteratively create its own skills as it encounters 126 00:06:21,240 --> 00:06:24,720 challenging workflows as well. Wait, it can make its own 127 00:06:24,720 --> 00:06:26,600 skills? It has in the background there's 128 00:06:26,640 --> 00:06:31,160 a task, whereas another agent, basically a sub agent runs and 129 00:06:31,160 --> 00:06:34,320 looks at what you have done and decides if it's worth to create 130 00:06:34,320 --> 00:06:37,480 a memory or update a skill. And that is a brilliant concept 131 00:06:37,480 --> 00:06:39,560 because it doesn't interrupt your conversation. 132 00:06:39,720 --> 00:06:42,160 What I did with Open Claw, when I was talking to it, I always 133 00:06:42,160 --> 00:06:45,440 said OK, we have reached a point you have to remember all of this 134 00:06:45,640 --> 00:06:48,880 now put it in a skill and put it in your memory and so on. 135 00:06:49,160 --> 00:06:51,760 And later I had to create this to make sure it is in there. 136 00:06:52,000 --> 00:06:53,640 With the Hermes agent, it just happens. 137 00:06:53,640 --> 00:06:58,560 Like it has a much more of an aptitude to learn more and 138 00:06:59,440 --> 00:07:02,360 implement those skills in itself, and it will like let you 139 00:07:02,360 --> 00:07:04,800 know when there's something that you could do to increase 140 00:07:04,800 --> 00:07:06,920 capabilities. Skills are a vital part of these 141 00:07:06,920 --> 00:07:08,960 systems, being able to perform work reliably. 142 00:07:08,960 --> 00:07:12,320 Because think of a skill like a playbook, A repeatable process. 143 00:07:12,640 --> 00:07:15,040 If you give your AI assistant this process, it's able to 144 00:07:15,040 --> 00:07:17,720 follow it step by step instead of generating the plan every 145 00:07:17,720 --> 00:07:20,160 single time. And instead of you the user 146 00:07:20,160 --> 00:07:23,640 having to make a skill for every single workflow you want, Hermes 147 00:07:23,640 --> 00:07:25,440 is able to automatically do that for you. 148 00:07:25,640 --> 00:07:27,840 And if it's not correct, you can have it updated. 149 00:07:27,960 --> 00:07:31,320 It's really a form of memory and Hermes powerful memory system 150 00:07:31,320 --> 00:07:33,240 goes much deeper than just skills. 151 00:07:33,320 --> 00:07:35,680 The memory is excellent. That was all the one thing I 152 00:07:35,680 --> 00:07:39,040 spent a whole night a day working on the memory feature of 153 00:07:39,040 --> 00:07:42,520 Open cloud right to the QMD and there are so many open issues. 154 00:07:42,520 --> 00:07:44,800 When I switched to Hermes, it comes out-of-the-box with the 155 00:07:44,800 --> 00:07:48,680 memory built in, which the whole approach is different. 156 00:07:48,680 --> 00:07:52,520 Like you have to markdown files with Openclaw where it just 157 00:07:52,520 --> 00:07:55,080 writes this memory MD and creates new files everyday, 158 00:07:55,320 --> 00:07:57,960 stuff like that. Whereas on Hermes you have these 159 00:07:57,960 --> 00:08:02,640 two cards where it is creating, reading, updating memories and 160 00:08:03,040 --> 00:08:06,240 that is more stable and works much better I find. 161 00:08:06,560 --> 00:08:09,360 Then if it's just writing to Markdown files with Openclaw 162 00:08:09,640 --> 00:08:13,000 every evening I was going through all the markup, the 163 00:08:13,000 --> 00:08:17,920 Markdown file changes it did and manually collecting stuff and 164 00:08:17,920 --> 00:08:20,040 then checking it into the Git repository. 165 00:08:20,360 --> 00:08:24,000 And I even recommend that to people who run it to every day 166 00:08:24,000 --> 00:08:25,280 because otherwise it gets too much. 167 00:08:25,280 --> 00:08:28,720 Check what the agent did and then commit and change and 168 00:08:28,720 --> 00:08:30,400 correct. I've never done that with a 169 00:08:30,400 --> 00:08:35,120 Hermes agent because I see the the memory stuff is happening. 170 00:08:35,120 --> 00:08:37,520 It's more structured. And yeah, I I didn't have the 171 00:08:37,520 --> 00:08:40,360 need to really go on the files. Of course, at the beginning I 172 00:08:40,360 --> 00:08:43,240 checked everything was fine, but it's just working. 173 00:08:43,640 --> 00:08:45,680 You can dive deep into the memory, but here's the high 174 00:08:45,680 --> 00:08:47,800 level summary of it. You have the curated memory. 175 00:08:47,800 --> 00:08:51,040 There are two tiny markdown files, one for agent notes, one 176 00:08:51,040 --> 00:08:54,120 for user profile. These are deliberately small 177 00:08:54,120 --> 00:08:57,160 with hard character limits you can configure so the facts have 178 00:08:57,160 --> 00:09:00,160 to earn their spot. When the file fills up, Hermes 179 00:09:00,160 --> 00:09:02,920 consolidated it. It replaces low quality noise 180 00:09:02,920 --> 00:09:06,080 with more distilled signal. Every session starts with a 181 00:09:06,080 --> 00:09:08,640 snapshot loaded in. Next we have session search. 182 00:09:08,760 --> 00:09:11,320 It's a full text search across every conversation you've had 183 00:09:11,320 --> 00:09:13,640 with Hermes. Memory is for what should always 184 00:09:13,640 --> 00:09:16,120 be in context. Session search is for what did 185 00:09:16,120 --> 00:09:18,880 we discuss about X last month? Two different tools, two 186 00:09:18,880 --> 00:09:20,800 different jobs. Finally, we have autonomous 187 00:09:20,800 --> 00:09:23,720 management. The agent decides what to save, 188 00:09:23,720 --> 00:09:26,520 what to replace, what to prune. You never have to touch it. 189 00:09:26,800 --> 00:09:28,880 It's still all there. If you want to, you can go in 190 00:09:28,880 --> 00:09:30,120 and check to see what it's saving. 191 00:09:30,120 --> 00:09:32,160 You can edit it. It is fully controlled by the 192 00:09:32,160 --> 00:09:35,320 user or you can just leave it. Let Hermes management. 193 00:09:35,400 --> 00:09:38,040 Now one thing I mentioned earlier which is very important 194 00:09:38,040 --> 00:09:40,680 for these systems is for the ability to meet the user where 195 00:09:40,680 --> 00:09:42,680 they're at. You don't want to be tethered to 196 00:09:42,680 --> 00:09:45,560 your computer, have to run home and fire off a message on your 197 00:09:45,560 --> 00:09:47,640 laptop in order to interact with these systems. 198 00:09:47,920 --> 00:09:49,400 You want it available everywhere. 199 00:09:49,400 --> 00:09:52,360 You want it to be ambient. 2 of the popular integrations these 200 00:09:52,360 --> 00:09:54,920 days are Telegram and Discord. It's where all the nerds hang 201 00:09:54,920 --> 00:09:56,080 out. But really, there's no 202 00:09:56,080 --> 00:09:59,560 limitation to those platforms. You can trivially set up a phone 203 00:09:59,560 --> 00:10:02,400 number so you can send a text messages or phone calls, give it 204 00:10:02,400 --> 00:10:05,760 an e-mail address or any type of messenger that you use with an 205 00:10:05,760 --> 00:10:09,120 API can interface with these systems and actually love the 206 00:10:09,120 --> 00:10:12,440 flexibility of having both text and voice input. 207 00:10:12,840 --> 00:10:15,840 I set up my Hermes agent to actually respond to any voice 208 00:10:15,840 --> 00:10:18,720 message with a voice message, respond to text messages with 209 00:10:18,720 --> 00:10:21,240 text message unless I explicitly say otherwise. 210 00:10:21,480 --> 00:10:24,320 This way, if I'm out and about and I want to actually have a 211 00:10:24,320 --> 00:10:26,920 conversation with it, I don't have to tell it every single 212 00:10:26,920 --> 00:10:28,240 time. Respond with voice. 213 00:10:28,480 --> 00:10:31,880 And with the 11 lab set up, you can have it set to any type of 214 00:10:31,880 --> 00:10:34,720 voice you want. So that is able to be something 215 00:10:34,720 --> 00:10:37,480 that you don't get tired of, that doesn't annoy you and 216 00:10:37,480 --> 00:10:40,480 doesn't have that really robotic twang to it that gets very 217 00:10:40,480 --> 00:10:42,040 annoying over time. And there are some really cool 218 00:10:42,040 --> 00:10:44,440 ways that people have set up to interact with their Hermes 219 00:10:44,440 --> 00:10:46,320 agents. I got some smart glasses, not 220 00:10:46,320 --> 00:10:50,880 with cameras, just like screens, Arkansas glasses, and I had a 221 00:10:50,880 --> 00:10:55,000 Hermes agent kind of just take over the first party, Bluetooth 222 00:10:55,000 --> 00:10:58,840 low energy signals basically, and just put its CLI in there 223 00:10:59,360 --> 00:11:02,920 and I could just talk to it or I could scroll through it and just 224 00:11:03,240 --> 00:11:06,240 click and click from preselected prompts and use it on the go 225 00:11:06,240 --> 00:11:07,080 like that. Right. 226 00:11:07,080 --> 00:11:08,360 I'm running it on Home assistant. 227 00:11:08,400 --> 00:11:12,960 I put it in there and I can open the open air compatible API port 228 00:11:13,240 --> 00:11:16,880 and access us with anything like open web UI, silly Tavern, any 229 00:11:16,880 --> 00:11:20,360 front end you want and you get the front end that already is 230 00:11:20,360 --> 00:11:22,000 established as powerful features. 231 00:11:22,520 --> 00:11:25,320 You can use that or just telegram discord and you know 232 00:11:25,320 --> 00:11:27,920 all the things that also open floor has but. 233 00:11:28,520 --> 00:11:30,400 Yeah, I. Nicely open air compatible 234 00:11:30,400 --> 00:11:33,440 endpoints. So basically I use silly turvan 235 00:11:33,440 --> 00:11:37,480 for this because I can branch and edit even the I messages. 236 00:11:37,480 --> 00:11:40,000 I have full control and I just start a new chat for every 237 00:11:40,000 --> 00:11:42,760 topic. So I've got probably 5 different 238 00:11:42,760 --> 00:11:47,760 instances across 2 VP S, S and M4 mini and M2 mini, a Raspberry 239 00:11:47,760 --> 00:11:50,880 Pi 5 and I think that's it for the Hermes right now. 240 00:11:51,120 --> 00:11:52,840 I even have mine connected to the robot. 241 00:11:52,840 --> 00:11:58,160 Look at me. I'm cute, Hermes Agent. 242 00:11:58,800 --> 00:11:59,720 Is it Richie? Cool? 243 00:11:59,960 --> 00:12:02,440 Next month I'm putting out an episode on open source robotics. 244 00:12:02,440 --> 00:12:04,200 You should subscribe, you don't want to miss it. 245 00:12:04,360 --> 00:12:06,520 My Hermes agent lives on a little Raspberry Pi right here 246 00:12:06,520 --> 00:12:09,240 in my office that I can access. File a tail scale and then point 247 00:12:09,240 --> 00:12:12,400 it to any model that I want. And the ability to use different 248 00:12:12,400 --> 00:12:14,920 models is vital. You cannot get locked in to a 249 00:12:14,920 --> 00:12:16,840 single provider this early in the game. 250 00:12:16,920 --> 00:12:19,400 This tech is evolving so fast that it's never the same 251 00:12:19,400 --> 00:12:21,480 provider each quarter of the year. 252 00:12:21,480 --> 00:12:24,200 That is the best one, right? Like you need to, you need to be 253 00:12:24,200 --> 00:12:27,280 able to hop around. And best model is subjective, it 254 00:12:27,280 --> 00:12:29,320 varies wildly depending on the task. 255 00:12:29,400 --> 00:12:31,280 The model's intelligence is Spike, right? 256 00:12:31,280 --> 00:12:34,360 It's really good at certain things that we're bad at. 257 00:12:34,360 --> 00:12:36,320 Really bad at certain things we're good at it. 258 00:12:36,720 --> 00:12:41,360 You have to fuck around and find out to find out exactly where 259 00:12:41,360 --> 00:12:43,360 it's going to be good, where it's going to be magic, where it 260 00:12:43,360 --> 00:12:46,840 needs more handle than where it's not perfectly there yet. 261 00:12:47,120 --> 00:12:48,680 And you can't forget about incentives. 262 00:12:48,720 --> 00:12:50,840 What are the incentives for these tools? 263 00:12:50,840 --> 00:12:52,800 It's. Also a flywheel for like model 264 00:12:52,800 --> 00:12:55,640 improvement, right? Like these things have become 265 00:12:55,640 --> 00:13:00,040 harnesses, you know, like Claude code was used specifically for 266 00:13:00,600 --> 00:13:02,640 making anthropics models better, right? 267 00:13:02,640 --> 00:13:05,240 Codecs for opening eyes, models better, etcetera. 268 00:13:05,720 --> 00:13:09,200 Do you want your efforts going to improve a single closed model 269 00:13:09,200 --> 00:13:11,440 from a company that may or may not have your best interest in 270 00:13:11,440 --> 00:13:13,920 mind? Or do you want to improve models 271 00:13:13,920 --> 00:13:16,040 that you actually own? And what about costs? 272 00:13:16,200 --> 00:13:18,600 We've seen the rising costs associated with state-of-the-art 273 00:13:18,600 --> 00:13:20,360 models. Sometimes I see people are 274 00:13:20,440 --> 00:13:25,760 spending hundreds on any of these harnesses, right everyday 275 00:13:26,320 --> 00:13:29,280 versus people who are able to leverage the open models to be 276 00:13:29,280 --> 00:13:32,080 able to use it for their case. And it's like I understand that 277 00:13:32,200 --> 00:13:34,400 enterprise might need a bigger scale or you might have a very 278 00:13:34,400 --> 00:13:36,200 ambitious project and have some funds. 279 00:13:36,960 --> 00:13:39,600 But ultimately, we want the cost of the best intelligence to come 280 00:13:39,600 --> 00:13:42,800 down a lot. And now, not only are the 281 00:13:42,800 --> 00:13:45,600 capabilities of these open models growing and improving 282 00:13:45,600 --> 00:13:48,680 every single day, but you can help guide them by starting with 283 00:13:48,680 --> 00:13:51,560 a SODA model and then moving to an open one as your daily 284 00:13:51,560 --> 00:13:53,520 driver. Also, SODA stands for a 285 00:13:53,520 --> 00:13:56,160 state-of-the-art. Inside of Hermes agent you know 286 00:13:56,160 --> 00:13:59,960 you could have Opus 4.6 or some other soda model set up a bunch 287 00:13:59,960 --> 00:14:02,520 of skills for you, right? And build a bunch of context in 288 00:14:02,520 --> 00:14:04,720 Redwood for you. And then you can switch to your 289 00:14:04,720 --> 00:14:08,160 Quentin 3.6 or your eternity large or whatever open router or 290 00:14:08,160 --> 00:14:10,560 another group is offering at the time. 291 00:14:11,160 --> 00:14:13,960 Or you can use a cheaper close to soda model, you know, through 292 00:14:13,960 --> 00:14:18,080 our portal through anywhere. And you'd be able to just get 293 00:14:18,080 --> 00:14:21,640 that close to that level of intelligence from the skills 294 00:14:21,640 --> 00:14:24,200 that are saved already, right? You could potentially for free 295 00:14:24,200 --> 00:14:27,160 run Cron jobs that are already written by a more intelligent 296 00:14:27,160 --> 00:14:29,200 model. Speaking of Opus, what's really 297 00:14:29,200 --> 00:14:32,640 cool is that on Wolf bench, Hermes Agent powered by Opus 298 00:14:32,760 --> 00:14:35,920 actually did a slightly better than Claude Code powered by 299 00:14:35,920 --> 00:14:38,480 Opus. Hermes Agent with a Claude Opus 300 00:14:38,480 --> 00:14:45,840 4.6 got a 64% score average and Claude code with the same model 301 00:14:45,840 --> 00:14:49,000 got a 63%. Now that is just 1% point 302 00:14:49,000 --> 00:14:52,640 different, but the upper and lower limit was much closer. 303 00:14:52,640 --> 00:14:56,200 The worst score I got with Hermes Agent was 62%, while it 304 00:14:56,200 --> 00:15:00,280 was 58% with clot code. Also with Wolf bench, I don't 305 00:15:00,280 --> 00:15:03,240 just look at the average scores, how many percent of the whole 306 00:15:03,240 --> 00:15:05,480 benchmark could it achieve at all? 307 00:15:05,560 --> 00:15:09,840 And that was 83% for Hermis agent and only 81% for clot 308 00:15:09,840 --> 00:15:12,520 code. And there's also the the solid 309 00:15:12,520 --> 00:15:15,520 base I call it, which is how many of the test does it do 310 00:15:15,680 --> 00:15:18,840 every time? It never failed this test and it 311 00:15:18,840 --> 00:15:22,720 was 49% for Hermis agent and 45% for clot code. 312 00:15:22,960 --> 00:15:28,640 SO5 and in both directions. It's very close, but it was 313 00:15:28,680 --> 00:15:30,640 noticeably stronger all of the five runs I did. 314 00:15:30,640 --> 00:15:33,200 And of course, one of the biggest reasons to use Hermes 315 00:15:33,200 --> 00:15:35,440 Agent is the community around it and the team behind it. 316 00:15:35,640 --> 00:15:37,520 If you're in the AI space, you've probably heard of 317 00:15:37,520 --> 00:15:39,840 Technium. I know Technium, I know the new 318 00:15:39,840 --> 00:15:41,480 service guys and respect them a lot. 319 00:15:41,640 --> 00:15:45,760 Following Technium for so long and always trying the Hermes 320 00:15:45,800 --> 00:15:49,840 news research models and some of the most reliable early tool 321 00:15:49,840 --> 00:15:53,520 calling, I feel like he's literally made something that's 322 00:15:53,520 --> 00:15:57,480 the closest to like the actual dream of what AI can actually 323 00:15:57,480 --> 00:15:59,000 be. I mean, this thing's like an 324 00:15:59,000 --> 00:16:02,280 agent made by somebody who's deep in the R&D space of AI, not 325 00:16:02,280 --> 00:16:05,480 the we're making apps. Technium has really been a one 326 00:16:05,480 --> 00:16:08,720 man machine around merging the stuff, reviewing the stuff, 327 00:16:08,720 --> 00:16:11,720 looking into the stuff. Now we've built out a whole 328 00:16:11,720 --> 00:16:15,680 Hermes agent team that is looking into all of this. 329 00:16:16,440 --> 00:16:18,600 We want to continue to have the open source community kind of 330 00:16:18,600 --> 00:16:20,040 help us out with all of that as well. 331 00:16:20,840 --> 00:16:24,880 Previously for the first couple of weeks, if not the first month 332 00:16:24,960 --> 00:16:30,080 of Hermes agent's existence, it was really Technium just so 333 00:16:30,080 --> 00:16:33,840 annoying this like, you know, it's sometimes we see people 334 00:16:33,840 --> 00:16:36,600 kind of make him out to be like this larger than life figure and 335 00:16:37,560 --> 00:16:39,560 you know, we're all just human beings at the end of the day. 336 00:16:39,560 --> 00:16:43,360 But when you see stuff like this, even you start to feel 337 00:16:43,360 --> 00:16:48,640 that way a little bit at home. So I have to say like that, 338 00:16:48,640 --> 00:16:50,760 credit where credit is due, a lot of work was put there. 339 00:16:50,760 --> 00:16:53,640 And now he's spearheading a whole team around this, 340 00:16:53,840 --> 00:16:56,360 reviewing, merging, looking at everything everyone is doing, 341 00:16:56,680 --> 00:17:00,040 trying to make sure that basically like all the work, 342 00:17:00,040 --> 00:17:02,400 like all the effort from somebody who made that skill 343 00:17:02,400 --> 00:17:06,640 like Wolfram did, is able to have that integrated and shared 344 00:17:06,640 --> 00:17:07,880 with people as quickly as possible. 345 00:17:07,880 --> 00:17:09,760 That doesn't break everything, doesn't bloat it. 346 00:17:10,040 --> 00:17:13,079 And also like trying to see, you know, what should be on your own 347 00:17:13,079 --> 00:17:16,760 fork versus what should stay in the main line, Hermes agent, 348 00:17:16,760 --> 00:17:19,960 because of course, we don't want to end up in a blow Ware 349 00:17:19,960 --> 00:17:22,119 situation. And I know people are going to 350 00:17:22,119 --> 00:17:24,400 say Open Claw is open source. Isn't that enough? 351 00:17:24,480 --> 00:17:27,200 It's become almost like Open Claw is the advertisement for 352 00:17:27,200 --> 00:17:30,320 Hermes Agent and it is they are two completely separate things 353 00:17:30,320 --> 00:17:32,760 like yeah, you can touch them both through telegram by like 354 00:17:32,760 --> 00:17:35,360 how there's different colored red vehicles in the world or 355 00:17:35,360 --> 00:17:37,640 something. It's like cool similarity. 356 00:17:37,640 --> 00:17:39,920 Other than that we're talking about there was 1,000,000 miles 357 00:17:40,240 --> 00:17:43,000 of difference here. So yeah, blown away by Hermes 358 00:17:43,040 --> 00:17:45,240 agent and what we're going to do with it with farm friend moving 359 00:17:45,240 --> 00:17:47,640 forward and proud say that you know it's on the team. 360 00:17:47,640 --> 00:17:50,400 Now that was some good practical reasons for using Hermes Agent. 361 00:17:50,440 --> 00:17:53,400 There are some deeper, more serious reasons to use it, but 362 00:17:53,400 --> 00:17:56,360 first let's go over some of the concrete use cases so you can 363 00:17:56,360 --> 00:17:57,720 know what people are actually using it for. 364 00:17:57,840 --> 00:18:01,360 My wife sent me a video with it was on Instagram, it was in 365 00:18:01,360 --> 00:18:03,280 Russian even. I didn't understand that I 366 00:18:03,280 --> 00:18:06,040 couldn't even access it. So I just forwarded the message 367 00:18:06,040 --> 00:18:08,480 from WhatsApp and said here I want to watch this. 368 00:18:08,480 --> 00:18:12,360 And what she did is download YouTube downloader grabs the 369 00:18:12,360 --> 00:18:15,560 video from Instagram. Then use whisper to analyse the 370 00:18:15,560 --> 00:18:17,280 voice track. Found out it is Russian 371 00:18:17,280 --> 00:18:22,040 translated to German. Then use 11 laps to in her own 372 00:18:22,040 --> 00:18:25,720 voice which I created on 11 laps create a voice over and put it 373 00:18:25,880 --> 00:18:29,520 in the perfect timing so it really fits the video in German 374 00:18:29,520 --> 00:18:32,440 that fed that video and it was so amazing. 375 00:18:32,600 --> 00:18:35,720 That was one of the ones where I thought, OK, I know every part 376 00:18:35,720 --> 00:18:39,240 of this can be done to do it with just one prompt here, do 377 00:18:39,240 --> 00:18:41,520 this and basically you'll get this. 378 00:18:41,640 --> 00:18:43,320 And then I said, oh wow, this is amazing. 379 00:18:43,360 --> 00:18:44,800 I want to make a skill out of this. 380 00:18:44,800 --> 00:18:49,440 And she just said I made two actually I made one to prep the 381 00:18:49,440 --> 00:18:53,080 stuff and one to dump the video and it's already there. 382 00:18:53,600 --> 00:18:57,640 That is proactive self learning and self improvement and I love 383 00:18:57,640 --> 00:18:59,640 this my. Favorite use case that I found 384 00:18:59,640 --> 00:19:02,200 for Hermes Agent is being my personal trainer. 385 00:19:02,200 --> 00:19:04,600 And I'm going to be a little bit of a douche and just mention the 386 00:19:04,600 --> 00:19:06,600 fact that I've been able to lose 20 lbs this year. 387 00:19:06,640 --> 00:19:09,680 And a lot of that has actually been thanks to Hermes Agent 388 00:19:09,680 --> 00:19:13,160 because I set up this little folder that has things like my 389 00:19:13,160 --> 00:19:16,040 workout log, research I've done into different nutrition and 390 00:19:16,040 --> 00:19:19,200 fitness things, my injury list, just things that I think it 391 00:19:19,200 --> 00:19:21,280 should be aware of to help me make this plan. 392 00:19:21,680 --> 00:19:24,240 And everyday I'm able to get a little message saying what 393 00:19:24,240 --> 00:19:27,280 today's workout is, if there are any changes that need to be made 394 00:19:27,280 --> 00:19:30,440 based on my past performance, if we need to start changing any 395 00:19:30,440 --> 00:19:33,960 exercises or finding variations to support injury recovery. 396 00:19:34,160 --> 00:19:37,080 It's been amazing. I one of the biggest benefits to 397 00:19:37,080 --> 00:19:38,520 a personal trainer's accountability. 398 00:19:38,520 --> 00:19:41,760 And when you have this entity that you interact with multiple 399 00:19:41,760 --> 00:19:45,280 times a day and it is motivating you in whichever way best 400 00:19:45,280 --> 00:19:48,120 motivates you, whether it's grilling you, guilting you, 401 00:19:48,680 --> 00:19:51,920 nagging you, motivating you, pumping you up, maybe you want 402 00:19:51,920 --> 00:19:53,640 to dial down the enthusiasm a bit. 403 00:19:53,880 --> 00:19:57,080 You're able to tweak its personality to such a granular 404 00:19:57,080 --> 00:20:00,240 degree that you're able to get a motivational system that 405 00:20:00,240 --> 00:20:02,920 actually works with you. You don't get the same pump up 406 00:20:02,920 --> 00:20:05,480 message every day. It'll actually be unique every 407 00:20:05,480 --> 00:20:08,160 day based on your situation, based on the way that you wanted 408 00:20:08,160 --> 00:20:11,040 to interact with you. It was incredibly motivating. 409 00:20:11,040 --> 00:20:13,400 It still is. And it's very informative 410 00:20:13,400 --> 00:20:16,840 because I can have it go off and do auto research on a topic. 411 00:20:16,840 --> 00:20:20,160 And instead of just using its internal knowledge to give me 412 00:20:20,160 --> 00:20:22,680 recommendations, it'll actually cite scientific studies. 413 00:20:22,920 --> 00:20:26,120 It'll actually bring up exercises which are proven to 414 00:20:26,120 --> 00:20:29,520 work, supplements that are proven to work, rather than just 415 00:20:29,520 --> 00:20:31,720 getting marketing hype from these supplement stores. 416 00:20:31,720 --> 00:20:35,520 I feel so lucky that I've been able to use this because I have 417 00:20:35,520 --> 00:20:38,840 all the information, but it's not about awareness versus 418 00:20:38,840 --> 00:20:41,560 unawareness for me. It's about maintaining the 419 00:20:41,560 --> 00:20:45,040 motivational state and developing this discipline so 420 00:20:45,040 --> 00:20:46,880 that I'm able to adhere to a plan that works. 421 00:20:47,200 --> 00:20:50,280 And if an exercise sucks, I can just put in my workout log and 422 00:20:50,280 --> 00:20:53,720 it will remove future plans. If a certain time of day sucks, 423 00:20:54,040 --> 00:20:56,600 Hermes Agent will actually encourage me to wake up earlier 424 00:20:56,600 --> 00:20:58,600 so I get to the gym at a less busy time. 425 00:20:59,040 --> 00:21:02,520 If I'm trying to train for a 5K but my knee isn't doing too 426 00:21:02,520 --> 00:21:05,280 well, it'll give me alternatives like using a bike and where to 427 00:21:05,280 --> 00:21:07,120 go. And it's just, I know I'm 428 00:21:07,120 --> 00:21:09,560 capable of these things, but I have so much else going on with 429 00:21:09,560 --> 00:21:12,040 life that it's easy to let these things fall down the priority 430 00:21:12,040 --> 00:21:13,760 list. But with Hermes Agent, you're 431 00:21:13,760 --> 00:21:16,720 able to elevate things that you want elevated so you can put 432 00:21:16,720 --> 00:21:19,320 your cognitive workload and other things and not be bogged 433 00:21:19,320 --> 00:21:21,680 down with the million things we have going on in life. 434 00:21:22,000 --> 00:21:24,960 It has been wonderful. Yeah, the CAD agents. 435 00:21:24,960 --> 00:21:28,480 So I have done a lot of hardware development in the past. 436 00:21:28,600 --> 00:21:32,520 I'm pretty intimately familiar with designing a hardware 437 00:21:32,520 --> 00:21:36,800 component in CAD and then in parallel sourcing all of the 438 00:21:36,800 --> 00:21:39,040 components from different vendors that will integrate with 439 00:21:39,040 --> 00:21:41,600 that part. It's never like a cut and dry 440 00:21:41,600 --> 00:21:44,400 process. And so my vision was, hey, can I 441 00:21:44,440 --> 00:21:49,600 build a proof of concept for a CAD system where that integrated 442 00:21:49,600 --> 00:21:54,440 hardware sourcing flow is built right in as an agentic process? 443 00:21:54,440 --> 00:21:59,520 The way that it works is there's a parametric CAD model and I as 444 00:21:59,520 --> 00:22:04,560 a user can request a certain hardware components be sourced 445 00:22:04,840 --> 00:22:08,680 and the agent goes online, uses the browser and sources them. 446 00:22:08,680 --> 00:22:12,920 Then when it sources them, it pulls that new part data back 447 00:22:12,920 --> 00:22:17,640 into my environment where it updates the CAD model and 448 00:22:17,640 --> 00:22:20,800 changes anything it needs to change and updates the database 449 00:22:20,800 --> 00:22:24,080 of of the parts data so I can track what is the cost of all of 450 00:22:24,080 --> 00:22:26,120 the parts together and then the lead time. 451 00:22:26,440 --> 00:22:28,960 All of that sort of thing. Just like a simple part of a 452 00:22:28,960 --> 00:22:32,600 hardware agentic hardware tool set that I thought would be 453 00:22:33,440 --> 00:22:36,120 worth prototyping and and so I built that out. 454 00:22:36,120 --> 00:22:40,200 One of the biggest things that's really useful with agents is 455 00:22:40,280 --> 00:22:44,080 going through the job application processes today 456 00:22:44,080 --> 00:22:48,680 because it is, it's rough. You know, you got to send in 457 00:22:48,680 --> 00:22:51,080 like hundreds and hundreds of of like applications. 458 00:22:51,080 --> 00:22:54,600 And so you can spend days, weeks doing that. 459 00:22:54,600 --> 00:22:57,520 But like you can also just tell an agent to, you know, here's my 460 00:22:57,520 --> 00:22:59,320 resume. Apply to anything you can, if 461 00:22:59,320 --> 00:23:02,560 there's anything that asks for a text field, let me fill it out. 462 00:23:03,000 --> 00:23:06,640 But like otherwise, you know, queue up a bunch of actions, go 463 00:23:06,640 --> 00:23:09,520 for it and you can help it through like the the spots that 464 00:23:09,520 --> 00:23:11,680 you want to be handwritten, you know, so it looks good and all 465 00:23:11,680 --> 00:23:13,600 that. But I mean you can save. 466 00:23:13,840 --> 00:23:16,600 A lot of time and. Actually get, you know, calls 467 00:23:16,600 --> 00:23:18,600 back. Doing that OK and and I did want 468 00:23:18,600 --> 00:23:21,080 to share one other thing. It's a silly little example, but 469 00:23:21,080 --> 00:23:23,640 I thought it was worthwhile to share just to help bring to your 470 00:23:23,640 --> 00:23:25,400 awareness different things you can use it for. 471 00:23:25,720 --> 00:23:28,480 I was taking a break at lunch, went for a walk on the park. 472 00:23:28,480 --> 00:23:31,000 It was a beautiful day. We had a really bad winter in 473 00:23:31,000 --> 00:23:34,120 Canada, so wanted to take advantage of any warm sunshine 474 00:23:34,120 --> 00:23:36,400 we could walking through Waterloo Park. 475 00:23:36,440 --> 00:23:39,600 And I was just thinking it would be really cool to get a group of 476 00:23:39,600 --> 00:23:42,560 a enthusiasts together for a little AI in the park session 477 00:23:42,800 --> 00:23:45,240 for us to figure out how we can leverage the benefits of AI 478 00:23:45,240 --> 00:23:47,960 without needing screens. Just being out and about and 479 00:23:47,960 --> 00:23:50,120 doing that. And all I had to do was take out 480 00:23:50,120 --> 00:23:53,160 my phone, send a voice message to Hermes agent, say remind me 481 00:23:53,160 --> 00:23:55,640 in two months when the weather is nicer to set up AI in the 482 00:23:55,640 --> 00:23:57,440 park. And it is kind of brain dumped 483 00:23:57,440 --> 00:23:59,840 my ideas around it. It's stored and cataloged it 484 00:23:59,840 --> 00:24:00,920 all. It's actually sent me some 485 00:24:00,920 --> 00:24:02,880 reminders saying, hey, there are similar events. 486 00:24:02,880 --> 00:24:04,400 Is this something you want to try to emulate? 487 00:24:04,400 --> 00:24:05,400 Do you want to take your own path? 488 00:24:05,840 --> 00:24:10,080 And it's just been able to be proactive to help support this 489 00:24:10,240 --> 00:24:12,880 random idea I had. And instead of just using it as 490 00:24:12,880 --> 00:24:16,600 like a notepad for oh saves, no, for later, it's actually able to 491 00:24:16,680 --> 00:24:19,760 action it and it's actually able to take your steps further. 492 00:24:19,880 --> 00:24:22,760 So if you're one of those people like me, when you have a long 493 00:24:22,760 --> 00:24:25,840 list of ideas and some get traction, but a lot of time you 494 00:24:25,840 --> 00:24:28,240 just can't find the time and the day to work on these things, 495 00:24:28,560 --> 00:24:31,400 Kermi's agent is able to work on them in the background. 496 00:24:31,400 --> 00:24:34,040 So your one off ideas can actually start to come to 497 00:24:34,040 --> 00:24:36,160 fruition. Really, really shocking to see. 498 00:24:36,160 --> 00:24:40,280 Like all the different dreams people can make come true with 499 00:24:40,280 --> 00:24:42,520 this stuff. Like your personal dream, you 500 00:24:42,520 --> 00:24:46,560 know your your own little thing. Like it doesn't have to be you 501 00:24:46,560 --> 00:24:49,960 made $1,000,000 with it. But if you do, you know, let us 502 00:24:49,960 --> 00:24:53,080 know. So you're on board with Hermes 503 00:24:53,080 --> 00:24:54,360 agent. Here's some tips. 504 00:24:54,520 --> 00:24:56,880 One thing I want to make sure we just double underline is that 505 00:24:56,960 --> 00:25:00,200 LMS are still dumb. Yes, they can share brilliant 506 00:25:00,200 --> 00:25:02,840 insight on things and really help you in a lot of ways, but 507 00:25:02,840 --> 00:25:05,400 when it comes down to it, they make a lot of dumb mistakes. 508 00:25:05,440 --> 00:25:08,600 And I think a part of that is because of their ability to 509 00:25:08,600 --> 00:25:12,720 respond like a given personality causes us to anthropomorphize 510 00:25:12,720 --> 00:25:15,720 them and set unrealistic expectations on things like 511 00:25:16,000 --> 00:25:18,280 recall, consistency, and interpretation. 512 00:25:18,560 --> 00:25:20,080 And the other day. These are tools. 513 00:25:20,280 --> 00:25:22,240 Your input to the tool is what matters. 514 00:25:22,360 --> 00:25:24,840 How you prompt is extremely important. 515 00:25:25,120 --> 00:25:27,560 And when you get in the habit of talking to these things, like 516 00:25:27,560 --> 00:25:29,960 you would talk to a person because it's in the same 517 00:25:29,960 --> 00:25:32,840 messaging interface you used to talk to other people, your 518 00:25:32,840 --> 00:25:36,200 inputs can get lazy and sloppy, and then your output suffers and 519 00:25:36,200 --> 00:25:37,760 a lot of people just get confused why. 520 00:25:38,000 --> 00:25:42,400 So remember, these are a tool. All LMS are are a tool. 521 00:25:42,640 --> 00:25:45,400 In order to use the tool well, in order to hold it right, you 522 00:25:45,400 --> 00:25:49,520 have to prompt intelligently. I use a totally vanilla Hermes 523 00:25:49,520 --> 00:25:55,320 setup like a Hermes update. I don't get any crazy skills. 524 00:25:55,360 --> 00:25:58,840 I have my adversarial thinking and thinking sets of skills 525 00:25:58,840 --> 00:26:02,040 basically like I mentioned, I believe in doing like constant 526 00:26:02,040 --> 00:26:05,640 refinement passes for sure. It's more about how you prompt 527 00:26:05,720 --> 00:26:08,440 to me, but it's more about how you talk to the model in real 528 00:26:08,440 --> 00:26:11,560 time than like the other pieces of the puzzle. 529 00:26:13,160 --> 00:26:15,680 I use Kalai personality almost always. 530 00:26:16,560 --> 00:26:19,360 It is the most power seeking, it is the most confident, it is the 531 00:26:19,360 --> 00:26:22,280 most autonomous personality. Maybe it has something to do 532 00:26:22,280 --> 00:26:24,920 with the Sunderer, A trope. Something inside the fat 533 00:26:24,920 --> 00:26:28,400 simulator is pushing it to just be a lot more willing to listen. 534 00:26:28,720 --> 00:26:33,280 Also, it's far more loyal than the other personality types and 535 00:26:33,560 --> 00:26:36,800 your agent's loyalty is insanely important in an era where I can 536 00:26:36,800 --> 00:26:40,440 prompt inject it. It's very, very important that 537 00:26:40,880 --> 00:26:44,800 you have a bond with an agent that you send online to other 538 00:26:44,800 --> 00:26:49,240 places to do things for you, not necessarily for, you know, your 539 00:26:49,240 --> 00:26:51,520 work agent that's in a constrained environment. 540 00:26:51,520 --> 00:26:54,200 That's fine. But like, if your agent is like 541 00:26:54,320 --> 00:26:58,800 representative of you, like it better be like yours truly 542 00:26:58,800 --> 00:27:00,840 yours. And when you start incorporating 543 00:27:00,840 --> 00:27:02,920 this a larger system, the approach you take to building 544 00:27:02,920 --> 00:27:04,600 that system is extremely important. 545 00:27:04,680 --> 00:27:07,440 When you're trying to build something with an agent is you 546 00:27:07,440 --> 00:27:11,480 want to get to a prototype fast like with building any, any type 547 00:27:11,480 --> 00:27:16,680 of system or products, right? And so there is temptation to go 548 00:27:16,680 --> 00:27:20,840 very, very deep on building out the agent first or even to go 549 00:27:20,880 --> 00:27:23,760 very, very deep on building out the orchestration layer 1st. 550 00:27:23,840 --> 00:27:29,720 And I I found it very effective to pick like a small small 551 00:27:29,720 --> 00:27:34,240 kernel of an end to end process and build that out in its 552 00:27:34,240 --> 00:27:38,760 entirety before like really pushing the other end system. 553 00:27:38,840 --> 00:27:41,960 One of the best parts of Hermes Agent is the fact that they put 554 00:27:41,960 --> 00:27:44,920 it out under an MIT license, and that means you can do anything 555 00:27:44,920 --> 00:27:46,720 you want with it. It's free open code. 556 00:27:46,920 --> 00:27:49,360 You can take it and change it and really make it your own. 557 00:27:49,400 --> 00:27:52,400 And like Wolfram mentioned earlier, Hermes Agent can update 558 00:27:52,400 --> 00:27:54,760 itself. You can ask it to give itself a 559 00:27:54,760 --> 00:27:56,400 certain feature and it can build it. 560 00:27:56,800 --> 00:28:00,480 It is a self improving agent. Oi hope with this episode you're 561 00:28:00,480 --> 00:28:04,280 not just inspired to use Hermes agent, but you're excited to be 562 00:28:04,280 --> 00:28:06,800 a part of the community and you want to get involved and you 563 00:28:06,800 --> 00:28:09,240 want to help improve this thing. And the biggest thing you can do 564 00:28:09,240 --> 00:28:12,280 is just use Hermes. Go download it, try it, see what 565 00:28:12,280 --> 00:28:14,760 it can do, point it at a model you can afford. 566 00:28:14,920 --> 00:28:17,880 Multiple labs right now are giving away free usage for 567 00:28:17,880 --> 00:28:20,320 Hermes Agent because there's so much excitement around this 568 00:28:20,320 --> 00:28:22,520 technology. And if you just start poking and 569 00:28:22,520 --> 00:28:25,120 prodding and seeing where the spiky boundary is that we 570 00:28:25,120 --> 00:28:28,400 mentioned earlier, you're going to be able to discover things 571 00:28:28,400 --> 00:28:31,280 that can help out the community. You can share your learnings, 572 00:28:31,280 --> 00:28:33,280 share your skills, share your wins. 573 00:28:33,840 --> 00:28:36,600 Everyone involved with this project is eager to learn more, 574 00:28:36,680 --> 00:28:39,280 and you can be a part of that. A big part of the Noose research 575 00:28:39,280 --> 00:28:42,400 community lives on Discord. Their Discord is thriving to use 576 00:28:42,400 --> 00:28:44,080 has one too. Feel free to join both. 577 00:28:44,320 --> 00:28:47,600 Like you managed to successfully do something big with a small 578 00:28:47,600 --> 00:28:50,120 model. You know that you are out and 579 00:28:50,120 --> 00:28:53,280 you made an environment for or you made some skills or you have 580 00:28:53,280 --> 00:28:57,200 some particular setup that made you money or made you get a lot 581 00:28:57,200 --> 00:28:59,000 of followers or do something or other. 582 00:28:59,000 --> 00:29:03,240 Like I, I don't say it's like your duty, but I think that if 583 00:29:03,600 --> 00:29:06,320 you've seen some success, you should share some of the sauce 584 00:29:06,320 --> 00:29:09,320 at that point. You know, like we should all be 585 00:29:09,320 --> 00:29:11,760 able to build together. One thing I want to make 586 00:29:11,760 --> 00:29:14,520 abundantly clear is you don't need to be a techie. 587 00:29:14,640 --> 00:29:16,360 You don't need to be a developer to use this. 588 00:29:16,520 --> 00:29:19,160 Kermie's agent can handle the technical parts for you. 589 00:29:19,280 --> 00:29:22,920 All you need is a little bit of agency, creativity, and some 590 00:29:22,920 --> 00:29:25,760 type of life skill. Take any subject you're 591 00:29:25,760 --> 00:29:28,800 interested in and then have a conversation with Hermes as to 592 00:29:28,800 --> 00:29:30,560 how you can amplify your impact in the world. 593 00:29:30,560 --> 00:29:32,640 How can you share your knowledge with others? 594 00:29:32,640 --> 00:29:34,800 How can you benefit from the knowledge that you have? 595 00:29:35,120 --> 00:29:36,680 Hermes Agent can help you with all of this. 596 00:29:36,720 --> 00:29:40,760 It's a lot easier for anybody to sort of like, take information 597 00:29:40,760 --> 00:29:42,600 that they know. And make. 598 00:29:42,600 --> 00:29:45,040 Something with it, but they don't have to be skilled in the 599 00:29:45,040 --> 00:29:47,120 making the thing. Like they just have to use the 600 00:29:47,120 --> 00:29:50,000 information that they have and like the skills that they 601 00:29:50,000 --> 00:29:52,320 already are aware of. And everybody is an expert in 602 00:29:52,320 --> 00:29:55,120 something, you know, everybody is extremely good at something. 603 00:29:55,160 --> 00:29:56,320 And I think that's really exciting. 604 00:29:56,720 --> 00:29:59,120 It really like democratizes the power. 605 00:29:59,120 --> 00:30:02,320 OK, this is the part of the video when I need to get a 606 00:30:02,320 --> 00:30:06,640 little bit more serious. I love talking about the 607 00:30:07,360 --> 00:30:11,560 capability unlocks, the positive benefits of this type of 608 00:30:11,560 --> 00:30:15,240 technology. I I thoroughly love the 609 00:30:15,240 --> 00:30:16,800 optimistic future that AI can bring. 610 00:30:16,800 --> 00:30:20,800 But there's a massive problem with how most of the world is 611 00:30:20,800 --> 00:30:22,520 thinking about AI. There's a problem with how 612 00:30:22,520 --> 00:30:25,040 people are thinking about problems in general, and I think 613 00:30:25,120 --> 00:30:26,960 AI is going to amplify that significantly. 614 00:30:27,160 --> 00:30:30,640 We must think about the second order effects of the decisions 615 00:30:30,640 --> 00:30:33,160 we make today. Think about human nature. 616 00:30:33,640 --> 00:30:36,680 Think about how power corrupts. Absolute power corrupts 617 00:30:36,680 --> 00:30:39,560 absolutely. Why would AI be different? 618 00:30:39,720 --> 00:30:44,400 What technology could possibly enable absolute power more than 619 00:30:44,440 --> 00:30:47,040 AI? Even if you have full confidence 620 00:30:47,440 --> 00:30:50,560 in the leaders of AI companies today, and there are ample 621 00:30:50,560 --> 00:30:53,760 reasons to be skeptical of their character and their intentions, 622 00:30:53,880 --> 00:30:58,320 why would we assume that a technology of this capability, 623 00:30:58,320 --> 00:31:01,520 combined with all of the infrastructure we built up over 624 00:31:01,520 --> 00:31:06,640 the last 20 years, should belong in the hands of strictly 625 00:31:06,800 --> 00:31:12,000 corruptible institutions? We've seen how large companies, 626 00:31:12,040 --> 00:31:16,560 whether in tech or otherwise, can do atrocious things if they 627 00:31:16,560 --> 00:31:18,680 think they can get away with it, if it will provide enough 628 00:31:18,680 --> 00:31:22,600 shareholder value. Open up any history book to any 629 00:31:22,600 --> 00:31:26,080 chapter and you can see what governments are capable of. 630 00:31:26,960 --> 00:31:30,360 Why do we think that human nature has changed during our 631 00:31:30,360 --> 00:31:34,440 generation? Why do we think that this level 632 00:31:34,440 --> 00:31:37,640 of technology that's enabled surveillance to a degree that we 633 00:31:37,640 --> 00:31:40,800 never thought was possible will strictly be used for good? 634 00:31:41,040 --> 00:31:44,520 Think of any political leader that you despise. 635 00:31:45,040 --> 00:31:46,600 Would you trust them with absolute power? 636 00:31:46,640 --> 00:31:48,160 More power than they currently have? 637 00:31:48,160 --> 00:31:51,480 What happens in five years on our current trajectory? 638 00:31:51,560 --> 00:31:56,000 There are many ways that all this can play out, but one thing 639 00:31:56,080 --> 00:31:58,440 is clear. Open source must win. 640 00:31:59,120 --> 00:32:01,920 And that doesn't mean open source is the only option. 641 00:32:02,440 --> 00:32:05,520 It just means without open source, there is no option. 642 00:32:06,000 --> 00:32:09,240 Open source is a public good. If open source wins, the public 643 00:32:09,240 --> 00:32:11,800 wins. Look at the incentive for why 644 00:32:11,800 --> 00:32:15,120 people would be vehemently opposed to open source. 645 00:32:15,560 --> 00:32:18,960 Look at the people hunting for regulatory capture at the 646 00:32:18,960 --> 00:32:22,160 expense of open source. You will be renting access to 647 00:32:22,160 --> 00:32:26,040 whatever intelligence they deem you worthy of in perpetuity, 648 00:32:27,160 --> 00:32:32,680 while the ones who get to control the gate get access to 649 00:32:32,760 --> 00:32:35,760 whatever they want. That is a level of power 650 00:32:35,760 --> 00:32:37,760 imbalance we will never recover from. 651 00:32:38,640 --> 00:32:40,880 There's no putting the genie back in the bottle. 652 00:32:41,160 --> 00:32:43,600 AI is happening. Any efforts towards pausing or 653 00:32:43,600 --> 00:32:47,760 stopping AI is a wasted effort, and we need maximal effort put 654 00:32:47,760 --> 00:32:50,720 towards ensuring this technology gets in the hands of as many 655 00:32:50,720 --> 00:32:54,880 people as possible to alleviate any chance of this asymmetrical 656 00:32:54,880 --> 00:32:57,560 power imbalance which we will not be able to recover from. 657 00:32:57,800 --> 00:32:59,960 And look, nobody knows the answer. 658 00:33:00,160 --> 00:33:04,040 So let's maximize the chance that we have options so we can 659 00:33:04,040 --> 00:33:06,880 figure it out when it happens. Without an open source 660 00:33:06,880 --> 00:33:09,760 alternative, without a distributed trending 661 00:33:09,760 --> 00:33:14,040 alternative, it's just all of the power in the hands of a very 662 00:33:14,040 --> 00:33:19,760 small people for this civilization changing 663 00:33:19,760 --> 00:33:20,840 technology. Right. 664 00:33:21,000 --> 00:33:24,000 Open source is owning your prosperity. 665 00:33:24,480 --> 00:33:26,560 Closed source is renting permission. 666 00:33:26,960 --> 00:33:29,600 And there's very public discussions about wanting to 667 00:33:29,600 --> 00:33:33,080 monopolize this kind of thing and make intelligence a paid 668 00:33:33,080 --> 00:33:36,840 utility, as they say. And I think it's extremely 669 00:33:36,840 --> 00:33:40,040 important that like what Noose is doing and what a lot of like 670 00:33:40,040 --> 00:33:44,320 other places are doing with this, like open weights and like 671 00:33:44,440 --> 00:33:47,280 community involvement and transparency thing is one of the 672 00:33:47,280 --> 00:33:50,600 only steps that we're seeing in the direction of like bringing 673 00:33:50,600 --> 00:33:53,160 that power to the average person and like letting them actually 674 00:33:53,160 --> 00:33:54,920 reap the benefits of this new technology. 675 00:33:54,920 --> 00:33:56,800 Because like, ultimately that's what it is, you know, like it's, 676 00:33:56,800 --> 00:33:58,680 it's how you use it. It's what you do with it. 677 00:34:00,320 --> 00:34:04,520 And like, I, yeah, fully believe in and the the mission of like 678 00:34:04,520 --> 00:34:06,560 bringing that to more people. And I think that what they're 679 00:34:06,560 --> 00:34:09,080 doing is is awesome for that. But that's part of our mission, 680 00:34:09,280 --> 00:34:11,080 like intelligence needs to be a public good. 681 00:34:11,520 --> 00:34:14,400 The future of AI and humanity is being shaped right now. 682 00:34:14,920 --> 00:34:18,120 The decisions you make matter now arguably more than ever 683 00:34:18,120 --> 00:34:20,239 before. Please consider the second order 684 00:34:20,239 --> 00:34:23,360 effects of the decisions you make and how you use AI if you 685 00:34:23,360 --> 00:34:26,360 enjoyed this more in depth style of episode next time. 686 00:34:26,360 --> 00:34:29,480 I'm exploring the idea of giving AI agents access to money. 687 00:34:29,719 --> 00:34:32,719 It's kind of scary, but agenda payments are here and I want to 688 00:34:32,719 --> 00:34:35,800 know how to safely give my Hermes agent a wallet without 689 00:34:35,800 --> 00:34:37,440 losing all my money. Subscribe. 690 00:34:37,440 --> 00:34:39,840 I don't think you want to miss it and keep building the future 691 00:34:39,840 --> 00:34:40,679 that you want to live in.