Pełna transkrypcja odcinka Setting Up Your Agents Brain Demo | Cold Wallets & Compute Records

general code cleanup to improve overall stability there. 56 00:04:37.535 --> 00:04:39.375 Back to the Morpheus UI, 57 00:04:40.330 --> 00:04:41.370 there was 58 00:04:41.530 --> 00:04:42.810 some additions 59 00:04:42.970 --> 00:04:44.010 to the 60 00:04:44.890 --> 00:04:45.690 multi 61 00:04:46.170 --> 00:04:48.410 multi wallet use. So 62 00:04:49.530 --> 00:04:57.255 just additional work. Right? We're focused on how do people use the product in different ways. So, like, this multi wallet use 63 00:04:57.815 --> 00:04:58.535 is 64 00:04:58.695 --> 00:05:20.800 is important on the UI side. We're also concurrently working on cold wallet use. And, you know, me and David could probably get into that in a little bit more detail in a second. But one is, like, how do we make some improvements now to help the user? And then two is how do we strengthen the system and and give users that, like, gold standard solution that they'd like? So incremental steps towards the goal there. And then 65 00:05:21.565 --> 00:05:26.845 there was one other security update that we made that just hardened the... 66 00:05:27.965 --> 00:05:28.685 Kind of... 67 00:05:30.205 --> 00:05:35.405 Not the private key management. That's the wrong way to phrase it, but to put in some more safeguards around 68 00:05:36.500 --> 00:05:38.340 what types of secrets 69 00:05:38.340 --> 00:05:45.860 are are known that you would never want to be sent through the router, how do we block some of those from being sent, right, 70 00:05:46.180 --> 00:05:53.245 as a a belt and suspenders approach. Right? We all have great or most of us have great operational security. 71 00:05:53.405 --> 00:05:56.125 Some of us can improve, including myself, on it. 72 00:05:56.525 --> 00:05:57.324 But 73 00:05:57.965 --> 00:05:59.245 sometimes having 74 00:05:59.245 --> 00:06:00.445 an extra 75 00:06:00.604 --> 00:06:09.169 fallback somewhere in the system to say, hey. Like, are are you sure you wanted to, you know, send this API key? Are you sure you wanted to send x y z? 76 00:06:09.810 --> 00:06:13.729 We just put some more backups in there that that will help with that, 77 00:06:14.210 --> 00:06:18.689 in the case of a mistake. Right? It's not something you can say, hey. I can 78 00:06:19.335 --> 00:06:21.495 go do some flagrant 79 00:06:21.655 --> 00:06:25.015 violations of of my own, you know, personal information. 80 00:06:26.775 --> 00:06:34.535 You know, not not to to stop that, but, hey. Just in case something happened, we're we're able to provide a little bit of backstop in in these scenarios. 81 00:06:36.990 --> 00:06:38.830 That, I believe, is the 82 00:06:38.990 --> 00:06:41.230 majority of it. Again, some other... 83 00:06:41.950 --> 00:06:54.295 I'm seeing in here there's some some course fixes in certain places and and so on, but I think the next round of updates that we make relating to the contracts are going to be much more impactful, 84 00:06:54.615 --> 00:07:02.215 for users. So, David, do do you want... I can explain it, or do you wanna explain it on the, the cold storage situation? Either or is good. 85 00:07:03.590 --> 00:07:07.750 No. I think you summarized it pretty well. You know, 86 00:07:08.310 --> 00:07:11.430 it's always great to see these cybersecurity upgrades. 87 00:07:11.590 --> 00:07:13.750 They continue to harden the system. 88 00:07:14.069 --> 00:07:15.270 You know, today, 89 00:07:16.175 --> 00:07:17.855 when people are 90 00:07:18.015 --> 00:07:21.055 running a node and they wanna stake, 91 00:07:21.215 --> 00:07:24.975 you know, there's an element of sort of using a hot wallet. 92 00:07:25.135 --> 00:07:26.575 Obviously, in cybersecurity, 93 00:07:26.575 --> 00:07:28.735 using a cold wallet is is a lot better. 94 00:07:29.310 --> 00:07:31.390 So, you know, getting towards, 95 00:07:31.790 --> 00:07:40.190 the ability for people to stake and not have to, you know, hold more in the wallet, I think is sort of like the gold standard that everybody 96 00:07:40.350 --> 00:07:46.905 would love to, to get to. So I know Alan and the other folks that are contributing have been sort of marching towards, 97 00:07:47.225 --> 00:07:52.505 that goal. Still, some upgrades to go. But, yeah, like this is part of, 98 00:07:52.985 --> 00:08:00.750 how the software matures is just to keep making it more secure over time. But, yeah. Appreciate Alan and everybody working on that. 99 00:08:04.350 --> 00:08:05.470 Yeah. Absolutely. 100 00:08:05.710 --> 00:08:06.430 The 101 00:08:07.150 --> 00:08:08.750 cold wallet stuff the 102 00:08:09.150 --> 00:08:11.070 cold wallet stuff, I think, is 103 00:08:12.095 --> 00:08:33.250 something that we've been talking about for a while. We've been trying to do a lot of rolling some... A a lot of different features into kind of the same contract updates here, but we decided that, you know, this is is really is really, really important for our users. So we're gonna kind of pull this up to the front and make sure that we get it out. And that's just allowing users... 104 00:08:33.330 --> 00:08:41.345 You know, we have the delegation in the contract now, but it maybe doesn't work as great as we love. So we're we're rolling out something a little bit more robust 105 00:08:41.585 --> 00:08:42.945 so that you can, 106 00:08:43.825 --> 00:08:46.545 essentially hold your Morpheus token in a 107 00:08:46.865 --> 00:08:49.825 cold wallet and then give specific authorization 108 00:08:49.825 --> 00:08:50.625 to your 109 00:08:51.069 --> 00:08:55.310 throwaway hot wallet to be used in the Morpheus consumer node, 110 00:08:55.709 --> 00:08:56.430 protecting... 111 00:08:56.670 --> 00:08:58.670 You know, reducing your your... 112 00:08:59.790 --> 00:09:04.029 You know, in the risk event, you're reducing the amount of more that, you know, that would 113 00:09:04.695 --> 00:09:07.815 potentially be compromised if you expose your private key. 114 00:09:09.015 --> 00:09:10.615 Yep. Good stuff. 115 00:09:15.655 -->

00:09:19.655 But what Cool. Go ahead. I mean, I I think it'd be great to 116 00:09:20.270 --> 00:09:21.870 talk about, you know, 117 00:09:22.670 --> 00:09:31.790 the demos the last few weeks have been awesome. It's been very cool to see what's getting built. And one of the most requested things I've heard 118 00:09:32.030 --> 00:09:32.750 is 119 00:09:33.035 --> 00:09:37.035 people wanna see how folks are building their agent, how they're improving. 120 00:09:38.635 --> 00:09:48.235 And so why don't we jump into that? Yeah. Absolutely. And and that's exactly right. I mean, people have been... And I know even just for myself, I'm curious, like how are people actually 121 00:09:48.879 --> 00:09:53.600 building what their agent looks like on the back end, the skills that it has, the things that it can do. 122 00:09:54.000 --> 00:10:02.639 A hot topic that's been brought up, I know you've talked about this, David, is the idea of recursive self improvement for agents, not just for models. 123 00:10:03.144 --> 00:10:08.105 And so what we thought would be a great idea is kind of showing and talking about 124 00:10:08.425 --> 00:10:10.825 some of the, like, steps to getting there. 125 00:10:11.305 --> 00:10:14.985 Those being the agent brain, looping, and then RSI. 126 00:10:14.985 --> 00:10:15.144 So 127 00:10:17.640 --> 00:10:19.560 the ability to have self 128 00:10:20.920 --> 00:10:26.200 recursive improvements. Sorry. Geez, my brain. My brain could use self recursive improvements. 129 00:10:26.440 --> 00:10:27.240 Now 130 00:10:27.480 --> 00:10:40.705 the challenge with this is that it's hard to show code. So we were talking about this, you know, as a group, like, hey, how do we visually show this as a concept? Because... And actually, Kyle, you brought up a great point. It would be somewhat like saying, 131 00:10:41.440 --> 00:10:42.720 look at this code. 132 00:10:42.960 --> 00:10:56.654 Okay. You know, it does it does... Hopefully, you can read it. Or, like, it'd be like showing somebody a piece of music and being like, hey. Look at this written down music. Isn't it great? You know? So somewhat of a challenge. But we thought a good place to start 133 00:10:57.695 --> 00:10:59.774 for this is to share 134 00:10:59.935 --> 00:11:03.055 the GitHub page for Bernardo, 135 00:11:03.454 --> 00:11:07.535 which for those who don't know, Bernardo is one of David's agents. 136 00:11:08.260 --> 00:11:12.580 Is this like kinda... This is kinda like alpha prime agent for you. Right? 137 00:11:13.140 --> 00:11:14.019 Bernardo 138 00:11:14.019 --> 00:11:15.380 is pretty prolific. 139 00:11:16.180 --> 00:11:17.779 I've been using him 140 00:11:18.100 --> 00:11:20.180 really since I got into 141 00:11:20.740 --> 00:11:21.620 Open Claw 142 00:11:21.935 --> 00:11:24.015 back in January, February, 143 00:11:24.334 --> 00:11:27.214 and he's just been doing a ton of coding. 144 00:11:27.855 --> 00:11:31.214 And, practically, I've open sourced it all. So, 145 00:11:31.535 --> 00:11:37.520 you know, everything from the Morpheus skill, where it makes it easy for somebody to use 146 00:11:38.160 --> 00:11:40.480 Morpheus with OpenClaw, 147 00:11:40.480 --> 00:11:43.279 like, here's the repo to do that. Right? 148 00:11:44.320 --> 00:11:50.399 If you wanna look at install OpenClaw where it makes it sort of one click, I've gotten a lot of people asking recently, 149 00:11:50.855 --> 00:11:57.175 hey. I wanna make a white label of install OpenClaw and do it in Chinese 150 00:11:57.175 --> 00:12:01.975 and do it in Spanish and do it in Portuguese and do it in, you know, 151 00:12:02.695 --> 00:12:05.495 whatever their local language is. I think that's really cool. 152 00:12:06.140 --> 00:12:07.180 And so, 153 00:12:07.420 --> 00:12:15.980 you know, that's one of the things I've I've seen a lot of the last week is all these people popping out of the community saying, hey. I wanna do that. I wanna run one of those. 154 00:12:16.220 --> 00:12:24.605 You know? And I was like, great. Here's the code. And so like you said, Bo, what a lot of this looks like is handing your agent a repo 155 00:12:24.764 --> 00:12:32.765 and just telling it, like, go crazy. Take this as as a starting point. So I... I'd really recommend for for folks 156 00:12:33.324 --> 00:12:38.410 if they're looking to... You know, the Morpheus skill, you know, has a lot of what you need 157 00:12:38.730 --> 00:12:42.250 if, you wanna run the, the Morpheus router, 158 00:12:42.810 --> 00:12:47.530 and it has a lot of the learnings that, you know, Bernard has had the last six months about how to do that, 159 00:12:47.850 --> 00:12:48.490 effectively. 160 00:12:49.355 --> 00:12:57.995 But when it comes to, RSI, if we wanna kinda focus in on that, yeah, I like this idea of let's talk a little bit about the brain today. 161 00:12:58.634 --> 00:13:00.795 Next week, let's talk about looping. 162 00:13:01.130 --> 00:13:06.730 And then the week after, let's really get into RSI because these things build on each other. 163 00:13:07.050 --> 00:13:10.170 Right? If you haven't set up a brain 164 00:13:11.050 --> 00:13:21.514 for your your agent, it's gonna be really tough to do the looping. And until you have looping, you can't really set up an RSI agent. So, you know, having 165 00:13:21.514 --> 00:13:24.075 gone through all those evolutions myself, 166 00:13:24.394 --> 00:13:28.315 I'm happy to, you know, have been the guinea pig and offer 167 00:13:28.640 --> 00:13:34.400 what I've learned in that process. So Which, let's by the way, really quick, David, like in practical terms, 168 00:13:34.480 --> 00:13:46.705 what I'm going to do and what I recommend other people do is download this video when we're done and transcribe it, and then give the transcript to your agent as well as access to Bernardo's GitHub page, 169 00:13:46.945 --> 00:13:47.505 and 170 00:13:48.305 --> 00:13:50.785 start there. But anyway, yeah, go ahead. 171 00:13:51.425 --> 00:13:55.585

Yeah. So, you know, let's start today on the brain side. 172 00:13:55.825 --> 00:13:56.305 And 173 00:13:56.630 --> 00:13:59.670 I looked at a lot of different potential solutions. 174 00:13:59.910 --> 00:14:01.350 You know, I'm not 175 00:14:03.029 --> 00:14:08.630 a maxi about one particular solution, but the best one, at least for OpenClaw users, 176 00:14:09.204 --> 00:14:11.045 that I found was 177 00:14:12.165 --> 00:14:15.685 OC Brain, short for Open Claw Brain. 178 00:14:16.005 --> 00:14:18.005 And this is a skill, 179 00:14:18.245 --> 00:14:20.645 right, just have your agent effectively, 180 00:14:20.805 --> 00:14:22.805 you know, Google OC Brain 181 00:14:22.885 --> 00:14:23.365 skill, 182 00:14:23.750 --> 00:14:25.430 and it's an MD file. 183 00:14:25.829 --> 00:14:30.550 And the approach that OC Brain takes is to create a database 184 00:14:31.110 --> 00:14:38.905 of all the sessions you've ever had with your agent. And so when I set this up, I had 6,000 185 00:14:38.905 --> 00:14:39.865 sessions 186 00:14:40.265 --> 00:14:42.985 with my agent since the beginning of the year. 187 00:14:43.385 --> 00:14:45.545 And what it's doing is effectively 188 00:14:45.945 --> 00:14:47.625 trying to find patterns. 189 00:14:48.105 --> 00:14:50.665 It's trying to take the learnings 190 00:14:51.089 --> 00:14:53.570 that it has from all those sessions 191 00:14:53.730 --> 00:14:54.690 and sorta 192 00:14:54.930 --> 00:14:56.130 build certain 193 00:14:56.529 --> 00:14:57.410 beliefs 194 00:14:57.410 --> 00:14:58.209 or, 195 00:14:58.290 --> 00:15:01.490 you know, things that it can extract and says, oh, 196 00:15:01.970 --> 00:15:02.769 when David, 197 00:15:03.415 --> 00:15:08.295 you know, gives the approval, it's only after it's passed its testing. 198 00:15:08.535 --> 00:15:12.055 Right? When when David, you know, wants to code, 199 00:15:12.375 --> 00:15:16.055 you know, he he always uses another model to do the audit. 200 00:15:16.569 --> 00:15:20.329 Right? And you just take all those learnings from the sessions, 201 00:15:20.889 --> 00:15:21.370 and, 202 00:15:21.690 --> 00:15:22.410 you know, 203 00:15:22.810 --> 00:15:26.649 and you can then have a really rich database 204 00:15:27.370 --> 00:15:29.690 in order to apply that brain. 205 00:15:29.930 --> 00:15:32.935 The way that I'm applying all of that is 206 00:15:33.175 --> 00:15:34.055 now 207 00:15:34.535 --> 00:15:37.895 instead of writing all the prompts by scratch. 208 00:15:38.375 --> 00:15:43.735 When I write a prompt and and I set it up to, like, hey. I want you to do x y z. 209 00:15:45.110 --> 00:15:49.110 Look at my... So let's say my cron jobs that I'm running. 210 00:15:49.910 --> 00:15:56.390 I had my brain now go back and look at those prompts for the cron jobs and enhance them 211 00:15:56.630 --> 00:15:57.830 with the context 212 00:15:58.505 --> 00:16:07.625 that's relevant to that job. So if I had 20 cron jobs doing all of these different things, all of a sudden they became a lot more accurate. 213 00:16:08.024 --> 00:16:09.145 They became 214 00:16:09.385 --> 00:16:10.345 a lot sharper. 215 00:16:11.570 --> 00:16:14.450 And honestly, what it did, it was just save a bunch of time. 216 00:16:14.930 --> 00:16:15.970 What your 217 00:16:16.050 --> 00:16:19.410 agent is often doing is kind of redoing 218 00:16:19.570 --> 00:16:24.210 all the work it did last time because it doesn't remember. It doesn't have all the context 219 00:16:24.290 --> 00:16:25.970 for how it got to the solution. 220 00:16:26.605 --> 00:16:29.965 So if you you take the time to set up 221 00:16:30.365 --> 00:16:31.645 the brain skill, 222 00:16:31.885 --> 00:16:34.845 and you've digested sort of all of those patterns, 223 00:16:35.245 --> 00:16:38.365 now all of a sudden your agent is a lot smarter. 224 00:16:38.810 --> 00:16:50.170 Right? I mean, that's that's why people call this having a second brain or or giving your AI, you know, its own its own brain skill is it's taking everything you've learned before, 225 00:16:50.490 --> 00:16:52.170 and it's putting it in 226 00:16:52.490 --> 00:16:53.290 a 227 00:16:53.370 --> 00:16:54.250 context 228 00:16:54.535 --> 00:16:57.015 that your agent can really quickly access. 229 00:16:57.255 --> 00:17:01.335 Right? And so now, you know, that's become just part of my process. 230 00:17:01.815 --> 00:17:02.615 Every... 231 00:17:02.775 --> 00:17:06.695 I think it's every day, I've set it up to effectively do a new scan 232 00:17:07.015 --> 00:17:09.095 of all the learnings from that day, 233 00:17:09.610 --> 00:17:12.570 add it to the database, and then the database, 234 00:17:12.970 --> 00:17:14.570 it has it as a tool 235 00:17:14.730 --> 00:17:16.969 to reference to sort of improve 236 00:17:17.370 --> 00:17:18.489 each one of 237 00:17:18.809 --> 00:17:23.374 these cron jobs or things that your your agent is doing. So my my 238 00:17:23.534 --> 00:17:26.014 prompts went from, like, a couple sentences 239 00:17:26.255 --> 00:17:29.134 to here's an enhanced pros... You know, 240 00:17:29.615 --> 00:17:31.455 you know, basically a one pager 241 00:17:31.695 --> 00:17:34.654 for the agent to be able to know everything 242 00:17:35.135 --> 00:17:36.575 that's in context. So 243 00:17:37.360 --> 00:17:39.279 that... That's a good sort of 244 00:17:39.600 --> 00:17:41.519 how to set up your your 245 00:17:41.840 --> 00:17:43.759 OpenCLaw brain in a nutshell. 246 00:17:44.080 --> 00:17:47.599 The other thing I would say that's been really powerful me... For me 247 00:17:48.080 --> 00:17:51.715 is to take those same type of lessons, 248 00:17:51.955 --> 00:17:53.955 and you could build them into formalized 249 00:17:53.955 --> 00:17:57.635 standard operating procedures. I've I've got, like, 42 250 00:17:58.035 --> 00:18:07.840 SOPs now. And I... And I've tried to open source a lot of these where I... I've sort of ripped out the private information, and I've just published a generic version. Some of

these are in the repo. 251 00:18:08.160 --> 00:18:11.520 And, like, okay. If I wanna do coding, 252 00:18:11.920 --> 00:18:19.075 here's how I could do that. Right? And always run this regression test, always, 253 00:18:19.075 --> 00:18:21.315 you know, do the cross model auditing, 254 00:18:21.475 --> 00:18:22.594 etcetera, etcetera. 255 00:18:22.835 --> 00:18:24.195 And now all of a sudden, 256 00:18:24.755 --> 00:18:37.210 your, you know, quality of your code goes from, you know, one agent that maybe has some gaps or errors to really, really solid code after it's been properly audited. So, you know, 257 00:18:37.850 --> 00:18:41.850 one of the things I would say is a lot of this this brain related 258 00:18:41.930 --> 00:18:42.809 activity, 259 00:18:43.210 --> 00:18:44.650 it's basically context management. 260 00:18:45.495 --> 00:18:47.655 Because, yeah, we got 200,000 261 00:18:47.655 --> 00:18:48.455 tokens, 262 00:18:48.615 --> 00:18:50.455 but man, if you try to shove 263 00:18:50.615 --> 00:18:54.455 all of your information in the context window, a lot of times it'll just fall over. 264 00:18:54.775 --> 00:18:58.375 Right? And and now I've got 42 SOPs. 265 00:18:58.455 --> 00:19:00.455 They don't all fit in the context window. 266 00:19:01.279 --> 00:19:06.479 And so I'd really recommend to people it's worth creating an index layer 267 00:19:06.799 --> 00:19:08.719 where you're indexing 268 00:19:08.880 --> 00:19:12.159 all of those 42, basically a small quick summary 269 00:19:12.399 --> 00:19:13.919 of every SOP. 270 00:19:14.425 --> 00:19:19.385 And just now now I've got it down to this point where I tell my agent, hey, 271 00:19:19.865 --> 00:19:21.784 check your SOP index file, 272 00:19:22.185 --> 00:19:25.945 find out which context is relevant to what you're working on, and go ahead. 273 00:19:26.345 --> 00:19:31.830 And now it can just read one short file. It sees, oh, this is SOP 39. 274 00:19:32.070 --> 00:19:34.950 Great. It only reads the context 275 00:19:35.190 --> 00:19:40.789 that is relevant to its task. So, you know, that's kinda what I would recommend is, you know, 276 00:19:41.535 --> 00:19:42.975 download the skill, 277 00:19:43.375 --> 00:19:44.735 create these patterns, 278 00:19:44.975 --> 00:19:47.295 start building out these SOPs, 279 00:19:47.295 --> 00:19:54.175 create an index file, and, you know, it's it's been amazing for sort of the accuracy and output 280 00:19:54.590 --> 00:20:04.110 of my agents went from, like, you know, having trouble most of the time to almost every time. Now they're able to get the task done because they have the information 281 00:20:04.190 --> 00:20:06.830 from the brain that they need to get it done. So, 282 00:20:07.470 --> 00:20:10.055 yeah, check it out. Give me feedback. 283 00:20:10.375 --> 00:20:14.295 I wanna open source more and more of my my SOPs, 284 00:20:14.775 --> 00:20:29.700 but, you know, this has been this has been a really big unlock. So if you spend the time to do that and sort of... You know, it took it almost a full day for my my agent to crank through using the skill to digest those 6,000 sessions. 285 00:20:30.820 --> 00:20:33.140 If you don't have so much data, it might be a lot faster. 286 00:20:33.675 --> 00:20:37.675 But for at least me, you know, it's it's worth the time though. It's paid off 287 00:20:38.315 --> 00:20:46.075 in a huge way the last few months that I've... Since I've set that up. So Basically. Happy to answer any question. Bo, you might you might have a good perspective, 288 00:20:46.549 --> 00:20:52.229 you know, coming fresh at this. Yeah. I do. I... I'm I'm curious. So, like, let's say 289 00:20:52.870 --> 00:20:55.269 I I share with my agent 290 00:20:56.150 --> 00:20:57.749 this GitHub repo. 291 00:20:57.830 --> 00:21:01.350 I tell it that, hey. I want to set up a brain using this skill. 292 00:21:02.495 --> 00:21:12.255 And you said that it it took about a day because it's going through all of the past sessions and kind of plucking the the important pieces to add to that brain and and actually, like, build out that database. 293 00:21:12.815 --> 00:21:20.170 Is this like a do it once and it's done type of upgrade to your agent, or is this something that you are continually 294 00:21:20.170 --> 00:21:22.410 monitoring and, like, remanually 295 00:21:22.410 --> 00:21:23.370 updating 296 00:21:23.690 --> 00:21:31.524 and adjusting the skills? That makes sense? I don't know if that question makes sense. Well, the nice thing is once you've got it set up, I just set up a cron job 297 00:21:31.845 --> 00:21:34.164 to update it every day. 298 00:21:34.485 --> 00:21:36.645 And so once I had that digested, 299 00:21:36.645 --> 00:21:38.005 you're maybe only doing 300 00:21:38.510 --> 00:21:43.870 10, most people, 10 or 20 sessions a day if your agent is cranking through different tasks. 301 00:21:44.110 --> 00:21:47.950 So it goes from, like, oh, I have a backlog of hours 302 00:21:47.950 --> 00:21:50.030 to do. Oh, it take a couple minutes. 303 00:21:50.190 --> 00:22:00.264 Right? And I don't have to do any of that manually. At this point, I just have a cron job, the agent looks at the sessions for the day, and just updates the brain automatically. 304 00:22:00.665 --> 00:22:04.664 Right? And then I have the brain also automatically looking 305 00:22:05.065 --> 00:22:06.345 at all of the things 306 00:22:06.970 --> 00:22:08.010 in my 307 00:22:08.570 --> 00:22:10.010 agentic backlog. 308 00:22:10.250 --> 00:22:15.690 This this is, you know, sort of a little preview of what I'll I'll cover next week in the looping. 309 00:22:16.090 --> 00:22:18.250 But you wanna get away 310 00:22:18.570 --> 00:22:19.450 generally 311 00:22:19.735 --> 00:22:20.455 from 312 00:22:20.855 --> 00:22:29.894 babysitting your agent and giving the tasks one by one. Eventually, I just created a backlog. Hey. Here's a backlog of everything that you need to work on. 313 00:22:30.455 --> 00:22:32.455 And and

once you have that backlog, 314 00:22:33.300 --> 00:22:36.340 as it's sitting there waiting to to be picked up, 315 00:22:36.580 --> 00:22:38.020 my brain is 316 00:22:38.500 --> 00:22:39.220 enhancing 317 00:22:39.860 --> 00:22:40.900 the context 318 00:22:41.380 --> 00:22:49.495 of of that prompt. So when the worker shows up, it has everything it needs to execute. So it's a really important 319 00:22:49.575 --> 00:22:50.775 prerequisite 320 00:22:50.935 --> 00:22:52.855 for setting up all the looping. 321 00:22:53.095 --> 00:22:59.575 Well, and I just wanna make sure, like, in in... Put in the most, like, basic terms too. Because I I already know, like, 322 00:23:00.250 --> 00:23:02.169 if I didn't already have you 323 00:23:02.490 --> 00:23:06.890 helping to explain to me how some of these things work, I wouldn't inherently 324 00:23:06.890 --> 00:23:10.249 understand the difference of just saying, okay. Well, 325 00:23:10.570 --> 00:23:17.195 I did my, you know, install openclaw dot x y z agent. Maybe I even took that next step of exporting 326 00:23:17.515 --> 00:23:32.429 it locally to my device, but I'm using Morpheus for the inference. Right? So now I have my agent that I'm talking to locally, but it's calling out not locally for the inference. But, hey. I have this MD file where I explain to it, this is what you are. This is what you do. 327 00:23:33.070 --> 00:23:35.790 Can we even, like, maybe even more simply 328 00:23:35.950 --> 00:23:37.549 just highlight again 329 00:23:37.710 --> 00:23:43.524 why the brain even matters in the first place relative to the MD file as it exists today? 330 00:23:44.085 --> 00:23:52.565 I know you you basically covered that in great detail just now, but, like, that's the question that I know I would have if I didn't... Hadn't already talked to you about this. 331 00:23:53.125 --> 00:23:56.885 So think about it like short term memory versus long term memory. 332 00:23:57.620 --> 00:23:58.260 Right? 333 00:23:58.500 --> 00:23:59.540 In 334 00:23:59.540 --> 00:24:08.420 previous discussions we've talked about with OpenClaw, for example, and I think it's very similar with Hermes and other Aejenta carnasses, you can set up a soul file. 335 00:24:09.060 --> 00:24:13.024 Right? And your agent's gonna read that soul file before 336 00:24:13.184 --> 00:24:14.304 every prompt. 337 00:24:14.865 --> 00:24:22.864 Right? And it's basically your system prompt for your agent is to say, hey. The soul is your name is Bernardo. 338 00:24:23.105 --> 00:24:24.544 You're helping me with coding. 339 00:24:25.190 --> 00:24:28.230 You know? You should always be frank about, 340 00:24:28.390 --> 00:24:38.470 you know, what you find and accurate in your outcomes, and here's your agent file. This is, you know, your list of skills and tools you have available. 341 00:24:38.790 --> 00:24:40.790 And, oh, here's the user file. Here's 342 00:24:41.135 --> 00:24:42.015 David, 343 00:24:42.015 --> 00:24:45.135 and he's working on these things, and he has these repos. 344 00:24:45.615 --> 00:24:47.615 But these are relatively small 345 00:24:48.015 --> 00:24:48.815 files, 346 00:24:49.135 --> 00:24:51.455 right, that you're putting in short term memory. 347 00:24:51.695 --> 00:24:59.570 Right? They're, you know, a a page each, basically, because you don't wanna overwhelm the context. If you literally try to shove a novel 348 00:25:00.050 --> 00:25:09.570 into the system prompt, you gotta take up a lot of the context window, and now you can't really get worked down because you've, you know, put all of this unrelated context 349 00:25:10.565 --> 00:25:13.445 into what it's trying to do, it can get confused. 350 00:25:13.605 --> 00:25:15.125 So Yeah. You have you 351 00:25:15.845 --> 00:25:20.085 you have the lost in the middle problem as well. Right? So if you... 352 00:25:20.725 --> 00:25:27.600 Right. The the more context you give, the agent is really good at remembering what happened in the beginning, 353 00:25:27.920 --> 00:25:30.240 what happened at the end, and what happened, 354 00:25:30.560 --> 00:25:37.440 I think, almost exactly in the middle. But in the in the the two con... The the two areas, 355 00:25:37.440 --> 00:25:38.559 you know, between 356 00:25:38.720 --> 00:25:40.560 beginning and middle and middle and end, 357 00:25:41.415 --> 00:25:42.535 it's kinda like, 358 00:25:43.175 --> 00:25:50.775 alright. I skimmed this, but I don't remember the specifics here, and and the specifics are important. So when you're creating that, 359 00:25:51.095 --> 00:25:53.015 let's call it a prompt pack. 360 00:25:53.095 --> 00:25:53.335 Right? 361 00:25:54.030 --> 00:25:56.750 When you're creating that pack, you're saying, 362 00:25:57.070 --> 00:26:00.830 you know, exactly what David said. You nailed it. It's who am I? 363 00:26:01.390 --> 00:26:09.024 What is... You know, what are the critical things that I need to know no matter what? You know? What are my goals 364 00:26:09.184 --> 00:26:27.879 for this, which which are generated not maybe at a global basis, but as a... I'm working on this right now, so this is the goal of that. And then what's my context of this? Right? Maybe it's the short term context where if it's your second message in a chain, it includes the first message. Right? But then also what's my... 365 00:26:28.120 --> 00:26:28.759 You know 366 00:26:29.720 --> 00:26:41.025 you know, let... Let's call chat history as an immediate context, short term memory as a, I've been working on this today, and then long term memory as, hey. These are global rules and learnings. And, like, how do I create 367 00:26:41.505 --> 00:26:45.265 that prompt pack to be both concise 368 00:26:45.265 --> 00:26:47.265 but but also comprehensive? 369 00:26:47.265 --> 00:26:51.380 Right? Concise in number of tokens, but comprehensive in being 370 00:26:51.700 --> 00:26:56.340 able to answer my question and and understanding enough what's going on. So I think that's... 371 00:26:57.940

--> 00:27:03.220 There there are a couple ways to do that, David. I'm I'm happy to... I'm interested in you explaining your implementation, 372 00:27:03.220 --> 00:27:08.955 but it's... This is something that we have all gone through many, many different times on 373 00:27:10.075 --> 00:27:16.155 how do you get these damn things to work well. And, you know, that's one of the the key triggers on 374 00:27:16.635 --> 00:27:18.155 garbage in, garbage out, if you will. 375 00:27:18.889 --> 00:27:23.450 Yeah. Yeah. For me, the unlock was... At first, I put too much 376 00:27:23.850 --> 00:27:31.210 into the sole file and the user, and it just kinda got confused. And it was it was too ethereal. My instructions were very philosophical 377 00:27:31.210 --> 00:27:35.604 at first. What I ended up with at the end of the day was much more practical, 378 00:27:36.164 --> 00:27:37.844 shorter descriptions 379 00:27:37.845 --> 00:27:41.205 of what I wanted. So that's again the short term memory, 380 00:27:41.445 --> 00:27:46.325 right, that you're putting in the system prompt with these things that go with every time 381 00:27:47.160 --> 00:27:54.680 it sends a prompt into to the agent. The brain is really your long term memory. Mhmm. Right? So it's everything 382 00:27:54.680 --> 00:27:59.000 you've ever talked about with your agent, but it's able to recall it 383 00:27:59.804 --> 00:28:01.565 from its short term memory 384 00:28:01.804 --> 00:28:05.565 and and access that long term memory only when it needs. 385 00:28:05.804 --> 00:28:14.524 Right? That's the key here. That's why I call it context management is the brain is basically giving you a tool for doing better context management 386 00:28:14.899 --> 00:28:16.099 at scale. 387 00:28:16.500 --> 00:28:27.940 Mhmm. Because these systems will work for a little while, and then as your context gets larger and larger, it has more and more trouble. So you're just basically giving yourself a structure 388 00:28:28.205 --> 00:28:34.205 to infinitely scale this with as many files as you want, and you're only calling in 389 00:28:34.365 --> 00:28:39.245 what you need for that particular task or session. So Yeah. I've always heard that 390 00:28:39.645 --> 00:28:42.765 that exact problem is kind of what stops 391 00:28:43.009 --> 00:28:43.570 the, 392 00:28:44.610 --> 00:28:46.209 replacement of humans, 393 00:28:46.370 --> 00:28:55.984 basically, is just the fact that our our context windows and ability to pull from the past in data that we took in, you know, twenty years ago is just so 394 00:28:56.304 --> 00:28:59.585 great, and in a lot of ways, not even understood. 395 00:29:00.144 --> 00:29:02.864 The example I always give is, like, it's crazy how 396 00:29:03.105 --> 00:29:13.099 you could smell something. Like, you could smell, like, a certain food Mhmm. And then it gives you a memory from, like, two decades ago of, like, a very particular time, place, and experience, 397 00:29:13.340 --> 00:29:19.179 and you have no idea, like, how your brain is necessarily storing that connection between the two. 398 00:29:20.059 --> 00:29:23.995 Yeah. We upgraded to JEV. We we got JEV and VectorDB 399 00:29:23.995 --> 00:29:25.034 up there. Yeah. 400 00:29:25.595 --> 00:29:27.674 Is Kyle's ready running JEV locally. 401 00:29:29.595 --> 00:29:31.674 Yeah. I guess, yeah, that's exactly right. 402 00:29:32.075 --> 00:29:36.235 No. That's a good explanation though. And I I always like going down to the... 403 00:29:37.010 --> 00:29:44.690 And I apologize, I have a cold. Going down to, like, the most layman terms that we can get to because I just know, like, personally, 404 00:29:45.010 --> 00:29:48.770 that's how I would experience it, you know, if I was listening to this. 405 00:29:49.410 --> 00:29:51.650 But, I think that's a good intro to 406 00:29:51.934 --> 00:29:59.614 to the brain as a concept, and we'll do it in small bites. Next next week, we'll talk about looping, and then we'll go into RSI, 407 00:30:00.095 --> 00:30:06.015 recursive self improvement, not self recursive improvement. I think I said that wrong about three times when we first jumped on. But, 408 00:30:07.090 --> 00:30:14.130 yeah, we we will, you know, continue down this thread, and awesome. Thank you for kind of running through that, David and Kyle. 409 00:30:15.490 --> 00:30:28.234 I did have... I mean, we had a handful of things in the news. We we are gonna try to make the stream a little less news heavy than it used to be, but I still think it's worth, you know, touching on some of the things for the people that that kinda stick around. 410 00:30:29.355 --> 00:30:31.835 The biggest thing that I wanted to mention was just that, 411 00:30:32.554 --> 00:30:33.354 obviously, 412 00:30:33.355 --> 00:30:36.715 all of the frontier models and open source 413 00:30:36.870 --> 00:30:48.070 creators are always trying to get to, like, the next level of intelligence, the next level of, you know, products and services that they're putting out. But it seems to be the case that competing on price has, like, really leapfrogged 414 00:30:48.070 --> 00:30:48.870 the narrative. 415 00:30:49.510 --> 00:30:51.510 Fortunately for everyone in our space, 416 00:30:52.605 --> 00:31:02.044 you can't really compete with the price of open source. I mean, it is still multiple orders of magnitude more cost efficient to be using open source 417 00:31:02.125 --> 00:31:03.005 models. 418 00:31:03.005 --> 00:31:08.949 And in some cases, you have, like, OpenAI changing the pro price of ChatGPT, 419 00:31:08.950 --> 00:31:15.109 but actually giving you less allowance. So they both increase the and decrease the allowance, 420 00:31:15.110 --> 00:31:16.309 which is hilarious. 421 00:31:17.535 --> 00:31:24.015 And a lot of corporate America, or at least an increasing number of corporate America, has moved to open weights as well. 422 00:31:25.375 --> 00:31:27.535 There was data from Alphasense 423 00:31:27.615 -->

00:31:36.510 that showed that AT and T currently runs about 40% of its AI workloads using open weight models, and they're aiming for 70% 424 00:31:36.510 --> 00:31:37.870 within the next year. 425 00:31:38.110 --> 00:31:46.794 And that's, I would suspect, one of many major companies who will be doing that. I don't know if you guys have a stance or any takes on on any of that. 426 00:31:48.075 --> 00:31:49.115 It's 427 00:31:49.115 --> 00:31:57.755 cool to see the move, you know. This is what Morpheus community has been predicting for a while. Right? That we're just gonna see ever more 428 00:31:58.440 --> 00:32:00.599 open weight model usage. 429 00:32:00.600 --> 00:32:03.080 It's what you see on a lot of the public statistics. 430 00:32:03.320 --> 00:32:06.600 You know? Last week, I think it crossed 50% 431 00:32:06.600 --> 00:32:09.960 generally in the industry are using open weight models. 432 00:32:11.065 --> 00:32:14.905 And, you know, god bless if this is part of what Jensen 433 00:32:14.985 --> 00:32:16.585 has been standing up 434 00:32:17.145 --> 00:32:21.065 for the open weight community because he wants to sell GPUs. 435 00:32:21.305 --> 00:32:31.479 Right? And he doesn't care if people are running open weights if they're running proprietary models. He just wants to sell more GPUs. And so he's really been a powerful advocate 436 00:32:31.640 --> 00:32:36.119 the last few weeks to avoid some of the regulatory capture stuff and everything else. 437 00:32:37.000 --> 00:32:41.565 The one thing I saw in the news that was pretty crazy was, I guess The UK 438 00:32:41.805 --> 00:32:44.445 is now formally blocking Venice. 439 00:32:45.165 --> 00:32:47.404 And Yes. I have that next on my list. Yes. 440 00:32:47.885 --> 00:32:50.764 If you're, you know, in The UK, 441 00:32:51.645 --> 00:32:53.725 now that your government wants KYC, 442 00:32:54.590 --> 00:32:56.830 right, to access AI, 443 00:32:57.470 --> 00:32:59.070 which which is pretty wild. 444 00:32:59.310 --> 00:33:01.230 And so they've just had to straight up 445 00:33:01.710 --> 00:33:02.350 block 446 00:33:02.590 --> 00:33:13.504 users unless they become a pro user and do the the, you know, basically, the equivalent of KYC, which is is prove you're you're an adult with a payment credential 447 00:33:13.664 --> 00:33:18.144 in order to access their... So it's really sort of limiting anybody that wants to use the free tier 448 00:33:18.385 --> 00:33:23.504 from accessing it. But, you know, it's a good reminder of why 449 00:33:23.985 --> 00:33:25.024 Morpheus is important. 450 00:33:25.720 --> 00:33:27.160 Right? Because 451 00:33:27.240 --> 00:33:27.960 any... 452 00:33:28.200 --> 00:33:31.080 You know, Venice is wonderful, but as a company, 453 00:33:31.640 --> 00:33:34.360 like, you know, they... They've got a deal 454 00:33:34.600 --> 00:33:36.840 with those type of of regulations, 455 00:33:36.840 --> 00:33:38.280 right, as opposed to 456 00:33:38.840 --> 00:33:40.200 an open source protocol, 457 00:33:40.675 --> 00:33:41.315 right, 458 00:33:41.555 --> 00:33:44.595 where there's there's nobody to run KYC. 459 00:33:44.835 --> 00:33:50.595 Right? It's a it's a peer to peer network. So... You know? But it's a good reminder. Right? And, 460 00:33:51.395 --> 00:33:53.395 hopefully, it brings awareness 461 00:33:53.395 --> 00:33:54.755 and, like, the 462 00:33:55.310 --> 00:33:57.390 level of Orwellian 463 00:33:57.390 --> 00:33:58.669 craziness 464 00:33:58.990 --> 00:34:01.389 that this is sort of exemplifying. 465 00:34:01.550 --> 00:34:04.430 Hopefully, you know, sort of wakes people up faster. 466 00:34:04.430 --> 00:34:06.269 But, anyway, I thought that was pretty wild. 467 00:34:06.510 --> 00:34:17.625 Yeah. I mean, if only there was a whole community of people across the globe who were building a way to access these things that are permissionless and perpetual and don't require, you know, 468 00:34:18.025 --> 00:34:23.385 any of these things that The UK wants to push. The UK's ability to just absolutely 469 00:34:23.385 --> 00:34:30.930 just shoot themselves... Cut themselves off at the kneecaps every time there's any type of technological advancement is quite quite incredible. 470 00:34:31.490 --> 00:34:50.785 It's it's truly amazing. It's truly amazing. And it's like, didn't you guys help give birth to the free speech thing? Like, what what happened to that UK? Like, you know, it's... I guess it's been a few hundred years, but... Yeah. They they officially gave up when the Internet came out, and now they're just, like, doubling down, tripling down, and it's 471 00:34:51.930 --> 00:34:56.010 the revolution really messed those guys up, man. People forget. Yeah. 472 00:34:56.410 --> 00:35:00.330 Yeah. It's, you know, the the freedom moved to a different jurisdiction, 473 00:35:00.730 --> 00:35:10.725 some... Somewhere in Texas or America or whatever. But, yeah, it's it's pretty wild. So, you know, we'll keep building open source and, you know, I think 474 00:35:11.205 --> 00:35:15.845 freedom of intelligence is a human right worth fighting for. And so 475 00:35:16.005 --> 00:35:20.405 here we are doing doing the good fight. It was it was a really amazing that day. 476 00:35:21.220 --> 00:35:23.860 I got to speak at the Astana hub 477 00:35:24.100 --> 00:35:26.420 and, you know, ton of 478 00:35:27.060 --> 00:35:29.380 AI experts and and 479 00:35:30.100 --> 00:35:30.900 local 480 00:35:31.700 --> 00:35:39.755 folks interested in AI showed up. And, yeah, I'll post some of the photos and stuff, but two hours. We spent two hours 481 00:35:39.994 --> 00:35:45.755 answering every possible question they can come up with on agents. It was it was really a blast. 482 00:35:45.994 --> 00:35:46.555 So, 483 00:35:46.795 --> 00:35:48.475 yeah, kudos to 484 00:35:48.795 --> 00:35:51.195 the network school on their new campus. 485 00:35:51.434 --> 00:35:51.755 You know? 486 00:35:52.500 --> 00:35:56.420 If I were to summarize

I've... Summarize it for people, it's kinda like Dubai North. 487 00:35:56.900 --> 00:35:59.540 Like, it's a colder version of Dubai, 488 00:35:59.700 --> 00:36:02.660 but it's very similar, like, fast growing 489 00:36:02.900 --> 00:36:04.260 neutral country 490 00:36:04.660 --> 00:36:07.220 with a a, you know, real 491 00:36:08.145 --> 00:36:09.025 friendly 492 00:36:09.665 --> 00:36:13.825 framework for for AI and crypto and and related projects. 493 00:36:13.985 --> 00:36:19.025 It's a kind of place where, like, oh, yeah. You can set that business up, and it'll take fifteen minutes to incorporate. 494 00:36:19.665 --> 00:36:35.105 And you can, you know, do all these things that would be impossible in most most countries. So, anyway, it's cool to see. It's it's a beautiful city. It's basically brand new, and a lot of people moving here. So, yeah, it's been fun to explore. Transparently, Transparently, I've never even considered 495 00:36:35.105 --> 00:36:37.585 wanting to go to Kazakhstan before, 496 00:36:37.985 --> 00:36:39.905 but I do now. So, 497 00:36:40.145 --> 00:36:52.120 you know, I I do... I would love to visit. I would love to see Network School and and Astana. So, yeah, it's great. I'm I'm very impressed at how quick they were able to just keep the ball rolling and and keep the mission alive. And 498 00:36:52.520 --> 00:36:54.120 it just shows that 499 00:36:54.920 --> 00:37:06.215 open... Like, as long as there is anywhere on Earth that wants to support open source and freedom of of, you know, freedom technologies and freedom of speech in the form of code, 500 00:37:06.535 --> 00:37:07.575 it will exist, 501 00:37:08.214 --> 00:37:20.250 you know? And so it is the gift that keeps on giving. It is it is the most for the people thing of all time, which is so funny to me because I, you know, people hear AI all the time and instantly assume 502 00:37:20.809 --> 00:37:22.650 evil empire centralized 503 00:37:22.650 --> 00:37:34.525 control of your life, and they don't realize that there is an entire, you know, thousands and thousands of people across the world who are, like, very much building the thing for them specifically, 504 00:37:34.765 --> 00:37:37.005 you know, for the individual. So, 505 00:37:37.405 --> 00:37:44.605 yeah, shout out shout out to them. Very cool that you're out there, David. I hope the talks and everything went well. It looked like it did, the one you gave the other day, so. 506 00:37:45.300 --> 00:37:49.300 Yeah. Last thing we're saying is whoever's going to token twenty forty nine 507 00:37:49.620 --> 00:37:54.100 is coming up, like, next week. So if you're gonna be out in Singapore, 508 00:37:54.420 --> 00:37:56.420 tons of Morpheus community there. 509 00:37:56.820 --> 00:38:02.115 They'll be at token twenty forty nine, be at the network state conference in Singapore. 510 00:38:02.275 --> 00:38:06.675 Should be a blast. So looking... Usually, the biggest conference in in 511 00:38:07.474 --> 00:38:10.515 crypto every year. So if you're coming out, come say hi. 512 00:38:11.390 --> 00:38:12.190 Awesome. 513 00:38:12.190 --> 00:38:22.110 Yeah. Please do, people. If you come out, you see the Morpheus shirt, I can guarantee that logo will be on David's chest, almost certainly. I think he's got a wardrobe of Morpheus t shirts. 514 00:38:22.670 --> 00:38:28.325 You will find them. But, yeah, absolutely go connect with the Morpheus community out there if you're token 2,049. 515 00:38:28.565 --> 00:38:29.365 Otherwise, 516 00:38:29.365 --> 00:38:32.325 thank you as always for joining. Hopefully you like the 517 00:38:32.645 --> 00:38:35.045 even further revised new format. 518 00:38:35.365 --> 00:38:50.960 Feel free to give us feedback. I mean, we really wanna provide, you know, information and things that are truly useful to the community and the people following the project who want to be involved in building on the project or a user of the project. Yeah. So just thank you to everybody. Kyle, thank you as always for hopping on. 519 00:38:51.280 --> 00:38:55.040 And with that, I think we will cut it and see everybody next week. 520 00:38:55.599 --> 00:38:56.480 Thank you, guys.