7/31/2026 - How to Build a Personal AI Brain Using Your ChatGPT History

Transcript

1. 0:03[music] [music] [music]
2. 0:25[music] Good evening, good people. Let me turn that thing off right there. Oops. Welcome to the AI learning lab in the AI
3. 0:41salon. It is fantastic to have you here. Um, let's see. I got to I'm I'm producerless, although the producer is
4. 0:48sitting in a Sheets gas station right now making sure we're live. Um, welcome, welcome, welcome. I can't see if anyone
5. 0:58I can't see the number of people here. So, right now it says there are no people here, but maybe there are people
6. 1:03here and they're just not showing. Um, okay. Do me a favor. If you're here, put a comment in the comments. So, so
7. 1:12there's a thing that happened. Okay. Digital gods, hello. There's a thing that happened. Streamyard bought Vimeo.
8. 1:20Remember when Vimeo was a big deal and it was like Vimeo's for business and YouTube's for consumers, right? That
9. 1:25thing. Um, so I can't see how many people are here. Um, all right. We'll figure that out. Anyway, they they bought them and we can
10. 1:35no longer live stream without using Vimeo. So, in theory, this might be a better experience. Um, it looks like
11. 1:44we've got hello saying oh hell [laughter] or or ole o. Um, and then we've got digital gods saying hello. So if anyone
12. 1:54else is here the Hey Claire the um popping out the chat is now in the lower in one of the corners either there or
13. 2:02there. It used to be up there. Now it's in one of those. So, I don't know like like let's see this one is the is the hand [laughter]
14. 2:12where these green notebooks are. Um, so is it is it on the guitar side or the green notebook side where the button
15. 2:21is to pop out the chat? You can you all can help me. It's ask me anything, but I'm going to need some help.
16. 2:32I do wish I could see how many people were in here. webinar settings. Uh the chat is below the video on a PC. It's
17. 2:40underneath. Oh, it's underneath. It's underneath the thing al together. Great. Pop the chat out there then. [laughter]
18. 2:52Um let me see. Webinar settings. This might have been a bad idea. Technical difficulties. Okay. So, we can't go to our webinar settings.
19. 3:03Show comments on stage. No. Share with viewers. Customize preview. All right. So, I have no idea to know
20. 3:10how many people are in here. Um, but [clears throat] if you're here, I think I think we're going I think we're live. [laughter]
21. 3:20So, so welcome. Let's see. Claire, I don't know where to pop it out on my phone. Um, I don't know. I don't know. [laughter]
22. 3:30And producer Brandon's on his phone in the car. Actually, I just opened the window and it's hot. Hopefully.
23. 3:46Oh, I hate this. Okay, tell me why. Charles Adams, what is it? So, so here's the thing. KK um what's it called? Streamyard bought
24. 3:59Vimeo and is now forcing us to use Vimeo as a webinar provider. Um, it sounds like you're not enjoying it. It will get
25. 4:08better literally one day. Streamyard was building it while I was talking to them. We are live. You have 10 people. How do
26. 4:18you know I have 10 people? And I don't know I have 10 people. Wait. Oh, here's people. Let me look at the people list.
27. 4:25No stage one person. backstage one person. No. All right. All right. Well, thank you, Dina. Okay. So, um if you're new to the AI salon,
28. 4:36welcome. Um we're working out the kinks with Streamyard. We'll figure it out. This is It's Friday night. We're hanging
29. 4:42out. Um here in Denver, it just rained. It's It's kind of a lovely evening. Uh it it took some of the heat out of the
30. 4:49air, but it is humid. So, I'm feel like I'm in I'm back on the East Coast all humid and moist. Um, and uh,
31. 4:59if you're new to the AI salon, be sure to check out all the different spaces down the side of the AI salon. Um, if
32. 5:05you didn't join, actually join. Oh, wait. We've got a we have a producer, producer Brandon. >> So, please share all your comments with
33. 5:12us uh, about the good, the bad, the ugly. We will give them to Streamyard. Streamyard is literally building this
34. 5:19boat as it's sailing down the canal. Like literally, I asked them like, "Hey, where can I put Kyle's picture?" And
35. 5:25they're like, "Hang on, the engineers are building that feature right now. Uh, give us 20 minutes and refresh."
36. 5:32>> Uh, so this is all very new. >> Um, so the good news is that when you click on webinar settings, I wouldn't
37. 5:41recommend doing this while you're live because it'll screw up your broadcast, but it will tell me how many people are
38. 5:47watching. So, right now, we have 12 people. Okay, cool. Yeah, Dina just says we have 12 people. If you click
39. 5:55>> just right in the upper leftand corner. >> I don't know if it'll work for us in our window, but in the player you can see it.
40. 6:03>> Yeah. Why would the people producing the show want to see how many people are in in the thing? Mary Carol, what's
41. 6:08happening? Um, okay. So, here's the deal. Um, Andy, >> I'm gonna go drive. >> Okay, you go drive. Thank you, producer
42. 6:16Brandon. Uh, beautiful. Uh, go God forbid. God, God forbid. Um, they just leave tech alone so we don't have to
43. 6:24spend yet more hours talking about tech on live broadcast. I know, right? Um, anyway, uh, get get that. It's
44. 6:30frustrating. Let us know things. We'll let Streamyard know. Um, please, if you're new to the AI salon, check out
45. 6:36the AI salon. Also, I can finally chat again. Yes, I haven't been able to chat in a live on my lap laptop in over a
46. 6:42week. Okay, one positive comment. Beautiful. If only there were a way to get chat to show transparently above the screen,
47. 6:50just like Tik Tok. All right, beautiful. All right, it's Friday night date night. Welcome everybody. Oh, look at that.
48. 6:57There's uh chat showing up on the screen. Hey, it's just like Tik Tok. That was cool. Oops. What did I do? Did I want to do
49. 7:09that? I didn't want to do that. They need to allow picturein picture so that phones are locked and forced into
50. 7:16watching and not even being able to check text without losing sound, etc. All right, beautiful. Linda Sherman's in
51. 7:25the house. All right, we got some good people here. All right, so it's Friday night date night. So, here's my request.
52. 7:30Let's not talk about the tech anymore. It's It's janky. It's a mess. I agree with you. We'll we'll do our best to fix
53. 7:37it and get this worked out. We were kind of forced down this path, so here we are. But it's Friday night date night.
54. 7:44It's ask me anything night. So, if you have any questions, please pop them in the comments. If you want to know where
55. 7:50the comments are, it's down below the video. Here is a link to pop out the comments. So, go ahead and do that. I
56. 7:56don't know if you can pop this out as a player anymore. I'm not sure how that works. Um, but uh but let's see. Try
57. 8:04Mighty Networks in a Chrome browser. All right. Beautiful, beautiful, beautiful, beautiful. Okay, so let's
58. 8:12see. We got a couple of questions here. Let's see what this one is. Hang on a sec.
59. 8:26Soggy Waffle Cat. Good name. Um, Buzz says, "Your people, your agents, your project, all in one place. Is that where
60. 8:35AI is heading?" heading less chatbot and more more shared work work space. Um, oh, you can pop it out. Okay, cool. This
61. 8:47chat is weird and not the same as before. Yes. Yes. This is an evening of who moved my cheese. Um, Streamyard
62. 8:55forced us to do this. [laughter] This is this is Vimeo technology um being rolled into StreamYard on our dime. Um [laughter] so okay,
63. 9:10[clears throat] so this question about Buzz, your people, your agents, your project all in one place. Is that where AI is headed?
64. 9:18Less chatbot and more work workspace. So Buzz is Jack Dorsey's new Slack like competitor. um that has a lot of advantages to it. Um
65. 9:34it treats humans and AI agents as sort of equal companions in in a workspace very much like Slack. It's got a lot of features around data
66. 9:46sovereignty. You can keep your identity. You can keep your data even if the buzz platform goes away. So it's it's quite
67. 9:56powerful in that way. You can collaborate with other people. You can bring in um compute resources from
68. 10:03people in your community. So there's a lot of things about it that are quite powerful. Is this is where AI is heading is less
69. 10:12chatbot and more workspace. I think that where AI is going is that AI is going to be everywhere in every
70. 10:23form factor. I think that buzz is one particular kind of thing that I'm personally excited about where you can
71. 10:31get multiple trusted people together kind of like you, you know, you do in a business like you would do in a Slack
72. 10:37channel um and you can collaborate on things and then as a group you can leverage lots of different AI platforms.
73. 10:45You can le leverage cloud code or codecs or your own models running locally or or just any combination of th those things
74. 10:54and you can easily create custom agents. But that's just one way to work, right? I think that that voice interfaces are
75. 11:02going to be huge. In fact, they just added that to Buzz. You're just going to be able to talk to your AI and have it
76. 11:06go do things. I think we're going to see form factors where we've got AI that just lives in physical devices. Um,
77. 11:15robots are coming. Even though the US just said, "We're going to ban any Chinese robots." At some point, they're
78. 11:21going to probably let that go. Maybe they won't. Who knows? Doesn't matter. US companies will start making robots.
79. 11:26Um, AI is going to be in physical form. It's going to be in chat bots. It's going to be in search engines. It's
80. 11:33going to be on websites. It's going to just be everywhere. So, I don't think that that any particular flavor of technological interface
81. 11:45uh is is going to be is is going to be the one. I think there's going to be lots and lots of them. Um we are on
82. 11:52Vimeo and not Streamyard. Well, Mary, [laughter] we are I am in Streamyard, but the
83. 12:07video stream is streaming through Vimeo because Streamyard bought Vimeo and they are feverishly integrating it. So,
84. 12:15that's what's going on. Kyle Brandon, I can still listen to the stream on my phone when I get a text. Oh, that's
85. 12:24good. Look, positive news. The the the Vimeo acquisition led to something positive. [laughter] That's awesome.
86. 12:34Okay, so this is good. We're I would say we're at about 7030 right now of this is awful and things are better. [laughter]
87. 12:42So, so, so let's see like Okay, let's see. Buzz looks like digital gods. Buzz Buzz looks really cool. have it on my
88. 12:49Mac Mini M2, but connecting with others is still a mystery to me. Only have had it for one day. So, apparently, if
89. 12:57someone sends you their public key, there's a way that you can invite them, but I haven't figured out how to make
90. 13:02that work. I can create an invite for someone and send it to them. And if they click on it quickly, it seems to work,
91. 13:10but if they wait too long, it times out. It's said it was supposed to be three days, it seems like it's more like three
92. 13:15hours. Um, I had no clue. Hey, remedial class here. [laughter] Um, it's not the worst experience. Worst
93. 13:23experience. It is working. So, a plus. Beautiful. Okay. Um, yeah, buzz looks interesting, but it is it is clearly very early software. It's
94. 13:35very early software. It's like 0.53 or something like that. So, you know, about the time a software gets to a 1.0
95. 13:43release, it'll probably be six months from now. um and uh and and it it it will be easier to set up, but right now it's it's just it's
96. 13:55kind of complicated. So, with that, okay. Um so, if anybody has any questions, again, I don't know if you could all just let
97. 14:06me know how many folks are here. Keep a keep a a tally and and uh if if new people come in, I will welcome them. uh
98. 14:13if if we jump up, you know, if all of a sudden 10 more people show up, just just let me know. Um if you have any questions about AI, if
99. 14:22you have any questions about what's going on, just pop them in the comments and I will do but my best to answer
100. 14:26them. We had a couple of comment a couple of questions come in earlier. So, let me let me pull another one up.
101. 14:40Butimus. All right. If AI agents can join a project like teammates, this is back to Buzz. How do we know know what to assign
102. 14:51the AI versus what still needs a human? That's a good question. Um, well, I think I think where the we're the the human's
103. 15:07job in in in a future that is increasingly AI amplified. The human's job is going to be to have ideas like what do we want to do? Why do we
104. 15:20want to do it? Like understand who they are, what they value. It's going to be the human's job to go, you know, this is a major thing I want
105. 15:32to work on. This is a minor thing. This is a thing I just don't care about. The agents, you give it to the agents to go do the
106. 15:40work. I need research done. I want something written. I want this. I want that. I want that. Um, you might have
107. 15:46humans write some things because AI writing sucks right now, right? So, so it's the human's job to have the
108. 15:54taste. It's it's the human's job to have the idea. It's the human's job to understand how and when to deploy the
109. 16:03agents. And then it's the human's job to go, is the thing that was created worth anything or is it a big old pile of poo? And if
110. 16:12it's a big old pile of poo, throw it out and go yell at the agent and say, "That was bad. Do another. Do another. Do
111. 16:18another. Thank you, sir. May I have another?" Um, Patty, I like this platform better. Very cool. All right.
112. 16:25Wait, 21 peeps now. Cool. Hover your mouse in the upper leftand corner to see the count on a PC. Let me try it on the
113. 16:33upper right hand corner. Wait, what's this? Should get No, not there. Yeah, that the the right right now the
114. 16:45viewers of this show have more access to what's [laughter] going on with the show than the people on this side of the camera. [laughter]
115. 16:57Human equals boss tells tells it what to do. AI do it do it if they can tell you what part the human has to do. Wait, AI
116. 17:08do it if they can tell you what part the human has to do. That's about it. Okay. What's this? I'm here, but I can't chat
117. 17:16correctly on my phone. Okay. So, for those of you that are new, um, service update, technology update,
118. 17:25StreamYard acquired Vimeo. Streamyard is now forcing this all channels to use Vimeo streaming technology and not their old streaming
119. 17:37technology. So, this is a new platform. Well, it's the old Streamyard platform, but it's a new streaming service going
120. 17:44through Vimeo. So, the features have changed. Um, it seems like it's better in some ways, worse in some ways. Um,
121. 17:50bear with us. Share any ideas you have. Brandon will we'll take those on and try to do things about it. I can still chat,
122. 17:57but I can't pop it out. It's a tad different on a mobile phone. This is much better not being able to use my phone
123. 18:06while something Wait, where did that go? while LLS is in. I don't know what that means. Hey, Lori. Charles Adams, I
124. 18:16reloaded. The population count is persistent now. Um, I know all that. Let's see. I know all that and have been
125. 18:23trying for 17 minutes. [laughter] It's hard to manage from your phone. Okay. It's Oh, it's hard to message from
126. 18:31your phone. Okay. All right. Well, we'll figure this out here. Here's here here's a thing I don't need to say, but
127. 18:37especially if you're new here, what's amazing about the people that show up to this show on a regular basis,
128. 18:44the irregulars, is that they will help you out as they figure out how to use this. They'll tell each other and
129. 18:50they'll tell you. So, if you have any questions about how to do this stuff, ask in the chat and and everyone's
130. 18:55really helpful. So, let me grab another question. If anyone has any questions, it's it's uh ask me anything. So, if you
131. 19:00have any questions about AI, just pop them in the comments and I will get to them as they come up. Um, wait, sorry, the text box drops while
132. 19:14trying to write from the phone. Can't see. Try it. Try if you're on a phone turning it horizontal or vertical. I
133. 19:21wonder if that helps because I know sometimes sometimes interfaces get wonky and and clearly it's from according to
134. 19:29Brandon they're they're sort of building this in real time as we talk. Okay. Um for someone brand new to AI
135. 19:38should they start by learning prompts trying agents or just use AI and tools they already know? Um honestly I I
136. 19:52You can learn prompts. I don't think that learning prompts unless you're going to be developing AI software or or AI enabled software. I
137. 20:03don't think you need to learn prompts. I think what you need to learn is to ask for what you want and have a
138. 20:12conversation. I think the biggest thing in learning AI is that what Google has taught us for the past 20 years is don't
139. 20:20put very much in the chat box. What we've learned with AI with chat GBT is it needs much more context about
140. 20:29whatever you're talking about. So if you say write me a business plan, it will write you a business plan and it will be
141. 20:36horrible and generic and you'll see like that's a shitty business plan. But if you say, "Hey, I'm thinking about this
142. 20:42kind of business and I like these kind of customers and I live in this kind of town and I want to charge these kind of prices and I
143. 20:49you can literally just talk. In fact, you can you can just go into chat GPT. Now, there's a little icon. I'll show it
144. 20:55to you. Hang on. Hold please." You go into the chat GPT and you see that little purple icon there? Your
145. 21:04color may be different, but see those little lines? It looks like a speaker. You can just do this. Hey Quinn, how's it going tonight?
146. 21:14>> Hey, I'm good. How are you? What's going on tonight? >> I'm just here with the AI learn lab and we're talking about, you know, AI and do
147. 21:22you need to learn prompting and how do you how do you do this AI stuff? And I said, you just talk to it, right? >> Right.
148. 21:28>> Yeah, pretty much. Right. >> Hey, I have a question for you. Tell me what the weather's going to be like this weekend in Denver. Okay,
149. 21:38checking mail. >> Yeah, thanks. You're the best. You know that >> looks hot and mostly sunny. Saturday is around 96 and sunny.
150. 21:54>> 102 on Sunday. Quinn, what are you doing to me? Fix the weather. >> Yeah, not I I know. >> Quinn. Quinn. Quinn. Quinn,
151. 22:07how did I get back to Quinn voice? Because Okay, so here's here's the thing. There's a bug in chat GPT right
152. 22:14now where if I go to my if I go to my voice settings, it's set for soul so which is Quinn. If I go into um
153. 22:26hang on a sec. If I go into remote, which is how I control codecs, and I put it in voice mode. Hang on. There's voice mode.
154. 22:49It's connecting to my my MacBook. Or is it? Maybe my MacBook's down. Anyway, this is a dude and I can't there's not a way to fix that. So, I
155. 23:01don't know if they fixed that yet. It doesn't look happy. Anyway, um so the question um the best way to learn this stuff is is
156. 23:16just start talking to it. Just learn to ask for what you want. Um, okay. Let me go grab another question. Question question.
157. 23:36Okay. Are openweight models becoming the people's version of AI or are governments and big labs eventually
158. 23:45going to lock them down? Hold please. Champy's warning. I will back.
159. 24:11Alrighty. Alrighty. Alrighty. All righty. So, here's what's happening. expect typos. Can't see my words. So,
160. 24:20asking [laughter] for understanding. All right. The people on mobile are struggling through through learning how
161. 24:25to comment. Uh, I appreciate all of your work. Uh, your hard work is is good. Who is asking Kyle those questions from
162. 24:32where? Are they AI generated questions? Danielle, how dare you? Do you think we would generate AI generated questions? [laughter]
163. 24:48Give me a good question, Danielle. Um, all right. The openweight model thing is really interesting. Here's what's going on.
164. 25:01China's strategy from day one has been to distill the frontier model. So, so, so here's what we've learned.
165. 25:13Open AI and Anthropic and Grock to a lesser degree. And Google, they're being quiet. They're just off in the corner
166. 25:19playing with their tools. They're spending billions and billions of dollars to make these models. And then they put these models out in the
167. 25:28world. And then the Chinese companies distill them. So they basically ask those models a bunch of questions, take
168. 25:35the answers and and reverse engineer the model and then they create other models and then they dump those lowerpriced
169. 25:42models as open-source models back to the US market. So you've got Anthropics Fable at $50 per million tokens. If you
170. 25:53if you think of you know this is a unit of AI compute this unit of AI compute from anthropics fable and mythos models is about 50
171. 26:05bucks and over here with Kim K3 it's about I don't know a dollar [laughter] right so or $25 cheaper um what's happening right now is that
172. 26:21anthropic and open AI I are [sighs] voicing their concerns about
173. 26:38how dangerous their own models are. And they're saying that the thing that's going to make it more dangerous is if
174. 26:46all these Chinese models come in and steal from them and compete with them and do that. And dear government,
175. 26:53you should you should these models are so dangerous. You should protect the American people by not releasing our
176. 27:02models without you reviewing them first. And then any other models, you should just shut them down. You should make
177. 27:09them. Chinese models illegal. Open weight models are horrible. And there's now an openweight safety coalition. Um,
178. 27:21and they're doing this to basically try to get the government to prevent anyone from competing to them, which is cynical
179. 27:28and capitalistic, and I I I don't like it. I don't approve. I understand it. I understand why they're doing it, but
180. 27:35they're kind of getting in bed with the government. It's like, why would you want to do that? Anyway, whatever. I
181. 27:40guess we've outgrown the boxes. We We've kind of outgrown the boxes, although we can go back to the boxes. All right. to
182. 27:46understand weights. So this is from digital gods LDG. I have word count envy now to understand weights. Okay. So when when you talk
183. 27:54about openw weight models so when when we get a a a chat GBT model it's a closed weight model meaning we get to
184. 28:03play with the model but we don't get to see how it was made and tweak it and be able to to do anything we want with it.
185. 28:12To understand weights, think of an AI model as a giant network of billions of light dimmer. During training, the
186. 28:19developer adjusts the dimmer until AI gives correct answers. Okay, cool. That makes sense. An openw weight release
187. 28:27means the developer gives you the exact blueprint of every single where every single dimmer is set. That's cool.
188. 28:36What you get the what you get the brain file. You download actual parameter files often gigabytes in size. Local
189. 28:45control. You can run the AI completely offline on your own computer versus servers. Customization. You can do some other things. That's a
190. 28:56monopoly and I thought not legal. Okay. So yes. So well it's a duopoly. therefore illegal. This is this So there's like this is I'm
191. 29:10I'm I'm a little steamy. Not not just because it's hot in Denver, but I'm a little steamy right now because OpenAI
192. 29:18and Anthrop Anthropic are clearly the leaders. Google I don't know what's going on with Google. I think Google is
193. 29:27farther ahead than they're letting on, but they're just kind of sitting in the background and letting these two do
194. 29:33whatever they're going to do. Um, both of them have offered the US government 5% stakes in their company. Both of them have basically been
195. 29:49traveling back and forth to Washington lots. Um, and both of them are now sort of fermenting fear and saying AI is
196. 29:58dangerous. And you know what's really dangerous? Chinese AI. [laughter] And so so I agree that what I think that what
197. 30:10they're doing, you know, borders on um unscrupulous business if not criminality, but we live in a we we we
198. 30:24currently have a very transactional government, right? So if you want to if you want to play, you pay. And that's
199. 30:30just kind of how it is. It's depressing. Um but that's what it is. I also I also believe long term this will just feel
200. 30:37like a ripple. Like I I don't know that this is a major inflection point. It's just annoying right now. And uh we'll
201. 30:45see if Elon Musk comes out with models that actually shake things up. We'll see if some non-Chinese company comes out
202. 30:52with models. It'd be interesting to see if anyone in Europe is doing anything. Um haven't seen much. There's there's
203. 30:58Mistrel out of France, but they've been kind of quiet for the past six months. They they made noise with one model and
204. 31:06then they they kind of went away. Um, so we'll see. We'll absolutely see. All right. Fant. Speaking of Kim, Kim K3. All right.
205. 31:22Kim K3 got a lot of buzz last week. Is it actually a big deal or just another new model that beats benchmarks headline? Well,
206. 31:30when a new model comes out that beats ben beats benchmarks, that just means things are progressing forward. Um,
207. 31:37Brent Peterson today in office hours said that his company and they use a lot of AI. He said he said he he he burned
208. 31:4819 billion tokens on Tuesday. So they are a they are a lab that is out there doing a bunch of this stuff and he said
209. 31:57that K3 is really quite powerful and they're using it to replace uh most of their use of anthropic. Um so Kimmy K3
210. 32:06is a big deal if if you have you know major players like Brent Peterson using it. It's a big deal. Um and you know
211. 32:14saying that it's good. Um it's a big deal. It is a big big deal. Um, I have a question for you. So, for those
212. 32:24of you who can can chat, if you if you can't chat, the chat's down below the uh below the thing here, below the video.
213. 32:32You can pop the chat out. I'd like to know from you, what is your go-to AI tool right now? Are you using one of the agentic
214. 32:43platforms? Is it just plain chat GBT plain clawed? Anyone still using Perplexity? Who's using what? I'd like
215. 32:51to just get a sense from you all of what you guys are using.
216. 33:08Buler, do you have any advice, do I have any advice for people learning to use Open Claw? Yes, I do. ChatGBT.
217. 33:25Chat GBT is still my go-to. Yep. I would say I would say for one-off things, it's Chat GBT. Claude, Lori Blair, Claude, Perplexity,
218. 33:37and Co-work mostly. Interesting. I'm currently 100% Codeex. I'm I'm probably 50/50 Codeex and Chat GBT right now. Occasionally, Claude
219. 33:53chat because of memory. Paying for Claude, but probably will not later. Love chat GBT. Oh, wait. Cindy I'm living inside
220. 34:03Notion, switching models as needed and midjourney. I That's kind of cool, Cindy. I That you're still using Notion is is powerful
221. 34:13because I never I never cracked the I never got over the hump of that one. Um still paying for Genesis Plus.
222. 34:23What is Genesis Plus? [laughter] Perplexity because of resources. Interesting. Um, do I have any advice for people
223. 34:35learning to use OpenClaw? Yes. So, whether you're using Codeex or Claude Co-work or Open Claw or Hermes or now Buzz,
224. 34:51the most important thing, well, there's a couple of them. One is dedicate a machine to running it and and
225. 35:02just do like a clean install of the operating system and then install OpenClaw or Hermes or whatever um what's it called? harness
226. 35:15on that local machine. And then also look up um agentic memory or or Andre Karpathy did a thing on on your local
227. 35:28brain like the brain for agentic systems for agentic harnesses. Um, you want to have your your local harness understand
228. 35:37who you are, understand your projects, and basically write all those to a lo local local to your local hard drive.
229. 35:44Um, and have it on a separate computer than all your main work. Because one of the things that's happening is that
230. 35:52these agents as they're getting stronger and more powerful, they also gain access to be able to write files to your local
231. 35:59computer, to your actual hard drive. And there have been reports of things like open uh OpenAI's uh chat GPT 5.6 erasing hard drives.
232. 36:15So put it on a machine that if it erases the hard drive, it's just erasing that hard drive. Now the other thing that you
233. 36:24can do is you can set up a GitHub repository. So on if you go to github.com you can create an account and then you
234. 36:32basically create a folder and then you tell openclaw or Hermes or codeex hey whenever you make changes to these local
235. 36:42files also make changes to GitHub that becomes your backup. So if in the event that your AI just destroys
236. 36:52everything you've got a backup for it. All right, let's see. Uh uh uh uh uh uh uh uh uh uh uh uh uh uh uh uh uh. Kelly Camp, sorry,
237. 37:04I'm on mobile and I can't see what I'm typing and can't correct [laughter] all. Okay, so here's the general thing
238. 37:11in the chat. People with typos are not just they haven't had too much coffee. They actually can't see their
239. 37:17letters. Um, if if you're new here, uh, this stream is using software called Streamyard. They bought Vimeo and
240. 37:27they've basically forced us to use new streaming technology, Vimeo. Um, and it's changed how things work. So, if you
241. 37:36see those comments, that's that's what those are about. Um, let's see. Used cloud on desktop mostly. Is that
242. 37:42something you can glean from chat GPT history? Yeah. So, so you can Mary, one of the things that you can do is
243. 37:57okay, let me go do it. Let me go do it because I This is something that's decently easy to explain, but it's not going to make
244. 38:07any sense. So, [laughter] so let's let's just let's just go do this. Okay, I'm going to hide that.
245. 38:23Move this over here. Okay, so I can see you all. Perfect. I have it on its own rig. Okay, good, Dina. So, yes, then you're then
246. 38:36you're doing good. Um, the other thing that you can do, Dena, is go to Grock. So, I I'll actually do it right now. So,
247. 38:44I'll do I'll do some of this stuff right now. So Grock, go to Grock gr.com and in Grock um ask it um what
248. 38:59are the latest best practices for um open claw
249. 39:15setup. up and um safety, we'll just call it that. Um distill it into
250. 39:30top 10 best practices in order of how I should set them up. The reason to
251. 39:45use Grock is that what Grock is doing is it's searching um X. You could also use something like Deep Research on um Chat
252. 39:55GBT or Claude or Perplexity to go search Reddit and do the same thing. Reddit and X are kind of the the the two places
253. 40:03where people are talking incessantly about the nitty-gritty details of the technology. So doing doing this search
254. 40:12on X through Grock will kind of give you the latest thing and then basically just follow that. Um we'll we'll see what it comes up with
255. 40:21here. Distilling the top 10 openclaw setup and safety practices like it went to docs.openclaw.ai. Lenny's newsletter. Openclaw the
256. 40:35complete guide to building. Okay, here's the 10 best practices. Isolate the environment. There you go. That's what
257. 40:41we just talked about. Use full fully dedicated identities and proper secrets handling. So, uh, for example, my open
258. 40:49claw and my codeex have has its own email. It's got its own Gmail email. Um, install cleanly via official paths.
259. 40:57That's a big one. Um, harden the gateway immediately. Bind off and network. So basically, when you're when you're
260. 41:05hooking up things like um uh chat GBT authorization to be able to use your 20 buck a month subscription. So anyway,
261. 41:13just go do this and then follow these instructions. Practice strict memory hygiene, add observability, human in the
262. 41:20loop, controlled autonomy, operational maturity, practical sequence setup. So just with these this how I started with
263. 41:29open claw was just I said give me the most conservative setup possible and then I just basically followed that and
264. 41:35then I ended up creating a really shitty agent out of it but that's how you learn you just learn by by going okay so now
265. 41:42the question is can you um can you use your chat GPT history to create the markdown files needed to create your own your own brain, right?
266. 41:58So, here's what I'm going to do. I'm in chat GPT. And so, it it knows all about me. And so, I think I'll I'll do this as a
267. 42:07microphone so you don't have to listen to me type. Okay. So, here's what I want you to do. I want you to go um find Andre Karpathy's
268. 42:17um agent brain post that he did or or you know an article or two about best practices for creating an agent brain
269. 42:28and I want you to tell me what are the most important files that I need to create and then I want you to create um
270. 42:37those files as markdown files that I can download based on everything you know about me. I want you to create
271. 42:46something where the system knows about me, something where it knows about my projects, and then something where um we
272. 42:54give it a role in how it's going to support me. Okay, so just tuck that in there and then bang. Now, this is going to go off
273. 43:05and do nothing because something went wrong. [laughter] Okay, let me go let me go copy this prompt. Reload the page and dang it. [laughter]
274. 43:18Oh, lordy lordy lordy. Okay, so chat GPT is off doing that. So, we'll come back to this in a bit. Uh, I'll find
275. 43:25Karpathy's originally po original post and a few strong primary sources. I'll distill the structure, then build a
276. 43:32personalized markdown agent brain. So, basically what this doing is it's taking what it what what it knows about me.
277. 43:39It's taking the the basically the inventor of this approach for how to use like a markdown file wiki to become your
278. 43:49you know your virtual you. Um and then it's going to you know create those files. So I found the original. Okay.
279. 43:59Let's see question isolated system. Yes. But when it links Google did I just unisolate it? Okay. There's two
280. 44:08different things there. That's a good question. Um, okay. This is a little bit apples and oranges, but it's a really good question. So, let
281. 44:21me I'm going to bring this back so I can I can look at look at my my hands moving. Okay. Um, we need boxes. Danielle, the boxes are
282. 44:34back. Okay. >> [laughter] >> Okay, let's see. All right, we we'll use this box. We'll use this Jurass Jurassic Park box. Okay,
283. 44:48this is your computer and you put Open Claw on it and then you're going to create these markdown files that are your brain. And
284. 44:59it's it's it's things like the way these agentic systems work is when they start up they're completely dumb. They don't
285. 45:05know anything. So, the first thing that they do is they go, "Hey, uh, who am I?" And it goes and it looks at the the um
286. 45:14the the file that's called uh I forget what it's called. It doesn't matter. Agent agent.md and and agent.md tells it who it is. And
287. 45:24then it goes and looks at the user file and that's me. And it it says, "Okay, I work for Kyle and here's how I act." And
288. 45:30then there's memory files which are like here's all the projects Kyle's working on. So, the first thing it does is it
289. 45:36goes and it figures out who it is and who I am and what we're working on and what we did yesterday or what we did an hour ago.
290. 45:44All of those files, all of that history, all of that memory is living here. Now, how it understands how to read those files.
291. 45:58So, so you type in a thing, you know, hey, you know, what's the latest update on my projects and it goes uh and it
292. 46:06goes and looks at these files and then what it does is it takes the the what it sees in those files and it sends that
293. 46:15to like Google Gemini or to Open AI or to Anthropic to to their GPT brain or their claude brain and it sends some of this context
294. 46:27up there with your question with your prompt and it says hey what's the latest status and then chatbt or cla or
295. 46:34whatever sends all that all the answers back and then this thing updates the files and it tells you okay the latest
296. 46:43on this project is this the latest on this project is this so while you're going out to do that compute out there
297. 46:50the actual retention of that is is happening here now here's what's really important If the if the information that you're
298. 47:00gathering is like what's the latest on my super sensitive medical problem, it's going to go look at your files
299. 47:09that's got all your medical records in it and it's going to send it to Open AI and then your medical records are now in
300. 47:16OpenAI's thing. And depending on how you've got your settings configured and depending on which way the wind blows,
301. 47:24OpenAI might go, "Hey, we're going to use that data today." Then your data is out there being used and trained and and things like that. Or
302. 47:34maybe you're trying to start a business that's got really proprietary information in it and you don't want
303. 47:38that to be in the next model that comes out. Well, if you're using the outside um processing service, the outside brain,
304. 47:48you just lost control of that. There's another thing you can do with an agentic platform, which is you run a local LLM,
305. 47:57and this is this is really important. Here's my computer. It's got all my storage in it. This is where my files are stored.
306. 48:05This this computer has some amount of RAM. Okay, you can get a Mac Mini with 16 gigabytes of RAM for like 800 bucks.
307. 48:15And everyone's like, "Oh, you can run OpenClaw for 800 bucks." But with 16 gigabytes of RAM, you can't run a very
308. 48:22powerful model locally. You have to use these outside services. So, the only thing the Mac Mini is doing is basically
309. 48:29the orchestration between writing things to my local files and then going and doing all the compute up at Chat GPT. But but
310. 48:40let's say you get yourself one of these boxes with 64 GB of RAM or 128 GB of RAM. Then you have enough RAM on this
311. 48:50machine that you can get a model like Quen 3.6 that is designed to run on a local computer and you can in in RAM
312. 48:59store a whole large language model. So now when you wake it up, it's still looking down here for the files, but
313. 49:05instead of going out to chat GBT or OpenAI to do the compute, it's all happening inside the RAM of your computer,
314. 49:13which seems bonkers. It doesn't make sense to me that I can download a single file or a series of files, whatever, that I can download
315. 49:25a package that I can stick on the hard drive here that it loads into memory here that contains all of the knowledge of humanity.
316. 49:38And then I can have that intelligence run locally without connection to the internet. But that exists today. That's
317. 49:44what I'm running over on my I've got a 64 gigabyte Mac that runs fast, has enough memory to run Quen 3.6.
318. 49:57And so I can ask any sort of questions I want without having to go out there. And then part of the art of running something
319. 50:05like um OpenClaw or Hermes or Buzz for example is there are sometimes when you want to have the real big compute of
320. 50:16clawed and open AI. There's other things where you don't need it at all and and and like getting smart about when you
321. 50:23process out there and when you process here is really important. But this idea that any Mac Mini can run um OpenClaw is
322. 50:32true except if you've got a skinny amount of RAM, you can't run very powerful models over here. So that's the
323. 50:38difference and and where this is going. Like right now I I bought a 64 gigabyte MacBook Pro. I'm kicking myself that I
324. 50:48didn't get a 128 gigabyte one. If you want to run the really big models, you need um a five 512 gigabytes.
325. 50:58And if you want to run something like GLM 5.2, which is a 1.5 trillion parameter model, that's like almost as
326. 51:07good as the Opus models. Then you need 1.5 terabytes of RAM, which is a lot. Um, so but just understanding that there's
327. 51:23compute in the cloud and there's compute on your local machine and if you start to invest in the hardware in your local
328. 51:29machine, you don't have to use the cloud. All right, make sense. It's a lot. I know it's an it's a lot. Okay,
329. 51:35let me see what other questions we have here. What exactly is good memory hygiene? What are those best practices? So, good
330. 51:44memory hygiene is things like Okay, well, let's let's go look at um let's go look at chat GPT for a second.
331. 51:58What I found, I believe what you meant is Andre Karpathy's LLM knowledge bases LLM wiki concept from April 2026. Okay.
332. 52:07A schema file, usually agents.mmd or claude.mmd defines how the agent operates. Um, the most important files, agent,
333. 52:17agents.mmd, who the agent is, user.md, who Kyle is, projects.md, what Kyle is building. I added four supporting
334. 52:27styles. Working style how to write research design challenge collaborate memory durable confined decisions and
335. 52:36preferences memory.mmd index.mmd navigation and routing and log MD append only history of meaningful changes
336. 52:43whatever so these files these are text files and you can go look at them one of the things that is tempting when
337. 52:54you start building these systems is that you're like, "Oh, I can go get everything I've ever written and
338. 53:00everything I've ever done and every blog post I've ever done and I can just dump them into this, you know, into this
339. 53:06memory." The problem with that is the way large language models work, if if you make these these files like huge and messy,
340. 53:16the quality of your work is going to get worse and worse and worse because it it it will basically lose context. And
341. 53:23it'll also be very expensive because you'll be sending massive amounts of context, you know, out to OpenAI or out
342. 53:29to Claude. So, good memory hygiene is being smart about what gets included and what doesn't. So, like right here it
343. 53:36says projects, you know, here's what Kyle is building. Storyvine, AI salon, AI learning lab, the great repurpose,
344. 53:43TGR Academy, content evolution, book speaking, port portfolio level priorities. Um, I think those are right,
345. 53:51but I might even want to say, "Oh, this one's not active right now." So, let's let's just clean that stuff up. So, you
346. 53:56can always change this. So, good memory hygiene is that. Um, let's see if this did it. I It probably should. Yeah,
347. 54:02look. So, here's um here's the complete package. Download the Kyle agent um brain.zip. And then it's got these files for me. So, agents.md
348. 54:15here you go. So agent agents tells openclaw or Hermes or codeex what it is mission you are Kyle Shannon's AI
349. 54:28chief of staff strategy partner creative director researcher editor and builder now you could read that and you could go
350. 54:37well that's not what I wanted to be I wanted to be you know someone who argues with me and someone who pushes back on
351. 54:44things right I I want an agent that that treats me like crap. So, you get to decide what's written in this. This is
352. 54:50just what what it recommended, but here's your role. Your chief of staff and what that means. You're a creative
353. 54:55partner. Here's what that means. Um, your default response pattern. Lead with an answer, recommendation, or the
354. 55:01strongest idea. Make the structure immediately visible. I like seeing things in like, you know, bulleted lists
355. 55:07and things like that. So, that's in here and it knows that about me. when helping Kyle choose a direction.
356. 55:14Does it advance human agency rather than mere automation? Right? That's big for me personally. So, you can write these
357. 55:21these documents manually if you want to, but because you've got all this history um in chat GPT, you can do it. So,
358. 55:30here's projects. This is another these are just text files. Here's storyvine. Here's my role. Here's the core promise
359. 55:36of storyvine. What differentiates us? Here's our strategic themes. Here's the here's how the agent is going to support
360. 55:42us. Here's what's going on with the AI salon. So, it's just, you know, what these files are are files that tell your your agents
361. 55:56what's the score, man. Wait, I can't move this file all of a sudden. Uh oh.
362. 56:11Okay, there we go. Um, makes sense. I love the box. The pizza box is smarter. MD finally found the chat. Hey, Cam Ken.
363. 56:24Oh, listen. Cam, sorry. Listen, everybody. Everybody is feeling the pain tonight. So Streamyard acquired Vimeo
364. 56:32and so Vimeo is now the the only way that we can do these streams through Streamyard is through Vimeo which
365. 56:40changed the chat. It changed everything. So if you're having trouble finding things, my apologies because you're not
366. 56:47the only one. We will get through this. U and apparently they're they are working to improve things like literally
367. 56:55as we speak. Uh that's what producer Brandon told us earlier. Okay. No laboxamies anymore. All right. I don't know why, but I can
368. 57:08now see what I'm typing. Been waiting two years to use it. What about locally on my own server? I don't quite know what this question
369. 57:20is. Well, what what about what about running a local model on my cloud server or ver virtual whatever secure network
370. 57:29whatever that is. Um, yes. So, the thing about having Okay, so let's go back to boxes.
371. 57:43Let's see. Okay, this is good. Okay, here's your laptop, right? Your laptop can has as much RAM as it has and it can do
372. 57:56what it can do and it's got as much storage as it has and it's generally kind of small. If you want to run a
373. 58:02larger model or if you want lots and lots of storage, you can get what's called, you know, a a a virtual machine or a virtual network
374. 58:12where where up in the cloud is a box, but all the contents of that box only you have access to. Chat GPT
375. 58:25doesn't, no one else does. and in this box. So, let's say let's say your laptop has 16 GB of RAM, but you're like, "Hey,
376. 58:33I want to run a big model." And instead of buying a new $7,000 128 gig laptop, I'm just going to pay 200 bucks a month
377. 58:44for a 340 gig virtual machine in the cloud. And then, so yes, Chef Kelly, you could in install local models there. You could
378. 58:55have your laptop be where you store your local files and do all the orchestration and then all your compute happens here
379. 59:03and then this box is chatgpt or claude or anthropic you know all these other servers and instead of having to go do
380. 59:11the compute over there you can do your compute and your local storage in the cloud. So yes, you can do that. It's a
381. 59:18more complicated setup, but it can be a lot more flexible and it is similar. It's secure because you own that virtual machine.
382. 59:27All right. [laughter] Okay. Is 1.5 terabytes even possible on a home computer? Cam Katkin, you are going to love this. So, Apple,
383. 59:43um, Tim Cook, who took over from Steve Jobs when when Steve was was on the way out. Tim Cook took over and he's been the CEO since,
384. 59:53he's out of here in September. And the new CEO comes from the hardware division. And the rumor is that the M7
385. 1:00:02chips, so right now you can buy M5 chips. The next set of chips that are coming out are the M6s. Apparently, the rumor of the M7 chips is
386. 1:00:13that they are configurable up to 1.5 terabytes on a single computer chip that will sit in a laptop or a desktop PC, a
387. 1:00:24Mac. So, yes, probably within I think they said by 2028. So, not next year, but probably the they'll probably announce it. When do
388. 1:00:35they announce their hardware? In September, right? So, in September of 27, a year from a year from now, they'll
389. 1:00:43probably announce the M7 that'll be available in the spring. And so, yeah, apparently 1.5 terabytes. And then if you had 1.5 terabytes on a
390. 1:00:55on a chip, you could run some of the biggest openweight models of today. So So things that are are really getting
391. 1:01:04quite close to the performance of the frontier model companies. So that's coming. Um 1.5 terabytes of RAM or VRAM. So the reason
392. 1:01:15Macintoshes are so popular is they use what's called unified memory. So the the 1.5 terabytes is available to both GPUs
393. 1:01:25and CPUs. And so you can basically all of that RAM is available to do all the AI inference stuff. When you're on a PC,
394. 1:01:33you have a CPU and then you have graphic cards that have VRAM in them. So if you're on a PC, that's a different
395. 1:01:40story. But but the fact that Apple is going here and they're going for high-end inference machines means that
396. 1:01:47other people will follow. So I would say probably within two years Cam we're going to be able to run like two things
397. 1:01:54are happening. The models are getting more powerful. They're also getting more efficient. So you're going to be able to
398. 1:02:00run more powerful models on increasingly smaller machines. But then the high-end machines will be able to run incredibly powerful AI. So
399. 1:02:11that's that's like two years out. Okay. I will add this to the community chat. It will be clickable. Brock Masaric, AI
400. 1:02:22for non- techies. Voice 2 plus codeex is insane. I agree with that. Hat tip from my friend Cass Phelps who I mentioned is
401. 1:02:31very excited about chat GPT work on desktop. Is doing this word his WordPress website in local files.
402. 1:02:37Endless possibilities. Oh, and there's a link. Great. Awesome. Thank you for that. Okay. Oh, the new chat makes it
403. 1:02:44clickable. Oh, that's cool. So apparently that URL is click clickable. That's great. Okay. Digital gods cam you can cluster max
404. 1:02:54with new protocols and aggregate memory. So yes, but it costs a fortune. And also there's another thing you can do. And
405. 1:03:01digital gods, I don't know if you know about this within buzz, the new slack-l like thing. You can use a thing called mesh dot mesh llm
406. 1:03:13which allows you to like I could take my 64 gigs and put it in a mesh and someone else could take their 128 gigs and put
407. 1:03:20it in a mesh. We could all be contributing our compute to a a single virtual uh machine. It's apparently slow
408. 1:03:28but that would allow multiple people to run a larger model than anyone could individually. So I don't know if that's
409. 1:03:35any good, but I know that's possible. Okay. Oh god, I got so much work to do. I know. So, here's the thing, Dina, and
410. 1:03:44for every for everyone, this is really important. I've been talking a lot about Buzz and how excited about it I am and all that
411. 1:03:53sort of It is incredibly complicated just architecturally because so, so think of it like this. ChatGpt. You log into a website, you put
412. 1:04:04in a prompt, you get an answer. That's really simple. codeex where you can save these files locally which is good
413. 1:04:12because chatb is simple but open AAI owns all of my chats I don't own them codecs or co-work or uh uh chat GPT work
414. 1:04:25all of those memory files are now stored locally that's good except I have to set those up and I have to set up the the
415. 1:04:32memory structure and I have to know how to do things like I just did with chat GPT about creating user files and agent files and all that
416. 1:04:41Um, and and I have to set all that up before it can be any good. I have to set all that stuff up and then I have to
417. 1:04:48turn on plugins and I have to give it access to computer use and I can and so it's much more powerful but it's much
418. 1:04:54more complicated. And then the next thing up this Buzz thing, I can connect multiple things like codecs sitting
419. 1:05:02underneath Buzz. So Buzz can can have agents that call codecs and claude co-work and my own local models and
420. 1:05:11digital gods compute and it can use all these things. All of that setup is incredibly complicated. So as you go
421. 1:05:18from just prompting chat GPT to agentic harnesses to multi-person multi- aent multi-harness factories like Buzz
422. 1:05:33powerful but incredibly complicated. Here's my strong strong recommendation. Do not focus on the harness. And like
423. 1:05:42like this was a mistake I made when I when I set up codeex. There's this button on the side that says plugins and
424. 1:05:50you can go and there's all sorts of plugins. There's plugins for Canva and for this and for Excel and for just
425. 1:05:57everything everything you can imagine computer use and this and that. So what I did is I went to plugins and I'm like
426. 1:06:03more is better and I just turned on all the plugins. Well, I don't know how to use the plugins. I haven't thought about how
427. 1:06:11like what so what I would strongly encourage you to do is if you're just starting out with this stuff, don't
428. 1:06:18start with what the technology is. Start with what are you trying to accomplish. Okay, I wanna I want to automate my
429. 1:06:29marketing because I've never been good at marketing and I don't post enough to social and I whatever it is and then research
430. 1:06:39what are the kinds of things you need to set up in codeex to be able to do that one task. And then that way you're turning on
431. 1:06:49plugins discreetly to be able to accomplish something or you're setting up instructions or agents or telling it
432. 1:06:56how to do stuff or automations to a purpose and it'll be much less frustrating because when you when you start with
433. 1:07:04something like Buzz and you just say, "Hey, the entire world can be connected to this thing and you start thinking
434. 1:07:10like, oh god, what do I connect and what do I not?" It's just completely overwhelming. All right, I got lost for a while. I kept looking
435. 1:07:17for you on Tik Tok and didn't read my emails. It's on me. I'm glad you found us again. Okay, let's see. We got lots
436. 1:07:25and lots of comments in here. This is great. I have five terabytes on my hard drives, but only 16 gigabytes
437. 1:07:31of RAM. So, you are meaning RAM correct? Yes, trying to run GLM5. No, you won't be able to run GLM5 on 16 gigabytes of
438. 1:07:40RAM. Here's the way this works. Um, okay.
439. 1:07:58Box metaphors. Yes. We got We got a Jeff. Oh, we got a Jeff Gold Bloom diarama. [laughter]
440. 1:08:18>> Okay, we got a Jeff Gold Bloom diarama. And I got this box [laughter] and then we got this box. Okay, this is
441. 1:08:29your hard drive. So, your hard drive stores files. Um, think of this literally like a book on your library shelf, right? like a like
442. 1:08:40one of like these notebooks over here, right? You can put files in here and if you want to learn from what's in here,
443. 1:08:47you pull them off the shelf and you go look at them. That's your your five terabytes of hard drive is like 5 terabytes of books of of files.
444. 1:08:59Um, this is your RAM. Your RAM is your brain, right? So you pull a book off the shelf, you you read some stuff, you put
445. 1:09:11it into your brain, and your brain has 16 gigabytes of RAM. So you have a you have a small brain. You have this is
446. 1:09:18this is your brain right now. So it doesn't hold very much. You can't put very many of the books into the brain.
447. 1:09:24And the other thing that needs to live in that brain is your operating system and any other apps you're running. So
448. 1:09:30you're running Mac OS, you're running um LM Studio, right? And then so maybe that takes up 8 gigabytes of just just to have your
449. 1:09:41computer running is 8 12 10 gigabytes of the 16 gigabytes of RAM. So you only have 4 to 8 available. GLM
450. 1:09:555.2 I think is a is a it's north of a terabyte in size. So like your five it would fit on your five terabyte drive. It would take up a
451. 1:10:10big chunk of this but it would fit on here but your your brain can't put it all in there. Okay. Okay. Now
452. 1:10:20if you have let let's say I I forget how big GLM 5.2 is. It might be 7 80 gigabytes. Let's say that you've got a 700 Let's
453. 1:10:32say you've got an 800 gigabyte RAM and you have a 750 gigabyte um model. You can technically you can technically run [laughter]
454. 1:10:49the the uh the model inside your RAM except it's all full up. There's no what's called headroom. What headroom is is
455. 1:11:01um I don't have another box, but whatever. Okay, I got this box. Will this fit in here? Not quite. Let's say
456. 1:11:07so let's say this fit in there. This fits in here. Let's say this is 800 gigabytes and this thing is 400 gigabytes.
457. 1:11:16Now I have 400 gigabytes of headroom. This is your context window, right? So everything talks about it's got a
458. 1:11:26million token context window or 200,000 context window. All of that context window has to be able to be processed in
459. 1:11:34your RAM above the the model. So when you run a um an AI model locally, you need to have lots of RAM and then you
460. 1:11:43need to have lots of RAM on top of the size of whatever model you want to run. Um to I think it was Silverf Fox's question. No, I I forget
461. 1:11:56who it was. Can you run or Chef Kelly's I think. Can you run this on a virtual machine in the cloud? if you have a
462. 1:12:02virtual machine in the cloud big enough, you could run GLM 5.2, but um Brent Peterson looked into that and to get a virtual machine in the
463. 1:12:14cloud big enough to to run GLM 5.2, he said it was about $200 an hour to run. $200 an hour to run. So, so when we get to pay $20 a month
464. 1:12:29and have these big frontier models that are like 3 to 10 trillion parameters and we get that for 20 bucks a month,
465. 1:12:39it's costing those labs anywhere between 8 and $20,000 for our stupid $20 a month subscription or a $200 a month subscription.
466. 1:12:53But investing in local compute, you know, big enough to run models, I think is actually really important because at the point at which the
467. 1:13:06government says they're going to they're going to monitor everything we do on those models or they just tell those
468. 1:13:14companies, hey, just shut off access to anyone who mentions the word orange. and you have a whole bunch of prompts in
469. 1:13:22there about oranges and all of a sudden you don't have access to your account, you're screwed, right? So that's why to
470. 1:13:28invest in local boxes. [laughter] Good lord. Did that make sense? Oh man, when you say almost as good as a
471. 1:13:41frontier model, what are we giving up when we use Hermes or OpenClaw? Well, so here's the thing, Chef Kelly.
472. 1:13:49If we're doing things like summarization and categorization and and just like writing emails and you know creating web
473. 1:14:00pages and things like that, relatively shitty models that you can run locally are actually more than enough for what most
474. 1:14:14of us need. They will run slower and they will be not quite as good. But but for most for like 90% of what we do,
475. 1:14:24probably 95 for most of us, you don't need the frontier model stuff. Cam Katkin, who's doing biology and DNA sequencing kind of stuff,
476. 1:14:38she needs models that are pushing the edge because she's pushing the edge of science. So, if you're doing things that
477. 1:14:43are pushing the edge of math or science or or require massive amounts of compute, you know, you're trying to come up with
478. 1:14:52cold fusion on a desk or, you know, you're trying to prove Tesla's theories and bring them to life, all that sort of
479. 1:14:59stuff, you're going to need all that that massive amount of compute. And if you're working on projects at that
480. 1:15:05scale, you've likely got money to invest in paying for big cycles and big things. Most of us aren't going to need that. I
481. 1:15:13would say we were we are safely in a place with with um probably on the low end a 32 gigabyte machine. I've got a 64 gig. I wish I had
482. 1:15:25a 128 gig. I feel like 128 gigs. The the the new AMD AI computers are 128 gigs. The Nvidia Spark is 128 gigs. Like that
483. 1:15:37seems to be the size that a lot of these models get get, you know, shrunk down to to be able to run on. But anything sort of 32, 64, 128
484. 1:15:47gigabyte of of unified RAM of usable RAM um will allow you to run powerful models and and like like surprisingly more
485. 1:15:58powerful than you would think. Um so so yes, you're giving them up, but like for most of us, I mean, I'm not gonna
486. 1:16:08I ain't going to tap them out. [laughter] I can't tap out what I'm using. Um, that's a good question, though. All
487. 1:16:16right, let's see. I was working on using Obsidian for that, giving the order to put in wiki link on each file in the
488. 1:16:26heartbeat to try to get this set up correct. That's my hope. Anyway, here's the thing about Obsidian. Obsidian is really more for us.
489. 1:16:38Your agent doesn't need it. But if you want to easily more easily see all of the relationships that the AI agent is
490. 1:16:49connecting together. So when it's writing all of these MD files within them, it's pointing to all the other
491. 1:16:55things in the other MD files. It's making all these connections. What Obsidian allows you to allows the agent to do is reveal all those
492. 1:17:05connections. And so it makes it easier for you to be able to navigate all the stuff that's in your brain. But for my
493. 1:17:12brain, I look at an obsidian graph and it just makes my head hurt. I'm like, wait, every one of these things I click
494. 1:17:19on is a whole markdown file and I just it just it makes me like sweat and just get like nauseous. And so I just turned
495. 1:17:28off obsidian because I'm not going to go look at it. And I can say to my machine, hey, what projects am I working on? and
496. 1:17:36it'll go figure it out and it'll tell me. So, um, but there's there's probably other ways to use Notion and Obsidian and those things
497. 1:17:46that I haven't tapped the surface on. It just doesn't work for my brain. I don't give a I I've got an AI
498. 1:17:52intelligence that can tell me what's going on. So, I'm going to let its brain do what it does better than my brain, which is make sense of
499. 1:18:02All right. I'm loving all the boxes. I know it's a we're we're back to boxes. This channel started with boxes. We're back. We're back, baby.
500. 1:18:12A gold. Everyone has a gold bloom diarama. [laughter] What if my brain is just a floppy disc? Remember that? I've got Where is it? Oh, this.
501. 1:18:29This is a CyQuest drive. This is a 45 megabyte hard drive platter in a in a cartridge
502. 1:18:45that you would plug into a machine and you could have up to 45 megabytes of information. We that that platter right
503. 1:18:55there holds the Urban Desires web server. [laughter] Oh man, [laughter] how we doing timewise? All right, I'm
504. 1:19:08trying to catch up on comments. You guys have a lot of questions. Okay, a second. That Tik Tok was better than
505. 1:19:15YouTube because you can just tap a comment to reply. Headless means running something without a visual interface. Very nice. You're
506. 1:19:24helping each other. I like that. All right. Oh, antiques. [laughter] Oh my god, you just explained obsidian
507. 1:19:32in a whole new way for me. I get it now. Oh, good. Yeah. Like, like this is the thing. Like, everyone's like, "You got
508. 1:19:39to have obsidian. You got to have obsidian." And like the more I'm looking at it, I'm like, "Wait, what is this?"
509. 1:19:44And it's like, yeah, like it is it's an interface for us. Here's the deal, people. the shittiest the shittiest local large
510. 1:19:57language model you could run even on your 16 gigabyte machine can make connections and think about relationships of information way better
511. 1:20:07than you can. I promise you now we as humans might have intuition and things that we can't we can do way
512. 1:20:16better than computers but like when it comes to things like categor categorization and relationships these
513. 1:20:24LLMs it's it's it's what they are made of they are made of relationships of thousanddimensional mathematical space of semantic meaning
514. 1:20:38a large language model is nothing but interconnections of fragments of knowledge because they were designed
515. 1:20:48architecturally to mimic our brain. They're not as sophisticated, but that's that's what their core design
516. 1:20:56is. And so something like Obsidian just allows them to reveal what they're thinking to us. So we can look at it and
517. 1:21:04go, "Oh yes, I see it now." M. Aha. That's fascinating. I don't know. I'm I don't I have surrendered. I have surrendered that
518. 1:21:18these things are smarter than me. And you can choose to keep fighting the battle. I'm smarter than AI. I'm smarter than AI.
519. 1:21:28Keep fighting it. Because here's the thing. At some point, you're going to get exhausted cuz they go 24/7 at unimaginable speed. [laughter]
520. 1:21:44It's It's quite liberating when you're just like, "Yep, the old wetw wear ain't what it used to be." [laughter]
521. 1:21:56formerly Lori, I had fun with Obsidian because of its visual, but I forgot about it after our FIRST DATE. THAT'S
522. 1:22:04EXACTLY what I did. I was I I told um I I had created my my memory files and then I had run read enough posts on X
523. 1:22:13that were like, "You got to use Obsidian. You got to use Obsidian." So, I'm like, "Okay, I'll never set it up, but an agent could
524. 1:22:21set it up." So, I said, "Hey, Codeex, could you go ahead and set up Obsidian for me?" He was like, "Sure." I
525. 1:22:26downloaded it, installed it. It set it up. It started writing all these things. I opened it. I looked at all of my
526. 1:22:33projects interconnected. I zoomed in on two or three of them and I was like, "Okay, [laughter] that was the last time I opened it."
527. 1:22:43[laughter] Okay, Dina, this is huge. LOL. So, I need to give up my control issues and let it do its thing for sure.
528. 1:22:55I think that one of the core existential crises of why people are freaked out about AI is exactly that.
529. 1:23:05If you talk to a film editor and you're like, "Hey, AI can make movies now." They'll watch it and they'll see like
530. 1:23:13one little spoke of one little wheel didn't rotate at quite the right speed and they're like, "Aha, see, see, it's
531. 1:23:22not as good as me." because I would get in there and I would Yeah. Well, that's why it takes you six months to make a
532. 1:23:27wheel roll, right? That human control thing about And listen, I'm not knocking anyone who loves craft. Um, but
533. 1:23:42these things allow us to actually give up that level of control where we have to pay attention to every switch and
534. 1:23:58everything. We can choose to not do that because here's my philosophy on this. In any given day, I've got some amount
535. 1:24:13of hours, some finite amount of hours where my brain is in its in in sort of genius mode, right? Where it's creative,
536. 1:24:25where it's I can problem solve, I can think about things, I can focus on things. For me, sometimes that's 30
537. 1:24:31minutes a day, sometimes that's four hours a day, occasionally it'll be like 10 hours a day, but rarely.
538. 1:24:41What I've experienced in my life is if I'm using that creative p power to problem solve little nuggly details like you know
539. 1:24:54formatting the outline. Oh, I should should these be capital letters or lowercase I letters or and I'm I'm using
540. 1:25:01that creative brain power for stupid little that doesn't matter. And so what you can do if you can if you
541. 1:25:11can convince yourself that that control it's okay to give it to the machine because the machine's going to be really
542. 1:25:19good at that. And where it's not good at that you can go ah okay so the machine can do 90% of my control freak stuff but
543. 1:25:28this 10% it actually still sucks at it. So I'm just going to I'm going to take my brain and put it there. But then what
544. 1:25:33that does is it frees up that creative energy that you have for actual ideas and creative thinking and strategy.
545. 1:25:43So So yes, I I know you said that kind of, you know, laughingly, but I I think this is actually a core battle, a core
546. 1:25:50human battle, right? I've got to be in charge. I've got to make sure everything's perfect. But what if you
547. 1:25:58don't for some things? and just let it go do that And then the things you really care about that that give you
548. 1:26:06energy to to obsess on, then just obsess on those. And then all of your good thinking is going to the that you love.
549. 1:26:16All right. But yes, Chef Kelly, I I wouldn't say sometimes AI has zero common sense. I think all the time AI
550. 1:26:26has zero common sense. But let me tell you a lesson I learned about common sense. I was in agency.com.
551. 1:26:34The president of North America is a man named Aaron Sugarman. Very sweet man. Still friends with him. And something
552. 1:26:41was happening in the company. And I walked into Aaron's office and I'm like, "These people are driving me nuts. It's
553. 1:26:47common sense. Whatever they weren't doing was common sense." And Aaron's reply to me, I'll never forget it, was common sense isn't common.
554. 1:27:00Common sense is like the human human judgment and its taste and it's not common because common sense means it's a
555. 1:27:10person that's thinking critically about a thing and if they're thinking critically about it, then it would be
556. 1:27:17common for them to do it this way. But that's actually not common. Most people are not thinking critically.
557. 1:27:25So, so I would argue that AI never has common sense. And it is our job to go, wait a minute, does this actually make
558. 1:27:32sense? Is this actually any good? We're the intuition piece. We're the taste makers. We're the ones that says, is it
559. 1:27:39good or is it And what most people are doing is they're just pushing the stupid button and publishing the results without looking at it or
560. 1:27:48thinking about it. And that's AI slop, right? It's a great comment. It's a great comment. Great point.
561. 1:27:56I have a question. Honestly, I hear at least one person a day tell me they didn't know they could change reasoning
562. 1:28:03in chat GBT and on Claude and they have multiple models that are good at different things. They both have
563. 1:28:08multiple models. So, so let me let me share that with people, Gareth, because I think it's important. Um,
564. 1:28:22so in chat GPT and you the same things in Claude. I'm going to just do a new chat here. So I have a new thing. You
565. 1:28:29know, here's where you chat, right? As you put you put it in the chat hole and then here's where you can do things like
566. 1:28:34choose deep research or I want to do a web search or create an image. By the way, you don't have to choose these. If
567. 1:28:40you just tell it to create an image, it will, but whatever. But over here to the right is this little model picker. And
568. 1:28:46this thing has changed over the over the years. Um, in chat mode, which we're in right here, which is the more simple mode, you
569. 1:28:57basically have instant, medium, and high. And it's it's basically saying, you know, we're using chat GPT 5.6 Soul,
570. 1:29:04which is their their biggest, latest model. Um, so you can do medium or you can do high. Now, why would you choose
571. 1:29:10to go down to medium? Well, you might choose to go down to medium because you've realized that you're doing enough
572. 1:29:18prompting that you're burning through your credits too quickly. And sometimes you don't need, you know, sometimes you
573. 1:29:24just want a quick answer to something. Both of these medium and high modes use what's called reasoning. So when I when
574. 1:29:33I go in here and I say um what is the average speed of an unladen
575. 1:29:48swallow and I hit return. See where it says thinking right here. This is a reasoning model. It's reasoning. It is saying to
576. 1:29:57itself, okay, he asked about this, right? And look, look what it did here. What do you mean African or European?
577. 1:30:03That's the joke is from Monty Python and it understood that. But it had to think about that answer. I can switch this
578. 1:30:10thing. I can go to a new chat. Switch this thing to instant. And notice it drops me down to the 5.5 model. But this
579. 1:30:16is a non-thinking model. So um why is Monty Python swallow joke funny? I just get a little blinking thing and
580. 1:30:30then it just answers. It's just an instant in instant answer. So, it's not thinking. If I if I switch over from chat to work, this selection
581. 1:30:42matrix becomes even more complicated. So now I can choose between GPT 5.6, Soul, Terra, or Luna. Basically, Soul is the big one. This is
582. 1:30:54big bear, mama bear, baby bear, right? Soul is Papa Bear, Ter Terra's mama bear, and Luna's baby bear. Um, and then
583. 1:31:045.5 is the old model. So, these are the ones ones you're going to want to use. And then within each So, I can go I I
584. 1:31:11want to I don't need a lot of thinking here. I'm going to go to Terra instead of soul. And then where the effort is,
585. 1:31:17that's now light, medium, high, extra high, and max. and then I can do it faster or slower. Now, I personally hate like despise with
586. 1:31:29a fa a passion that I have to spend any of my brain thinking about this. This is not my job. Gareth because Gareth likes to think
587. 1:31:42about all the different ways to do things and he actually thinks sorry I'm speaking about you as if you're not here Gareth but he he thinks
588. 1:31:50in systems thinking where he can think okay over here I'm doing tasks that require a lower level of stuff and
589. 1:31:58therefore I'm going to choose one of these lower models I don't care if it's fast or slow I don't care if it uses
590. 1:32:04high thinking or not so I'm going to use the cheapest one but over here I've got some real problems solving to do. So,
591. 1:32:10I'm going to do that there. Gareth shared with us today on on office hours. I hope this is okay for me to share that
592. 1:32:20he's got two $200 a month subscriptions. One for chat GBT, one for anthropic. And because of Agentic Harnesses where
593. 1:32:29they're doing way more prompting than you are, he's burning through those his weekly limits regularly. And you're like, "That doesn't make
594. 1:32:38sense." It does when you start using these agentic things where you say, "Hey, I want you to go off and think for
595. 1:32:43six hours." And if the thing goes off and thinks for six hours at that most expensive setting, you're going to blow
596. 1:32:49through tokens. You know, you're going to blow through your usage limits like that. And so that's why I'm saying we're we're
597. 1:32:59living in this world now where you could just go to basic chat GBT and you could prompt it as a human or you can go all
598. 1:33:04the way up to these really complicated things with agent harnesses and multiple things and tool use and and all this and
599. 1:33:12you're going to spend more money up here. And and as chips get more scarce and compute gets more scarce, your
600. 1:33:20ability to process tokens or buy them from OpenAI or Enthropic, they're going to get more and more and more expensive.
601. 1:33:28They're being heavily subsidized right now. We're out of time. Um, great questions tonight. Let me see if
602. 1:33:36there's in Codex CLI. You also have Max and Ultra. Yeah, there you go. >> [laughter] >> They also don't know about the models.
603. 1:33:45All right, let's see. Arg frustrating. Settings is your first lesson I teach in my my cohorts. That's really cool. Good.
604. 1:33:57It's okay if you don't. My wife didn't even know that. [laughter] Um Okay, cool. Listen. Um,
605. 1:34:12couple of things. First of all, thanks for hanging out tonight. Um, as we figure out this Vimeo stuff, this
606. 1:34:17transition to Vimeo, you know, please keep giving us feedback. We will give that feedback to Streamyard. We've got a
607. 1:34:26decent chance of them actually addressing it because they're integrating this right now. Their teams are working on it actively. Um, so we
608. 1:34:34may end up with with, you know, a better user experience than uh than we currently have. We might have a
609. 1:34:43good one. Also, if you're new to the salon, please sign up. Like, put your email address, sign up, get an account,
610. 1:34:51you know, for the AI salon, and it'll walk you through like a sevenstep welcome process. Do all that. Do things
611. 1:34:56like introduce yourself. Also, if you haven't checked out the the AI salon mastermind, please go check that out.
612. 1:35:03It's the subscription area of the salon. Most of the salon is free. This show is free. Most of the salon is free. Um, the
613. 1:35:11mastermind is for folks that are really leveling up their game and for like it doesn't it's it is ridiculously
614. 1:35:18reasonable for what you get. Um, so so please go check that out if you haven't done it. I really appreciate you all
615. 1:35:25being here. Um, I love doing this. Thanks for being so engaged tonight. Um, thanks for being patient. Um, and have a great weekend. I
616. 1:35:37If you're all hot and bothered and excited about agentic harnesses, play with aic harnesses all weekend. If
617. 1:35:44you're not, go walk in the grass. Go sit on a bench. [laughter] Okay. Get in touch with Oh, I met I met
618. 1:35:53uh Teton Skier. I met Todd uh tonight. He was in Denver and and came to Catalyst and we talked for a while and
619. 1:36:00he was just talking about maybe I just need to go sit on a bench for a while. So if you need to go sit on a bench uh
620. 1:36:05this weekend, go sit on a bench. Just do it in the shade. Um all right everybody, thanks a lot and uh I really appreciate
621. 1:36:12everything. Cool. Peace out.