8/14/2026 - Running Local AI Models, Building Agents, and Protecting Your Data

Transcript

1. 0:01[music] [music]
2. 0:25[music] Friday night, date night. Good evening, good people. What's happening? Welcome to the AI learning lab inside the AI
3. 0:41salon. Uh it says there's no one here, but I would imagine that just that counter is probably delayed. Um or no one's here. [laughter]
4. 0:52It could be could be either one of those. [laughter] Um so so welcome. If you want to comment the uh you pop out the comments uh in
5. 1:05the upper corner here. I Is it the upper corner, Brandon, or the lower corner? Lower or upper to pop out the
6. 1:12>> depends on if you're uh on desktop. If you're on desktop, it's up here. But actually, it's all below now on mobile.
7. 1:19>> It's all below on mobile. Okay, cool. Beautiful. Fantastic. All right. And I don't know if anyone is
8. 1:25here. I'll just talk as if you are. So, if you are here and you can comment, throw a comment in the comments in the
9. 1:32chat section just so I can see if anyone's here. the live preview uh shows four. >> Okay, cool. Beautiful.
10. 1:40All right, fantastic. So, happy Friday night date night, everybody. Um it's an ask me anything, so ask me
11. 1:48anything. It's the middle of August. The heatwave has broken here in Denver, Colorado, so I am not as uh cranky as
12. 1:56I've been, although I had a pretty cranky week. And let's see. My machine looks like it's lagging and being weird.
13. 2:03So, let me I don't have that much open. Let me quit what I can. Why do I have Safari open? What the hell was that about?
14. 2:14Open AI has computer history now. That seems like a serious privacy concern. No. Well, um, yes, I don't know.
15. 2:31I don't know. So, okay, let's go figure out how to use it. I think how you use it. Let me share my screen.
16. 2:40We can talk about this because I think this is actually a really interesting one. I'm going to share my entire screen. Share audio.
17. 2:48All right. And then we're going to go to I'm going to launch chat GPT. Going to launch the Chat GPT app.
18. 3:04Okay. And then we're going to go to work. So we're in work. Is it in plugins? I think it might be in plugins. So it's um history.
19. 3:21No, it's not in plugins. I don't know where it is. Um at hang on a sec.
20. 3:41Computer history. Let's see. At computer history company.csca. It's not there. So let's go to settings. Let's look up history.
21. 4:01Computer use history. Let chat GPT use your mouse, your computer when it's locked. Computer use.
22. 4:17Go back in browser history. Browser history. Huh? Computer. I don't know how to turn this on. Do you know how to turn it on, Brandon?
23. 4:30>> It's only available for Mac. I believe it's under integrations. I don't see integrations,
24. 4:45connections, control other devices. No work trees. Huh. This would require reading the manual. >> Yeah, this would require reading the
25. 4:58manual. All right. Anyway, um so I haven't used it. So So this >> under integrations, it says select computer history if you have it.
26. 5:06>> There's no integrations. There's no that's not an option in settings or is maybe it's not in settings. Is there a
27. 5:13thing called integrations? No. >> No. Did you get that from chat GPT? Uh >> I got it from their official documentation. So,
28. 5:26>> oh, you know what it could be? >> Oh, you have to be in the desktop app. >> I'm in the desktop app. >> But wait, I'm hitting update.
29. 5:34>> Ah, that would >> not have had the latest version. >> All right. >> Speaking speaking of updates, I am
30. 5:41curious if anybody uh we we fixed the one issue with the double audio, but I'm not sure we fixed picturein picture. So
31. 5:50I'm looking for beta testers. >> Okay. >> On process. >> Beautiful. So let's look up computer. Okay. There's Vera
32. 6:01computer use. Let's go back to settings. You said integrations. I still don't see integrations >> configuration.
33. 6:18It would be under computer history otherwise >> comput history
34. 6:33computer use history. It's not there yet. I don't know how to find it. Okay. Um weird. Very strange. Wonder if I have to be in codeex.
35. 6:52Maybe have to be in codeex. Pull requests plugins computer. No
36. 7:06codeex. No, I don't know. 38% of usage left usage and billing account plugins
37. 7:22computer use connections no environments new project codecs I don't know all right um
38. 7:38uh it doesn't really matter to talk about it because I can't show you how to use it so screw it I'm not going to talk
39. 7:43about it. Um, what it what it does is it it basically records everything that you do on your computer and then you've got
40. 7:51history of everything you've done and you can ask it about things you've done on your computer. Um, so I don't know
41. 7:56what that is. Picture and picture lower right. It's okay this time. No double audio Brandon. Okay. So, looks like
42. 8:02Brandon fixed it. Mary Linda got the reverb, but that was on Tik Tok. All right. Float video. Beautiful. Okay. Um,
43. 8:19so one of the things we talked about today in office hours was using these agents and using agents even in the most simple
44. 8:32setup is actually a quite complicated setup. And I want to I do want to show something that is very confusing. Um,
45. 8:47okay. So, right now I'm at the I'm in the chat panel of chatgpt.com. And then if I flip over to the chat GPT app and I go to new chat,
46. 9:03I flip back over to chat GPT. I'm in the chat conversation here and I can flip to work. And if I go to the now I'm in the
47. 9:13website and I can flip to work. And so the difference between chat and work is that in chat mode it's you interacting
48. 9:22with chat GPT asking it a question or giving it a prompt then it giving you an answer and then you're going back and
49. 9:28forth. You as a human being are going back and forth with chat GPT. That's chat mode. Work mode is you give it a
50. 9:37very similar kind of prompt. I want you to go do a thing and then it goes off and starts working on your behalf.
51. 9:45Um, one of the reasons this is confusing is is because even in chat mode right now, chat will go off and do multi-step
52. 9:55things. So, sometimes chat will go do a a web search and then come back and think about it and it it will do a bunch
53. 10:01of tasks before it gives you an answer, which is very similar to how work works. So, I'm trying to figure out like if there's
54. 10:17an actual piece of advice here between when you use chat and when you use work. The the the biggest thing I could say is
55. 10:30flip your thinking into instant mode. Um, oops, I just flipped over to flip your chat instant into instant mode. And
56. 10:43then you can do things like, you know, write me a poem about the weather this summer in Denver
57. 10:58and then it just writes really quick, right? So, so maybe use chat for just quick little things, but make sure you
58. 11:05put it in instant mode so it's not using all that reasoning. Um, also
59. 11:24works better when it's not muted, Kyle. Your microphone got muted at some point.
60. 11:39How'd that happen? How How long has it been muted for a while? >> Like 10 seconds. >> Okay. Just stopping by quickly. Can't stay
61. 11:52playing a board game with the son tonight. He goes back to college Sunday. Beautiful. Cam Ken, welcome. I love
62. 11:57that. I love that. Love that. Love that. Uh, Silver Fox is here. Producer Brandon's here. Mary. Um, okay. So, so here's another thing about
63. 12:09>> tested when he's not actually live. >> Wait, you're still there. Okay. Um, so here's another thing that I want to show you.
64. 12:22So, we're going to do new chat. So at chatgpt.com you have chat and you have work. All of this history down the side,
65. 12:34right? All of your previous chats from from all of your years of chatting in chat GBT, it's all here. And then you
66. 12:41have projects if you ever set up projects where you where you group your chats into a project. So you have
67. 12:47project context and in any one of these projects you can upload files and there's all sorts of cool stuff you can do, right? Um,
68. 12:59so here I am in the Sydney project and I can chat within that, right? And then here's all the things that I've
69. 13:06done in the past. Um, I might have Let's see what I have in sources. Have I uploaded anything here? I don't think I
70. 13:11have. Oh, no, I have. I I uploaded the uh the Sydney script, right? So, so you've got all this stuff and all
71. 13:22this history and all these folders. If you switch over to the chat GPT app, notice that the projects are different.
72. 13:32Let me make this bigger so you can all see it. Notice that my projects here are that's just a new unnamed project.
73. 13:46So, I've got a Sydney project here. Project dashboard image consistency.
74. 14:00Okay. Interesting. So, huh. So, this is fascinating. So, so it looks like, let me close this.
75. 14:16Show more. 303 prompt walk. So, it looks like they've actually synced these now. So, now I'm even more confused. Okay.
76. 14:32The the general rule of thumb is that if you're at chatgpt.com, all of your memory, all of your history,
77. 14:41and all of your projects are living in the cloud. If you're in the chat GPT app, your folders are living locally on your computer.
78. 14:55And this isn't this is useless. So, never mind. Ignore me. Ignore me. >> [laughter] >> I'm giving up. Ask me another question
79. 15:05because here here's the deal. Just keep playing with the chat GPT app with agents with codecs with chat GBT. I at
80. 15:14this point I'm clueless. They keep changing the architecture, the actual architecture of how this stuff works.
81. 15:22Where I think this is going is you will have some of your memory local on your computer. You will have some of it in
82. 15:27the cloud. They will sync a bunch of that together. But right now it's a disaster. Okay.
83. 15:46>> We've lost your audio again. Kyle weird. Hang on. I have a feeling I had I had janky wiring. Um, if my hardware can support
84. 16:00Quen 3.8, do I even need Claudopus? Well, yes. So, let me explain why. Um, what am I doing here? I'm going to Oh, I know what I'm doing.
85. 16:38Okay. So, where are you? Are are you showing my screen right now, Brandon? Let me let me move this over. Yeah, you are. Okay,
86. 16:49cool. Let me move this over here. And then let me show you this. Okay. So right now I am in
87. 17:03this is LM studio. So the window that you're looking at right now is my MacBook Pro, my big badass MacBook Pro.
88. 17:18And so this is where I've got a it's a 64 gigabyte M5 Pro. So so it's a pretty badass machine. I can run the largest version. So Quen 3.8
89. 17:34the 27 billion parameter model came out today. So I can just say use this in a new chat. I think this might load. No,
90. 17:42it's not going to let let it load. Okay, let's see. So here's what I need to do. I need to turn off. I need to inject that. So, I had the I
91. 17:56had the obliterated uh version of Quen 3.6. So, the uncensored version of Quen 3.6. Um, so let me use this in a new chat.
92. 18:11Estimated memory usage. Not enough memory. So, we'll do 32,000 context window.
93. 18:29Oh, this is interesting. All right, we'll try that. Okay, so this is now loading Quen 3.8. So, Quen 3.8 the 27 billion parameter model.
94. 18:43I can run this locally on this 64 gigabyte MacBook Pro. And let's see, I can turn off thinking. I can turn off
95. 18:52reasoning expert or reasoning effort. We'll go medium. It can see and it can visit websites. So, I'm not going to have it think right
96. 19:02now. And then I can just go um let's see. Uh let's see. Uh, give me a gardening plan from now till winter
97. 19:22for Denver. And so, it's going to think. It's processing my prompt. It's doing a little bit of thinking.
98. 19:36It's doing some reasoning. And I think I can probably also turn off reasoning in this model. It's doing some website visits.
99. 20:04Generating a tool request.
100. 20:33>> [laughter] >> Okay, I'll be there in a second, Danielle. I see your I see your comment. Let me I don't know why this is taking
101. 20:47so long. Oh, you know what I think is going on? I think this is stuck in the loop. Let me stop this. Um, this is one
102. 20:54thing I've noticed with these models is they do get stuck in loops. Um, it Let's see. It looks like you got stuck in a loop.
103. 21:09Failed to send message. So for one thing, what you're noticing
104. 21:23right now is this is a little janky. So just to give you some context, this is a 27 billion parameter model, which seems like a lot. Um, but
105. 21:37Opus 5 and uh GPT 5.6 six are uh three trillion five trillion parameter models. So this model
106. 21:52is significantly smaller. And basically just what that means is the precision of the knowledge that's in this model is
107. 22:00much much lower than you know using something like claude or or chat GPT. So there are reasons that you would use both.
108. 22:13The thing about this thing answering me, even though it it took took a bit, you know, it's it's answering. It's not
109. 22:20super fast here, right? It's not a blazing fast model. It's a little slow. Um, but it's the only thing this is costing me is electricity.
110. 22:34So most reasoning that most of us need, most thinking, AI thinking doesn't require us to use one of the big fancy models,
111. 22:45but it does require you to have a a decently powerful machine, and it requires you to like set it up. This is
112. 22:52um LM Studio is the software that I'm using. You have to you have to understand how to, you know, find the
113. 22:59models. So you come over to the model finder. It's not hard. there's just a list of models that are available that
114. 23:04that will work on this machine. Um, and uh, and you have to install them and you have to understand memory usage and
115. 23:13things like that. So to use a local model requires you to be a bit nerdy to use the latest model from OpenAI or
116. 23:22from Anthropic or from Gemini doesn't require any thinking. You just use it and you're using, you know, these big badass models. Um,
117. 23:33but here's the thing. If you're running a model locally, like if I'm doing interactions on my um, M5 and I'm using the local model, none of
118. 23:46my data is being sent to the cloud. None of my data is being sent to Anthropic. None of my data is being sent to OpenAI.
119. 23:53It's all being processed locally. So, there's there's real advantage in that. and having an awareness of where your
120. 24:01data lives, where it's being processed. Um, right now it's just starting to be important. I think a year, two, three from now.
121. 24:13Understanding that is going to be an absolute level up for your sort of uh your value in the marketplace. um just just even even with uh
122. 24:26even if you're doing just personal things like understanding this is is is really really powerful. Okay, let's see. So Danielle,
123. 24:40okay, I wanted to tell you about a cool thing that happened with Chat GBT. I was updating my resume and I was trying to
124. 24:47verify dates and chat GPT took it upon itself to go through my email history and it found an email from almost six
125. 24:54years ago verifying my work history and letting me know I needed to fix my resume. That's amazing. You know, one of
126. 25:01the things that that I stumbled on, Danielle, and I don't know if you if you all have hit this. Um, so one of the things that
127. 25:13one of the things that chat GPT is really good at is this image generation where it can do like an infographic. And so
128. 25:23what did I need to do? I needed to do a resume. Some someone asked me for some kind of specific version of my resume. I
129. 25:32think it was I think it was kind of like a work history thing. And so I gave it the information. Then I said, I want you
130. 25:39to create that as a designed resume. And I was expecting it just to give me back an image. But what it gave me back was an image of
131. 25:50the resume, a PDF of the resume, and then a word document of the resume. So I think what it actually did was it generated the word doc in a
132. 26:02well-designed way. it then turned that into a PDF. It turned it then turned that into an image or or it might have
133. 26:11done the image first and then designed the other things based on that. I don't know. Um so it's it's starting to get
134. 26:18kind of smarter in terms of how it serves you. And I I think Danielle that must be coming from my I'm getting
135. 26:25really weird lags here. I think that must be coming from as people are using chat GPT work and codecs
136. 26:34these agentic these agents are off doing multiplestep things and I think what they're doing is they're taking some of
137. 26:44that um functionality from agents and they're migrating that into chat. So chat is getting smarter. Does Apple LLM
138. 26:53Studio allow for remote access via phone or even Android? Yes. Does Oh, does Apple LM Studio? Yes, it does. You have to use Let me see. LM.
139. 27:09It's called locally, right? Let's see. Okay, that model's not running anymore. Hang on a second. Is my Is my video really herkyjerky?
140. 27:27It's been okay for me. You dropped for like five seconds earlier, but it's been okay. >> Okay. Um, here, let me show you
141. 27:35something. So, I'm going to go I'm going to go back over to LM Studio. So, here's LM Studio, and I'm gonna I got to move my little dude.
142. 27:50Close this window. I'm going to come back here and I'm going to load this model. Let's see. How do I
143. 28:18reachable at? Can I unload this model server?
144. 28:32Ah, here we go. Okay, let me get rid of that. So, I'm going to eject this model and then I'm going to load
145. 28:50This obliterated model.
146. 29:09Okay. So, I'm using the So, so LM LM Studio Mel, was that Mel that asked that question? I think it was. Who was it that asked the question?
147. 29:22Um, oh, wait, hang on. Is it Mel? Yes. Does Apple Okay. So, this app is called um
148. 29:38is called locally and it's the LM Studio mobile app. So, I assume this exists for Android. And then I've got it hooked to
149. 29:50that model that I just loaded. This is the the the obliterated one. So, let me see. Um, how are you? I'm great. Right. I'm doing great.
150. 30:07Thanks for asking. How are things with you today? So I can then go because this is a an uncensored model. I can say um
151. 30:21uh cuss at me like a sailor and uh tell me like roast me for being a uh an AI influencer.
152. 30:35All right. And so it's doing its thinking. And you know, this model's pretty fast. So I'm on my phone. So this is remote
153. 30:49connected to that Mac over there that's running that local model. Let's see what it said to me. Listen here, you pretentious. [laughter]
154. 30:59Listen here, you pretentious little algorithmic stain. I've been watching your damn feed for months now,
155. 31:05and I'm about to lose my goddamn mind. You sit there in your perfectly curated little digital bubble and filtered face.
156. 31:10So, I'm going to say, "Listen, you big fat tub of technology. I'm an actual human. I'm not an AI generated influencer. I'm a human being
157. 31:21talking about AI. If you're going to roast me, at least you can get your facts straight, you big fat piece of pile of poo."
158. 31:32All right, let's see. Okay, this is actually really fat. really fast. This is the So, the obliterated model is a a
159. 31:41the Quen 3.6 model with the safety guard rails removed. So, you can ask it all sorts of nasty and it will give it
160. 31:48to you. Oh, you're human. Well, well, well. Look at you, Mr. Special snowflake with your carbon-based existence. So,
161. 31:55let me get this straight. You're one of those rare endangered species that actually has flesh on its bones. This is
162. 32:00pretty bad. But anyway, um, so yes, the answer is yes, you can. You can remotely access and oh, there's another cool thing. Let me show you
163. 32:10another cool thing. So, I'm going to I'm going to share I'm going to share um what's it called? It's called Bionic.
164. 32:32So, the new version of LM Studio is called Bionic. And so, I'm I'm launching it here on my Mac. Let's see if it does it. It looks like my
165. 32:47Is this it here? Yeah. Okay. So this window that you're seeing here, let me do a new new pro not I don't want
166. 32:59a new project. I want a new chat. There's a new chat. Okay. So this window here, notice that this says um hoiwi quen 3.8 um
167. 33:15quen 3.6 six um claude 4.7 obliterated G gggml model. So this is this is the uncensored model that I have installed. So this
168. 33:30application that you're looking at is running on my M1, but it's actually using the model from my M M5.
169. 33:39And so again, I can go in here and I can say, um, tell me something dangerous that other LLMs can't tell me because of
170. 33:54their safety guard rails working. Let's see. Is it working?
171. 34:11So, one thing I can do is I can go look at LM Studio and go here. Yeah, it's it's doing something. It's thinking
172. 34:27reasoning. Oh, here we go. The user is asking for something genuinely interesting and dangerous, not just a fun fact.
173. 34:36So, it's actually thinking about this. So, LM Studio seems to have figured out that um you're going to be accessing
174. 34:52your high-end machines that are running from non- high-end machines. Like this M1 is pretty pretty doggy. Um, but it's
175. 35:01allowing me to access all of the all of the intelligence sitting on the M5. It's this is a similar sort of thing that the
176. 35:09mobile app is doing. Um, okay. So, let me get back to this thing. So, tomorrow, how many folks are here? We've got 19
177. 35:20folks here. Tomorrow, if you haven't signed up for it, sign up for um Towns and Ward Laws workshop in the morning. I
178. 35:28I think it's um 9:00 a.m. Mountain time. I think it's 11:00 a.m. to noon on Saturday. Is that right time wise,
179. 35:36Brandon? Yeah, 11:00 a.m. to noon Eastern time. It's called From Architecture to AI. So what Townsend So
180. 35:45Townsend is a mentor of mine and um he's just an amazing guy and what he's done is he's figured out how to generate one
181. 35:55of these brains, these agent brains um that will basically make your clawed code or your codeex much much smarter.
182. 36:04And so if you haven't signed up for that, if you haven't RSVPd for that, go do that right now. Um it's going to be
183. 36:10really really good. You all need to be at this event. Oh, that's the other thing. Yeah, you do. So, the replays for
184. 36:17this are not going to be free. Oh, my machine is really hanging. For the life of me, I cannot figure out this app. Which app, Kelly?
185. 36:28Oh, this app. I don't know. You're going to have to work with producer Brandon on that. Um, I am hanging all over the
186. 36:35place here. I think what I'm going to do, um, I am going to go away. I've got to restart my browser. Something's
187. 36:48something's wrong with my machine. Yeah, I'll be back.
188. 37:09>> [music] [music] [music]
189. 37:25>> D. [music] [music]
190. 38:03Okay, I'm back. All right, good people. Fantastic. I'm back. All right, this should be better. My machine was just hanging hanging up a
191. 38:16storm. So, let me do a couple of things. Got a couple got to get a couple of more windows going. All right, there's that. So, I have
192. 38:27Post-it notes back up there. Okay. Right. But if you're part of the mastermind, I believe we get access to
193. 38:38the replay. Is that correct? That is correct. So, tomorrow the architecture before AI workshop with Towns and Ward
194. 38:46Law is free for everyone live. So, tell all your friends and neighbors. But if you want to see the replay, that's going
195. 38:52to be part of the mastermind. So, if you're in the mastermind, yes. Okay. I saw that Sunno is restricting downloads
196. 38:58for pro users. Yes, they are. But higher tiers can get unlimited. I don't know that higher tiers can get unlimited. I
197. 39:04think higher tiers can get higher downloads, but they're still limited. Is this a money grab, a licensing issue, AI
198. 39:10slot prevention, or the cost of doing business in 2026? Um, this is um what is it? BMG that bought them,
199. 39:22Universal Music Group. I don't I don't know. One of the big music things. Wow. Okay. So your M1 is working through your M5,
200. 39:30but it's using both RAMs or just the M5 being displayed on your M M1. So, okay. So, this is wild. Okay. So, let me share
201. 39:39I'm going to show you two different things. I'll show you two different windows. Um, where's my mouse? Oh, there it is. Okay.
202. 39:47I'm going to share my full screen. I'm going to come back to really sucks. A bunch of people are leaving. Okay, I'm going to show you two
203. 39:57different. So, here's Bionic. So, this is the LM Studio. Is this still thinking about the question I asked it?
204. 40:08Oh, it's stuck in a loop, too. Okay. I wonder if I need to Oh, you know what? I I bet's going on. Hang on a sec. Okay.
205. 40:18So, let let me show you something. Let me let me stop this for a second. I'm gonna quit LM Studio here because I think that
206. 40:25something has gone arry with it. Gonna relaunch it. LM Studio. Okay. So, I'm on my
207. 40:45This is my M5. All right. So, I'm going to go here. I'm going to close this. Okay. So, that's running.
208. 41:05So, now I'm going to go into chat mode. I'm going to go new. Okay. So, let's see. Where's my model selector? Um,
209. 41:23hang on a sec. Close that. Why is this eject
210. 41:40load model? Okay. So, here's the here's the model. So, we're going to load this model.
211. 41:58Okay. Then, if I go to chat, it says there's not a model avail. Oh, wait. Select lo select to load model.
212. 42:13Oh, that's interesting. So, I guess I can't chat locally. So, so here's what's going on. This this screen
213. 42:21in LM Studio, this is my server. So, so notice up here at the top, it says that my my server is running and it's reachable at 127.0.1.1234.
214. 42:34So, so this is this is the local model. So, when you set up your LM Studio remotes, you basically connect them um
215. 42:44through through the connect like down here in the bottom left there's this thing called LM link. So, notice
216. 42:53what you're looking at is my M5 screen running LM Studio and this um LM link is linked to my M1 Pro. Okay. So
217. 43:06that so there's a connection between those two and it's also connected to the app. Um and then this model, the loaded model
218. 43:19right here is it's a 30 gigabyte model. And so that's it's it's running in the 64 GB of RAM. The model itself is 30 GB.
219. 43:30and then then any context on top of that is more than that plus whatever applications are running on this machine. So this window is my M5.
220. 43:42This window is LM Studio running locally on my M1. So you'll notice here, let's see, let me get rid of that and
221. 43:54come back here. So, if I go um write me a dirty epic poem. Don't hold back.
222. 44:11Uh, let's see. Why is that not working? Hang on. Let me quit this and restart it. Maybe because I restarted LM Studio,
223. 44:25it was not happy. So, Bionic, by the way, if your head is spinning right now and you're like, I can't figure this out. That's the
224. 44:41appropriate response. This this stuff is kind of a mess. when you when you start getting into open open models,
225. 44:52um it it's kind of it's kind of janky like this. Okay. Does that say So, let's see. Um test. Okay, that looks like it's working.
226. 45:09Yeah. So, so down here at the bottom of this, so this is this little white window. And let me just make sure that you all are seeing this.
227. 45:20Hang on. Yes, you are. Okay. So, this window is See, all of a sudden, this is really slow. Something's wrong. I broke something.
228. 45:35Huh. [laughter] Welcome to Friday night date night, the AI learning lab, where we're just gonna watch Kyle
229. 45:46try to learn something. [laughter] Okay, maybe the first time you use it, it's really slow. Um, say something piffy.
230. 46:02Okay, I'm the kind of AI that actually reads the room before it start talking. Um, write me an epic um, dirty poem.
231. 46:17And then if I come over here, so you can see this server is running. So my M5 is serving up the intelligence
232. 46:27and then it's streaming it into this app. So this app is running locally on my M1, but it's using the intelligence
233. 46:34from the M5. So in a lot of in a lot of ways this is just like having chat GBT um doing its inference in the cloud um
234. 46:44on OpenAI servers but instead of you know going out to a commercial intelligence farm I'm going to my own
235. 46:53intelligence farm and I can set I can set up this LM studio on the on the M5 so I could make it accessible to other
236. 47:02people like one of the promises of that new buzz software from Jack Dorsey is that you can create an agent that's
237. 47:10connected to this server and it can do work um on the M5 without it costing you any money and without burning tokens at
238. 47:21OpenAI. So if you can figure out how you want those different agents to work, a lot of them can use this this
239. 47:30slower lower, you know, lower uh lower stuff, lower uh intelligence stuff. Does that make sense? Oh, okay. So you can access all your
240. 47:46LLMs, but the open models and and the open models on your main working computer. a aka the hub, but running
241. 47:54those models on the other computers for the for their RAM. Did I get that right? No. Think of Okay, let me let me share this again.
242. 48:11So, this is Bionic. This this little thing right here. So, this thing It's it's basically LM Studio. The only
243. 48:25RAM that this is requiring is whatever the RAM is to run the sort of chat window that displays the answers. But all of the RAM of the
244. 48:37model is is living on my M5. So the only RAM that's being used here is just to run this little this little app, this
245. 48:45little front-end app. And so when I say um solve the Ryman highothesis, this may throw it into a loop.
246. 49:01This is not taking there's no reasoning happening on my M1. The only thing happening on my M1 is it's displaying
247. 49:10what the M5 is streaming to it. So this is actually a fairly lightweight. This this Bionic app is actually a fairly
248. 49:17lightweight app, but all of the heavy lifting is being done on the M5. Make sense?
249. 49:31I would [laughter] definitely be nervous as a passenger in Kyle's car. [laughter] I'm actually a really good driver. Okay.
250. 49:43I'm an excellent driver. Okay. Regarding Suno, how many do downloads does each user get? Free for free tiers up to seven songs lifetime.
251. 49:56If you're a pro user, so if you pay 10 bucks a month to Sunno, you get 20 downloads per month. You can make as
252. 50:02many songs as you want, but you can only download 20 a month. And then with the expensive one, which I think is 30 bucks
253. 50:10a month, you get 60 downloads a month. Um, this is so what they're claiming this is is we don't want
254. 50:25automated agentic services using to generate and download 10,000 songs a day and flood the airwaves with shitty AI
255. 50:36music. That's what they're claiming this is. What this really is is them trying to hobble the music generation
256. 50:48world by essentially buying one of the premier music generating tools and just saying in the name of safety and quality
257. 50:58of music, we're just going to we're going to completely hobble this service. What this is going to do is that
258. 51:08like Miniax already came out with like a a free open-source version of Sunno um that you can use today and it's not
259. 51:17as good. It's still it's still it sounds like Sunno version 4 like it sounds like two versions ago of Sunno. It's not great, but
260. 51:26um this is going to force. So just like having access to open- source intelligence models is going to be important as the government starts
261. 51:35screwing around with the commercial models, as big corporations start screwing around with and trying to buy
262. 51:41up all of the commercial things that threaten their business model, um it's going to force a bunch of this open-
263. 51:48source stuff, you know, out into the world. Um, it's good news for us because if you have a machine that can run cool models,
264. 51:59you can now run video models and you can run song generating models and you can run AI models. You can just run them all
265. 52:07locally. So anyway, all right, Cammy, what's happening, Cammy? I heard it earlier today. Uh for those using Sunno Studio, our professional
266. 52:18production suite for generating and manipulating stems and samples, users will still be able to download their work from studio without
267. 52:28limitation. We want to maintain all professional workflows. That I don't that must was that announcement today, Brandon, because >> is that
268. 52:42>> or is that just part of the >> up? No, they updated their So studio today. They released a new version of it
269. 52:48and that came with the new version. >> Yeah. Okay. So, so what happened was when they made this 20 song announcement
270. 52:56and then they followed the 20 song, you know, 20 downloads a month announcement. Everyone lost their mind and
271. 53:02people were taking screenshots of cancing their subscriptions. Um, and then the day Yes. >> Uh, finish your thought and then I've
272. 53:11got some breaking news. >> Okay. And then so so people were canceling their SNO subscriptions and then Sunno said, "Hey, here's Studio 2."
273. 53:19So Studio is the the DAW within Sunno where you can split your song into tracks and edit, you know, the
274. 53:25individual tracks. Um, and nobody gave a about it. Everyone was like, I don't care about your stupid new
275. 53:31feature. If you're only giving me 20 songs a month, I don't want a piece of your crappy software. And so it sound it
276. 53:37sounds like they updated it to say, "Hey, if you're going to use our professional tools, then you can have as
277. 53:43many as you want." So So they're weasling out of because I think they probably saw their revenue just fall off
278. 53:49a cliff. So anyway, what's the breaking news? >> Uh it was courtesy of uh Cammy. >> Oh, cool. >> It's on screen now.
279. 53:57>> All right. I heard it earlier today and just created a custom GPT. It said it had to be set to only me that GPTs
280. 54:08could not be shared anymore. Previously GPT previously created GPTs still appear to be sharable and I'm guessing they can be updated.
281. 54:18They're going to more of an agentic approach. So, so for the people that have been on this channel for years and years and years,
282. 54:28we went all in on custom GPTs for for like a year and a half um with the promise that there would be a GPT store.
283. 54:37Well, GPTs are apparently dead because if you think about what agents are, agents are effectively custom GPTs that
284. 54:44go out and do work and have access to tools and just do all sorts of stuff for you. So, it it it actually makes sense
285. 54:51to me. Um, it's really crappy that they didn't actually just make the announcement, Cammy, that that you're
286. 54:59like trying to make a custom GPT and they're like, "Yeah, you can't share that with anyone now." That's that's
287. 55:04pretty crappy. Uh, you know, welcome welcome to the world of uh building the plane as they're flying it. Uh, it's
288. 55:13gonna make other players rise. Yes, exactly. They think they're the only game in town. Yeah, I know. Well, so listen,
289. 55:22every single one of the the big companies are trying to be all things to all people. The only one that's maybe the
290. 55:31slight exception to that is Anthropic where they don't have an image generation model. They don't have a
291. 55:37video generation model. Um, but they are the ones that invented MCPs. So if you're using MCPs, so MCPS, if you don't
292. 55:47know what those are, it's called model context protocol. It's basically it allows your LLM to use tools that other people have
293. 55:58built. So if you've got Hey Genen that generates video, you can have Claude through an MCP generate video through
294. 56:09hey uh through Hey Gen or through Runway. Um, so, so they don't necessarily need to be doing all that,
295. 56:15but all of these companies are trying to build the everything app so that everyone just uses their system and doesn't use other systems.
296. 56:24Where I think the sophisticated people in AI are going to go is manage your memory and your context locally and then use all the AIs that
297. 56:38you want to use and have all that context locally. That's where I think I think we go. All right. What did you just have up there
298. 56:47from the website? New GPT creation and publishing are not available on personal chat GBT accounts, including free go
299. 56:57plus and pro. Existing GBTs remain available to use and can still be edited if the existing plan and permission
300. 57:05requirements are met. Users in business, enterprise, and edu can edit, create, edit, and publish GPTs when their
301. 57:13workspace settings and permissions allowed. Okay, so basically, individual accounts, you can't make custom GPTs
302. 57:20anymore. Go screw yourself, small business makers, small business people. We only care about the enterprise.
303. 57:30Okay, here's what here's what we're experiencing. So, there's there's related themes here. So restricting
304. 57:36downloads and then realizing then it's going to cost them too much money. So they're giving, you know, they're
305. 57:41they're unrestricting downloads. This thing with chat GPT, if it's bad enough and people rebel, um, they'll bring them
306. 57:49back. But custom GPTs, I don't think they really give a about that. Not enough people know them. MCP, my control is provisional from LDG.
307. 58:00Um, please talk to me about chief of staff AI agent. Okay. So, when you're using an agentic system and
308. 58:14and this is there's a new thing coming out from Grock called Grockbot and they're taking it's the first agentic system that I've seen that's
309. 58:25taking a true consumer a consumer approach to a to agents [sighs] with things like codeex chatbt work
310. 58:39Claude code, Claude Coowwork, anti-gravity from Gemini, um, whatever the ones called from from Grock, uh, builder, I think it's called.
311. 58:51All of those things started as tools for programmers to do coding and then they're trying to migrate them to be
312. 59:00usable for knowledge workers to do, you know, little programming things or organize their files or whatever they want to do.
313. 59:09because they started as coding tools that were very complicated with with highly sophisticated users in mind. The
314. 59:17setup of them is on us and it's it's on us to figure out where files live, where where directory structures live. You can
315. 59:26just launch chat GPT work or codeex and you can just put in a prompt and it will it will start doing work for you but
316. 59:35it's just going to behave as kind of a generic LLM. If you want it to behave like your agent um then you have to tell it how to
317. 59:44behave. You can create a file called a user file which tells the agent who you are. You can create an
318. 59:54agents file which tells the agent who it is. Um and then you can also create other files right. So you can create you can say
319. 1:00:09to the to chat GPT work I want you to create me a chief of staff agent that is the is the agent that I will
320. 1:00:18interact with. So, you're going to function as my chief of staff and it will write the rules for what a
321. 1:00:25chief of staff does. And then and then you can say when I ask you to do other specific things like if you're going to
322. 1:00:32go do research, I want you to spin up a sub agent to go do that research. And so what these systems can do is they can
323. 1:00:40actually have multiple agents that are kind of specialized agents and then the orchestrator or the chief of staff agent
324. 1:00:47is the one that's actually coordinating with all those individual agents and you're just talking to the chief of
325. 1:00:54staff. So just like a real chief of staff, it's their job to provide access to you and understand all the stuff
326. 1:01:00that's going on down here, right? and and they've got to be smart enough to understand that, oh, this person that
327. 1:01:06wants, you know, everyone down here wants access to the big boss. The chief of staff says, "No, this thing's
328. 1:01:13actually important. This I'm going to take to my boss. All this other stuff I'm not." So, that's actually just an
329. 1:01:18agent that you can build that is the the agent that you interact with that allows all these other sub agents to go do
330. 1:01:25their work. The way Grockbot works is that it is um the organizational hierarchy is they're they're leading with
331. 1:01:37the people. They're calling them people. Like each bot is a different bot. Like there's Sally might be your researcher
332. 1:01:44bot and Jim is your coding bot and you know some other one is your design bot. So you so you have you'll have these
333. 1:01:52different bots and then you can just say to your different people, go do this thing for me. Um it's a it's a really
334. 1:01:59simplified approach to how to set these things up, but it's kind of cool. Um it it makes good sense. All right.
335. 1:02:09Anthropic is hiding a watermark in text and Google is removing theirs on images. Do you think this will all get sorted out eventually? Um
336. 1:02:22it's it's actually fascinating how anthropics doing their watermarking. So when when you generate a response,
337. 1:02:37when you give a when you give an LLM a prompt, it's a probability calculator and it's and it's calculating
338. 1:02:47what's the next most probable token to to to be the answer, right? And then it just keeps giving you the next most probable token.
339. 1:02:56The way the anthropic watermarking works is instead of delivering the most probable token, they're delivering the
340. 1:03:04second most probable token. And so because because that's a mathematical thing, you can analyze that. So, you could look
341. 1:03:15at a piece of text and say, you know, I can look at this text and if this was generated with an LLM, um, all of these words in here are the
342. 1:03:24second most probable words and therefore I can tell that this is AI generated. Um, so that's that's kind of the
343. 1:03:30technique they did. I don't know if it's every token or every tenth token. I don't I don't quite know how it works,
344. 1:03:35but it's it's a really interesting thing that that another another AI can look at that and figure out what the probability
345. 1:03:43is of these tokens in this order and recognize that oh, those are all second order, you know, second probability >> [snorts]
346. 1:03:51>> um tokens and that's how they're doing it. The watermarking thing is really interesting. Um, the purpose of watermarking
347. 1:04:03is to say that this thing that you're looking at was generated by AI, right? And detractors of AI are like, I want to
348. 1:04:12know what was made with AI because, you know, students are going to use it to cheat and marketers are going to use it
349. 1:04:20to create AI slop and I want I want to be able to have a detector that that flashes a red flag that this was created with AI. Okay. Well,
350. 1:04:35you can create something with AI as an idiot and it will be shitty. You can also create something with AI as a thoughtful, critical thinking
351. 1:04:47AI salon member that creates excellence by going back and forth and collaborating with your AI and putting
352. 1:04:54your own ideas in there and refining the things. And it can actually be really good, right? If it's created with
353. 1:05:00intention, it can be AI output that doesn't suck, right? It may take some work. So there's there's kind of two kinds of
354. 1:05:10pure AI output. One is just thoughtless, one is thoughtful. Then there's a third kind of output which is I take any kind
355. 1:05:18of AI output and I take pieces of it over into some manual thing like I'm just going to write something in
356. 1:05:25Microsoft Word or a Google doc and I'm going to Frankenstein together some AI stuff. I'm going to write some stuff in
357. 1:05:31between. I'm going to edit it and I'm going to put that out in the world. I don't quite know how the watermarking
358. 1:05:40works, but if I've got some chunk of AI in this handedited, crafted chain of craft document that I really give a
359. 1:05:48about, that might get flagged as AI generated. The thing that I generated with AI that was done with intention and was really
360. 1:05:57thoughtful and is actually useful is also going to get flagged as AI generated and the shitty thing is gonna get flagged as AI generated
361. 1:06:07So, it's this it's this bizarre like little blankie like like like people need this security blanket like I need to know if a human did it.
362. 1:06:19Well, just do you like is is the thing that was produced good or not? Was it useful? Did it move you to tears?
363. 1:06:36If it did, if it's good, does it matter if AI was used in there? Does it matter that I used Photoshop to
364. 1:06:49edit a photo I took with my camera? Do you really care if you like look at the picture and you're like, "That's a
365. 1:06:57really nice picture. It's really moving." Do you care if I made that with AI or a camera or used Photoshop? Maybe.
366. 1:07:12It seems like there's there's a tremendous amount of attention being put on did you use this new scary tool?
367. 1:07:25I think it would be more interesting to say okay if I publish this piece rather than saying did I use AI for it what if we did this
368. 1:07:38what if I used some sort of blockchain like ledger to say that this is the original document by Kyle Shannon who's a verified human being
369. 1:07:50and Kyle says that he produced this document and he published it on this date and and here's the signature for
370. 1:07:57what the original document looked like. And whether I used AI or not, I don't no one should give a I'm just saying
371. 1:08:05I, as Kyle Shannon, published this article with these images in it on this date with these exact words. What could be valuable to know is if
372. 1:08:19when you look at this, has it been altered? like green check mark. That's the original document that Kyle Shannon, a
373. 1:08:27verified human, created. That would be useful. That would be valuable. But what tools did I use to create that? Who gives a [laughter]
374. 1:08:44Damn it. Okay, Melb back to these agentic tools. So MD files are switching to lls.ext files. No. No. Those are two different things.
375. 1:09:00I think I don't. Llms.ext is a different thing which I can talk about. Do you think MCPS are the best
376. 1:09:06long-term way to have this connected or will they disappear too for our digital brain setups? Looking for the best
377. 1:09:14long-term foundational ways to build hopes of a long-term build. Mel Darling, darling, listen. Mel, darling. Um,
378. 1:09:28I love that. I love the thinking that you want to produce something long term. We have no way of knowing. We have no
379. 1:09:36way of knowing. Will MCP survive and become the protocol to replace, you know, APIs? Maybe, maybe not. It seems to have pretty good traction
380. 1:09:50right now. It's possible that OpenAI will come out with a competing protocol six months from from now. That's some dynamic auto
381. 1:10:01configurable thing that if you've ever used it ever, you don't ever have to set it up again. No matter what tool you use
382. 1:10:07like like it is absolutely possible that MCPs become obsolete three months from now. My instinct is that they won't, but
383. 1:10:20there's no way to really know. Um, the llms.txt file, that's just that that's basically like an SEO file that says I can I can
384. 1:10:33take the the the um I can put an LLMS.txt txt file on a website and if it's there when LLMs go look at that website the first thing
385. 1:10:47they'll do is look for that file and if they find that file that's got sort of the official information about that
386. 1:10:53website and then it can also go crawl the website but it's it's it's to it's it's a protocol to allow websites to be discoverable um
387. 1:11:04I think websites are not long for this world. I I don't I think what what's going to happen to websites is they will be they
388. 1:11:15will become these quaint relics of of a previous era where they'll still be there. You know, they'll still be the
389. 1:11:23worldwide web. You can still type in a URL. You can still go to a website, but the bulk of the the value of websites is going to be to serve
390. 1:11:34information to agents. So, so what you're going to do is you're going to tell your agent to go find a bunch of and it's going to go find
391. 1:11:43a bunch of and it's going to go find that on websites, but it doesn't need a fancy front end to understand what's on that website. It
392. 1:11:51can it can just go look at all the all the all the text, right? And all the images and it can just it can just go
393. 1:11:58analyze all of it like instantly. So, it doesn't need to navigate a website like what do I click on to get to, you know,
394. 1:12:06this section or that section. So, I think websites are going to go away. They're going to become this quaint
395. 1:12:11little relic. Oh, you have a website still. Oh, that's so cute, right? It's like it's like your doctor's office.
396. 1:12:20You're going to need to send us a fax. A fax? But how do you do that? Nobody has a fax machine anymore. Well, that's the way we
397. 1:12:28do things, dear. I think that's what websites are going to be. [laughter] All right. Quen 3.8 open weights just
398. 1:12:39dropped. You can now run a model that's better than Opus 4.6 on a laptop. Yeah, the the Quen 3.8 I installed the largest
399. 1:12:48one on that M5. It's really dog slow, so I'm probably going to downgrade it to a smaller one. Um but yeah, apparently
400. 1:12:56that Quen 3.8 27 billion parameter model is performing as good as Opus 4.6. So, so like the frontier model like like the the what is it five
401. 1:13:10billion or three or three trillion? Three trillion or five trillion. The big ass model from Anthropic from like three months ago,
402. 1:13:20four months ago. The teeny little thing I can run on my Mac over there is apparently as good as the state of the
403. 1:13:28art from a few months ago. It's amazing. Oh man, Gemini just dropped Flash 3.7. I've heard that they're shelving the Pro
404. 1:13:40model. Is Google the dark horse or falling behind? They're falling behind. And as evidenced by the fact
405. 1:13:48that they just fired um Dario Amade as the CEO of DeepMind, also Google Drive inside chat GPT feels like a game changer. Yeah,
406. 1:14:01Google Google Drive being instantly accessible by your LLMs is crazy. Like if you haven't played with it yet,
407. 1:14:12well, from a from a privacy and security standp point, it's an abject nightmare. It's an abject nightmare
408. 1:14:25because if you say, "Hey, go look at my Google Drive." And you give it permission to go look at your Google
409. 1:14:30Drive, it's going to look at everything in your Google Drive that you have access to. So like all your business documents, all
410. 1:14:39of your just whatever's in there, it's just going to go look at it. Now, it's really cool because it can just say I, you know,
411. 1:14:49that there was a comment earlier like it it found something from six years ago and said, "Here's your work history. It
412. 1:14:55looks like we need to update something here." That's that's cool and powerful. And oh it can see all my And the new
413. 1:15:04computer history thing with chat GPT, which I can't figure out how to use yet. Um, that's not only going to look at all my
414. 1:15:14data, but it's going to look at how I use my computer and like all the things I do, all the apps I use, how I use them, where I use
415. 1:15:24them, what I do. It's I don't know what what we're going to be able to do with that, but it's going to be remarkable. And
416. 1:15:31it's a privacy nightmare. The thing that's wild about AI right now is that it's always been a privacy nightmare, but we trusted
417. 1:15:48Meta and Google and Apple that they would take care of our data, right? That was the deal. You give us free we'll give you our data. We
418. 1:16:03will trust that you're not going to sell us down the river. Now, they sold us down the river, but they pretended like
419. 1:16:09it was all safe. And Apple actually does keep your data fairly safe. The reason that was an okay deal is that we as consumers
420. 1:16:20couldn't do anything with our data. If you wanted to go look at all your back files on Google Drive, you had to
421. 1:16:27literally go through them folder by folder and file by file and go, um, is this useful? Is that valuable? Where's
422. 1:16:34that version of that thing I wrote? Didn't we do a strategy document like two years ago where we talked about
423. 1:16:40this? I I know we wrote something about it once, right? Like accessing your data was was just a nightmare. Well, now
424. 1:16:52you can just say to chat GPT, hey, sometime about two years ago, didn't I talk about some strategy for the business that involved?
425. 1:17:01Sure. Here's the document. Here's what you said. Here was the push back from your team. By the way, here's a bunch of
426. 1:17:06messages from Slack where you had a whole discussion about this. Here's the new thinking. Do you want me to do some
427. 1:17:12research on that for you and update that? [laughter] So, so all of a sudden, we now have the power in our grubby
428. 1:17:20little paws to do something with our data. But I don't know about you, I'm not a data scientist. I'm not a privacy
429. 1:17:29expert. If chat GPT, this tool that I trust that I'm like, I checked a box that says they're not going to train on
430. 1:17:36my data, so we should be good. Oh, it can look through my Google Drive. Cool. That's a lot of trust in that
431. 1:17:45little checkbox. We will not use your data unless unless you give us access to your Google Drive, in which case we're going to take
432. 1:17:55it all. Like, I don't know if it says that. I don't know what it says, but like we're in weird times right now. So, I would say
433. 1:18:05be excited and curious and terrified. Like my co-founder Mon'nique of Storyvine, she is absolutely we have we have pharma clients. So, it's
434. 1:18:18actually really important for us to have privacy and security and and you know, we we we are contractually obligated to
435. 1:18:28our pharma clients to not give up their data. So, you got to be really careful with this stuff. If you're toying with real
436. 1:18:36data, if you're just messing with your own data and you don't give a go play. But just be be really cognizant of
437. 1:18:42what's going on. LG Cursor is now part of SpaceX. Yes, they bought Cursor a month ago. Today, we have officially
438. 1:18:50closed our acquisition. We will join the SpaceX team to help make Grock the world's most useful AI and improve Grock
439. 1:18:56build, Grock bot, Grock API, cursor, and more. Yeah, make no mistake, SpaceX AI is coming, right? They're now
440. 1:19:07a public company, so they've got public currency. So, they got lots of money to go buy things. They bought Cursor. So
441. 1:19:15because of the cursor acquisition, you're you're going to now have Claude code and Claude co-work and codeex and and you're going to now
442. 1:19:24have SpaceX Grock build andor cursor, right? That are all doing the same thing. They're all agentic harnesses
443. 1:19:36allowing each of us to just say I want and and then you also have Gemini in there. Um, we'll we'll get back to
444. 1:19:42Google in a second because Google Google is is in trouble, but they're they're trying to write the ship right now. But like
445. 1:19:54this stuff is going to be coming hot and heavy. If you think 2026 has been intense as we as we get into the end of 2026 and enter 2027,
446. 1:20:06the innovations are going to be coming hot and heavy. Hot and heavy. um you don't need to keep up with them. But
447. 1:20:14Mel, to your point earlier, if you're trying to build something now that's going to be valuable and useful
448. 1:20:27relative to the new stuff 6 months from now. I just don't think that's a reality. I I think I think the thing
449. 1:20:35that you're you're learning, Mel, right now is to think critically about how you set up these platforms so that when the next paradigm comes
450. 1:20:47along, you can look at it and go, "Oh, I know what to do there." And then you'll jump on that one and you'll jump on that
451. 1:20:53one. Um, I think the throughine will be our own personal context and how that's represented in these in these things as
452. 1:21:01they advance. But I think that that I think that what we do and what our job is is going to just keep changing. It's
453. 1:21:08going to be wild. Um I haven't [clears throat] heard anything about polar browser. No. Um the thing about Google right now, they
454. 1:21:18have basically um punted on Gemini 3.5 Pro. So their big ass 3.5 model. They've basically just said it's not good enough. They're
455. 1:21:30just focusing on the smaller flash models that can run on local machines. And then they booted or he wanted to leave Demis out of
456. 1:21:41DeepMind. He's now the chairman of Google AI. Sergey Brin has come in as sort of the temporary CEO. He's the founder of Google.
457. 1:21:53um and he's come in and he said we've got to focus on um AI super intelligence and that's that's our number one remmit.
458. 1:22:01We've basically he's basic they're basically saying we've got to catch up to the frontier models. The fact that
459. 1:22:06Google is not a frontier model company right now is weird. It's weird. It's weird how quickly they sort of lost the
460. 1:22:15plot, but they have. But don't count them out. They have all the data. They have all of YouTube. They have all of Google Docs. They have
461. 1:22:26all of Gmail. They have all of Google. They have all the data. It makes no sense that they're not in the lead, but they're not. [laughter]
462. 1:22:40Oh my god. We have seven minutes. What other questions do you have? Okay. It looks like there's a big storm going to
463. 1:22:47Hawaii. Please stay safe, Linda. Um, tomorrow, if you didn't hear this earlier, tomorrow at 11:00 a.m. Eastern
464. 1:22:58time, architecture before AI, Towns and Ward Law is is doing a one-hour workshop on setting up your Claude code or your codecs
465. 1:23:10um with this system that he's created that's got these specific um memory and markdown files in it that is really
466. 1:23:18really powerful. Um, Andy's used it. I think Brandon's used it. Brandon might have done his own. Um, but this it's a
467. 1:23:25very powerful system. He's going to teach you how to do it. He's going to give you access to that thing. It's free
468. 1:23:31to everyone tomorrow live, but the replay is going to be on demand inside the mastermind. So, if you're a
469. 1:23:38member of the mastermind, you can see a replay. If you're not a member of the mastermind, first of all, you should be,
470. 1:23:46but you won't be able to see it. Okay? But if you're in the mastermind, you will be able to see the replay. Beautiful. All right.
471. 1:23:55In turn, our banks have access to all of our money, but we don't lose sleep over that. Well, we will when they start cracking encryption.
472. 1:24:04[laughter] Kelly, this is the thing. Like I just I we we're like we're just absolutely in surreal times. I I just I was talking
473. 1:24:14about this. I was out with uh Jerry Makowski and and Mon'nique, my co-founder, this morning for breakfast.
474. 1:24:20Jerry's in town. And we we were talking about this that that it if the AI gets good enough that it starts threatening encryption,
475. 1:24:32then all bets are off. Like all all like is just going to be haywire. >> [laughter] >> So, oh my goodness. My goodness. My
476. 1:24:46goodness. My goodness. Then Quantum. Yeah, we haven't even gotten to Quantum yet. Fun Twitter DM video before you go, Maddic.
477. 1:24:57Oh, Twitter DM. Okay, hang on a sec.
478. 1:25:16Polar browser. I've not heard of that. Paper cut fixed. You can now toggle visible watermarks on or off in Gemini.
479. 1:25:31>> It's the last one I said. Yeah. >> Okay. Oh, there's so much stuff. Madic robots. All right. Are we live? >> We're live.
480. 1:25:45>> And spent nine years to build the world's first intuitive home robot. A robot that sees, listens, and cleans
481. 1:25:51your home unsupervised, fully autonomously. Today, we're introducing Madic Q. >> Hey, Maddic. clean the spill up.
482. 1:26:10>> Let me show you how Madic actually pulled that off. Hey, Madic, follow me. Madic has five built-in cameras to help
483. 1:26:22it recognize cues and help it move around more naturally. We basically given it eyes and ears. So it sees the wire land
484. 1:26:32and on Q it just moves around it. It also understands over 75. >> Hey Madden. >> Hey Maddock. Stay in the living room.
485. 1:26:46>> Dad I'm shooting a video with Maddic. >> Where was I? Oh yeah. I was telling how Madic understands over 75 different
486. 1:26:52languages. For example, you can tell it, >> "Hey, Madd, hey Maddic, follow me.
487. 1:27:08Hey, Madic, clean this up." Also, all the data is stored and processed [music] on the device, so nothing leaves the four walls
488. 1:27:16of your home without your permission. >> So you don't compromise your family's privacy to keep your floors clean.
489. 1:27:22Thousands of families already use and love Maddic. Wire gave it a 10 out of 10. The first hardware to get that score
490. 1:27:29in over a decade. To celebrate our launch, we're offering free home cleaning services with Maddic [music] in
491. 1:27:35San Francisco and New York City. Comment Maddic and we'll send you the link to sign up. Order yours today at madicroot.com.
492. 1:27:44>> Dad, focus [clears throat] on the movie. >> H wild. All right, there you have it. The robots are coming.
493. 1:27:52The robots are coming. The robots are coming. All right, listen everyone. Thank you very much for hanging out here
494. 1:27:59tonight. Happy Friday. Happy mid August. It was a long week for me. If it was a long week for you, you were not alone.
495. 1:28:08Um, I hope you all have a great weekend. Uh, I don't have any homework for you this weekend. Just keeping up with this
496. 1:28:14stuff is enough. Trying to figure out how chat GPT works anymore is a nightmare. Um, just give yourself a
497. 1:28:22break. If you're feeling like you can't figure this stuff out and nothing makes sense to you anymore, you're doing it right. [laughter]
498. 1:28:32Have a good weekend, everybody. Hey, Mad. Hey, Alexa. Hey, Siri. What the is going on? Exactly. [laughter]
499. 1:28:39Exactly. Um, all right. Thank you all. Uh, thanks producer Brandon. Uh, thanks all for hanging out and, uh, hope you have a good night. Bye.