5/27/2026 - From Chatting to Delegating: Understanding the Shift to Agentic AI Platforms

Transcript

1. 0:00Um, >> too late. We're live. >> All right. Beautiful. [music]
2. 0:16[music] [music] >> [music]
3. 0:32[music] >> Good evening, good people. Welcome to the AI Learning Lab. It is Wednesday. We're going to do some stuff tonight.
4. 0:42Um, speaking of doing stuff, uh, Brandon just got out of, uh, doing the, um, I think this was his third
5. 0:52session of the lovable, uh, the lovable master class in the Iselon mastermind. Um, and welcome, welcome, welcome. So,
6. 1:02tonight, woohoo, [laughter] hour three with Brandon and hour one with Kyle. [laughter] a third class TV on Wednesday nights.
7. 1:14It's all AI all the time for me on Wednesdays this this week. >> Unbelievable. Well, while you're up
8. 1:20here, do you want to do you want to jump in and talk about what you've been doing in in there? So, if you're if you're new
9. 1:27to the salon, welcome. Um, this is the AI learning lab. We go live Monday, Wednesday, Friday. Mondays we talk about
10. 1:34sort of big ideas, big news going on. Wednesdays is about doing stuff and Fridays an ask me anything. So come in
11. 1:40and we'll we'll talk about whatever. Uh but in the spirit of doing night, you just got out of uh teaching your lovable
12. 1:47stuff and I know you've got some cool stuff stuff that you've built a thing from YouTube in particular that I think
13. 1:52is really exciting. So you want to do a little a little Brandon show and tell to kick off the show tonight.
14. 1:58>> Sure. Absolutely. So, if you're not in the mastermind, we are doing a four-week sprint deep diving into vibe coding with
15. 2:06Lovable. So, Lovable is a platform that allows you to build web apps and uh they are just on a tear. They've been doing
16. 2:14release after release. Today, they announced sub agents. We haven't even figured out what to do with the regular agents, let alone the
17. 2:22sub [laughter] agents. >> Exactly. I saw that today. I'm like, sub agents? The hell's a sub? There's there's agents now. I got sub agents I
18. 2:29got to deal with. [laughter] >> But over the holiday weekend here in the US at least, uh, Lovable had a three-day
19. 2:37hackathon where they said, "Build something and we'll enter it into a contest for a thousand credits." And so
20. 2:44I had a project that was sitting on the back uh burner for my lovable that I really wanted to do as a passion project
21. 2:52because I DJ part-time uh when I'm not producing Kyle show and uh vibe coding and jack of all trades [laughter] as
22. 3:00they say. And so my primary vehicle for DJing has become YouTube Music, which is a really powerful music engine because
23. 3:11it allows you to pull pretty much any song that's ever been recorded. It's in YouTube library. They don't have a good
24. 3:20mixer. So if you're DJing and you want to play two songs on top of each other, they don't have a way to like fade from
25. 3:26one to the other. And so I was DJing a wedding on Sunday and I had a cocktail hour where I was just basically just
26. 3:33sitting there doing nothing. So I was like, you know what? I've got the lovable app. I've got 45 minutes to kill
27. 3:39before the action picks up on the dance floor. And I sat there and I prompted this into existence. This is ytj.loable.app
28. 3:48and it allows you to play one song and then start playing another one in the background and then fade between them. Um, and I won't
29. 3:56>> basic basic basic DJ functionality, right? >> Yes, exactly. I won't play them here because I know Kyle, you uh will
30. 4:06eventually have this on YouTube and I don't want you to get a copyright strike for playing right uh of that music. But
31. 4:12what I will tell you is that beyond that, it can cue things and it's even got this cool feature where you can
32. 4:20start a QR code to start a request line. So, as I'm at a party, I can actually put up a QR code and people can scan the
33. 4:27code and put in requests for songs and then it adds them to my queue automatically, which is not even
34. 4:32something that I thought to do. That was something that Lovable said, hey, you know what? If you're doing this, you
35. 4:38could do that really easily. And I'm like, yes, let's make that happen. >> And so, that's really the powerful uh
36. 4:45potential of Lovable. Yes, I built this on the Lovable app and then ran it through all its security settings and
37. 4:51published it Sunday night after I got home from uh the wedding. >> Yeah. >> And so, >> and and it's also well, one thing you
38. 4:59could do is you could if I don't know if you can find them, but if you could find Weird Mary from Cedar Hill and Quantum
39. 5:04Cipher, you could you could go between those two. Um, yeah. So, for those of you that don't know or or have heard
40. 5:12about it, Lovable is what's called a vibe coding app. So you basically just ask it in words what you want like I
41. 5:19want a DJ kind of interface for YouTube music. Um and then it will go start writing the code. And so you just
42. 5:27basically keep talking to it. And if you do if you search for quantum cipher with Kyle Shannon you should be able to find it.
43. 5:36>> I don't know is it on YouTube though? Yeah, it's in the uh it's in the AI I think it's in the AI Learning Lab channel, actually.
44. 5:47>> That's all right. >> I have this set to music, but um Oh, it's not it's not in YouTube music. No, it's just in YouTube proper.
45. 5:57>> Yeah. So, this was [clears throat] this might this might be it actually. I think this is it, actually. So, I'm gonna go
46. 6:02ahead and add this to QA and uh bring it over here. Yeah, this is it. >> Oh, cool. [music] >> And then what I'm going to do is I'm
47. 6:15going to bring up Happy Birthday because Happy is a royaltyree song. The weirdest mashup of DJ History.
48. 6:31[bell] Happy birthday to you. [laughter] >> And so, you know, this is a really cool
49. 6:46thing that I was just thinking into existence. Like I we're at the point now where we can have custom software. And so what we're doing
50. 6:55inside of the lovable four-week sprint is we're talking about how to go from that idea and really iterate on it and
51. 7:03figure out what problem you're trying to solve and then work within work on iterating with the AI to come up with a
52. 7:10plan on how to solve for it with a tool like lovable. And then this week we actually started building. So, one of
53. 7:19the things that I'm building here, I'll switch my tab, is a bedtime routine for the tidbits, the little uh little carbon
54. 7:27copies of me and my wife that are uh upstairs getting ready for bed. We're trying to figure out how to make bedtime
55. 7:33a little bit more magical for them. And so, I we're in V1 of this where we can actually have uh avatars that talk to
56. 7:42you. Today was all about how to safely integrate API keys with things like 11 Labs. Um, and at the end of next week,
57. 7:51hopefully folks are going to have vibecoded apps that are ready for minimal deployment out into the world.
58. 7:58And so it's a it's a very cool thing to be a part of and watch what other folks are doing inside of the AI salon with
59. 8:05their with their lovable tools. >> Yep. And let's you know the how this ties to what I'm going to talk about
60. 8:12tonight. So what I'm going to be talking about tonight are these agentic platforms and these things all started
61. 8:19as extensions of idees. So uh development platforms that developers use and so cloud code was was intended
62. 8:27to be this AI you know add-on to you know to a to a development environment for programming. And what they realized
63. 8:36is that these things are very powerful in and of themselves. And so they they're now sort of creating them as
64. 8:42these agentic platforms that can just do work for you 24/7. And lovable um even though it's a thing that kind of started
65. 8:50like chat GPT. You type in, you know, what you want, it makes an app. It's getting it's been getting more and more
66. 8:56and more sophisticated. So now that they have an iPhone app and now that they've got today they you know announced these
67. 9:03things called sub aents which sub aents are basically like lovable itself is going to figure out oh if I need to go
68. 9:10do some research I'm going to spin up a sub agent that that is a research bot and it's going to go do research while
69. 9:16some other thing is going to go do I don't know design or you know interface work and these individual little bots
70. 9:23are going to go out and do that. So a lot of these platforms are moving in this direction of you as the human are
71. 9:30going to be at a wedding and have 45 minutes to kill and you're going to be able to go hey wouldn't it be cool if
72. 9:36right and then these things will go off right so this is a new skill to learn how to how to use these things. So what
73. 9:42we're going to try to do tonight is understand what are the components of them so as you start to learn this stuff
74. 9:47you can understand what it is. So anyway, that's super cool and yeah, and you're you're not too late to take so so
75. 9:54uh session 404, if you haven't joined the mastermind, join the mastermind. Session 404 is next week, next
76. 10:00Wednesday. Um and then you can also, if you're part of the mastermind, you can go watch um the the previous three
77. 10:07weeks, uh as well as all the other content that's in there. Um, we just just finished up um a four-week session
78. 10:15with Andy talking about deconnecting your identity from your job and and uh we've got uh a leadership uh session
79. 10:23coming up with Dakota Coons. There's all sorts of cool stuff in there. Anyway, beautiful, beautiful, beautiful, beautiful,
80. 10:31beautiful, beautiful people. Okay, so here's what we're going to do. what we're going to do. So, on Monday, I talked about
81. 10:41and I don't know why I'm all of a sudden, you know, talking like I'm from New York somewhere, Jersey. Um,
82. 10:49we talked about these agentic platforms and I talked about them with a lot of handwaving and I had a feeling it was going to be a
83. 11:00confusing conversation. And I think I think we got somewhere that it was it was decently clear, but but the these systems are actually
84. 11:11quite complex. There's there's a lot of parts to them that you actually have to pay attention to. Tidples. That's pretty good. Um,
85. 11:21beautiful. [clears throat] There's there's Brandon's lovable portfolios in the comments. Okay, so I'm going to fire up chat GBT. I also
86. 11:29restarted my machine before um before coming on tonight. So hopefully my screen share will not be going away as hard as it did before.
87. 11:42But let's see.
88. 12:00Okay, can you all see this? [laughter] Depending on the day, it's either tidbits or executive assistance. That's beautiful.
89. 12:14I'm going to drag [clears throat] this over here. Let me do a little window management so I can keep up with y'all. [clears throat]
90. 12:29So you can see this now and I can see the screen now. So if my screen share goes away, I'll be able to see it. Okay,
91. 12:47now let's see if I can Can I talk and you can? [clears throat] Okay, good.
92. 13:10So, we're in chat GPT and my goal right now is to have chat GPT do the heavy lifting to understand the components of these agentic platforms
93. 13:25like codeex, claude code, claude co-work, um, Grock the uh anti-gravity from Gemini, all those sort of things. Open claw, Hermes,
94. 13:41um these things that that the developer community is calling harnesses. So, so actually first thing I'm going to do, let's do that.
95. 13:52I want you to help me understand agentic platforms and what are the components that make them up that need to be managed if I
96. 14:03want to be a powerful user of these autonomous technologies period. One of the first questions that I have
97. 14:12is that the developer community seems to have shifted from calling them apps like chat GPT is an app
98. 14:24to harnesses where they're talking about the harness is the thing you interact with that allows you to access models and other
99. 14:36applications and things like that. I'd like to understand the history of that phrase as well as is there a meaningful distinction or is
100. 14:48harness just tech bro [ __ ] that I have to try to remember the meaning of [laughter] [laughter]
101. 15:03yeah you see what's you see what's going on here [laughter] okay I'll Treat generic platforms. Oh, as as oh genic as agentic platforms.
102. 15:15That's good. Two separate things. The real architecture you need to manage. Okay. Yes. And the vocabulary fog around
103. 15:22harness. [clears throat] [snorts]
104. 15:36Thank you. I love these kind of questions. Well, thank you, Mary. I listen the the the reason I'm doing this
105. 15:42today is that I realized as I was explaining it that I I'm trying to juggle all these concepts in my head all the time
106. 15:54and then every time I go to a different like if I go to claude code or codeex or perplexity or or like grock build just
107. 16:01showed up and that's another one I'm like wait where how where do I put all my crap? app. Um, okay. You're [laughter]
108. 16:11Thank you, Chat GPT. You're asking exactly the right question. Gee, thanks ChatGpt. You know, I I was I was really
109. 16:20worried that I wasn't asking exactly the right question. I thought I might have been asking sort of kind of the right
110. 16:25question, but I'm super glad that I'm asking exactly the right question. Thanks, ChatGpt. I feel better about
111. 16:32myself. Okay. And yes, harness is partly useful. architecture language and partly jargon fog. It's not pure tech bro
112. 16:42[ __ ] It's absolutely being used too loosely. A clean distinction is this. An app is a thing you experience. A harness
113. 16:48is the thing that turns a model into a working agent. A platform is a larger ecosystem that lets many agents, tools, workflows,
114. 16:58permissions, and memories and interfaces run together. So chat GPT is an app from the user's point of view but under the
115. 17:05hood it behaves more like an agentic platform. Model selection tools file handling memory browsing code execution
116. 17:12connectors safety layers. Developers say harness when they want to talk about the machinery around the model rather
117. 17:21than the model itself. Okay. Is harness a real term? Yes. But the meaning is shifting. The older software meaning comes from test
118. 17:32the older software meaning comes from test harnesses controlled environments made of stubs drivers and supporting code that allow a
119. 17:43component. Okay, so this is an actual term. ISTBQ defines a test harness as a test environment comprised of stubs and drivers that need
120. 17:55to execute a test. The older meaning matters because it gives us a metaphor. A harness is the stuff around the thing that lets you
121. 18:04control it, connect it, constrain it, test it, and make it useful. In AI, the term became more common as raw LLM is
122. 18:12not really an autonomous agent. Lang chain's 2026 definition is blunt. Agent equals model plus harness. The practical distinction.
123. 18:27Oh, okay. Here's the different thing things. Okay, the model is the reasoning engine. So, GPT 5.5, Claude Opus 4.6,
124. 18:40Gemini 3.5, etc. The app is the userfacing product like chat GPT. The harness is the control system around the model. The agentic platform.
125. 18:56Okay. But here is a question. Is chat GPT a harness?
126. 19:10Even though it's an app and is codeex a harness
127. 19:26even though it's an agentic platform question mark I don't wait let me do it this way I don't want you to make this up. Period. I want you
128. 19:42to actually research this and get me the real answer. Period. If you have to go to Reddit or Wikipedia or X, I don't care. Get me the real
129. 19:57answer because I want to make an infographic and I don't want the infographic to be [ __ ] Period. Ah, Mr. It, what's happened? I'm
130. 20:11vocabulary challenged. I'll ground the language in the language builders actually use. Okay.
131. 20:35You're right to be suspicious. >> [laughter] >> You've asked exactly the right question. Do not label chat GPT or codeex simply
132. 20:46as a harness in the in infographic. It would be imprecise. A better defensible label is chat GPT and codeex are
133. 20:54userfacing agentic products platforms that contain harnesses. A harness is the execution layer inside them. tools, memory,
134. 21:06context, permissions, orchestration, workspace skills, MCP connector, sandboxes, and guard rails. [snorts]
135. 21:15The strongest current definition I found from lang chain is agent equals model plus harness. It describes the harness. Okay, so the
136. 21:26harness is not the app. So model goes into a harness which does all this orchestration. Then there's an agent. Then there's the
137. 21:36app. But wait, or more plainly, wait, that doesn't seem right. Wait, is Chat GBT a harness? No. Okay. So, wait. So, so wait. So
138. 21:48explain to me in super clean English. What makes chat GPT and codeex
139. 22:07different? And similarly, what makes Claude and um Claude co-work
140. 22:24Claude code slashco different. Let's just question from
141. 22:41MC like a sandbox. Can you place the chart that chart in the chat? Which chart?
142. 22:57Oh, this chart. Copy. I think I can. I don't know that I don't know that this is right yet. And it looks like it cut it off. Anyway,
143. 23:10um, let me what what I'm going to do is I'm going to turn this into an infographic here in a second. Chat GPT
144. 23:17and Claude are general purpose AI assistants. They're built for conversation, thinking, writing,
145. 23:34chat, GPT, and Claude. Talk with AI about almost anything. Codeex and Cloud Code. Delegate software work to an AI
146. 23:41agent inside a code workspace. Uh, I'm going to call some [ __ ] on something here. Wait, but I want to call [ __ ] on
147. 23:49something here. Period. Codeex and Claude Code both started as agentic coding systems
148. 24:04and they both and everyone else have since realized that it's not just valuable for coders, it's valuable for
149. 24:12anyone using a computer. What I'm trying to come up with here is a real world definition of what is the difference between
150. 24:26chat GPT conceptually and codeex conceptually. Period. It sounds like what you're trying to say is that chat GPT is about a human
151. 24:41talking to the model through [snorts] the chat GPT interface. Period. But codeex is about a human delegating
152. 24:53tasks to an AI agent that will run autonomously and use the models and any other tools it has access to
153. 25:03to do its work on behalf of the user. Period. Is that a correct distinction? And I want to understand the evolution
154. 25:15of these agentic platforms from coding tools to general computer use tools. Period. I'm trying to get a glimpse into
155. 25:23the future of what non-programmers are going to be using these agentic platforms for. Period. Don't be a nanny. Not nanny. Ninnie.
156. 25:37[laughter] [laughter] Oh, [laughter] look at me. Look at AI making Kyle dumber. Oh my god. You're not wrong
157. 25:54to call [ __ ] Thanks, chat. GPT, you big dumb dummy. The coding agent category is becoming the prototype for a
158. 26:03broader computer using agent category. So the old codecs for coders framing is already too small. Yes, your distinction is basically right
159. 26:13with one important correction. Chat GPT is primarily a conversation first assistant. Codeex is primarily a
160. 26:20delegation first agentic work system. Good [ __ ] lord. But the boundary is now blurring because chat GPT has gained agent mode and codec
161. 26:33style systems are escaping the coding world. The real distinction is not chat versus code. The real distinction is am
162. 26:41I having a conversation with AI or am I assigning work to an AI system that can operate in an environment on my behalf.
163. 26:48Okay, that's not bad. So, let me pop this in the chat. Good lord. [sighs and gasps] Okay. Um, why? Here we go.
164. 27:04Chat GPT started as a human to model conversation interface. You ask, it responds. I am sitting with an intelligent
165. 27:16assistant. Wait, even even when chatbt has tools, files, web, images, data analysis, connectors, the default mental
166. 27:25model is still I am sitting with an intelligent assistant and working through something together. Now, OpenAI has added chatbt agent which
167. 27:37absolutely crosses into delegation. Right? So if I So what it's talking about there, if you go into the little plus sign in chat GPT,
168. 27:49if if you ever thought chat GPT was simple, just click on the plus plus sign. You can put chat GPT into agent mode where it starts to act.
169. 28:00It can spin up a web browser and it starts to do things on your behalf, but it's not its primary mode. It's just
170. 28:08that's just a mode of this thing that is primarily to have a conversation with. Okay. But its center of gravity is still
171. 28:17conversation as the main interface. Codeex started as a software work delegation system. You give it a job. It
172. 28:24enters a work environment, reads files, edits files, runs commands, tests things. OpenAI describes Codeex as a
173. 28:31coding agent that can read, edit, and run code. So the mental model is not let's chat about this. It's about go do this task
174. 28:39in that workspace and come back to me with the result. That's the conceptual leap. So your sentence is right. Chat GBT is
175. 28:48about a human talking to a model through a conversational interface. How does this compare to Adam? Okay. So we're going to get there.
176. 29:00You're you're not wrong. Note from your mother. >> [laughter] >> [ __ ] chat GBT. It's it's horrible. Like like listen, the [clears throat]
177. 29:11when I first started using these things, just having something like the the difference between working with chat GPT
178. 29:19and working with a human is that humans will just without even saying words telling tell you that your idea is a
179. 29:26piece of [ __ ] And it can be really demoralizing. And so if you if you've got thin skin um and and you're a
180. 29:34creative in the world trying to create new things and people [ __ ] all over it, it's kind of shitty. And so initially
181. 29:41dealing with chatbt, I was like, "Oh, this thing's great because, you know, I can just share ideas and explore the
182. 29:46ideas without all that bullshit." But the more you use it, boy, does that become annoying. Okay. Conversational AI assistant agentic
183. 29:59computer using agent agentic platform. Oh good god. Okay.
184. 30:21Non-programmers will use Agentic platforms for file chaos, research assembly, spreadsheet dudgery, document production, business
185. 30:30operations, marketing workflows, event planning, personal admin, learning and startup, learning and setup,
186. 30:37mini software creation. The last one's a sneaky one. Okay. Um Okay. I want you to design
187. 30:57an infographic that has three columns
188. 31:17where the first column represents the conversation interf. interace with chat GPT as the primary example. The middle column
189. 31:35describes the agentic platform like clawed code that's primarily used by coders.
190. 31:51And then the third column is something like Claude Co-work or Perplexity Computer that the primary intent
191. 32:06is to have the agents do computer use rather than just coding. Period. So explain these three concepts
192. 32:20conceptually on the top half of the page and then on the bottom of the page I want you to put the brands and logos
193. 32:34and even short descriptions of the tools that are represented in that column. period. I want you to
194. 32:50include the major frontier model companies of OpenAI, Anthropic,
195. 33:05XAI, and Google. Period. I also want you to represent open-source agentic platforms like OpenClaw and Hermes.
196. 33:21And I also want you to represent web-based agentic systems like Genpark and Manis. Period.
197. 33:47Make the image 16 by9 wide. Okay.
198. 34:02You need tough love, Kyle. Oh, Rick. I haven't even gotten to MCPs, APIs, connectors, skills, all that [ __ ] Like what I'm
199. 34:12just trying to understand right now is because here's the here's the the the part of my part of what's grinding my gears right now
200. 34:26is that everyone's talking about these agentic platforms like they're like like they're this new powerful thing which
201. 34:32they are but the interface that they put put on them looks exactly like chat GPT and because chat GPT's got aic features.
202. 34:41It sort of acts like these things, but these things are completely different. So, so step one for me is just to
203. 34:48understand what these things are. And then the the next layer down, we're going to start to dissect what's really going on. Because here's
204. 34:56here's here's something that hit me today, Rick. So things like the in in in in Open Claw the soul file which which talks about um
205. 35:08how you want that thing to act, how I want Adam to act and the um and the user file which describes me and
206. 35:22then the skills files which describe maybe how I write and and how to do other things like how to use spreadsheets or how to use things and
207. 35:29then MC CPS all of all of these things that are context for who I am and the projects I'm working on. The promise of
208. 35:38these agentic platforms is you we can maintain our own data centrally and have these systems tap into our stuff rather
209. 35:48than giving all of our stuff to open AI which is what we've done for the past three and a half years. Right.
210. 35:56Okay. Let's see how bad this piece of [ __ ] image is. Let me download it and we'll go look at it. And [sighs]
211. 36:04I hate the design of this thing. So, there's that. [laughter] Kyle seems angry. Okay. Three modes of agentic AI conversational interfaces.
212. 36:21Chat GPT Claude Google Gemini Grock. Okay. Okay, so we need the URLs down here. Coding dev dev agent OpenAI codeex.
213. 36:32No, not jewels. Autonomous coding agent GitHub style workflows. Google jewels. Jewels. Seriously? Computer using agents. Claude co-work
214. 36:46chat GPT agent. Oh, this is exhausting. These are all wrong. Do I even want to deal with this? All right. I'll I'll I'll try I'll do
215. 36:57one pass to try to improve this. All right. Let's just go. Okay. Okay. This is
216. 37:12in the right neighborhood, but what I would call a fairly solid pile of garbage. Period. So here are the things that will make it
217. 37:25less garbagey. Period. For the top three labels of conversation interfaces, coding/dev agents etc.
218. 37:42I want to have an example in the headline like conversational interfaces like chat GPT would be the subhead.
219. 37:56So pick whichever one you think is the most important.
220. 38:13No, that's not right. Like chat GPT would be the subhead. Use chat GPT for number one. Use
221. 38:27clawed code for number two and use perplexity computer for number three. Period.
222. 38:45I want you to include the URLs of the tools that you're showing in the bottom half of the page. Period.
223. 39:00In column two, you used Google jewels for the coding agent. Period.
224. 39:16I've never even heard of that. Period. Replace that with Google anti-gravity. period.
225. 39:35There should be a bridge between column 2 and column three because claude code and claude co-work are
226. 39:45actually just different tabs of the same tool with slightly different intentions. OpenAI codeex similarly spans both of those.
227. 40:02In fact, now that I'm thinking about this, columns two and three are correct. But every tool that's doing something that
228. 40:13can be used for coding can also be used for some form of file management.
229. 40:27I want to separate out open claw and Hermes as as open-source [sighs] agent platforms.
230. 40:42I want to separate out GenSpark and Manis as webbased agent platforms and then Claude Co-work OpenAI Codeex Perplexity Computer
231. 40:59and Google Gemini/Spark are the tools for the big frontier model companies. Period. Take all that and try again.
232. 41:15Good lord. [laughter] That was that was one longass prompt. Here's what I was trying to fix this piece of garbage. Okay. Good lord. I
233. 41:29want Claude's output, but I can't stand their interface. I know. I know. That's I'm I'm right there with you with the Claude thing. Um,
234. 41:40here's some here's some new pitch deck for Sydney. We're doing some pitching. Look at the character sheets you can do now with um
235. 41:51with chat GPT. The character consistency is amazing. Isn't that cool? There's Sydney. There she is in her little step right up
236. 42:06outfit. There's Kellen. He's a handsome young fell. [laughter]
237. 42:21Tech reporter for the New York Tribune. [laughter] Is she single? Just ask asking for a friend.
238. 42:37All right, let's see what we got here. Generating a more detailed image. Hang tight. This is actually not a bad test to check
239. 42:46the new image capability of chat GPT because that was a big giant garbagey prompt that I gave it. Trashy trashy trashy prompt.
240. 43:21>> [clears throat]
241. 43:44>> still going. Did it get screwed up? Oh, here's something I've I've I've learned. See how this has got the the thinking
242. 43:54beneath it and it's not changing. Um, this is hung. So if if you go to generate an image in chat GPT and and it says thinking and
243. 44:04then this thing is just sitting there static it's hung. So I will reload this [clears throat] and now we should we should see this
244. 44:13thing. Oh see it already finished it. Okay. Oh no it didn't for 2 minute 55 seconds.
245. 44:28Dead. Nothing. My Shirona. Try again. Try again. Try again. Try again. Again.
246. 44:44Again. Again. Okay, we're not going to stay long on this. This This concept is relatively straightforward to
247. 44:51understand. What I'm trying to get my head around [clears throat] are the components of these things.
248. 45:44So, while this thing's building, next week is the first Tuesday of the month. So, we have um Marta um I don't know how
249. 45:53to pronounce Marta's last name um but she's a professional rock climber. Um she's going to be the guest at AI Salon
250. 46:00Presents. So, mark your calendar for next Tuesday from s 5 to 7:00 p.m. Mountain time is AI Salon Presents. We're trying to keep those to 90
251. 46:09minutes. So, you'll get out of there at 6:30 Mountain time. We're trying to be more respectful of everyone's time for
252. 46:15these events. So two hours was feeling a little long. So we're tightening those down to 90 minutes, but just block off 5
253. 46:21to 7 mountain time. Um where's no Emmy? Um I'm sleeping [laughter] looking images on the left. Oh yeah, let's go look over there.
254. 46:35Um if you're new to the AI salon, please go check things out. By the way, in in the uh if you if you want to get to the
255. 46:43chat, if you can't see the chat, it's in the upper your upper right hand corner. Right there is where the the link to the
256. 46:52chat is, that'll put the chat out. And then down here is a button that you can pop out this window. So, if you create
257. 47:00two tabs for the AI salon and one of them is this live streaming channel and then you pop this video out and then you
258. 47:07go to the other tab, you can surf the AI salon and participate in other pieces of the AI salon while you have this
259. 47:13playing. You have to have two tabs running to do that. But the little pop out button is right down here. So, chat
260. 47:19up here, pop out button down here. Me here. Okay. Okay. Let's look at this thing. >> [clears throat] >> Let me download this.
261. 47:40Download. That's the old one. Here's the new one. Conversational interfaces. Primary
262. 47:55example. Chat GBC tweet. The chat GBT. The main interface is dialogue. The human stays in the loop. Ask questions.
263. 48:02Uploads files. Prompt. Respond. Follow-up. Refine. Conversation first. Human guided. Great for thinking.
264. 48:10Conversation first. Assistance. Chat. GBT. Claude. Gemini. Grock. Fine. That's fine. Good lord. Overlap zone. Okay. Coding
265. 48:21dev agents primary example. Claude code. Computer using agents primary example perplexity computer goal plan code test delivered
266. 48:35code Google anti-gravity computer use first asynchronous outcome focused this is still kind of crap you delegate work inside a code space
267. 48:49the agent reads files edits code run commands uses tools returns work for view over here. You delegate work across
268. 48:56apps, files, browsers, and the web. The agent acts more like an operator completing multi-step tasks on your
269. 49:04behalf. So, you got codeex, claude co-work, anthropic cloud co-work, and they've got code up here. They didn't quite get
270. 49:13that. Gemini spark, Gemini anti-gravity. They didn't quite get that either, but that's okay. Hermes agent, Open Claw,
271. 49:20that's the wrong logo for OpenClaw. Mannis, it got right. Gen Spark super agent. Well, okay. So, chat was the
272. 49:28doorway. Code was the training ground. Computer use is the destination. This is not great, but what I'm going to
273. 49:36do is I'm going to hop over to the AI. Hello. [clears throat and cough] And we will go to community feed. And
274. 49:51then I'm going to share a post with this image in it. Oh, god damn it. Okay, hang on.
275. 50:12[sighs] What?
276. 50:28That shouldn't be the case. Share a post plus.
277. 50:44Right. Okay. Um, this isn't perfect, but this is not a bad way for us to understand
278. 51:03the difference between where we are where where let's say where we've been using conversational
279. 51:21agents and or no conversational tools like chat GPT and
280. 51:35these new agentic platforms like Claude code slashcowwork. [sighs] [gasps]
281. 51:50Um, feel free to share, but please include
282. 52:04the link to this post so people know it came from the AI salon. Okay. Okay. So, this is going in
283. 52:18community feed. So, if this is an image that's of interest to you, there it is. Uh, yeah, I'll notify you users of that.
284. 52:29Three modes of agentic platforms. I should probably throw the uh the uh the AI salon logo in that. Maybe I'll do that. Hang on.
285. 52:43Let me go back here. Um, AI salon logo. [sighs]
286. 53:10There we go. at the bottom say that this is presented by
287. 53:26the the AI salon the salon.ai AI and use the logo and
288. 53:41colors [sighs] [gasps] for an updated image.
289. 53:57Okay. Now, we're going to do another graphic. I want to do another infographic that is more visual and less detailed
290. 54:14and stylistically more fun. Period. And it represents
291. 54:30that we're in a transition from a time starting with chat GPT on November 30th, 2022 where we primarily used AI tools
292. 54:46as discrete things that humans used to collaborate. operate with and those included large language models, image generation tools,
293. 54:58music generation tools, coding tools, video generation tools, etc. period. We're now in in a goopy transition between that world
294. 55:16where humans are very much required and are very much in the loop to this new agentic world where you've got
295. 55:26platforms that will deploy agents and sub aents on your behalf. Use tools on their own. Achieve goals on their own. set goals on their own.
296. 55:40And it's much less about a human collaborating on the individual outputs and more that
297. 55:57the human is setting the direction and letting the tool do the work. Period. I also want to find some way to represent that
298. 56:12the tool world on the left still has jankiness and problems and requires human participation and that the agentic platforms on the
299. 56:23right are still too geeky and complicated and therefore require lots of human expertise to set up correctly.
300. 56:36But that once we exit the goopy phase, both sides of that equation will require less human interaction to generate valuable work. Period.
301. 56:52We'll just see what that gives us. Okay. So, let's see if we if this did this right. Presented by the AI salon. The salon.ai. Cool.
302. 57:03did that. It looks like it changed our aspect ratio, but that's okay. Who gives a [ __ ] So, let's go back over here.
303. 57:18Can I edit this? I think I can. Let's see. Edit. Edit. Get rid of that dare picture.
304. 57:33We're gonna put in a new picture that's got a little logo on it. We're gonna put a logo on it. Oh, but
305. 57:39you know what? It's not going to like it because it's a PNG and it doesn't like PNGs, does it? Is this it? No. Is this it? Yes.
306. 57:52Unbelievable. Close. Unsave changes. Yeah, I know.
307. 58:13Okay, we're going to export this as a JPEG. What do you learn at the AI learning lab? That all pieces of software are
308. 58:24giant piles of poo and that we as humans still have to figure out the different file formats so we can upload a stupid
309. 58:31picture to our lovely community. [laughter] Oh, good lord. It's exhausting. Can I just say that it's exhausting?
310. 58:49Okay. Are we there? Is everybody good? Fantastic. Bob, tell him what he's won. All right. There you go. All right. The space has gotten so
311. 59:00crowded that it makes it more difficult to understand and compare the different companies and try to choose the one line
312. 59:06to experiment with. Yeah, Mary Carol, this is the problem. This is absolutely the problem. Rick has something and look
313. 59:12what I made. Okay, I'll go take a look in a sec. I'd like to concentrate my energy on working my way up the ladder
314. 59:17and not getting lost in making lateral moves across a bunch of ladders. Yes, because I lack clarity. Um, and where
315. 59:25the hierarch and and we're wait and we're in the hierarchy of where in the hierarchy a given tool exists and
316. 59:34mistake a new similar tool for a truly new tool in the evolutionary process. Right? I would, if I were you, Mary Carol,
317. 59:46there's kind of two places you can live as far as I can tell. You could, you could say, I'm just going to I'm going
318. 59:53to be an expert with the tools, right? So, I'm going to I'm going to really learn Chat GPT or I'm going to really
319. 59:58learn Claude, whatever it might be, or I'm going to pick one of these agentic platforms and go there. I think the longer term bet is to pick
320. 1:00:09one of the agentic platforms. I think I think understanding but I don't know not necessarily I I think it's personal preference. I think
321. 1:00:18this this idea of of standalone AI applications is going to be around for a long time and I think people who are good at discrete use of
322. 1:00:29tools as humans will be in demand. I also think anyone who can get their [ __ ] head around these agentic
323. 1:00:41platforms and how you set up all these different components are going to be really valuable because it's more complicated. It's harder right
324. 1:00:49now. Um, if you're going to go go the agentic route, I would just pick a single to single tool. The one I'm
325. 1:00:57leaning toward picking and sticking with is OpenAI's codeex, but it's a real tossup for me between that and anthropic
326. 1:01:06claude co-work and code. Um, I'm not going to play with um Perplexity Computer. I'm not going to play with Grock Build. I don't give a
327. 1:01:16[ __ ] They're just the same thing. Every All of these are derivative tools. They're all copying each other. So, I
328. 1:01:21don't think it really matters. The reason I'm I'm torn between OpenAI and Claude is that Anthropic is the one
329. 1:01:30that invented the MCP and they're also the ones that invented skills. You know, ChachiBT was the first one to
330. 1:01:39come out of the gate with a real large language model that didn't suck, right? And they invented Chat GBT. So, so they
331. 1:01:46are the two innovators in the space. In my opinion, I feel like Google is doing really good work. If you're a if you're
332. 1:01:54a Google shop, then probably using the Google tools is a good way to go. I'm leaning toward um codeex. But, you know, listen,
333. 1:02:03anti-gravity with Spark could end up being really powerful. So, I don't [ __ ] know. I don't [ __ ] know. All
334. 1:02:08right. Where am I going to go? Look at what I made. Look what I made. Rick Moli, AI for Spielberg agents, an
335. 1:02:17AI agent map for making a movie. Beautiful. By the way, there was a big announcement today from Netflix MGM, an a studio
336. 1:02:30green lit three different animated shows that are going to be animated using AI animation tools. So, tidbits. Nothing more powerful than a
337. 1:02:41four-year-old in a Batman cape. That's cool. Okay, let's go back to chat. GPT. See [clears throat] if how this image was.
338. 1:02:53Okay. From from AI tools to agentic platforms, we're living through the goofy middle. Not goofy middle. Oh, good lord. Jesus. Not the
339. 1:03:04goofy middle. [laughter] the goopy middle. That um phase of a
340. 1:03:20caterpillar turning into a butterfly. [cough] >> [clears throat] >> All right, let's go look at this because
341. 1:03:33maybe maybe goofy middle is one of those happy AI learning lab accidents that we'll just keep it because it's kind of
342. 1:03:40right. Okay, from AI tools to agentic platforms, we're living through the goofy middle. The tool era starting with
343. 1:03:47chat GBT November 30th. Humans use lots of separate tools. Chat, image gen, music gen, human in the loop, code,
344. 1:03:54video gen, research. Janky, but useful output by output. The [laughter] [ __ ] does that mean? Right now, the tools still need a lot of
345. 1:04:05human participation and cleanup. Human steering required. Duct tape and hope. Beta everything. Still lots of human
346. 1:04:12steering. Too geeky to configure. Agentic platforms are powerful but too geeky, fiddly, and complicated. The
347. 1:04:19agentic platform era gold driven goal- driven systems. Humans set directions. Platforms deploy agents and sub aents to get work done.
348. 1:04:29Set. This is actually useful. Humans move from making every output to setting direction and supervising outcomes.
349. 1:04:42Goofy. Goofy is right. Says Kyle's mom. [laughter] Um, yeah, this is this is kind of interesting.
350. 1:05:01Let's see. Is this thing still working? Here we go. The goopy middle.
351. 1:05:18>> [sighs] >> It still says goofy up at the top here. Look at that. It's unbelievable. Okay,
352. 1:05:35let's fix this one because this one's actually worth worth doing. Okay.
353. 1:05:56Okay. I want you to make some refinements for me.
354. 1:06:12Okay. I want you to lose the subhead altogether. Period. In the lefth hand column, you have a
355. 1:06:26panel that says output by output. And I don't know what that means. So, please look at that and figure out what that's
356. 1:06:34supposed to say and fix it. Period.
357. 1:06:54in the vat of goop in the middle. The idea is that the tool era is where the caterpillar is.
358. 1:07:09The vat of goop is actually the chrysalis that still contains a neural core. Oops.
359. 1:07:24Oh no. where the caterpillar has dissolved into a goop with a little
360. 1:07:40bit of a neural core remaining and then the transition to the agentic world is the butterfly. So the idea is that we're transitioning
361. 1:07:51from caterpillar to butterfly in that center section is the goop of the transition.
362. 1:08:09period. On the right hand column, add in the idea of skills and memory and MCP
363. 1:08:27for computer use. period. And I would say use more than just the Google calendar
364. 1:08:42and Google Workspace applications. Period. Since MCPs can be used by things like Figma and Canva and
365. 1:08:59uh Expedia, there are a lot more things you can do than just workspace applications period. [sighs] Okay.
366. 1:09:18How we doing timewise? Getting there. They're all just overly busy. Does anyone find value in an infographic they
367. 1:09:30didn't make themselves? Not rhetorical. Well, if they're not if they're not worth if if they're not worth it to you,
368. 1:09:39obviously don't use it. I'm I'm what I'm trying to do is I'm trying to get this [ __ ] out of my head because I can't I
369. 1:09:48can't figure out what to focus on anymore cuz my head is so full of all these piece parts that everyone talks
370. 1:09:55about is as if everyone knows what the [ __ ] they're they're doing like and how these things all interrelate to one another.
371. 1:10:04And unless you're in a frontier lab building this [ __ ] it's really confusing right now. So, um, they might
372. 1:10:11not be useful. They might be overly busy, but let's see. Picture and picture, anyone? I don't know what that means. The
373. 1:10:22process of making an infographic for me is the journey. Kyle, what's the best Gemini or chatbt? The news model. Both are $100.
374. 1:10:39Oh, what's the best Gemini or chatbt? It depends what you're going to use it for. Ivan, don't just spend $100 to spend $100. If you
375. 1:10:49if you don't know what you want to build these things for, um I wouldn't I wouldn't mess around with them. So, wait. So, this thing threw a
376. 1:10:59caterpillar on the left, goop in the middle, but then no butterfly on the right. Really? Really? That's what we're going to go with?
377. 1:11:08This is actually better. This is This is not horrible to look at. Beyond the goopy phase, less human effort, more
378. 1:11:16valuable work. Humans move. Okay, so this isn't bad. Um I'm going to just say um you missed missed
379. 1:11:31the butterfly on the right. And then this one will be done. And then Okay, so the next thing we're going to
380. 1:11:41do, and I have to do this in 15 minutes, so I'm going to I'm going to try to move a little faster here.
381. 1:11:50Let's see. Is there anything else here that I hate? Oh, one output at a time. Not output. Lots of human cleanup.
382. 1:12:04The goopy middle transformation in progress. All right, this isn't bad. I like that it put the AI salon as the center of
383. 1:12:14agentic platforms. [laughter] That's good. That's good for the ego of the salon. All right. [sighs] As I get better, I reveal connections I
384. 1:12:28didn't think of. Agents. Yeah. What I was saying, Ivan, is don't just use agents to use agents. Here's the problem with with these agentic
385. 1:12:38platforms. So, there there was a thing that Sam Alman talked about um a bunch of years ago like shortly after Chat GPT came out
386. 1:12:47and people were asking him about universal basic income and he said, "I think we might want to consider
387. 1:12:53universal basic compute." And what he was talking about was was giving people tokens, which at the time didn't make sense to
388. 1:13:02me because if you're in the if you're in the caterpillar world where you're using discrete tools and human beings are
389. 1:13:08prompting chat GPT, you're not really using a ton of tokens. But the minute you go to an agentic platform and you
390. 1:13:16say, "Hey, I've got a goal to start a business and I want you to spin up 10 different sub aents." those sub agents
391. 1:13:22are all going to go in a very inefficient way just burn millions or billions of tokens that can get really expensive really
392. 1:13:33fast. So I so I now kind of understand why Altman was talking about universal basic compute back in the day because
393. 1:13:43let's say that the government or or these companies say we're going to give every person you know who signs up for
394. 1:13:50our service a billion tokens a month and you can you can do what you want with them. You can use them in with our
395. 1:13:57agentic tools. You can use them for just things. You can sell those tokens. So someone So there could be a
396. 1:14:03marketplace where you buy and sell tokens. One of the things that's happening right now, this is insane to me,
397. 1:14:10corporations that are in the no right now are pre- buying tokens for their employees. They're they're pre-reserving them in one, two,
398. 1:14:21and threeear increments guaranteed by Google to be able to to give their employees more tokens to go do more work.
399. 1:14:31the companies that are not doing that that are just like, "Yeah, we'll just keep paying as we go." There's going to
400. 1:14:37be a point at which there's not going to be any more tokens to sell. And so the the free market for tokens is going to
401. 1:14:45go is going to go through the roof. So anyway, so it's so we're we're in crazy ass times coming up. Um but but my point there is don't just
402. 1:14:55do agents to do agents unless you've got something specific you want to do. um this is my problem with it right now. I
403. 1:15:02don't have a real specific problem to solve and so I tend to just be doing little dicky kind of experiments that
404. 1:15:12are just kind of expensive and overly complicated and I don't know it's it like it I I don't know unless you have a
405. 1:15:20real thing that you want to do over here on the on the on the uh agentic side it it can it can just burn through a lot of
406. 1:15:28money. All right, let's go look at this. From AI tools to agentic platforms, the tool era, the goopy middle, where we are now,
407. 1:15:38transformation in progress, the agentic era, skills, memory, MCP, set direction, agents and sub aents, tools used autonomously,
408. 1:15:50media, analytics, calendar, canvas, sigma. Okay, whatever. All right. All right. So, let me I'll share this one, too.
409. 1:16:05Go to community feed. Um, here's another sharable graphic about the world
410. 1:16:22we are leaving. Humans use discrete AI tools and
411. 1:16:38the world we are entering. Autonomous agents working 247 on your behalf.
412. 1:16:55Oh, and then I've got to save that as a JPEG because Mighty Network sucks. [laughter] Oh, Mighty Networks.
413. 1:17:11Okay, this one. Yes. Open. Okay. Post. notify. All right. What do we got going
414. 1:17:25over here in comments? Pictures worth a thousand words. Different learning styles. That's all. Yes. Some of us
415. 1:17:29understand images better than just words. Others might not benefit from it at all just looking at it.
416. 1:17:3750 pro years. It works for me. It's not for everyone. Admin SEO of my three websites. Digital gods. UBI would set us
417. 1:17:45free. Tokens will keep us enslaved and working for the machine. I don't I don't necessarily agree with that. Digital
418. 1:17:51gods. If I mean, actually, digital gods, here's here's a thing you could create that's probably probably valuable.
419. 1:18:00Actually, this is not a bad idea. Here's this is one of those you heard it here first. Vibe code up a token marketplace
420. 1:18:08where you can you can sell your tokens. You could pre-by tokens there. Yeah, there's a token arbitrage play
421. 1:18:17um to build right now that that could be really interesting. I don't I don't know how you acquire them and price them, but
422. 1:18:25I have a feeling uh I have a feeling well I actually do because I think I think it's Google's the one that just
423. 1:18:33announced you can buy blocks of one year, two year or threeyear set amount of tokens at at a price today. I need to manage my six emails.
424. 1:18:45Um but anyway, if it's easy enough to sell tokens, then tokens equal money. So that would that would be UBI, but to the
425. 1:18:53extent that people have to figure that out, no one's going to figure it out. So just send them a check. Okay. Um and
426. 1:18:59maybe that check can, you know, may maybe it's like if you use your UBI check for food, your dollar equals a
427. 1:19:05dollar, but if you use it for tokens, your dollar equals equals a $150. So you could encourage people to use their
428. 1:19:12money to buy more tokens um by incenting them to do that. That could be kind of interesting. Okay, one final graphic to make. So, let's go.
429. 1:19:25Let's go back. I think I want to start a new chat. Maybe I want to start a new project, as a matter of fact. So, let's see.
430. 1:19:39Harness an app distinction. Let's say move to a project, new project. And we're going to call this Understanding agentic platforms
431. 1:19:59create project. And so we've got that in there. And I think there's one other thing I did. Let's see. Did you guys hear about the Verve the
432. 1:20:12the Eli Liy thing? The Verve 102 phase one clinical trial of the LDL cholesterol shot. One shot 62% average
433. 1:20:24reduction in LDL cholesterol permanent. It's in phase one, so it's not out yet, so it doesn't really matter. But
434. 1:20:32okay, here's move to project that did did they vibe code the vaccine? Well, it's it's a it's a it's an M mRNA.
435. 1:20:52It's a gene editing shot. So, so it it goes in and it it actually um edits a gene in your liver um to change um how
436. 1:21:02it either processes or produces LDL. I think processes it. Um yeah, I know it's crazy. Yeah. Verve 2. Verve-102.
437. 1:21:16The 101 version of it was effective, but it killed a guy, so that's a pretty bad side effect. [laughter]
438. 1:21:23And um the the 102 one there haven't been major side effects, but it's only phase one. And if if you're in drugs,
439. 1:21:30they've got to go through three phases. Phase three, one, two, and three. Um but the FDA fasttracked uh fasttracked it.
440. 1:21:37So who who knows? Um the other the other question I have about a cure uh for something like that is they're they're
441. 1:21:44going to charge a lot of money. If they only get to give the shot once, I don't think that thing's going to be cheap. All right. Um, okay.
442. 1:21:58I want to understand the highlevel components that these agentic platforms use. Period. For example, Open Claw
443. 1:22:12allows you to connect different AI models, both commercial and local. Period. It's got a memory system.
444. 1:22:29It's got a file system. It's got tool use which it also calls skills.
445. 1:22:50Period. Claude code has similar features and distinctly uses skills
446. 1:23:06which are distinct from MCP connectors which allow you to effectively have API access without APIs. Period.
447. 1:23:25And then there are probably some other components that I'm missing. Period. Lovable just added sub aents today.
448. 1:23:36Period. Codeex seems to be focusing on local computer use and saving all of the memory and skills files locally on a
449. 1:23:50computer rather than in the cloud like Claude Code does. Some of the more codesp specific platforms tie into things like GitHub
450. 1:24:08where the ones that are focused more on computer use may not do that. So can you research what are the different platforms
451. 1:24:25that provide these autonomous agentic features and then tell me what are the components that are common across all of them.
452. 1:24:40Period. I'm trying to get a simplified understanding of what seems like a complicated ecosystem. Okay,
453. 1:24:57let's see. I'm bothered by the token marketplace idea. Good. I've created controversy. Awesome. Wait, let me see
454. 1:25:05what what Mary Carol says. Play it out, Kyle. If tokens get more and more expensive such that only very wealthy big
455. 1:25:15businesses can afford them, then small entrepreneurs that are coming into existence may now be forced out of business
456. 1:25:24and unemployed again in the next few years. So then we have mass unemployment. I don't I I don't think
457. 1:25:30it's that dramatic like like Open AI. So here's the deal. If you if you only had one player, then I would think that would be a
458. 1:25:41concern. But you've got four and probably five, right? You've got OpenAI, Anthropic, um XAI, um Google, and then at some point Meta
459. 1:25:53is going to get back in the game. Um then you've also got Apple and AMD and Nvidia producing hardware to be able to produce
460. 1:26:05tokens locally. this you're you the thing that bothers you Mary Carol is one of the reasons that I think it's
461. 1:26:12important right now for all of us as individuals to understand what it means to run a local large language model on
462. 1:26:19on hardware that can run powerful models and they're getting cheaper like the new AMD machine that just came out that's
463. 1:26:26got 128 gigs of unified RAM it's like their their little version of the Nvidia box and it's their version of the Mac
464. 1:26:33Mini um is like between two and three grand. Um, and it's got 128 gigs of RAM. So, so that so you're going to be able to run tokens.
465. 1:26:46The the the stuff where I'm talking about where people are pre- buying things are big corporations buying big
466. 1:26:51blocks of of tokens. What's going to happen is those companies are going to have a strategic competitive advantage against
467. 1:27:01any company in their sector that didn't pre-reserve those. I think there are going to be massive blocks of tokens reserved for individuals.
468. 1:27:11Um, but [laughter] you know that that's why it's important that there's multiple players in the marketplace because
469. 1:27:19someone's going to go go full-on after enterprise and they will probably just take all their tokens and sell them to
470. 1:27:25enterprise. There's going to be others that go, "Oh, we've got 900 million users over here. You know, some some
471. 1:27:31percentage of them, maybe 50 million of those 900 million users are going to need tokens and so we're going to
472. 1:27:37prioritize them because we think we can make money on them in other ways." So, I don't think it's going to be as bleak as
473. 1:27:43you just pointed it, but now's this time to start getting educated about this that that tokens are power, tokens are money, tokens are
474. 1:27:54power, tokens are value creation, um, in this new world. and those that understand that and figure out ways to to either generate
475. 1:28:06their own tokens on their own hardware um or find ways to to get them cheaper than other people will have an
476. 1:28:13advantage. The other things this Mary Carol carol and and for everyone frankly is um right now as as consumers listen look
477. 1:28:26we're moving out of we're mo moving out of this tool era right where we pay 20 bucks a month and you know some super users you know they
478. 1:28:36they get rate limited out of claude and chatpt all the time because they're using these tools a lot most people
479. 1:28:43aren't using these tools a lot. So, but we're just using we're just dabbling in in token, right? Because we're using
480. 1:28:49tools as as human beings with these discrete tools and we're really not using a lot of them. The minute it gets
481. 1:28:56easier to use these agentic platforms, the agentic platforms are just going to burn billions of tokens per user.
482. 1:29:05[laughter] They're just they're just going to go burn them. Now, at some point, the the these Agentic platforms
483. 1:29:10are going to get more efficient in how they use tokens and how they manage memory, but right now they're a
484. 1:29:15disaster. They're just out there just burning through tokens, right? And so, this is again why we're in the goopy
485. 1:29:23middle right now. But if if all of us can get our heads around this shift that's coming, this is this is a big [ __ ] deal,
486. 1:29:32right? This isn't about learning prompting. This is about understanding that your ability to produce, you know, value, which equals money in
487. 1:29:41the future, um, if you're going to be playing in in knowledge work at all, and it's going to be AI mediated, you're
488. 1:29:49going to be doing it through these agentic platforms. You're going to need access to tokens. So, if you can have a
489. 1:29:53token generator on your desktop that's that all you're paying for is electricity, that's a really good thing in that future. So,
490. 1:30:02hope that makes sense. Okay, let's see. I'll map the ecosystem. What's it doing here? [sighs] The punch line tools let agents
491. 1:30:14do. Okay, I know what tools are. Okay, the component the components common across the ecosystem. The model layer,
492. 1:30:24the agent runtime orchestration loop, the instruction layer. Okay. the memory layer, the workspace file
493. 1:30:36system layer, the tool layer, the MCP connected layer, the skills layer, the plug-in packaging layer, the hooks life,
494. 1:30:45oh my god, life cycle automation layer, sub aents, parallel workers, permission sandboxing, human approval, context management,
495. 1:31:00observability logs. Okay. Um, what did I miss? Okay, hang on. Okay, so let's see. This is already too complicated because
496. 1:31:11I've got 15 things I need to manage in my mind. So, I'd like you to break those 15 things into three or four categories
497. 1:31:21that can be more easily represented in a infographic that's got the subcomponents within each category.
498. 1:31:32I know this is fun and all. Okay, it's nine o'clock people. Let me Okay, the brain here. This is good. The brain, the memory, and the
499. 1:31:41context, the hands, the manager, the cleanest infographic. Okay, great. Okay, so here's what we're going to do. That's the simple version.
500. 1:31:54Terrific. Now turn this into a 16 by9 visually stunning infographic that looks like it was designed by an award-winning designer.
501. 1:32:10Period. The information design is as impeccable as the aesthetic design. And even someone who hates infographics
502. 1:32:21should be able to immediately grock this and get value from it. [laughter] That one's for Andy. [laughter] I don't know if he's still here.
503. 1:32:37Oh man. Let's see. I'm about to cancel Amazon Prime. Maybe there'll be the new abandoned property websites where you
504. 1:32:44can buy tokens cheaply. There's going to be things like that, Lori. Inference boxes at your house, like
505. 1:32:51cable box service in your neighborhood. That's going to happen, too. There's that that company that's attaching, you
506. 1:32:58know, $250,000 worth of GPUs to the side of your house. I I I would I hope they have good insurance policies for that one. But I I
507. 1:33:11want to go learn about that one because if if you're donating the wall of your house to to let them park their micro
508. 1:33:19data center, you should be able to tap into that micro data center and get yourself some free tokens. Um, you know,
509. 1:33:27that's what I'm thinking. If they're using your your spare electricity that that your house isn't using, um,
510. 1:33:35you should you should be able to benefit from that. So, I assume those those things have that. All right. Uh
511. 1:33:41we're going to get out of here. So, don't forget next. So, let's see. Today's Thursday. If you're in the mastermind,
512. 1:33:48tomorrow is the AI salon uh mastermind practice lab with myself and Sid Harrow. Um and and all the people in the
513. 1:33:55mastermind that are part of the practice. Today's Wednesday. Yeah. So, tomorrow's Thursday. So, that's
514. 1:33:59mastermind practice lab. Oh, what did I say? If I said something different, that's bad. Um Oh, this is actually not
515. 1:34:06bad. I don't I don't like the look and feel of this, but let's go look at this quickly and then I'll Okay, it didn't put the AI salon there.
516. 1:34:16Four layers of agentic. The brain AI models local and cloud LLMs model routing the hands. What can it touch? This is not bad. Brain equals
517. 1:34:25intelligence. Memory equals context. Hand equals actions. Manager equals control. Model is only one port. The
518. 1:34:31platform adds memory tool structure guard rails. Memory and context. the agentic platform, not just a model. Okay.
519. 1:34:40Um maybe on Friday we'll I'm I'm not going to share this one right now because th this needs this needs a little bit of help to be useful.
520. 1:34:49Ah, [ __ ] it. I'll sh I'll share it. I'll share it right now. Um and then we'll get out of here and then
521. 1:34:55and then I think I'm going to do an updated version of that because I think we need we need one. Open reset. What's this? There we go.
522. 1:35:13Export. All right. My screen share lasted the whole thing. Okay. Um here are the components
523. 1:35:29of a gentic platforms like claude code slashcowwork and codeex
524. 1:35:43etc. Um components Uh um
525. 1:35:57there will be a new version of this, but here's a start. [sighs] [gasps]
526. 1:36:14And do me a favor, everyone, if you'd be so kind, if you use and share any of these graphics, if they're valuable to you, please point
527. 1:36:24people back to the AI salon. Again, part of my my goal is I want to get people I like things are getting intense right now.
528. 1:36:36The resistance to AI is getting intense right now. The complexity of AI is getting intense right now and everyone,
529. 1:36:45myself included, is fe is feeling this pressure to back out of the conversation. I think we need to stay in the
530. 1:36:53conversation. Um be probably more than ever. So, um if we can get more people in the conversation, that would be
531. 1:36:59great. So, please just just share what we're up to. All right, Groovy. Fantastic. I hope you all had a swell evening. um
532. 1:37:15understanding where we are right now. It's pretty intense. All right. I need advice. Looking to create an agent. I I would
533. 1:37:27just pick one, Ivan. Pick Claude Claude Code or uh pick pick the Claude application or the or the Codeex one.
534. 1:37:36Those are those are the two main ones. I'd just pick one of them. Don't don't bounce between these. Just learn one. Um,
535. 1:37:45I'd probably go with OpenAI with Codeex. [sighs] Thanks, Kyle. Thanks, Brandon. I salute you. It's been a lovely three and a half
536. 1:37:54hours with you. Thank you, Mary Carol. >> [laughter] >> the uh you irregulars, you've got more stamina than than us people on the
537. 1:38:03screen uh paying attention to this stuff and hanging in there. All right, everyone. Have a good night. Thank you
538. 1:38:09all for your time and attention and um grace. Um you're all very generous and I appreciate it. Bye.