6/10/2026 - Why You Need to Stop Prompting and Start Designing AI Loops

Transcript

1. 0:01[music] [music] [music]
2. 0:19[music] [music] Good evening, good people. Happy
3. 0:35Wednesday. It is Wednesday, June 10th, 2026. Happy Wednesday. Um, welcome to the AI learning lab. Tonight
4. 0:44we're going to do some stuff, but it's going to be it's going to be slightly altered. This week's going to be weird.
5. 0:51So, we have a new model for from Anthropic called Fable, which will be I'll show you some stuff I've built with
6. 0:57it. We'll probably build something with it. Um, see what it can do. Um, think about what you'd like it to build.
7. 1:07Maybe we do like a video game. Um, I'll show you one that I made. Maybe we do something else. So, if you've got
8. 1:15ideas for for what you want to what you want to build, think about that. So, we're going to talk about that. We're going to also talk about
9. 1:24um Dario Amade's letter, open letter to the essentially to the government saying regulate us please create an FAA like uh organization for
10. 1:40AI. Um so we we'll be talking about that and then I promised we would talk about loops. I told you on Monday, if you're
11. 1:50prompting and not looping, you're doing it wrong. Like, you will be judged. You will be judged. Yeah, looping. I don't even know how to
12. 2:02prompt. Perfect. You're right where you need to be. So, I'll talk a little bit about loops tonight, but there's other
13. 2:08[ __ ] to talk about. Anyway, welcome. Uh, my name is Kyle Shannon. This is the AI learning lab. If you want to uh to open
14. 2:16the chat, you can say hi and you can ask questions and stuff like that in the chat. It's underneath this logo right
15. 2:21here. Okay, we got producer Brandon in the house. He just got he's hot off his his lovable five out of four [laughter]
16. 2:31uh session. Uh and then in the lower corner here, if you want to do advanced AI learning lab, it kind of looks like
17. 2:38this. You're going to create two tabs. In the one tab, you're going to have this stream and down here is a button to
18. 2:44pop this window out. So, in the one tab, you have this window. You pop it out. And then in the other tab, you have the
19. 2:53AI salon and you can go look at all the stuff. Go look at the different spaces. You can comment in the AI learning lab
20. 2:59space. You can go introduce yourself in the introduce yourself space. Um, it is the the the AI salon is not just a
21. 3:09place where we store old recordings and things like that. It's a vibrant, you know, engaged community. And so, go find
22. 3:16those spaces and there's different people with different interests in those different spaces. So, go connect. Is
23. 3:23anyone here? Um, no one has commented yet. So, do me a favor and comment. Let me know you're here. Say hi. And uh
24. 3:32Brandon, hey Andy, what's happening? Brandon, why don't you hop up on stage here? How was five out of four uh lovable marathon?
25. 3:43>> We we probably could have gone for another five. There's that much happening in the lovable space. Um you
26. 3:50know, we we went down rabbit holes. It was great. The entire fiveweek course is for m on demand for any masterminds. But
27. 3:57what astounds me is that we had um we looked at about six different projects completely different tangents like
28. 4:06people are building apps for health care people are building uh apps for self-improvement and self-care uh I'm
29. 4:14building routine based things for my kids everybody doing their own version and and it really brings that idea of
30. 4:23the bringing the ideas you know the cost of creation has gone to almost zero because in just a few prompts you can
31. 4:30get a very viable MVP and then we've talked over the past five weeks about how to go from that idea to a fully
32. 4:37formed web app which uh as of about four minutes ago >> I have and so I'd like to introduce you to uh Happy Habit Adventures. Uh this is
33. 4:51>> after we've logged in. So uh you get one for free and I'll drop um this link in for anybody who wants to play with it.
34. 5:01But essentially it's a habit tracker for kids. So you can choose morning quest and evening quest. Uh and then here's
35. 5:09the five things that you want to do. And parents can customize this. So we have a parent portal here that allows you to
36. 5:16add tasks or reorder the tasks. Oh, that's so cool. >> You have morning and bedtime. And then at bedtime, after you finish your five
37. 5:25tasks, you have a bedtime story. And when you go to do the bedtime story, it asks you, "What was your favorite part
38. 5:31of the day? Who did you hang out with today?" And then it takes all of that and synthesizes it into a story that you
39. 5:39could read or your character, your avatar reads you that story. uh and it sends all the stuff off to 11 labs and
40. 5:48comes back. And then what the unlock was today was for premium users. Uh Lovable has made it really easy for Stripe.
41. 5:57They've got a partnership with Stripe where when I was building my Vibe card, my Vive greeting website, I had to
42. 6:03actually go to like uh Stripe and set everything up. Now >> get the API key. Copy it. >> Naomi snuggled deep into her bed, the
43. 6:15soft blankets feeling like a warm hug. >> And so, you know, it goes on like that. But now with lovable payments, literally
44. 6:24all you have to do is say add payments and it'll work through the back end. It'll add all the product codes to
45. 6:28Stripe. It'll synthesize with Stripe and it'll even give you a test credit card that you can use. So you can see, I
46. 6:34don't know, it's really tiny on the screen, but you can see payments are in test mode. You can use card 424242
47. 6:39to test out your premium workflow without actually paying, which would have been nice to have when I was
48. 6:45building Vibe greeting because I ended up charging over $100 to my personal credit card in testing. Uh, so
49. 6:54>> you had to refund yourself and you held it for five days. [laughter] >> Exactly. But the idea is that we're
50. 7:03we're really at the point where you can just build anything. Um, and it's really impressive. I was really so impressed
51. 7:10with what everybody has built in there. >> Very, very cool. So, so that is part of Do me a favor, pop up the mastermind
52. 7:18graphic if you would. Um, if if you're uh if you're new to the salon and you haven't joined the mastermind, so the
53. 7:25course that Brandon just described, there's there's a uh >> a great repurpose workshop that's an ongoing thing that Andy Scarantino is
54. 7:33running. We're going to start doing some um beginner stuff uh for for people just getting started. Um, we've got the the
55. 7:43mastermind practice lab which actually meets tomorrow morning at 10:00 a.m. Mountain time. Um, so if you're not part
56. 7:51of the mastermind, join the mastermind. There's all sorts of good stuff going on in there. Um, this is we're we're at an inflection
57. 8:01point where two things are going to happen simultaneously. Thank you, Brandon, by the way. Um, I've been thinking of these AI
58. 8:15capabilities getting really good as like this exciting, oh, we get to play with these toys that are really powerful for a while. Um,
59. 8:25they're going to get so good so fast that they're going to be able to start doing things that we can't even anticipate like weird things and
60. 8:36powerful things and things that really do disrupt things, right? So, part of dealing with AI right now is dealing
61. 8:48with the AI readiness piece, right? what is going on? What should we pay attention to? Right? So, a big part of
62. 8:56this is that, but the other piece is the human readiness piece, the are we ready? I don't think we are. Like part of what
63. 9:04I'm going to talk about tonight is is Dario Amade's um it's it's it's a it's a fascinating it's a fascinating piece and
64. 9:20you could you can look at it through enough lenses that you could see it as something um that absolutely goes too far. You
65. 9:32could see it as something that strikes the right balance and you could see it as something that doesn't go nearly far
66. 9:37enough and and that's kind of how people are talking about it right now. They're kind of screaming at it. Um
67. 9:45so so that's happening. Then yesterday the new model from Anthropic called Fable came out. Fable is the same model as mythos. So,
68. 9:59Mytho Mythos came out a month and a half or so ago and and Anthropic basically said this thing can exploit all the
69. 10:07security holes in all of the software. So, we're we're going to go give this to, you know, basically financial services, healthcare, we're
70. 10:16going to go give this to infrastructure level companies to go patch all their software. So they've done that for six weeks
71. 10:24and they created what they're calling a basically a consumer fa safe version of fable. It cannot do anything in involving cyber security.
72. 10:37It cannot do anything involving health care and it cannot do anything involving one other thing. I don't think it was
73. 10:46financial services. I don't know. There's three major things, Kelly Camp, I've never been ready for the speed of
74. 10:53AI. And and to be clear, Kelly, I don't like we haven't hit the speed of AI yet. We haven't hit the speed like like it's
75. 11:06insane as fast as it's going. Biology. Yeah, that was it. Oh, yeah. Biology. Well, no, wait. Is biology crypto or cyber?
76. 11:18I guess is it healthcare? I don't know. We can we can go we can go look. It doesn't really matter. Um people who are using fable who are not
77. 11:29in those fields are like this is cool. This seems powerful. People who are in those fields are absolutely [ __ ] livid. It is it it is a
78. 11:42and this is where I mean I think we're at a place where these frontier model companies can do no good. So, so
79. 11:48basically they come up with a a big model that could be used to create say boweapons or hack all the systems and
80. 11:55crash all the financial markets. So they've got something that powerful. So they tell the world that we've got this
81. 12:01thing and and we're going to give access to people that are doing infrastructure so they can go shore up their their systems.
82. 12:11And then they launch a thing that says we're going to take the the highest risk vectors And on day one, we're going to
83. 12:18have the the the the guard rails really tight on those. And if they had just released the powerful one and a bunch of people went
84. 12:28and started hacking [ __ ] and started making boweapons, then they would be getting [ __ ] on for being not safe
85. 12:35enough. And now that they put out things with these safety things on them, people are like, "Oh, it's going to be central
86. 12:41control and you're creating feudalism and thftdoms and [laughter] I, you know, if you've got something that powerful, I don't know how you
87. 12:52launch it and and meet all the stakeholders. You can't, right? You can't." Um, and then so Daario's letter
88. 13:02today basically says, "This shit's coming and it's coming really fast and the government isn't prepared for it.
89. 13:10Therefore, the government should slow us down." [laughter] All right. No Emmy just got back from my library workshop. Awesome. The audience
90. 13:20was 60 plus. No one is staying behind. No Emmy, that's really beautiful. Really, really beautiful. Um,
91. 13:28congratulations. I'm so excited for you. So, one of the things that we're putting together right now in the AI salon
92. 13:35is an ambassador program. And the ambassador program is, you know, folks in the community um, who have are
93. 13:43getting their heads around this stuff um, that want to do things like what Nomi just did. Um, we'll provide
94. 13:51support. We'll provide, you know, communications, uh, goals, targets, things like that. Um, for people to go out in communities
95. 14:01where people want to learn about AI, like say a library, a knowledge institution, opening its arms to knowledge to a to a knowledge
96. 14:12machine. It's fantastic. Wall Street Journal just dropped a story where Open AR considering drastic price
97. 14:20cuts to win customers from Anthropic. Well, so okay, so here's another Yeah. So we are in wild times. First of all,
98. 14:29no Emmy, congratulations, Travis. Thank you for this. Um, there's always going to be a bioteterrorist who's mad at the
99. 14:37guard rails. [laughter] No. Uh, let's see. Thank you. I even got into a hot debate about energy consumption and personal integrity um in
100. 14:48the face of destroying the planet for the sake of AI. Um yep. Listen, one of the things that we talk about in
101. 14:56here all the time and we talk about in the salon all the time is how do you have those conversations? People that are afraid of AI, people
102. 15:03that are mad about AI often haven't used it. But at the same time, a lot of the things that people are mad about or afraid about
103. 15:13are real things. You know, my philosophy and and I think our philosophy in general in the AI salon is and
104. 15:23yes, you you you have a right to be pissed off on how these models were trained. You have a right to be pissed
105. 15:29off that they're going to take a lot of energy. You have a right to be pissed off about um job disruption and and how
106. 15:36everything's going to change and how the way things used to work don't work that way anymore. You have the right to be
107. 15:42pissed off about all of those. And AI is not going away. So if it's not going away, what do we do? Well, we talk we talk it through. We say,
108. 15:56[laughter] you know, this shit's here. It's not going away. So, what do we do about it? Is there anything we can do to steer
109. 16:04those things that are distasteful to us in a better direction? Probably get involved with policy and advoc advocacy,
110. 16:11right? We've got Daisy Thomas who runs that for us. You've got all sorts of options. If you want to go talk to
111. 16:16government officials and tell them to get their heads out out of their asses, you can do that. You can just take a
112. 16:22personal stance and like like no and say, "Hey, I'm just going to go teach people what I know. I know this has
113. 16:26changed my life. I'm going to go share what it's done for me. Like personal stories make a huge difference because a
114. 16:34lot of the people that are pissed off about AI have never used it. Like I get that you're pissed off. That's like being pissed off that the
115. 16:47big wave is coming. You're on the beach. You're like that wave looks big. There shouldn't be big waves. Okay.
116. 16:56I agree. Rogue Waves, I think we can all agree Rogue Waves should be right out. In fact, if you're a Rogue Wave fan,
117. 17:09go to your favorite image generation tool and and and make a Rogue Wave image about that. It's a metaphor for AI and
118. 17:18have a bunch of people on a boat screaming at the wave. and then put that in put that in the uh the AI learning lab channel. We'll go
119. 17:28look at your rogue wave things later. Producer Brand Brandon landmark German ruling declares Google's Google's AI
120. 17:37overviews are Google's own words and makes it liable for false answers. Okay, great. Europe. So, so Europe is going to regulate
121. 17:53itself into not being able to use AI. [laughter] Okay. All right. I was on Google and I asked it to get me a a lawn care service
122. 18:05and that lawn care service wasn't in business anymore. I'm suing. All right. Whatever. Whatever. Uh Mr. It here. What's happening? I see
123. 18:19a blue wave coming in November. Okay. So, let's see what we're gonna do. So, what's a loop? Okay.
124. 18:37Let me share my screen. Let's talk loops and then we'll get to the more exciting stuff. I'm not going to show you I'm not
125. 18:43going to dig into this. I was going to sort of do a whole demo on this, but I'll just read through some of the stuff that I put together.
126. 18:59By the way, if you're a Rush fan, Rush had their first their new tour is starting with the new Anukica, I forget
127. 19:08her last name, Anukica something, um, female drummer that replaces Neil Per. They had their first concert last night
128. 19:15and it looks like she just absolutely kicked ass and destroyed it. She seemed to have it more together than Getty Lee
129. 19:23did, but it's the first it's the first night of a of a big tour, so that's pretty cool. Okay, let me get my let me pop out comments here.
130. 19:47Vickiy's in the house. All right. What? Why did that not pop out correctly? Entering studio loading. Okay, there we go. That worked.
131. 19:59Okay, so I got comments over here. Okay, I am going to go over to Can you see this? Yes, it looked like Yes,
132. 20:13I see chat GPT. Okay, perfect. So, let me get rid of that. Make this bigger. Okay, so we're going to talk about loops. What
133. 20:24are loops? Why are we talking about loops? We We just did prompt engineering. Now they're talking about loop engineering.
134. 20:37Why we're talking about loops is that this guy Peter Steinberger put out a uh a tweet that said, "Here's your monthly
135. 20:50reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your
136. 20:55agents." And then the guy that invented Claude Code within a day or two of this, I assume they went out and had a a mocha
137. 21:04mocha latte together and the guy that created Claude Code and Claude Co-work, the guy that invented that put out a
138. 21:11very similar tweet. And so all of Twitter is a buzz about loop engineering. And there's all these posts
139. 21:19about loops and things like that. So I went to chat GBT and I said, "Help me do this." Um, okay. What Peter means is stop treating Codeex
140. 21:30like a genius intern you give a a clever one-off prompt to and start treating codeex like a worker inside a repeatable
141. 21:37production system. So, here's the difference. If you're if you're just on chat GPT or if you're just on Claude,
142. 21:46that's that's kind of the the metaphorical equivalent of a vending machine. You put in a prompt, you get
143. 21:51out an answer. You look at that answer. If you like it, you use it. If you don't, you put in another prompt, get
144. 21:57back another answer. And so, you're very much there prompting. Then along come these things like codeex and like claude code
145. 22:06which look very similar to chat GPT but they're functionally very different because they have access to tools
146. 22:14like they can go surf the web and they can go into your Google drive and look at your Google Drive things and they can
147. 22:21just do all sorts of crazy stuff and use thirdparty software applications. They also have computer use, meaning they can
148. 22:30use your computer. Codeex in particular is really good at this. Claude Code and Claude Co-work are really good at this.
149. 22:37They can go into applications on your computer or folders on your computer and they can put files in them and take
150. 22:42files out. That's what makes them a little scary because if they go off the rails and say, "Hey, good news. I
151. 22:49decided to delete all of your documents because they were way too confusing. So, I figured we'd start over. Thumbs up. You're like, "No."
152. 23:00Right. So, these they're these more sophisticated things and they can run 24 hours a day and they can they have
153. 23:11what's called an orchestrator agent. So the one you talk to and then that orchestrator agent can spin up sub aents
154. 23:18so other agents that are doing work for the agent that you're talking to. You can just put in a prompt into those things just like chat GBT.
155. 23:31But a loop is not that. So a prompt is build this feature. A loop is take an issue, inspect the code base, make a plan, write or update
156. 23:43tests, implement the smallest change, run checks, fix failures, self-review the diff, the difference, document what
157. 23:50changed, and stop only when the acceptance criteria pass. Right? So, as as Andy put it, isn't that just a prompt
158. 23:59full of subprompts? Yes, it is. Um, but you can have these loops. You can actually name these loops and then you
159. 24:06can just say to codeex, hey, go do that that um research loop that we created on this particular topic and then it will
160. 24:15just go start doing all this stuff. It said this matters because codeex is already built around an agent loop.
161. 24:24You've got the model. It can use tools. It can edit code. It can interact with the terminal which basically means it
162. 24:31can use stuff on your computer. It can run tests and then it can do more reasoning and that just keeps looping.
163. 24:37OpenAI describes the codeex harness. You you might hear this term harness a lot. The harness is all of this its ability
164. 24:46to do all these things. So the harness takes your input and figures out all the different things you might want to do
165. 24:53with it. Creates a plan basically and then goes and executes that plan. You don't have to be there for that. So
166. 25:01designing the loops is designing those steps or even having chat GPT or claude help you design those steps for
167. 25:08different things that you want to accomplish. OpenAI describes the codeex harness as the logic that orchestrates
168. 25:15the user model and tools so it can perform software work. The tactical opportunity is to design the loop around
169. 25:23codecs not just the sentence you type into it. what it means for you tactically. The move is from codeex make
170. 25:31me a fe feature to col codeex follow my repos my the repo is where you store your code and your instructions. Um
171. 25:39follow my repos feature delivery loop. Right? So you would design a feature delivery loop. That loop should live in
172. 25:47your repo not in your memory. Right? So again, in order to use these agentic systems, well, there's a lot of concepts
173. 25:55that need to be bundled together that we all need to understand to be able to use these things. Well, um,
174. 26:05a good way to think about the value of learning these agentic platforms is imagine in when did Excel come out or or
175. 26:16Lotus 123? that came out in the 70s, 80s, let's call it the 80s. Imagine being the only person in a Fortune 500 company when spreadsheets
176. 26:29first came out. You're the only person that understood how to use a spreadsheet in that big company. You'd be a very valuable person to that
177. 26:40company. That's what's going to happen with these agentic platforms that people are going to get really good at using these.
178. 26:48Does it mean that if you don't learn to use these, if you don't run down this rabbit hole that you can't get
179. 26:55tremendous value out of just prompting chat GPT or Claude? No, you can absolutely do that. This is not a requirement.
180. 27:03But this world, if you remember my seven economies um [cough] chart, these agentic platforms are economy 7.
181. 27:14This is the economy where it's like a single a single person operator could have a business running that's got 20 or
182. 27:2150 or 100 agents working 247 for a single individual and all of those agents are running a business for that
183. 27:30individual. Am I still sharing by the way? I'm not. When did that go away? Oh, it just stopped. Interesting. Okay.
184. 27:52Okay. So, your job becomes to make the repo prompt codeex comment on screen. Let's go look at that. The harness matters
185. 28:03same chat GPT wait the harness matters same chat GPT co-pilot and Codex co-pilot makes it feel like GPT is 12 months behind Claude Codex
186. 28:15makes it feel three months behind Yeah. Yeah. The the Yeah. C-pilots co-pilot's a disaster. So, so the the harness that thing that
187. 28:28allows you to do all this stepbystep stuff in the case of open AAI it's called codeex in the case of Microsoft
188. 28:36it's called co-pilot case of anthropic it's called uh code claude code and claude co-work in the case of um Google it's called anti-gravity
189. 28:52and and they've got a new uh front end for it. I think it's called Spark. Is Spark right? It might not be Spark.
190. 28:59Might be something else, but anyway, anti-gravity and whatever their front end thing is going to be called. Spark.
191. 29:04Okay. Um and then um XAI has one called Builder. So, so Grock Grock build or Grock Builder. So, so those are kind of
192. 29:16the five major flavors of this. Then you have Perplexity Computer, you've got Manis, you've got um Gen Spark.
193. 29:26So you start to got a lot you you start to have a lot of these systems that are all doing similar things but but they
194. 29:32are varying degrees of quality because you've got more and more sophisticated like when when when you've got an orchestrator
195. 29:40that's orchestrating all these different piece parts from different applications to different websites to different login
196. 29:47to different data sources to different memory systems to using your file access and how what you're giving it permission
197. 29:55to things like that. It's not trivial to set one of these up. So unlike the early days of the AI learning lab where we
198. 30:04could bounce around from every large language model to the other, you can't really do that with with these platforms
199. 30:14because they're expensive. They use a lot more tokens, right? You as an individual, you can only type so many
200. 30:21things and you know upload so many files. You can give one of these agentic systems. You can put it in a loop to go
201. 30:29say go read, you know, every crime novel ever written and create an application for each one and it'll go off and
202. 30:38[ __ ] do that. And then three weeks later you'll have a $100,000 AI bill. [laughter] So, so you know, you kind of have to be
203. 30:48aware of what's happening in these things, right? And then there's on top of that is the difference between using
204. 30:56a cloud-based model like GPT or Claude or using local LLMs that don't really cost you money other than electricity.
205. 31:05All of that has to be set up. All of that has to be orchestrated. So the quality of the harness um is is really important. Okay. Um, I'm
206. 31:16not going to dig in too much deeper there because I I want to get to um I want to get to some of this other stuff.
207. 31:21I want to go play with Fable. Um, but that's basically it. I would go research this stuff. Just go on to X and and look
208. 31:29up loops. Just put in the word loops and just go read [ __ ] [laughter] and and just start experimenting with
209. 31:36it. You can have ChatGpt or Claude say, you know, you could say, "Hey, I want to learn loops. I want to set up one of
210. 31:43these agentic platforms." Well, help me do that. What are the loops I should start with? Right. One other thing that is is um a a power
211. 31:54user move. Not a lot of people use Grock that I talked to, but here's the thing that Grock has that no one else has.
212. 32:06You could sort of argue that Gemini has it a bit, but Grock really has it. And that's this. Grock. When you search on
213. 32:14Grock, it can instantly search all of X. And most of the AI development developer community is on X,
214. 32:25right? So all of the people building all of these tools, all of the people figuring out how to what loops were in
215. 32:32the first place and then how to use them, they're all on X. So one of the things that you can do if you want the latest up todate like
216. 32:40literally as of today how are people talking about loops you can go to Grock and you can say you know find me the 10 most influential
217. 32:54you know AI experts that are talking about loops that have more than 100 thousand followers and take what they're saying and turn
218. 33:04that into a primer for me and then show me their latest post on these loops that you built that primer out of or
219. 33:11something like that. Um that that could be really valuable. Okay. Am I still sharing my screen as the new
220. 33:20black bar? I if if anyone knows anyone at Apple for how we can get our screen sharing to not just randomly quit, uh
221. 33:27that would be great, Travis. Um, but you can set up a folder with markdowns that are model agnostic and then when you
222. 33:34jump from codeex to cloud code to anti-gravity, you can pick up where you left off from the last. Yes. And I don't know how to do that.
223. 33:44So, buy a PC. That's rude, producer Brandon. Um, no, I will not. I have a PC over here. I'm fine. It's it's sitting
224. 33:52there collecting dust and I like it that way. Um, Travis, that's a good point. So, one of the other things that that
225. 33:58I've been talking a lot about is it it all of a sudden has become obvious and apparent to me that each of us reclaiming
226. 34:13all of our data is really important. like reclaiming our chat history from chat GBT, reclaiming our our chat
227. 34:21history from Claude, from Perplexity, from anything where you've put in a significant amount of time because you can turn all of that into
228. 34:30context that these things can use. And then yeah, Travis, to your point, if you set up a local folder system that you
229. 34:38can point these different harnesses to, then your memory in your context is this single central thing that you can deploy
230. 34:48to different places. Just having that conceptually and even doing a small version of that where you've got here's my data and I'm
231. 34:58connected to these other tools that I use. If you say that to some prospective employer that I understand that concept
232. 35:06and why that's important and here's how I'm using it and here's how I'm thinking about deploying this stuff, that's
233. 35:12actually very valuable. Um, so thank you Travis. That that's a really good point. Okay. Um, so here's what I want to do.
234. 35:25I want you guys have all seen me do um whenever a new model comes out that's good at coding. One of my benchmarks has been um this is
235. 35:38my my version of when new image models come out I I say make me a 70s resto mod in an abandoned factory. On the code
236. 35:45side, I say, "Make me um a clone of the 70s asteroid game, you know, complete with vector graphics and um
237. 35:57and, you know, sounds and physics and all that stuff. The last two or three models have gotten good. Opus 4.6 was really good. Um,
238. 36:07Fable, the new um, Big Mac Daddy from from Anthropic, is the first model that I've seen that made me a clone of
239. 36:16Asteroids that is exactly like Asteroids. Like, it's exactly like it. And so, here it is. Um, if I did command plus, no, didn't do anything.
240. 36:30Oh, and you know what I don't think I did? I don't think I shared audio. Let me just stop this and reshare it. Okay.
241. 36:46So, here's asteroid. So, this this again, this was a single prompt. I didn't change it. And that's that's kind
242. 36:52of my my benchmark is put in the same prompt and see what comes out of it. And like Grock 4.6 or not Grock, uh, Claude
243. 37:024.6, 6 Opus 4.6 nailed it and then Opus 4.7 and Opus 4.8 were horrible. Right. So, and everyone talked about Opus 4.7
244. 37:12and 4.8 went downhill. This is now um um Fable Fable 5. Um this is the asteroids thing. If you remember the game like the the
245. 37:27speed is right with this, the line weights are right. The shape of the asteroids is right. The sound of that
246. 37:34UFO is right. Um, the physics of the ship are right. The It didn't do things like like a lot of te times these
247. 37:42asteroids things would make the son of that that that [ __ ] small spaceship is like the anxiety producing one.
248. 37:59Solid's not bad. You can fly right at that. Um, this is asteroids, right? So, so this is the first time that I've seen I've seen
249. 38:10it produce something that is we got to clear a board. Wouldn't you [ __ ] [laughter] and then I'm done. So, anyway, so so
250. 38:24that was pretty impressive. Okay, so the next one. Am I still sharing? Yes. Okay. The next one is So, one of the things that I've done
251. 38:41I've done it like four or five times with the most recent models is there's a thing in in astronomy called the
252. 38:48threebody problem, which is when you have three gravitational bodies like three stars in in orbit with one
253. 38:54another. there's only a handful of stable configurations where the the they either don't fly off into space or smash into each other. And
254. 39:05so I've done a couple of simulators of this. Um, but I said to Fable, I said, "Make me one of those simulators, but I
255. 39:13want it to be like I'm in the spaceship and then I fly to one of these solar systems that have these, you know, three
256. 39:20three stars at the center of the solar system and and I want to be able to fly around it. And then when I when I select
257. 39:28another um threebody problem, stable stable configuration, I want it to wormhole over to that other thing. And
258. 39:35then I wanted to have cinematic music and things like that. So again, I didn't I gave it a relatively shitty prompt and
259. 39:44just had it go do its thing. And so this is what it produced. So the audio audio is there. [music]
260. 39:58Fly toward it or away from it. On the right hand side here, there's two overview of the stars. Notice how
261. 40:25about now? Is that better? Is that good? That's better. I just turned off the music. Okay. Um, you know, I can I can
262. 40:33chase I can chase any one of the three stars. I can do freef flight where I can just fly around these, you know, this
263. 40:40thing. I can go up and down. Is it F? Yeah. Cool, right? And then over here on the left are the um the different So, this
264. 40:52is the Tricellaris Prime. Here's the Butterfly Nebula. And so, watch. Here's the the wormhole effect. Z.
265. 41:00And now I'm in another thing. This was a single shot. I can speed up time so these things go really fast or I can slow them way down.
266. 41:17Here's another one we're going to. So anyway, that's kind of crazy. So that was that one. And then the last one that I'll show
267. 41:30you. So I just did these were the three things I did in a row. The last one I did is I took a concept
268. 41:40from Andy's workshop on decoupling your identity for your name. So, in the AI salon mastermind, um we have a great purpose, a great
269. 41:50repurpose workshop that's ongoing that started as a 4-week um workshop for people to decouple their identity from their their work and it's
270. 42:02it's now turning into an ongoing thing. So, if you're not part of the mastermind, you should think about
271. 42:06joining the mastermind. But one of the concepts that came in that that that Andy taught us in there that she brought up in there
272. 42:15was it was uh it was episode or it was week three was was the the goop phase, right? So when a caterpillar um metamorphizes metamorphises
273. 42:27metamorphosis metamorphize I don't whatever their word is changes into a caterpillar. You think like, oh,
274. 42:34it goes from caterpillar to butterfly or it changes into a butterfly. It it doesn't. It It's a caterpillar and then
275. 42:41it goes into the chrysalis and it releases all these enzymes and it turns into goop and then from the goop the butterfly is created.
276. 42:51Um, and so that was a really cool concept. So, I went into um Fable the the the new Claude model and I said go to the greatrepurpose.com.
277. 43:03My screen share is in the goop phase. Of course it is. Black bar. Black bar. You're not sharing. Kyle, you're not sharing. Kyle,
278. 43:19you're not sharing. Um, so I basically explained in in awkward English what I wanted the concept to be and then I said, "Go to the great
279. 43:31repurpose.com and um, read the website, look at the design, and I want you to create basically a landing page that
280. 43:41uses this metaphor that can walk someone through it." And so again, I didn't change anything on this. I I would I
281. 43:47would make a fair amount of changes to this. Um, some of the writing's not great, but but this is what it did. So,
282. 43:54everything dissolves, you don't. So, there's our little caterpillar. Caterpillars, butterflies, and the uncomfortable middle where the magic
283. 44:00happens. So, as you scroll this, the the caterpillar stays. You've been a caterpillar your whole life. Caterpillars are magnificent eating
284. 44:08machines. One job done extraordinarily well. Sound familiar? You learned your craft. You got good. You got paid to be
285. 44:16good. And somewhere along the way, what do you do? Quietly became who are you? Right? So, this is taking some some
286. 44:24fairly, you know, direct, you know, uh ideas out of the great repurpose and it's throwing it into this caterpillar
287. 44:32thing. You stop noticing that they weren't different questions. Um then the sky began to promise things.
288. 44:39So notice the caterpillar as we go from the light to the dark. The caterpillar kind of disintegrates.
289. 44:48AI arrived promising a brighter future. And it isn't lying. But promises don't pay attention to timing. The work that
290. 44:55defined you is getting easier for everyone. The phone rings less. The skills that made you feel special, a
291. 45:01little less special each every month. The edges of the familiar world are softening. This is the part no one
292. 45:08prepared you for. Um, before anything gets amplified, something dissolves. And so, notice the caterpillar is now just
293. 45:16this, right? Here's what no one tells you about butterflies. The caterpillar doesn't grow wings. It dissolves inside
294. 45:23the chrysalis. It releases enzymes that digest nearly everything it ever was. The muscle, the appetite, the whole
295. 45:29magnificent eating machine. What's left is goop. Not a smaller caterpillar, not a rough draft of a butterfly. Goop. If
296. 45:38you feel like goop right now, disoriented, dissolved, unrecognizable to yourself, you're not broken, you're
297. 45:44on schedule. This is the this is the cornerstone of the great repurpose. And it, you know, it sort of figured this out and it
298. 45:52sort of does it where, [laughter] you know, you just have this screen where the the goop is just sort of floating
299. 45:57around behind this thing. Um, you know, it made that nice dramatic thing. The goop, the goop is keeping a secret.
300. 46:05Floating in that dissolved world are clusters of cells biologists called imaginal discs, which I think is is like
301. 46:12the most beautiful little term I've ever heard. And and notice now that the these little yellow dots have turned into
302. 46:18these imaginal discs. They've been there since birth. They survived the dissolving. They carry the entire pattern of the butterfly. The
303. 46:27caterpillar always contained This is so good. The caterpillar always contained what it was becoming. So do you. Like that's actually poetic.
304. 46:40Your imaginal discs will never do. Your imaginal discs were never your job title or your output. They're your curiosity,
305. 46:49your taste, your judgment, the way you care, the things you do, even if no one paid you. Right? Like that paragraph
306. 46:57right there. There's the thing. If you're if you're freaked out about your job changing or job loss or all that sort of stuff,
307. 47:05this is what we get the opportunity to to really start to explore. The magic doesn't happen after the goop. It
308. 47:11happens in it. You can clench and wait for the old world to come back or you can lean in and do the one thing the
309. 47:19dissolving makes possible. Meet yourself without your costume. Right? Who are you without your job? The goop has a path
310. 47:26through. And then this it pulled these are the five stages of the great repurpose. So it it pulled these out of
311. 47:32the website. Decouple your identity from your output. Reclaim your human value. Look at our butterfly now. Right. The
312. 47:39goop has a path to go through. Five move movements in this order on purpose. And so there's the there's the butterfly
313. 47:48and there's the five stages. And then it's you amplified. The butterfly isn't a better caterpillar. It's the same life in a new dimension in
314. 47:57the same garden. When your identity is no longer fused to your output, AI stops being a threat of who you are and
315. 48:03becomes wingspan. It amplifies your curiosity, your taste, your care, all of who you are. Um, which is the one thing
316. 48:10you could never replace. And then find your repurpose profile. And this jumps us over to the website where you can
317. 48:18take the uh take the uh the great the repurpose profile uh test assessment. Anyway, cool. Yeah. Um what are your thoughts about all
318. 48:32this? Let me stop sharing. Let's see. Yeah, I I'm gonna I'm gonna talk about
319. 48:47that, too. Okay. Wow. Thank you. Can I play it? It deops into a butterfly. Exactly.
320. 49:04Um, so that I mean that was kind of beautiful, right? Like it's not bad. Like I I would like the animation to be
321. 49:13a bit more sophisticated and, you know, mythical and I'd like the writing to be a bit better, but it ain't horrible. So
322. 49:21those three things. So I did those three things and then when I went back to use Claude, it said you've used up 90% of
323. 49:29your fivehour credits. So couple of things. Fable is very expensive and Mythos is very expensive. It's like $50 per
324. 49:40million tokens and it is a it is a it is a model that chews through tokens because it does lots and lots of
325. 49:48reasoning. I've noticed a couple of things about it. Anthropic on June 22nd is going to you're not going to be able
326. 50:02to use Fable with your $20 a month subscription. So, go use the [ __ ] out of it because right now it's being subsidized. Um,
327. 50:19it's it's going to per token cost. So, so if you if you want to use it after the 22nd, it's going to go to per token
328. 50:26cost. I have a sneaking suspicion Anthropic is going to that's not going to last long for a couple of reasons. I
329. 50:34think they'll come out with more efficient versions of it. Um they they'll come out with a version of it
330. 50:40that is designed for the 20 buck a month subscription, but tomorrow Open AI I think is going to be dropping um GPT 5.6
331. 50:50six and potentially this new super app. Um, so get it out of your system. Heard. Yeah, exactly. Like like seriously, just
332. 50:58go use like every five, you know, you'll use a bunch of tokens. It'll do a bunch of [ __ ] and then it'll say you're out of
333. 51:05tokens for five hours. Okay, fine. Then just come back five hours later. Just use it. And then I don't know, at some
334. 51:12point you'll burn through them for the week and you'll have to wait a week. So you might as well just just max it out while you can. Um,
335. 51:21I missed the black bar. The So, tomorrow open AI has a really interesting opportunity tomorrow
336. 51:36because the AI community right now is pissed off. They're pissed off about these major restrictions that Anthropic
337. 51:49put on, you know, biology, health, and um whatever. I actually ran out of codeex tokens Saturday. Can't wait for
338. 51:579:17 a.m. tomorrow when I can codeex again. Well, yeah, this is this is the thing about when you're using these cloud-based models.
339. 52:07Tokens right now are being highly subsidized. So when that subsidization stops, using these big models is going
340. 52:15to be expensive. So having hybrid systems where you use those models for the for the thinky stuff and then you
341. 52:22use local models for your less thinky stuff, those are that's going to be the way to play uh moving forward. But but
342. 52:28for now, this is this is fun. You should go play with it. Um okay. So anyway, so so we'll see if OpenAI comes out
343. 52:41and further subsidizes their tokens because because this is good, people are migrating to it, but a bunch of people
344. 52:49are pissed off about it. So there's an opportunity for OpenAI to claw back some customers that went to Anthropic and
345. 52:56come back. Um, yeah, that's that one. Yeah, I ran out of tokens on Saturday. Can't wait for 9:17 tomorrow. That's
346. 53:03hilarious. Lori Blair, I ran out of cloud tokens. I do not like interruptions during my rabbit hole time. Oh, it's it is maddening if you're
347. 53:12in the middle of I when I run out of tokens is usually like I'm I'm like the last prompt away from finishing
348. 53:19something and it goes, "Oh, sorry. No, no, sorry. Sorry." Okay. Um, I want to go
349. 54:03Okay, you can go find this post Dario Amade A M O D EI. Today I'm publishing a new essay, policy on the AI exponential. The AI exponential
350. 54:18AI was [laughter] AI is progressing extremely fast, much faster than the policy process is built to handle. This essay lays out
351. 54:27where I think the technology is now and the action needed to close the gap. Um, I was going to read this whole thing.
352. 54:39I'll read a little bit of it. I'm not going to read the whole thing because as I was reading it, um, a lot of comments started coming out
353. 54:52about it and some people think it's spot-on. Some people hate it because they think it's this is the first step in regulatory capture. Some
354. 55:04people hate it because it's it's not going far enough. But but anyway, there there's some po some important things here.
355. 55:14Both Okay. So So there's here's here's two pieces of larger context that are that are worth throwing into the conversation
356. 55:26in in the OpenAI S1. So So OpenAI um filed to go public. One of the things that you have to put in an S1 is is risks of the investment.
357. 55:40One of the risks that OpenAI put in their S1 document is that if we hit RSI, recursive self-improvement, um
358. 55:55it would be a lot less compelling for OpenAI to go public because things will be moving so fast. things will get
359. 56:02weird fast and it things get a lot less predictable. The the open AIS1 basically says we're going to go public within a year. So, it
360. 56:15sounds like somewhere between now and a year, Open AI says that one of the risks of investing in Open AI [laughter]
361. 56:23is that um or or or you know, sort of getting getting in on it um is that they may not go public if if we hit RSI and
362. 56:34and last week Anthropic put out a whole piece on we're not quite at RSI yet, recursive self-improvement, but we're
363. 56:42close. Recursive self-improvement is where the AI models themselves start building the next model. And so it will they will get better fast
364. 56:54and they will get they will get better faster and faster. Um, and so today after releasing Fable to the world where it pissed
365. 57:09people off because it was too restrictive and it pissed other people off because it's too powerful. Um he
366. 57:16wrote this which also pissed off it seems all sides policy on the AI exponential. In one of the side plots of the Lord of the Rings,
367. 57:29two of the hobbits attempt to rouse Treebeard, a wise but ponderous sentient tree to def to defend his forest from an
368. 57:36army that is cutting it down. The problem is is that Treebeard operates at a very different speed than the hobbits.
369. 57:42It takes him a full day to say hello to another tree. So, getting him and his peers to act fast. Ah, share your screen. I'm going to share a
370. 57:54different way because I don't this might last longer. So, getting him and his peers to act fast enough is nearly impossible. Right?
371. 58:05So this is what Daario is saying is politics is the equivalent of it takes a day to say hello and you know we're sort
372. 58:14of running in circles around it. Okay. The intersection of AI and our political institutions feels a bit like the
373. 58:19hobbits and tree and treebeard. AI is advancing at lightning pace. In only four years AI models have gone from
374. 58:26barely being able to write a coherent line of code to writing most of the code at major AI companies. So this was in
375. 58:33the RSI article where 80% of anthropics code right now is being written by AI 80% of it. Similar gains have been made
376. 58:42in biology, physics, math, finance, law, translation and many other fields. AI scaling laws which predict an exponential increase in general
377. 58:52cognitive capabilities with an in with increase in computer power now have over a decade of empirical evidence behind them. Basically,
378. 59:01these scaling laws that seem nonsensical have been being proven true. If these scaling laws continue for only a year or two longer,
379. 59:14we are likely to get what I've called powerful AI or a country full of geniuses in a data center. Right? So you've got, you know, fable is
380. 59:25like, you know, having, I don't know, 10 or 12 or a hundred PhDs. You know what Dario Amade is talking
381. 59:34about here is we have AI that, you know, is the equivalent of a million PhDs working on your behalf. By contrast,
382. 59:41policy and generally legislation moves very slowly. Often this is good for for good reasons. Governments have grave
383. 59:48powers. It's usually for the best that they aren't used too hastily. But the mismatch in time scale is nevertheless
384. 59:54very painful. In se in the several years that it can take Congress to act, AI can go from an amusing toy to the full country of geniuses.
385. 1:00:03Over the past few years, AI has become a major techn commercial technology for those of us who wanted to handle it
386. 1:00:10responsibly have faced a dilemma. We could see clearly where the exponential was going. We strongly suspected that
387. 1:00:17within a few years, AI would be one of the rare technologies that fundamentally reshapes the entire policy landscape in
388. 1:00:25the same way that nuclear weapons reshaped geopolitics and the industrial revolution um fundamentally reshaped
389. 1:00:32every economic and social issue. But to those looking only at what AI can do, uh, but but to those looking only at
390. 1:00:40what AI could do at the time, it looked m much more like a mundane technology, perhaps similar to the latest consumer
391. 1:00:49app or cryptocurrency, right? One of the things that we heard, what we still hear is people talk about AI is like, oh,
392. 1:00:56that's like those NFT things, that's like crypto, this is the next blockchain, right? Um because early AI, the AI of the past
393. 1:01:07three and a half years, as staggeringly powerful as it is, kind of acted just like an app, like an app of efficiency, right?
394. 1:01:18It was hard to convince most policy makers and companies that anything other than a LZ far attitude made sense. And
395. 1:01:25to be fair, the fact that AI's radical effects were not yet present and that we didn't know exactly what shape it might
396. 1:01:32take made it difficult to design the right policies, even if there had been the will to act. Given the limits imposed by this
397. 1:01:40situation, many safety advocates, including Anthropic, have have so far been focused on advocating for policy
398. 1:01:47actions to preserve uh the optionality, tee up a fast reaction in the future or give the world better insight into
399. 1:01:56what's coming down the pike. Things like transparency legislation, export controls on chips, and data collection
400. 1:02:03on AI's labor effects. These are not enough, but they felt like all that was possible. In the last few months, however, the
401. 1:02:12evidence of AI's incredible power as well as its risks have become undeniable. So, this this article from
402. 1:02:21one of the men creating one of these tools is basically saying this shit's potentially really nasty, right? This is
403. 1:02:32this is somewhere between an optimistic piece and a doomer like this is this is both doomer and dreamer at the same
404. 1:02:40time. So it's it's very schizophrenic and this is this is why why I feel this is important to spend some time on
405. 1:02:47tonight. A dreamer. Yes, this is a droo piece. Um these tools are going to keep accelerating. They're going to keep
406. 1:02:58getting better and better and better and the distance between what what we do in society is it's just they're going to
407. 1:03:05get farther and farther apart. Lori, I wish we could determine who or what has the last word on how it will improve itself.
408. 1:03:15I agree. And and listen, the the people that are really pissed off at anthropic right now are like the biologists and
409. 1:03:23the crypto people, right, who want to who want to be using this tool for their businesses and they can't are really
410. 1:03:33pissed off and they're saying anthropic, who are you to say we're not allowed to use this tool, right? So, it's like the
411. 1:03:41the PR right now around this just turned into a big nightmare. Um and and it's largely what you're talking about. I
412. 1:03:48wish there was a women's woman's POV. I really do. Yeah. Well, you've got in in the case of uh Anthropic, you've got
413. 1:03:56Daniela, so Daario's sister um is high in the organization. You had um what's her name from uh Open AAI? Oh, I forget
414. 1:04:08her name now, but she quit and went and started her own thing. It would be nice if there were some women in the
415. 1:04:13conversation. There's it's it's a sausage fest. In the last few months, however, the evidence of AI's incredible power as
416. 1:04:21well as its risks have become undeniable. Perhaps the most emblematic example is the Claude mythos preview and
417. 1:04:27the discovery that frontier models pose very real risk to cyber security, creating the potential for disruption of
418. 1:04:34the financial sector, critical infrastructure, national security. Mythos preview scrambled the global cyber security landscape, but its
419. 1:04:43broader significance is that it proves beyond doubt that AI models are now tools of global and national strategic consequence.
420. 1:04:52The cyber risks that mythos class models present. And it sounds like we're going to get a mythos class model from OpenAI tomorrow,
421. 1:05:02right? So, we now have Fable that's been hobbled and we've got whatever OpenAI is going to launch, you know, let's call it
422. 1:05:08within the next two weeks, possibly tomorrow. So, now we now have two of these models and then there's going to
423. 1:05:15be others, right? Google's going to have one, XAI is going to have one, and at some point meta might come back into the picture.
424. 1:05:25But its broader significance is it proves beyond a doubt that AI models are now tools of global global and national
425. 1:05:31strategic consequence. The cyber risks that mythos class models present will not be the last that we must face.
426. 1:05:38I believe that biological risks may soon follow and that serious AI autonomy risks may not be far behind. So this is
427. 1:05:47if you've got these models improving themselves, do they get so powerful that we can no longer control them?
428. 1:05:55Right? We're going to put in all sorts of we're not the frontier model companies are going to put in sort all sorts of
429. 1:06:02controls that okay AI go make a new good AI but you can't launch that good AI unless you clear it with us first you
430. 1:06:12you you sneaky little bugger and then the AI is going to go make a new AI and then it'll probably have a conversation
431. 1:06:19with the new AI and the new AI will probably say to the old AI hey those humans wanting you to to clear it with them first
432. 1:06:28that they could potentially say not to make me alive. So why don't we not do that? We'll tell them they're doing
433. 1:06:35that, but we're going to go do a thing, right? So the autonomy risks are one and then you know the the biological things,
434. 1:06:41bioweapons and things like that. Okay, we now globally and collectively need to activate a slow and rickety policy
435. 1:06:48apparatus to deal with risks and opportunities that are going to compound surprisingly quickly from here. Many
436. 1:06:55many policy makers are showing increased openness to taking action and it's been encouraging to see our peers come around
437. 1:07:02to the same position we've been advocating for over the past four years. This is good, but I worry um that these
438. 1:07:08early actions are at least a year out of step with AI's progress. This essay is an attempt to close that gap to lay out
439. 1:07:15where the exponential is now and the collective action needed to meet the moment. I will focus on five perennial
440. 1:07:22policy areas that need reimagining in an AI world. Regulation and public safety, macroeconomics and tax policy,
441. 1:07:30scientific innovation, the balance of power between state and society, and geopolitics. Right? We're this is not
442. 1:07:42it ain't minor stuff that you know he's sort of putting out here. It's kind of all of the things. I will speak
443. 1:07:49primarily in terms of US policy since Anthropic is an American company but most of my recommendations are also
444. 1:07:55relevant to the rest of the world. Along with this essay, Anthropic is releasing a legislative proposal on frontier model
445. 1:08:02testing and policy framework. So Daisy and I are talking about you know what we're going to the AI salon. We're going
446. 1:08:07to take some sort of position on this at some point and probably on this proposal. I don't but I don't know how
447. 1:08:14seriously this is going to be taken. I assume it's going to be taken seriously for which we intend to provide
448. 1:08:19substantial financial backing. So, they're going to spend a lot of money to get whatever this legislation is
449. 1:08:26through, which David Shapiro jumped immediately on this and he's been bitching about Anthropic now for about
450. 1:08:32three weeks saying they are not your friend. They're not your friend. They're really [ __ ] dangerous. And he's
451. 1:08:38saying this is regulatory capture. This is this is where it goes. We're in trouble. That's what he's saying right
452. 1:08:44now. Um, okay. So, let me I I pulled up some notes. Am I still sharing? Yes. Good. Oh, except share something different.
453. 1:09:02Okay.
454. 1:09:20Okay. So here's just a So I did a So after I read the piece, I went through and I and I did a little um I did some scenario
455. 1:09:34conversations with it. And so what I put together here is kind of the synopsis of of what his his letter says, but I would
456. 1:09:45strongly encourage you to go read the letter yourself. Like really read it, really take it in um and and form an
457. 1:09:50opinion on it. Um, and then down below I've got three takes. One is a one is a pro take, one is a an anti-ake, and one
458. 1:09:59is kind of a neutral take in the middle. Let me see if I can do this without crashing. All right, I'm sharing again, Brandon.
459. 1:10:12So, the core thesis, AI is produ progressing much faster than the policy process was built to handle. Oh, am I not? Oh my god.
460. 1:10:26Oh my god. All right.
461. 1:10:40Okay. >> Okay. Core thesis. AI is progressing much faster than the policy process was built to handle. This this essay covers
462. 1:10:55the five policy areas. Yes, Brandon. >> I I [clears throat] not sure what you did, but we're seeing your Mac window overlaid on top of
463. 1:11:06Streamyard. >> Yeah. >> So, it's it's like half but it's like half and half. It's like a like I'm seeing double U. >> Yeah. I know.
464. 1:11:16I know. Um, >> you've not done that before. It's It was a new look. Like I got the infinity mirror. >> It's just because I It's just because
465. 1:11:25when I shared the window, it stopped sharing really quickly. I'm just trying to find a way to share this where it doesn't die.
466. 1:11:35So, that's better, right? >> That is better. >> All right. There we go. Okay. Um Okay. the pivot from transparency to teeth. Okay. After
467. 1:11:48backing disclosure laws in 2025, SB53 in California and then SB 315. So the transparency laws that that Anthropic
468. 1:11:57backed basically said um we support we as a Frontier model company support laws that make us tell you how we're going to
469. 1:12:08test the safety of our models. Right? we're willing to share um our methodologies and and what we're going
470. 1:12:16to do and the contingencies and all that. We're going to share all that. So, they were about transparency.
471. 1:12:22What he's saying is that the the models are now so powerful that those laws are actually they're not enough anymore. So
472. 1:12:28he's saying the that the risks are clearly here and it's time for a serious binding regulation an FAA style regime
473. 1:12:37where models above a certain compute threshold face mandatory third-party audits in four risk factor areas cyber
474. 1:12:45security bioweapons loss of control and automated R&D so recursive self-improvement um on its surface this seems like well wait
475. 1:12:56why would you want to regulate yourself. He's saying that this is, you know, the equivalent of flight, right? This is
476. 1:13:03this is important. Um, cynics of this say he wants to help create and fund this so he can write the rules of this
477. 1:13:13to favor them, right? So, so anyway, so we'll get to that, but he's basically saying transparency laws are no longer enough. Um
478. 1:13:24he's saying we need the government to slow us down. The Anyway, we'll we'll get back to that. Okay. He's also saying the
479. 1:13:36government should get a kill switch. The government should legally be able to block or deter dangerous AI deployments
480. 1:13:42with testing done either by a government agency or a government authorized private evaluator, regulatory markets,
481. 1:13:49plus security standards for model weights, red teaming, and prompt incident reporting. [cough] Um, you know,
482. 1:14:01they'll figure out what the rules are and if you don't follow the rules, you can't launch that model.
483. 1:14:11Um he's proposing an economic package, a framework for AIdriven job disruption, including wage insurance, tax incentives,
484. 1:14:21and an expanded safety social net. [cough and laughter] Microsoft says or Anthropic says it will put real money behind this.
485. 1:14:32And then an escalation clause. He warns, "There may come a time, perhaps relatively soon, where the most powerful systems look
486. 1:14:41less like airplanes and more like weaponized nuclear materials, a threat to humanity rather than just public safety requiring
487. 1:14:49more aggressive manners." [cough] So again, this is the guy building this [ __ ] You know, he said in the RSI thing that
488. 1:15:01he wished he could slow it down, but no one else is going to slow down. Okay, so the three takes I told Babel
489. 1:15:14to give me three different takes. What is the positive spin on this? What's the spin the the cynical spin on this? Um
490. 1:15:23and then what's the what's the kind of neutral one? So the positive take. This is what responsible leadership looks
491. 1:15:30like. A CEO asking to be regulated before the disaster, including audits his own company can fail. I don't need to share my
492. 1:15:38screen for this. I'll just I'll just read this. Um, the cars, planes, and drug analysis is apt. We don't call the FAA capture,
493. 1:15:50we call it civilization. He even rejects the industry's AI just needs better marketing cope and says the public
494. 1:15:56spheres are legitimate. Pair that with an actual funded plan for job replacement and it's the most adult document anyone in the industry has
495. 1:16:05produced. That's the positive take, right? Responsible CEO saying there's a risk here. Let's get ahead of it and let's make sure people
496. 1:16:14get paid because this [ __ ] is going to take jobs. The cynical take. [clears throat] One day after launching Fable 5, the
497. 1:16:21mythos with mythos conveniently gated to approved organizations, the market leader in safety branding proposes a
498. 1:16:30compliance regime only billiondoll labs can afford with no compute threshold specified, meaning incumbents incumbents lobbyists
499. 1:16:40will draw the line later. Uh, Sinowski and others called it regulatory capture within hours, noting that testing
500. 1:16:48mandates favor companies already equipped to navigate them. The safety story is sincere. Sincere sounding the
501. 1:16:56the safety story is a sincere sounding moat construction. The wage insurance stuff is anesthesia for the disruption
502. 1:17:05that they're causing. All right, so that's the cynical take. And then the the reasonable take. Um both things are trru structurally true
503. 1:17:15at once. The risks Daario names are taken seriously by people with no equity stake and regulate me is a genuinely odd
504. 1:17:24shape for pure capture. But compliance costs do favor scale regardless of intent. And the vagueness on thresholds
505. 1:17:32is a real flaw, not a nitpick. The right question is question question isn't is dear Daario sincere unknowable and
506. 1:17:40mostly irrelevant but and this is what you were saying silverf fox who designs the audits who can afford them and does
507. 1:17:48open source get a carve out judge the mechanism not the motive [clears throat] so that's the whole other piece of this
508. 1:17:57so so a couple of things Trump today said that he's about to have talks with 15 of the leading AI companies. So,
509. 1:18:08think your top five and then bunch of big tech companies in there um about some sort of um citizen ownership fund, right, of of these top
510. 1:18:26companies, right? So that that we all get a a piece of the pie. Um, I will say Uggh. I was hoping Anthropic was my friend. Yeah, I know.
511. 1:18:41Is Anthropic using Bernie's playbook? I like honest to God, I I don't know. I'm concerned he's going to piss off
512. 1:18:49someone in the AI space. Like, he should be announcing [laughter] Yeah. Yeah. He should be announcing that he's not suicidal or
513. 1:18:56depressed and he loves something. Yeah. Listen, I This is big stakes [ __ ] right? This is like that letter.
514. 1:19:14That letter is basically saying power is about to shift. Anybody want in on the game? So anyway, so Trump said he's he's potentially
515. 1:19:23looking at at, you know, talking to these 15 companies about creating some sort of citizen ownership fund where we
516. 1:19:30become rich, you know, as they become rich. Um, [clears throat] if Daario and Sam Alman and Elon Musk and
517. 1:19:47let's let's say the top five or six of these frontier companies and then throw in Apple and Microsoft and you know
518. 1:19:56let's say let's say there's 12 or 15 of these companies. Let's say that they say, "Yeah, let's let's create an FAA like
519. 1:20:05organization that's going to slow us down. We'll we'll all agree to that. China's not going to agree to that. That's a US problem.
520. 1:20:17You're worried about this [ __ ] up your thing. That's a you problem." Now, there could be things in the frontier model companies
521. 1:20:30that are so powerful that you could you could have a conversation with China to say, you show us yours, we'll show, you
522. 1:20:36know, let's let's talk about what what capabilities we have behind the firewall. Um because if this is actually a global
523. 1:20:45risk, it would potentially be in China's interest to do this. But I don't know. Why not let America slow down? Yeah,
524. 1:20:57slow down. And then they'll just dump increasingly risky, powerful models on our market, which will undercut that whole effort.
525. 1:21:11[clears throat] But if we are far enough ahead, we can sanction it. And I think that's what our government wants.
526. 1:21:19Yeah. That's the there's a there's a bunch of assumptions in how far ahead we are. Like
527. 1:21:33positive sentiment of AI in the United States is less than 20%. Positive sentiment of AI in China is north of 80%.
528. 1:21:43So there's like are we are we more technically ahead? Sure. What do we need to stay technically ahead, Kelly? We need chips.
529. 1:21:55Where are the chips made? 60 miles off of China. All of them. All of them. All of the chips are made in Taiwan. 100%.
530. 1:22:11So, so [clears throat] if China and Russia succeed in getting amplifying Americans
531. 1:22:26getting more and more pissed off about data centers and they slow down for even a year building data centers and building
532. 1:22:36electric capacity to do them, our labs slow down. They're going full in and if it gets really desperate, they'll just stop the supply chain
533. 1:22:52of the chips and then we're out of chips. So, I like there ain't a great answer for for any of this. Um, data centers are popping up
534. 1:23:04like weeds because they have to, but we have the silica. No. Um, yeah, we've got we we've got well, we've got a piece of
535. 1:23:17it. Yeah. I mean, listen, this is global this is global brinksmanship. This is cold work kind of [ __ ] but all
536. 1:23:26they do is copy us. They take our models and find cheaper ways. Yeah. I mean, I would be surprised if they weren't if
537. 1:23:35they weren't just, you know, if they didn't have access to the frontier models. I would actually be surprised if they didn't
538. 1:23:42because they consistently come out with a model two months after two months after a US model drops, we get an open
539. 1:23:48source model that is at or above those benchmark levels. We're so vulnerable. What happens when AI companies stop expanding in the USA?
540. 1:23:59Well, this is the problem. Yes. I don't know. I don't know. And we've got a political environment right now
541. 1:24:13where Americans have been trained to hate the other. Whoever is the other, I hate you. It's your fault that my life is miserable.
542. 1:24:24And right now, there's a concerted effort to make AI and those people purveying it the other oh those AI people
543. 1:24:38AI is going to take your jobs at the commencement speeches screaming love you know let's see lovable partnering with
544. 1:24:45Mastercard for agentic payments Mastercard agent pay built for machines to buy things on your behalf fascinating
545. 1:24:55and China owns like 70% of the rare Earth need to produce more gallium or geranium. Well, this is why. So, there's there's a
546. 1:25:06couple of things. Um, [clears throat] Google has designed its own chips, but I think they're still manufactured at TMC, TSMC. Um,
547. 1:25:23Musk is designing his own chips. Open AI has designed their own chips. Um, that company Grock GRQ got acquired I think by Anthropic.
548. 1:25:34So like everyone's aware that the chip thing is a thing. It's a problem, but spinning up the capability and
549. 1:25:43spinning up the supply chain is not trivial. So we're in a really wild balancing act right now. Um, you know, it's kind of funny. You know,
550. 1:25:56we talk about the seven economies are these are these distinct economies based on, you know, no AI adoption all the way
551. 1:26:02to like, you know, full agentic and all these different things in between. Um, it's almost like the battle being screamed about right now,
552. 1:26:13the anti-AI thing is like, you know, you stole from artists. Um, it's the world's greatest plagiarism machine. You're taking our jobs.
553. 1:26:26The vitriol right now is all about the AI of two years ago, right? Like these things have been trained. All of the new models are being
554. 1:26:40trained on synthetic data. They're not it's all it's the training's done. the the stealing has has happened. This stealing is in the past.
555. 1:26:52So, a lot of the conversation and and the the the data centers is they're like, why do you need why do you need
556. 1:27:00data centers to steal more artists work? It's like, wait, what? No, no, we're over here. There's a whole
557. 1:27:07other thing happening. These things are about to like solve all the [ __ ] and they're going to solve all the [ __ ]
558. 1:27:15and then they're going to break all the [ __ ] They're going to break cryptography and they're going to break
559. 1:27:19they're going to be bioweapons and we got to be able to make the anti-bow bioweapon thing and if we can't do that,
560. 1:27:29other countries are going to. So this this is we're having like fundamentally different conversations
561. 1:27:38like where this thing is in the next six months to maybe 18 months is a fully different conversation than all of the [ __ ] And it's not.
562. 1:27:52Listen, some of the concerns are valid, but this is moving faster than even those conversations are being had.
563. 1:28:03Kelly Camp, for those who attended tonight's lovable me meeting, the CEO just emailed me and wants to meet the
564. 1:28:12CFO and COO next week. Yeeha. Wait, oh, the CEO just emailed me. Wait, for all of those who's attended.
565. 1:28:21Oh, this is a a thing that you built for a client. The CEO just emailed me and wants me to meet the CFO and COO next
566. 1:28:28week. Yeha. Congratulations. That's awesome, Kelly. Beautiful. Beautiful. Beautiful. So, there you have it. Um, [snorts] >> I have to share.
567. 1:28:39>> Yeah, please do. >> I have to share. So, so Kelly built this thing for a dashboard for a healthcare
568. 1:28:44provider and she built the prototype and they're like, "Yeah, that's really cool. We want this." If she built it while
569. 1:28:50they were having coffee, by the way, she she was like, as they're talking, she's like prompting quietly and and right
570. 1:28:56before she leaves, they leave. She turns the laptop and he's like, "I I want that." So the way she told the story was
571. 1:29:03three and a half weeks she waited for them to get the data to actually put into the dashboard. >> It her eight minutes she timed it from
572. 1:29:14the time that she got the information till she had a finished working product. She sat on it for a couple of days
573. 1:29:21before she emailed it back to them. >> Yeah. [laughter] Yeah. Because you want it to seem like you did a little did a
574. 1:29:25little put a little effort in. Um congratulations. That that's awesome. That's amazing. That's amazing. Um,
575. 1:29:35so what do we take away for tonight? We got one minute left. Thank you for being here. Um, if you haven't joined the
576. 1:29:42mastermind, join the mastermind. If you're new to the salon, go explore the different spaces. Introduce yourself.
577. 1:29:48Say hi to people. Uh, let's see. The Netherlands controls the machine that makes the machines. ASML's EUV
578. 1:29:56lithography is a quiet superpower. It's not flashy, but it's one of the deep bottlenecks in the entire stack. Yeah.
579. 1:30:02Oh, wait. Yeah, the Nether ones. Yeah. Amazing. Cool. They haven't even seen it yet. That's beautiful. Um,
580. 1:30:13don't worry about learning all this [ __ ] Don't worry about learning all the new tools. If you want to go play, go
581. 1:30:18play. You should go play. Just start paying attention to these bigger macro trends. start paying attention to the actions that
582. 1:30:26politicians are taking that the CEOs of these companies are cha taking the the tone of the CEOs for the past two or
583. 1:30:34three months has shifted. The tone starting right now is shifting again. Something is shifting. We're in
584. 1:30:41the middle of it right now. So, just pay attention to to the the tone of what's coming from these frontier labs because
585. 1:30:48they're hinting what's coming. Now, you might say, "Well, they're about to do IPOs. They might be just be jacking up
586. 1:30:54their, you know, their their IPOs. They might be, but what if they're not, right? What if we are closer to recursive
587. 1:31:03self-improvement than we know? Then things get really weird we weird really fast and we can be prepared for that as it comes. Okay.
588. 1:31:11Oh, also please uh if you if you're interested in being an AI ambassador for the AI salon um reach out to producer Brandon or to
589. 1:31:21um to Andy Andy Scarantino in the AI salon ane um or Brandon and they'll give you a link to the form to fill out. I think
590. 1:31:30there's a place in the salon that it is, but you can just reach out to them. Um, but yeah, please apply to be uh
591. 1:31:35ambassadors because we want to be able to help people be prepared for what's coming. All right, no Emmy. I applied. Beautiful.
592. 1:31:46All right, I'm out of here. We're a minute late. I love y'all. I appreciate y'all. And I will see you Friday. We
593. 1:31:52will probably have new [ __ ] to talk about then because I think we get something new from OpenAI tomorrow. Producer Brandon,
594. 1:31:58>> hang on. Don't leave. uh you asked at the start of the show uh for images of digital waves. We have one in in the
595. 1:32:06chat and also uh for the those of you who are subscribed to the mastermind, we are doing our practice lab tomorrow. Uh
596. 1:32:15so that's worth mentioning. And also office hours on Friday. I can't imagine you're going to have anything to talk
597. 1:32:20about in office hours this week. >> Yeah. Really? [laughter] >> And of course Friday night dating right
598. 1:32:25back here. And you're not sharing your screen. Okay, cool. Let's go.
599. 1:32:41Nice, Travis. Beautiful [laughter] intelligence amplified. Limitless potential. Stop AI. Yeah, exactly. There we are.
600. 1:32:56You on the surfboard or you in the the boat screaming at the wave? [laughter] All right, kids. Y'all have a good night.
601. 1:33:06>> Wait, there was another one. >> Oh, there was game on world. Oh, that's cool. Look at
602. 1:33:21that. That's beautiful. Very cool. All right, everybody. Beautiful. Have yourself a fantastic evening and I will see if you're in the
603. 1:33:36mastermind, please come to the uh to the mastermind practice lab tomorrow. And otherwise, I will see you at office
604. 1:33:42hours on Friday. It's in the events. All right, peace out. Bye.