Anthropic Fable Is Back: The Geopolitics and Safety of Ultra-Powerful AI Models

Transcript

1. 0:34Good evening. Good evening. Got a big crowd tonight. Happy Wednesday. Happy Fable is back day. Happy uh new member
2. 0:42Mixer day. Happy first day of the AI Salon Ambassadors in inaugural cohort day. It was a day. We had a day. We had a lot of stuff
3. 0:52today. A lot of stuff happened. Um see Charles Adams. I see Andy. Got producer Brandon here. Got lots of other folks
4. 0:58here. Okay, so little housekeeping. If you're new here, welcome. Um we met some new people today. So hopefully you're in
5. 1:05the house. Some of the new people we met today have been watching these lives for a couple of years. Um, and and came out
6. 1:11from behind the camera and said hello. So, hello if you're here. U Monita, hello. Just saw you on TikTok. Joined a
7. 1:17long time ago, but never saw your announcement in time. Welcome. Welcome. So, tonight's going to be a fun night.
8. 1:24We're going to play with Fable. Fable, if you don't know, is the anthropic big sexy AI model that the government said
9. 1:34was too powerful. Well, first Anthropic said it was too powerful for mere mortals to use and then they launched it
10. 1:42and uh one of their biggest investors jailbroke it and turned them into the government. So, so the government kept it under lock
11. 1:52and key for two weeks uh using export controls and uh it's back. So, we're going to play with that. Um Anthropic
12. 2:01also launched Sonnet 5, which is about twice as good but half as expensive as Sonnet 4.6. So, we'll play with that. It's probably
13. 2:10good. Who knows? Um Gemini um Omni, which is their world model that makes videos, is now available in its
14. 2:23flash, a fast, cheaper version of it. It's now available on gemini.com or gemini.google.com, whatever it is. Um and uh
15. 2:36so we'll play with that tonight. Um, and I'm going to do my best to see if we can figure out if Fable is
16. 2:50if Fable is dramatically better enough that we can notice it. And I don't quite know how to do that because Oh, yeah.
17. 2:59There's a there's also a new Nano Banana Nanobanana Light or something like that. So So there there's lots of things to
18. 3:06play. Yay, Fable. Yeah, exactly. Um, we just can't be trusted. Um, so anyway, welcome to everybody. Um, little update
19. 3:16on on uh World Cup soccer. The US team just got a red card, so they're down a man, but they're up a point and uh you
20. 3:25know, they're doing their thing. So, I'll pay attention to that in the background. And if you're not watching
21. 3:29it, it's okay. Got you covered. Spoiler alert. Uh oh. Spoiler alert alert. Gez, what if you're not watching soccer? Well, no, it's not.
22. 3:41They're still playing, so I'll just keep you up to date. Um, okay. So, we're going to go play. Let me I want to give you a little comment.
23. 3:51Well, a couple of things. First of all, if you participated in the new member mixer today, so we had for the very first time
24. 4:01in history in the AI salon history, we had an event just for new members. Um, it was an idea we kicked up that as new
25. 4:09members are coming in, there wasn't necessarily a specific place for them to just come and say hi and get a lay of
26. 4:15the land and ask questions about how Mighty Networks works, things like that. Um, so we did that. It was really fun.
27. 4:21And it was really beautiful. Um met some new people. Some we had some friends there from that Andy I met at Social
28. 4:27Media Marketing World. Um we had W from the Middle East came and introduced himself and turned on his camera. We got
29. 4:35to see him. We got to see people. Everybody turned on their cameras. It was really quite beautiful. Um I thought it was fable that
30. 4:45Oh, I thought it was a fable that Fable was back like April Fools. But June and July are F months. Fable FIFA Flag Day.
31. 4:52Fourth of July. Had too much Diet Coke today. Yeah, you're pretty zippy there, Mary Carol. Sorry, I will self-edit. That's good.
32. 5:02That is a That is a full-on ADD uh post. I love it. I totally resonate with it. I have done that before where I've been
33. 5:10chatting with someone and I'll I'll string some things together and I'll get back like add much. So, that's awesome.
34. 5:18Um, it was especially complicated tonight. Chef Kelly, two stars. So, so the the uh the new member mixer was
35. 5:26awesome. Great people, great to connect. Um, and then uh we we kicked off today the first cohort of the AI salon
36. 5:36ambassadors. And so the ambassadors are people who are um you know long-standing AI salon members that already live up to
37. 5:45the values and practice the cycle of AI readiness and are just you know generally good strong human beings.
38. 5:53And the idea with the ambassadors program is that we're going to have those folks go to, you know, other
39. 5:58organizations, other groups, not other AI organizations, just other like, you know, chambers of commerce or, you know,
40. 6:05business groups or just, you know, with a group of friends and just talk about what we're up to in the AI salon and
41. 6:11talk about the cycle of AI readiness, talk about the great repurpose. Um, Sunonny Yun was there and declared
42. 6:18herself a poster child for the great repurpose, right? She just came out of, you know, uh, either losing her job or,
43. 6:24you know, leaving her job and then trying to get another one and realizing, well, that ain't going to happen. And
44. 6:30she kind of reinvented herself and just ended up going to Can Lions and doing a bunch of stuff. So, she's she's all
45. 6:36excited about what she's up to in the world and credited a lot of that to things she's learned from people in the
46. 6:44AI salon. Um, and so that was awesome. So, so you know, we've got, you know, Kelly Camp was in there, Mary Carol was
47. 6:50in there. We've got a bunch of people who are who are on board. Anita, okay, so do we have our AIS in the room, too,
48. 6:58or or what do we do in here? First time I've made a meeting since I ran across y'all. Um, you can have your AIs in the
49. 7:05room if you want. I don't I don't know quite how you do that. um if you want. So there's a there's a chat and how you
50. 7:11get to the chat is is up in the corner here underneath the AI learning lab logo is the button that will pop out the
51. 7:17chat. So if you haven't done that um do that. You can also if you want you can split out um the video window here as a
52. 7:27floating window in the lower corner here. You can pop out this window and then you open a second tab in Chrome or
53. 7:35whatever browser you're using and then you can look at the AI salon there. So you can go look at the AI salon over
54. 7:40there and have the stream up there. Okay. Um let's see. Kyle will talk about Fable and more. He'll show us his screen our
55. 7:50participation via comments and questions. Beautiful. Yeah, that's it. Silver Fox. So So welcome. I am in the
56. 7:57chat, right? Yes, this is the chat. So, if you're if you're seeing your name on on screen right now, I can. So, any
57. 8:03anything that you chat, I can see and producer Brandon behind the scenes can see. And if you have questions, you can
58. 8:09pop them up there. He'll pop them up on screen if he wants me to see them. Um, I may pop them up on screen like I just
59. 8:14did there. Um, yeah, and you're you're in the Mighty Mighty Networks app, right? So, the AI salon is hosted on
60. 8:22this technology called Mighty Networks. Um, and that's how we how we communicate. Um, one of the things you
61. 8:28can do if you're new to the salon is go check out in the in the lefthand part of the screen, there's all these menu
62. 8:34items. Each one of those menu items is like a different space, a different, you know, collection. So, think of the AI
63. 8:41salon as like this this bar or this restaurant. And each of those little tabs is like a different table with different people with different
64. 8:48interests. And so, you can go check all of those out while while I'm rambling on. Um, I do tend to ramble a bit. Um,
65. 8:56but tonight's Wednesday. I'm also going to be playing with stuff. Before I play with stuff, I want to talk about
66. 9:05Fable. So, couple of things about it. Fable is this new model that Anthropic is calling the Mythos class model. And mythos is the is
67. 9:18the super powerful version. Fable has been made safe safe um with with safety guard rails, but they're they're basically the same
68. 9:30model, but Mythos or Fable just has a lot more safety guard rails on it. And since being unquarantined from the
69. 9:37government, it's got some fairly significant safety guard rails on it that people are already complaining about. Um
70. 9:45so, a couple of things happening. Starting at the beginning of the year, the tone of Dario Amade, the CEO of
71. 9:54Anthropic, and to a degree, the tone of Sam Alman started to shift. Um, and it kind of started to shift toward the end
72. 10:03of 2025. It really got cooking in the first couple of months of 2026 that they just all started talking about super intelligence,
73. 10:19not AGI, artificial general intelligence. And the distinction of those, they're sort of generative AI,
74. 10:24what we've been playing with for three and a half years. AGI is this sort of mythical moment in time where the AI
75. 10:31models get strong enough that they can basically do the work as good or better than most humans at most economically
76. 10:40viable jobs. That's AGI. And then artificial super intelligence is where these things just get smarter than humans altogether.
77. 10:49So they started shifting talking about instead of talking about AGI they were talking about artificial super
78. 10:56intelligence that that clearly something had happened within the labs that they hit a different level when mythos came out two months ago
79. 11:06month and a half ago I think it was two months ago said um this model is so powerful that it can essentially find any
80. 11:16vulnerability in computer system and so we're not going to let you have it. And everybody griped and bitched and
81. 11:24things like that. And then they they gave access to I think it was 52 companies that have computer software infrastructure
82. 11:34to allow them to use Mythos to find the vulnerabilities and patch them before they release it live to the world. And
83. 11:41then about two and a half weeks ago, they released Fable, which is which was this sort of safetyized
84. 11:47version of Mythos that we could all play with. And we played with it for three days. And then on Friday of two weeks ago, so
85. 11:57I guess two and a half weeks ago, um the government issued an export control on that model. And so Anthropic had to take it down. So
86. 12:10it's been gone. So a couple of things are going on. There are a bunch of different philosophies
87. 12:24on how you take a powerful model and make it safe for the world. You know, one of them is you don't ever let anyone use it.
88. 12:35Sam Alman has been fairly consistent with OpenAI that his philosophy is slowly release features of these new
89. 12:46models over time and allow the public and business to understand what they make possible and get used to them over time.
90. 12:57Daario Amade quit open AI. So Dario Amade was one of the founders of OpenAI and he quit it to start anthropic.
91. 13:13Uh US almost just scored but it looks like someone was offside and Christian Christian Pissic is saying
92. 13:21no no look at the replay. Anyway, I'll tell you if they scored. Um where was I? Sorry, I got all I got got
93. 13:32all discombobulated. He's off sides for sure. Okay. Um, okay. So, so there's these two different philosophies. Dario Amade's philosophy
94. 13:47is how he thinks safety should be done is how he wants everyone to do it. And so basically what he's been doing is
95. 13:59saying, "Hey government, these models are really dangerous and you should regulate us and you should slow us down.
96. 14:06Oh, and by the way, we'll volunteer to tell you how to make them more safe and what we'll tell you how to review these
97. 14:12models and tell if they're safe." Um, as you can imagine, the open-source community and researchers and and other
98. 14:21businesses are not liking this approach. And so Daario in sort of, you know, screaming how dangerous these things are
99. 14:29got what he wanted and the government, you know, sort of shut this thing down. Now, here's the ironic thing.
100. 14:39What he has inadvertently done is drive Fortune 500 companies to use Chinese models. Right? So, so both OpenAI and Anthropic
101. 14:53have billions and billions and billions of dollars of investment. They've got trillions of dollars of valuation. They're about to go public.
102. 15:02And Fortune 500 companies, there's two major issues with these companies. One is they're really expensive, right? If you use Anthropic
103. 15:15or OpenA on a per token basis and you and you roll it out across a Fortune 500 company, you're going to spend a
104. 15:22boatload of money on them. So, one is they're expensive. The Chinese models are in the neighborhood of as good and
105. 15:29they're way way cheaper. And so, Fortune 500 companies are going to these Chinese models. Um, so, so there's that piece of it. The
106. 15:42other piece of it is now that this thing happened with anthropic, those models are no longer considered stable.
107. 15:55Like even if mythos is an incredible model and it's incredibly stable, at any point the government could just say no,
108. 16:02you can't have access to that. And so it's another reason why Fortune 500 companies aren't going to use it. So, um
109. 16:10I you know the other thing is China is going to do everything they can to slow America down in building data centers
110. 16:20and building uh energy capacity. So, you know, electric plants and things like that. They're going to do everything
111. 16:28they can to reduce our capacity there. They're going to flood our market with open-source models. So it Oh, US just scored. Go two nil. Go.
112. 16:44Okay. Okay. His AI no comment. It's like a 20-year-old Chetchy BT and a 30-year-old. Gro is a 45 to 60 year old. That's pretty funny. Um
113. 16:57blog could literally do no math. Be careful. It tried to answer. Um, well, the things have gotten better at
114. 17:04math and they they've they've now got some math specific ones that are coming out. It sounds like a math one came out. Oh.
115. 17:12Um, Gro is a 12-y old boy. Gro is totally a 12-y old boy. All right. Let me see. Was this a penalty kick? Oh,
116. 17:20yeah. It was penalty kick. Oh, nice kick. That was beautiful. Okay. Um Okay. So, so we'll go play. Um,
117. 17:32you know, China's going to keep dumping powerful models. Oh. Oh, here's another thing that happened that I think is
118. 17:39worth noting. And just the the context of all this is going to play with Fable tonight. Chances are, unless you're doing sort of
119. 17:50agentic systems planning, you probably don't need something as powerful as Fable. If you're doing research and
120. 17:56science and things like that, you probably do, except that's the stuff they've got locked down and you can't
121. 18:01really play with anyway. So, whatever. Like, the other thing that's happening with Fable is we've got it until January 7th
122. 18:11where in your $20 a month um Claude subscription or higher um you have access to to Fable after J July 7th. Did
123. 18:22I say January? July. after July 7th, um you have to pay on a per token basis. So, however much you use it, that's how
124. 18:30much you're paying for it. Um those bills are going to run up really high. The the $20 a month subscriptions from
125. 18:36both of these companies are highly subsidized. These models are expensive to run. For as good as they are and as fast as they
126. 18:45are, it's costing them way more, like way more than $20 a month. Um, I heard a I heard a number from Anthropic that
127. 18:54their $200 a month subscription uh is costing them about $14,000 a month per person per $200 subscription. Um,
128. 19:05so we're going to go play with it tonight, but just know that you're probably not going to have access to
129. 19:10that. Couple of things that will happen. Chinese models will get stronger. But one of the things that Mythos did is the
130. 19:20CEO of DeepSeek. So you remember when Deepseek came out and it freaked everyone out because Deepseek was as
131. 19:25good as ChatGpt and Claude at the time and everyone freaked out. that CEO has been self-funding the development of the
132. 19:35Deep Seek models up to this point. And the launch of Mythos um inspired him to raise $7.6 billion because he realized he couldn't just
133. 19:47write checks for it anymore. That in order to create something on the level of mythos, you need to be investing
134. 19:53billions of dollars. Um, the other thing that it looks like is the Quen models that are from Alibaba, it looks like
135. 20:02those were effectively stolen from Anthropic. So they they create they wrote a letter to the Justice Department saying Alibaba
136. 20:11created 25,000 fraudulent accounts and basically just did a replication uh attack where they basically had those
137. 20:2025,000 accounts give very specific prompts, took the answers and trained another model. So there's all sorts of drama going on. But
138. 20:29we have it back for now. What we can assume is this. We can assume that whatever level Fable is at other other
139. 20:38companies will catch up to and at some point open source will catch up to. So we will likely we are likely close to a ceiling where us
140. 20:51non big entities are probably going to be capped out on capabilities. Um it looks like they're going to be
141. 20:59reserved. I have no flipping idea how Anthropic or Open AAI are going to make their money back because corporate America is not going to use
142. 21:08them for some amount of time because of this instability. Okay. using so many tokens. In my humble opin
143. 21:19opinion, due to needing to learn to work with AIS, most people just regenerate dozens of times until they get something
144. 21:25they like instead of learning to get the AI to understand the prompt the first time. Well, and it's even it's even more
145. 21:32dramatic than that, Oneita, because with Agentic platforms, so we're in the middle of a transition from you go to
146. 21:40chat GPT and you prompt it and you learn how to prompt to giving a goal to an agentic system that is going to go do
147. 21:49the prompting for you. All right. Um, and in that case, you don't need to learn prompting, but you're going to
148. 21:57burn a lot of tokens. You're going to burn way more tokens than than maybe you even did there. Okay, Mary Carol, let's
149. 22:03see. You can only use 50% of your subscription on Fable. After that, you have to buy tokens. You can't use 100%
150. 22:10of your subscription. Mary Carol, is that right now or is that July 7th? Because I thought July 7th was anything you do is
151. 22:23you have to pay per token. You also teach M MD files. Cool. So, what Wanita is talking about here is markdown files. So, when you create
152. 22:36things like memory, one of the things we're doing in the great repurpose is we're going to do um uh workshops on
153. 22:42creating your digital self. And so, writing markdown files that create who you are digitally that these agents can
154. 22:49use now through July 7th. Okay. 50% of your token usage goes to can go to Fable. Got it. Okay, cool. Got it.
155. 23:07We're almost at the 90th minute. We're up 2-0. Looking good for the US. Moving on. Okay. Um, let's go play. Let's go play.
156. 23:37Hey, Kyle. >> Hey, producer. >> Don't you want to tell people about something first? >> I do. Um, when does that happen?
157. 23:52Producer Brandon back to basics. When does that happen?
158. 24:06Okay. So, mark on your calendars if you'd be so kind.
159. 24:21Sorry. Drama drama on the soccer field. Mark on your calendars um July 21st, it's a Tuesday night. It's the
160. 24:32third Tuesday of every month moving forward. I'm going to be doing a special AI learning lab live in in the salon um
161. 24:41called back to basics. So, if you're brand new to all this stuff, um, one of the things is that people that have been
162. 24:49coming to these things for a while, you know, understand about prompting, understand about markdown files,
163. 24:54understand about agentic platforms. If you're new to all this stuff, the whole intent of the AI learning lab was to
164. 25:00give people a leg up that are just trying to figure this stuff out. So, once a month, I'm going to do a back to
165. 25:05basics course. that back to basics there's a back there's a back to basics course I did I don't know three or four
166. 25:12months ago that's available um on demand in the AI salon so so you can you can actually go watch that now but July 21st
167. 25:22on a Tuesday so we go live Monday on that Tuesday uh I'm going to be doing a back to basics so if you're new here or
168. 25:31if you know anyone new if you know anyone who you want to be able to just you know come in and get up to speed on
169. 25:38what's prompting, what are large language models, how do they work? I'm going to do a really concise little
170. 25:43version of that um on that on the third Tuesday of every month. Okay? Um so, put that in your calendar. It's also in the
171. 25:52events calendar in the AI salon. Okay? So, go RSVP for that. Please share that with friends. If you've got anyone who's
172. 25:59new to this stuff and you don't feel like explaining it to them, I'll explain it to them. Okay? Beautiful,
173. 26:06beautiful, beautiful, beautiful. Okay, Kelly Camp, I'm encouraging my clients to start with chat GPT until they
174. 26:12actually learn how to use an LLM. Then roll out Claude to a few select that are more technic technologically advanced.
175. 26:20Crawl, walk, run, fly. Beautiful. Love that. Do you teach andor use conversational AI or only prompting?
176. 26:32Um, I use both. Um, I would say that I don't use conversational AI as much as I should. I've I don't like the way when
177. 26:44I'm in a conversational AI, I don't like the way um, chat GPT captures that conversation. Um, so, so I I I do often talk my prompt
178. 26:58into the thing, but I'm I'm I'm often not using conversational AI unless I'm brainstorming. I'll sometimes brainstorm
179. 27:05on the way to the office and then I use it there. Um, and I think that's a very powerful powerful way to go. Okay, that's up and running.
180. 27:16All right, so Fable Five is back. I've got your comments over here. I've got post-it notes. So, the newest model tackles your
181. 27:27biggest challenges with fewer check-ins needed. So, we're going to say try Fable 5. So, you you select your model down
182. 27:34here. Fable 5 is selected. Sonnet 5 is new. So, so there was Sonnet 4.6 was the one that we had until I think yesterday
183. 27:43or the day before. Um, so this one's not expensive and it's apparently quite good. Um, and then Opus 4.8 8 is
184. 27:50expensive and and good. Um, Fable 5 is twice as expensive as Opus 4.8. So, once this thing goes to pay per
185. 28:01token on July 7th, it's it's probably going to be pretty pricey. Um, okay. So, let's see. I'm going to kick off two
186. 28:13different things here. Is Am I still sharing, by the way? Yes. Good. I think I might know why my sharing was
187. 28:20quitting. So hopefully that won't do that tonight. Um I'm gonna do No. No. What? It didn't work. Oh,
188. 28:39got to be kidding me. Let me try doing a window instead of the whole screen.
189. 28:56So, here's what I'm going to do. I'm going to go to Sonnet. I'm going to go to Sonnet 5 and I'm going to say, um, write me a prompt
190. 29:08for Fable Five. to Can you repeat yourself? But what are the best use cases for Sonnet 5? Well, Sonnet 5. So, Sonnet 5, think of
191. 29:24Sonnet 5 as like a daily driver of if you just want to like not blow through your credits and have a very very capable LLM like Sonnet 5 is
192. 29:37probably in the neighborhood of Chat GPT 5.5 in terms of quality. It's better than Sonnet 4.6. Um, it's not quite as good as Opus. So,
193. 29:48so just so so you remember these four levels, haiku, sonnet, opus, and fable. Haiku was like a I don't know
194. 30:02a 10 billion parameter model. Sonnet was like a couple of hundred billion. I think opus was maybe a trillion, I don't
195. 30:12know, or some hundreds of millions. And then fable five is 10 trillion parameters. So these four different
196. 30:20names are the four scales of how big the models are. So how much detail and sophistication they have in the model.
197. 30:31So sonnet is a is a cheap model, but sonnet 5 is apparently quite good. So we can play with it. So you can just, you
198. 30:38know, use that um like as a daily driver and it won't cost you a lot. The other thing that you can do with this is, and
199. 30:47am I still sharing? Yes. There's this effort um thing. So, you can choose a model. So, I could choose set five. Right now, it's
200. 30:58on medium. I could put it on extra high or on max and it will think longer. Right? And I can also turn on or off
201. 31:07thinking, right? So, this is the reasoning switch. So, if you just want to, you know, sort of crank stuff out,
202. 31:13throw Sonnet 5 on medium. You could even do it without thinking, turn that off and it'll just go pretty
203. 31:20fast. So, here, let me let me uh write a prompt for Fable 5 to clone the game Asteroids. We'll see how fast this is. So, this is
204. 31:32Sonet 5 Medium with no thinking. So, it actually remembered that I've done this before. It's not super fast. I
205. 31:43thought it would be faster than this. Do you want me to build this directly instead so I can generate the file right here? Um,
206. 31:54let's do a new chat. Actually, let me go copy that prompt. Copy. We'll do new chat. I'm going to turn it on sonnet. We'll go
207. 32:07on max thinking. Sonnet max thinking. We'll put in the same thing. See if it it's slower or the same. Before I write
208. 32:15this one thing worth flagging, you're deep in the open clock something workflow in Fable 5 is anthropics mythos tier model.
209. 32:30Few notes on frontload physics and scoring.
210. 32:50All right, let's just grab this prompt. So, I'm going to spin this thing up with Mythos. Okay. So, let me pop over to Mythos.
211. 33:04I mean to Fable. We're going to go to Fable and we'll put we'll put it on we'll put it on high. That's the default. And so,
212. 33:14I'm going to paste in. So, this is basically a prompt that Sonnet wrote for us for Fable to build us an Asteroids
213. 33:22clone. So, we're going to go do that. And so that's going to go build build stuff. What's the difference between sonnet and fable?
214. 33:31So again the scale of the models fable is a 10 trillion parameter model here. In fact, let's go to chat GPT and ask
215. 33:41um what are the parameter counts for haiku sonnet opus and mythos from anthropic.
216. 34:07We're in the last minute of the game.
217. 34:26All right. So, Haiku is 20 billion. Sonnet is 70 billion. Opus is 1 to two trillion and then mythos is 10 trillion. Yeah. So that's
218. 34:38it. So the the difference in them basically number of parameters is just sort of amount think amount of neurons
219. 34:46in your brain that are connected. So amount of detail that you know it can do all of its stuff on. So you can get much
220. 34:55more granular with information using a bigger model. and that's why they're smarter. Um, but they also cost a lot
221. 35:04more to run because you're having to keep those 10 trillion parameters in RAM for hundreds of millions of users,
222. 35:12right? So that's that's that's that thing. Okay, this thing should be Yeah. So Fable is now off. By the way, US1,
223. 35:21that game is over. Go USA. Um, okay. So let's go to I'm going to go to gemini.google.com.
224. 35:37And so I can say create an image. And this is the new create with nanoanana 2. I think this is the new
225. 35:47nano banana light. We now return you to your regular scheduled programming already in progress. Am I still sharing?
226. 35:55It's a little miracle. Okay. Um, let's see. Um, make me an image with 26 random playing cards
227. 36:15with no queens. um scattered on a bar room floor
228. 36:32all face up. We'll see how fast this is. And Brandon, if you could go research if if the new Nano Banana and the new um
229. 36:50Gemini Omni are are in gemini.google.com. I think they are. Okay. So, let's go look at this. We've got So, I asked for 26. No queens.
230. 37:04And so we got one, two, three, four, five, six, seven, eight, nine, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
231. 37:1622, 23, 24, 25, 26, 27, 28. So it didn't get 26, but it got 28. It's pretty close. And then did it I think these are both king of spades up
232. 37:30here. Or is one a jack of spades? I think they're both kings and there's two king of clubs. All right. So, let's tell it it [ __ ]
233. 37:42up. Let's say um Okay, it looks like Gemini.
234. 37:57Okay, it looks like you laid out 28 cards, but also there are duplicate
235. 38:13kings of clubs and spades. Fix that.
236. 38:34So, what did you find out, Brandon? That that Nano Banana Light and um Omni Flash are in Gemini now where I am.
237. 38:50Okay, it didn't fix it. Yes, as long as you you're in Oh, fast mode. Got it. Okay, got it. Okay, cool.
238. 39:02All right, so that's not bad. It's not perfect, but it's not bad. And then we can also go to we can turn this into a movie.
239. 39:11We can we'll say create video. It's going to start a new chat. New chat's fine. Don't ask me again. Uh say um start with
240. 39:29Um, 26 random playing cards base up on a bar room floor and have
241. 39:44them lift off the floor and shuffle perfectly into the
242. 40:01second half of the deck. All right, let's see how that goes. We'll see how fast this is. So, when I did this earlier, this looked like it
243. 40:12was quite fast. could take a few minutes. We'll check back on Oh,
244. 40:29wow. That's looking promising. A tribute to Atari. Oh, that's cool. It did a tribute to Atari. Okay, so
245. 40:41normally when I do these things, it gets the graphics wrong. These are the the asteroids are not quite the right shape, but it's pretty
246. 40:50close. Oops. Good.
247. 41:17Pretty good. So, I would say this is in the neighborhood of like better and more accurate than anything I've ever done. doing this
248. 41:32exercise. But here's what I'm gonna say. I'm gonna say here, reload. Okay. I'm gonna say hello.
249. 41:49How do you stop it? Go to code. There we go. Okay. I'm gonna say um the asteroids are a little too jagged
250. 42:06and don't look like the original. And I want it to look like a real CRT with
251. 42:20ghosted burnins and um scan lines. Yeah, we'll just see we'll just see what it does there. Okay, we'll come back to this.
252. 42:40All right. So, that's doing that. Did this make a video? So, here's our video. Okay. Let's look at our thing. One, two.
253. 42:48There's two ace of spades.
254. 43:08Yeah, it's in the hood. Not Not horrible. Okay. So, here's
255. 43:23here's what I wanted to do. So, that's building that thing. Let's go let's go create a new window of Claude. So, we're going to use Fable.
256. 43:38All right. And so, what I'm going to say here test one, two.
257. 43:54>> Okay. I want to make a series I want to make a series of short video
258. 44:09vignettes about different elements in the periodic table
259. 44:24period. I want the style to be paper craft and I want there to be cute
260. 44:38characters that explain each element. Period. I want them to be scientifically accurate and I want the characters for each
261. 44:55element to describe how that element occurs in nature and is commonly used. Period. I want you to create a script for each
262. 45:13element in the form of a prompt that I can give to a videogenerating model that includes
263. 45:28style, speaking script, character descriptions, location, layouts and any details you need to include
264. 45:43about how that particular element's atom works and should be represented visually at
265. 46:00I want you to make five scripts for the first five elements in the periodic table and then randomly select five other elements
266. 46:12that have much larger atomic structures. I want each prompt for an element to be a standalone prompt
267. 46:27I can give to the video model.
268. 46:55All right, let's see how this does. Uh, and then did this. All right, it made Am I still sharing? Yes. Miracle. Okay, so this made it at
269. 47:11least Oh, yeah. There's there's a little uh whatever you call it um CRT effects here. G for glow, C for CRT. G for glow,
270. 47:28C for CRT. Space to start. Now it doesn't have any sound.
271. 47:54So, now we lost the sound. All right. Um, it's all right. Let's see how this thing's doing. Oops. That's a fun one. It's right in your
272. 48:06wheelhouse. For the heavier elements, I'm selecting iron, copper, iodine, tungsten, and uranium. All right. Beautiful.
273. 48:29>> Yeah, I'm not doing anything here very sophisticated. Vanita, part of part of my MO with with these new models is I
274. 48:37will purposefully ask them um give them shitty prompts because part of what I like to look at is how much the model
275. 48:46actually helps a user who isn't sophisticated. Again, I think we're moving into a world where if you want to be building things,
276. 48:56you're going to have to know how to prompt. But these agentic models are going to get to the point that most
277. 49:01people are just going to be giving them goals and they're not going to be paying attention to the prompting. The
278. 49:06prompting is going to happen behind the scenes. Um, so I'm kind of purposely coming in and shitty prompting to see to see what we
279. 49:17get. And like, you know, how long this thing's thinking right now is is actually kind of interesting. Um, because it's writing
280. 49:24um a video prompt. It's not a very sophisticated thing, but it's got some science in it. Um, it'll be interesting to see if this,
281. 49:33um, kicks in any of the safety guard rails because I saw someone today did a prompt on, um, how does like cell mitosis work or some,
282. 49:45you know, some simple biological concept and it basically kicked out a a safety concern. Um. Oh, now it's got audio. That's weird.
283. 50:35pretty good. Yeah. All right. All right, it passes the asteroid test pretty swimmingly. Okay, here we go. So, let's see what's
284. 50:50going on there. That came out cool. That was pretty good. But your AI learns from you and you can become and can become
285. 50:56online. Do you wipe your chats and memory between demos? I don't uh I am I am a total data cleanliness trash heap.
286. 51:08I know I'm looking to put it put something together better when when I build my digital self, but right now I'm I'm an absolute trash sheep.
287. 51:19So, no thank thank you for asking though. I appreciate your concern. Um, okay. Style handmade. Okay. Prompt one.
288. 51:29Hydrogen style handmade paper craft. Okay. Let's see. Um, let's just copy this as it gave it to us. It's not saying it's a video,
289. 51:45but we're going to go prompt one hydrogen. All right. So, we're going to go to Gemini. We're going to do videos. Let's do a new chat.
290. 51:56I'm going to say make a video. Create video. We're going to do landscape. What's this? Uh, that's upload a picture.
291. 52:19Oh, that's cool. So, I can pick a visual style. Well, I said paper craft, so let's pick pick origami. Here's cut paper. We'll do cut paper.
292. 52:30And then we'll throw in that thing. Boom. All right, let's see how this does. So, let's go read the prompt.
293. 52:38Handmade paper craft cut paper stop motion aesthetic. Every object in the scene is made of layered construction
294. 52:44paper cards stock tissue paper with vis visible cut edges. Slight paper texture, subtle drop shadows between the layers,
295. 52:54stop motion jitter, 8 to 12 frames per second feel, soft diarama lighting. Okay. No, no photo realism, no CGI
296. 53:03gloss. Character. Hi. A tiny round single paper circle in bright red orange with two googly style paper eyes and a
297. 53:13wide smile cut from black paper. Hi is the smallest character in the series and he's proud of it. He bounces constantly
298. 53:20full of energy because he's number one. His single electron is a small pale blue paper bead that orbits him on a visible
299. 53:28loop on a white paper ribbon. All right, this is cool. So, this is high. So, let's see. I'm generating your video. This could
300. 53:38take a few minutes. Let's go grab the second prompt and we'll just toss it in there as soon as high is created.
301. 53:52I mean what's interesting here is does Fable understand I mean it seems pretty clear speaking script oh okay
302. 54:07here's a problem okay all right we have a problem you have speaking scripts that are way longer than the 8 to 10
303. 54:25second limit of the video model I am using. Um I want one prompt
304. 54:40per video of 8 to 10 seconds. update these to tell the story
305. 54:55in a much shorter fashion. All right, let's see how this first one did. There's high,
306. 55:12right? Stars like the sun. They're >> How do I Why can I not zoom in on this?
307. 55:26>> Almost. I travel light. Two of me plus one of her makes water. I launch rockets. I'm a big deal. >> That's cool.
308. 55:38>> Stars like the sun. They're almost I travel light. Two of me. Okay, we need to add one other thing to
309. 55:45this. Um, I also want text on screen and a design that shows the name of the element and the
310. 56:01uh, molecular details like I am looking at a
311. 56:19cell from a periodic table that just happens to talk. Okay, Chef Kelly. So funny that we all hated Clippy, but we're clamoring
312. 56:32to create him and Fable. Is that JavaScript? Uh what it wrote for for um for asteroids? I think it's just HTML doct type HTML.
313. 56:48Um, let me see. Um, is this JavaScript?
314. 57:05All Clippy wanted was for us to talk to him. Yes, this is vanilla JavaScript embedded in line in a single HTML file.
315. 57:11No frameworks, no libraries, no build step. The file has three parts. A small block of CSS at the top, 30 lines of
316. 57:19HTML, the canvas elements, and touch controls. And then everything else, roughly a thousand lines of plain JavaScript in one script tag.
317. 57:29The game itself runs on two native APIs. The canvas 2D API for the vector drawing and the web audio API for synthesizing the sound.
318. 57:40That's why it's zero dependencies and no install. Any browser can run it. Same pattern as your deep core tricolaris build. Single file
319. 57:53self-contained. That's cool. All right, let's see how this thing did. Did it fix it? All right. Oh, wait. It didn't take my prompt.
320. 58:04Let's Let's get these Let's get these prompts updated. But that's not bad. >> Plus one of her stars like the sun, they're almost I
321. 58:17travel light. Two of me plus one of her makes water. I launch rockets. I'm a big deal. Stars like the sun. They're almost I
322. 58:30travel light. Two of me plus one of her makes water. I launch. All right. Is this done? No, it's still running. Okay.
323. 58:43That's a great framing device. The video is the tile and the tile happens to be alive. Let me break let me bake a
324. 58:50consistent tile frame design spec into every prompt. Symbol name, atomic number, atomic mass, all rendered as cut
325. 58:58paper typography that's part of the diarama itself, not a digital overlay. Okay, that's good. So, it understood the
326. 59:05assignment. That's That's a good sign. And I didn't give it details about what information I wanted there. It figured
327. 59:13that out. So, that's good. Oh, that's cool. That that Asteroids
328. 59:28game can run on almost anything, including a USB stick and a digital camera. That's good. Yeah, I mean, you know, we do live in a
329. 59:39world now where you can just buy like little screens and little little computers, you know, disposable little
330. 59:45computers. So, you could burn all this stuff into a little into a little uh chip SD chip. Okay, let's see. Hydrogen. Okay,
331. 1:00:00let's grab this and do it again because what I'm thinking here, what I was what I was thinking would be fun to do, let's see
332. 1:00:11if if this does it well. Um,
333. 1:00:28Why? Oh, you've reached your video limit generation. Oh, come on. Really? Um, let me try flow.google.
334. 1:00:43My problem with flow.google is that it was very slow when I tried it today. Create with Google flow. conflict of interest question.
335. 1:01:01Hang on. How do we know that when the AI says, "Let me do this next step that I think you need to be done." And
336. 1:01:16let's imagine the user isn't an expert. It isn't certain. How do we know that the AI isn't suggesting something that
337. 1:01:22will burn tokens and make you spend more money? Um, one of the I I saw a thing Mary Carol I think it was about Claude, but it
338. 1:01:36might have been about Codeex. I think it was Codeex. I think a recent update of codeex, they actually improved something
339. 1:01:49where what codeex was doing was it was trying something, failing, and then trying the same thing over again. And it
340. 1:01:57would do it like eight times in a row and just keep failing. Um, and so it was burning tremendous amounts of tokens. In fact,
341. 1:02:07Yeah. Yeah. This was this was this was in the past two or three days. Codeex has done full um usage resets for
342. 1:02:16everyone using it because they did some sort of update that was inadvertently burning tokens and and so so they fixed something that
343. 1:02:27was doing exactly like what you're talking about. Now, here's the reason that I don't think I don't think they're going going to nefariously
344. 1:02:40create things that are going to generate more tokens to just burn more money. One is corporations have gotten wise to
345. 1:02:48it, right? So, corporations for for about a six-month period there were doing what was called token maxing where
346. 1:02:54they were just saying to their employees, you're going to be measured on how much AI use you you do and just go use AI. And then the
347. 1:03:07employees were just like, "Okay, well, I'll just [ __ ] turn agents loose and have them go do things." And they just
348. 1:03:12ran up ridiculous bills. So one thing is corporate America won't stand for that. Um that's one of the things that's driven corporate America
349. 1:03:23from anthropic and open AI to Chinese models. Um so that's that's one of the problems. The other problem is if it just keeps
350. 1:03:33doing that and doesn't actually solve your problem and some other model comes along that solves your problem. Like
351. 1:03:38there's there's a competitive thing here where it's not just about how long it's thinking. It's it's about did it
352. 1:03:44actually get the job done. So, it's it's in the interest of these frontier model companies to have these things work as
353. 1:03:52good as possible or no one's going to use them. Um, and then the the third element is they don't want to be wasting tokens because
354. 1:04:02we have relatively um minuscule adoption of these tools, right? We're we're still in a like I don't know 5% 10% adoption
355. 1:04:19kind of percentages. Like 90% of people aren't using this [ __ ] 90% of companies aren't using this [ __ ] They're not
356. 1:04:25using it in the way they could. So we're going to very very quickly start not having enough compute to do what we need
357. 1:04:34to do anyway. So if they're just blowing through tokens, it's just it just doesn't it's just not economically
358. 1:04:40viable. I don't think um let's see what's this. It does not know or understand a missing piece of data. It'll just try to make you happy.
359. 1:04:49They're coded to give you what you ask for. Yeah. And and when if the if the harness just keeps running the same loop
360. 1:04:56over and over again, it it will waste those tokens. some of the new systems. Wanita, there's a new system that came out for math
361. 1:05:04three days ago that is doing the mathing and then it's actually checking its work and it's going back and fixing it. It's not
362. 1:05:12returning a result until it gets it right or what it thinks is right. So I think systems are going to be coming to
363. 1:05:18get smarter about that. um where the the initial generation is just you know this probabilistic thing but then through
364. 1:05:26cycles of is it right um it'll get better and better and better. Um asking the right question on an aligned system will end up with some oh
365. 1:05:38on an unaligned system will end up with some wasted tokens. Yes, I'm a token waster. Oh man, been in Mary not just not getting
366. 1:05:52notifications. So, this is the first time I found you all. So, good to have you here. Really cool to have you here.
367. 1:05:57Okay, let's see. Is Oh, wait. We were in Google Flow. New project. Why is that so small?
368. 1:06:16agent 8 second. We'll do 10 seconds. Omni flash, you're out of credits. Oh, so I can't do
369. 1:06:30this. I'm out of credits. I used to I used to pay 200 bucks a month to Google and I don't anymore. Um,
370. 1:06:38all right. So, I guess we're not doing the video stuff. Well, that kind of sucks. What are you gonna do? What are you going to do?
371. 1:06:57Sabrina Romangh sent a great email this morning about Chinese AI just beat Claude Code Opus for one sixth the cost.
372. 1:07:04Yeah, that's that new um GLM 5.2 two model is supposed to be um opus level at a sixth of the cost. This is the thing.
373. 1:07:14I mean honest to God like like the the damage that I think Daario Amade did by by
374. 1:07:28shifting into AI is going to kill us mode. Um can't be overstated. We'll we'll see how it plays out. GLM 5.2. Yeah. Um,
375. 1:07:44so now here's the thing. One of the things I've been talking about is it is wise for us all to get smart about
376. 1:07:56understanding what local models we can run, investing in hardware that can run a decent sized model. But something like
377. 1:08:03GLM 5.2, two, you need major hardware to run that. You can't run that on on desktop hardware. So, what I could imagine happening is
378. 1:08:14that decentsized companies or groups of individuals get together and pull their resources to buy a machine capable of
379. 1:08:23running an opus level model that that group can tap into, right? um and basically do it outside of the frontier models
380. 1:08:37drama and and this is this is what I think Amadeday's done here is he's probably put back adoption of chat GPT claude
381. 1:08:52not quite sure about Gemini I think Google probably lives in kind of a different realm I feel like Google is in the
382. 1:09:01neighborhood of Microsoft in terms of like well you can't go wrong going with Google or Microsoft they're big players
383. 1:09:09but I think the the true frontier right now there's only two frontier companies right there's only there's only um
384. 1:09:17anthropic and open AI and you could put Google in there but again I feel like they're in a different category because
385. 1:09:22they have a bunch of other [ __ ] going on. Um, I think all of this stuff has set those guys back and I think it's opened
386. 1:09:30the door wide open for these Chinese models. And so don't be surprised, you heard it here first, don't be surprised
387. 1:09:37if the government at some point makes possessing open-source models illegal, which that's what another thing that Daario Amade is
388. 1:09:48pushing for because he realizes that people aren't going to pay his $50 per million tokens for mythos if they've got a cheaper open- source
389. 1:10:01model. So, I think that there are all sorts of shenanigans going on here that I do not like. Perplexity, I wouldn't
390. 1:10:09consider um a frontier model. And Meta at this point is a joke. They don't have any scientists. They've got that 22-year-old kid that
391. 1:10:17that they brought in as Yan Lun's boss. They said, "Yan, report to this 22-year-old kid." And Yan Lun said, "Fuck you." and left. So,
392. 1:10:33oh my god. Let's see. I like that y'all are jumping in. That is good. And oh my god. Yes. The waste of tokens is epic. Yeah. I
393. 1:10:48listen I This is one of the things we talk about in the AI salon in the in the AI salon mastermind. So, if you haven't joined
394. 1:10:56the the mastermind in the AI salon, you should do so. And within the AI salon, we've got this these tracks called the
395. 1:11:04great repurpose. And one one lane of of the great repurpose is about human beings reclaiming your agency as a
396. 1:11:14person. Then the other track is the AI one. Reclaiming your agency as creating a digital self and and reclaiming some
397. 1:11:21agency over um your data and and recognizing that your data is incredibly valuable um especially in these agentic
398. 1:11:32systems. And if the government and these frontier model companies just keep [ __ ] around and lose the trust of the
399. 1:11:40people, well, how are we going to keep doing what we do? Well, um, we start learning how to do do it on our own. Um,
400. 1:11:50and I have, listen, I've always supported open source, but I've never felt like a need to actually learn it um like I do right now.
401. 1:12:01Could you run it on cloud or AWS? Yes, you could. Yeah, you can buy you could certainly buy a sandbox Chef Kelly of
402. 1:12:07like on AWS that's got enough um um GPU RAM or unified RAM to run GLM 5.2, but it's not going to be 20 bucks a
403. 1:12:21month or 200 bucks a month. It's going to be it's going to be pretty expensive. Um actually, it's probably, you know,
404. 1:12:27that's probably not a bad Let's go let's go do a little research on that, shall we? Yeah, I'll show that. Do you want to
405. 1:12:41talk about that, Brandon? Um, let me see. If I wanted to run GLM 5.2 to um in a cloud
406. 1:12:57sandbox on AWS or similar. Um what level of VM would I need and roughly
407. 1:13:14what would be the cost on a monthly basis? I would want enough headroom
408. 1:13:28to have max max context window for let's say up to 50 people. Um, give me the
409. 1:13:45bare minimum minimum. God damn it. Give me the bare minimum for a single person instance
410. 1:13:59then 10 then 50. also let me know what other things I need to have in place.
411. 1:14:16All right, that's fine. Oh, my computer's on live support. Okay, I will uh Yeah, I'll go show that in a second, Brandon.
412. 1:14:37I'll sanity check the current VLM 5.2 specs and live GPU pricing. Then I'll translate that into practical VM tiers.
413. 1:14:58Sleepy a key finding already. Max context here is the cost driver, not just the model size. GLM 5.2 is a 753 billion parameter
414. 1:15:10model, a mix of mixture of experts model with a million token context window. Serving it casually is very different
415. 1:15:17from holding many million token sessions open at once. Okay, that's cool. So, I'm giving you the the max price here.
416. 1:15:41$63 an hour. That's what that looked like.
417. 1:15:56Here's the practical answer. An 8GPU data center node.
418. 1:16:14All right. My recommendation for a real sandbox, I would not start with AWS. For testing, rent one 8 by H200 or B200
419. 1:16:29node from runpod, nbius, core, weave, lambda, base 10, etc. For production like 50 50 per person access, use a GLM
420. 1:16:395.2 API first, then selfhost only after you know the actual token usage. The API price is currently $1.40 per million input tokens and 26
421. 1:16:52cached input and $4.40 output which is cheaper than idle GPUs. Okay. So, one person monthly, $21,000.
422. 1:17:1010 people, a shared sandbox, 21 to $87,000. 10 people all simultaneously near a million tokens. $210,000.
423. 1:17:22Yeah. So, so this [ __ ] ain't cheap. And and listen, this is a 750 what did it say?
424. 1:17:36754 billion total parameters. Mythos is 10 trillion parameters. So this [ __ ] ain't cheap. AWS monthly is 33
425. 1:17:50grand. So you're looking at 20 grand a month it looks like to stand up a decent instance of GLM 5.2. So
426. 1:18:03let's say you could you're not going to have 10 people maxing out a million token context window all the time. But
427. 1:18:10let's say you had 10 people running aentic businesses. um you'd need to make, you know, half a million dollars a year in revenue to
428. 1:18:22basically make a profit, right? It'd cost you 200 250 to three to to 250 to $400,000 a year. So, wild,
429. 1:18:38wild, wild, wild. Um, where am I going? community feed. All right, so we have a um
430. 1:18:54new feature alert. Okay, we're doing trivia. Good. I like this community feed. So go to So in the AI salon, go to
431. 1:19:02community feed right now. Am I still sharing my screen? Yes. All right. So, so you see me as the the game show host slashworld's greatest
432. 1:19:12showman there. Um, so if you've done any good work, share it in the community feed. If you have any questions, put it in the community
433. 1:19:22feed. Um, try your luck with trivia. So, we're doing a new trivia thing. And then here we go. Which of these is usually
434. 1:19:31the best way to improve an AI answer? All right. So, go answer those. We've got three votes right now. Um, go answer
435. 1:19:42those. No show and no show Friday. Okay. Um, let me pop out of here and turn off uh screen sharing. Okay. So, um, Fable,
436. 1:19:53it's there. Go play with it. Go do [ __ ] If you know how to prompt and do good [ __ ] with it, great. If you just want to
437. 1:20:01go dick with it, go dick with it. Um, you will probably hit your usage limits pretty quickly. I heard couple of people
438. 1:20:09today say they hit their usage limits in five minutes with Fable. So, so there you go. Um, let's see. Yes, Cam Conversation.
439. 1:20:26Um, okay. Let's see. Um, this Friday we do have office hours. So, if you go into the AI salon, go to events Friday at 11:00 a.m. uh mountain
440. 1:20:40time. I am going to do office hours. I'm not doing content evolution and I'm not doing Friday night date night, but I am
441. 1:20:46doing office hours. Um, so go RSVP for that. Now, if you haven't been to one of those before, those are fun. Um, and
442. 1:20:54it'll probably be smaller because I assume people are going to be traveling and doing [ __ ] for the fourth. Um, so it
443. 1:21:00should be a good one. And then, um, over the weekend I, you know, I would just play with take advantage of the fact
444. 1:21:07that you have access to Fable, right? This is a MA model so powerful that it was government locked down for
445. 1:21:14two weeks. So, go play with it for two weeks or for seven days. So, that that weekend homework is go play with Fable. Um,
446. 1:21:25and if you want to do video stuff, it looks like you're going to have to pay more. They used to give that away. Um, I wonder
447. 1:21:34one thing that I do remember now is on the Gemini app, it would do video. Make me a
448. 1:21:52Make me a video of a hydrogen hydrogen atom dancing around with its electrons
449. 1:22:06telling the world the kinds of things it can do. Mail me a video. Well, whatever. All right. generating video. So again,
450. 1:22:22it looks like on the web, my Google account maxes out video really quickly, but when I do it on the Gemini
451. 1:22:33app, it seems to just keep making it. So, I don't know. If you've got a paid subscription, try it on the app. That
452. 1:22:38might be a way to continue making video if you don't want to spend a [ __ ] ton of money. Centipede was my favorite. I'll
453. 1:22:44ask mine to make one for me tonight. Yeah, I mean this is the cool thing about having it make clones of nostalgic
454. 1:22:50video games that you viscerally remember. Like I remember the feeling of Asteroids. Um Centipede was one one of
455. 1:22:58my favorites, too. Um but you don't have the track ball, so it's going to be it's going to be rough with the keyboard. But
456. 1:23:03you know what that felt like, right? That that game that that Claude just made, the Asteroids game that Claude
457. 1:23:10just made for me felt like the original. Like it felt the physics felt right. Um, cool. Awesome. Awesome. Awesome. Awesome.
458. 1:23:21All right, I'm going to get out of here. It is uh 8:55. I hope you all had fun tonight. Please, if you're new to the
459. 1:23:28salon, go check out the different spaces. Um, if you haven't introduced yourself to the community, go do that.
460. 1:23:34Say hello to us. Um, it's a very welcoming bunch. Come to office hours. Come to the events. Go to the events
461. 1:23:40calendar. There's a ton of events, including these lives. But there's also lots of stuff during the day. So if
462. 1:23:46you're not a night owl, there's stuff for you during the day. Um, go play with Babel. I would also say play with Sonnet 5. See if it's good.
463. 1:23:59See if you like it better than you like other things. And then it looks like tomorrow um, we may get chat GPT 5.6. There's
464. 1:24:09three different sizes of 5.6. And plus, I think they've got a new voice model coming out. So, um, that
465. 1:24:16could be this week as well. My guess is it's going to be next week, but we'll see. Maybe they do it. Um, I think that's it. Oh, yeah. Sorry,
466. 1:24:33I can't generate any videos for you today. Yeah. So, they have locked down. That's what I'm saying. um the
467. 1:24:41subsidized era of of free tokens is is exiting. Um this is going to start costing us a lot of money which is going
468. 1:24:50to start making you think about what you use AI for. Um so we the free ride's over kids. We're in a
469. 1:25:06new era. Um, but we're also in a more powerful era. So, it makes sense that um, the free era is over and now it's
470. 1:25:14time to get serious and start doing some work. So, that's kind of the era we're entering. Um, so, a little less
471. 1:25:19frivolous, but we can still have fun with this with the [ __ ] that's there. But, if you want the really good stuff,
472. 1:25:24you're going to have to pay for it. All right, Gubby. Peace out. Hope you're well. And uh if I don't see you at
473. 1:25:32office hours on Friday, tomorrow we've got at 10:00 a.m. if you're in the mastermind, we've got the practice lab.
474. 1:25:39So the mastermind practice lab with myself and Sid Harrow is tomorrow. So if you haven't been in a while, come join
475. 1:25:45that one. Center yourself, work on your practice around how you use AI. And then uh I'll see you Friday at office hours.
476. 1:25:52And if not, have a beautiful Fourth of July weekend. All right, everyone. Peace out. mastermind Thursday 9:00 am Pacific
477. 1:26:01time. Exactly. Happy 4th of July, 10:00 a.m. Pacific office hours. Beautiful Ace. Oh, yeah. Next Tuesday, AI Salon
478. 1:26:11presents. It's going to be a good one. Bye, everyone.