9/18/2026 - Exploring Open Source Fruit Fly Brain Simulations and Neural Networks

Transcript

1. 0:00Heat. Heat. [music] [music] [music]
2. 0:20[music] >> [music] >> Good evening, good people. Happy Friday
3. 0:34night. Friday night date night. AI Salon is back. We took two days off. Monday and Wednesday, I was in Philadelphia,
4. 0:42city of brotherly love. I was there uh doing uh work for Storyvine. It was a fierce pharma week. I was doing pharma
5. 0:53marketing with all those fierce pharma marketers. [laughter] But uh it was a really good it was a really good trip. Uh and uh and got some
6. 1:04good work done and hopefully made some uh some connections to to sell some stuff. Dr. Jay's in the house. Yeah, if
7. 1:13you're here, the the uh the comments are down below the live stream. I'll be able to see them here in the live. Please say
8. 1:21hello. If you got any questions, ask them. Um, normally on a on a normal week, we would play on a Monday, we
9. 1:28would build on a Wednesday, and we would probably interview someone or talk about leadership on a Friday. Since this week
10. 1:34got mucked up, we're just going to sort of catch up. Um, a lot of stuff happened this week. So, I think tonight's going
11. 1:39to be a bit about uh surfing around uh on on the uh the X. I got access to it this new model called Jev JV um [snorts]
12. 1:52which looks pretty remarkable but it does require programming skills which I don't have but I do have codeex which
13. 1:59has them so we can probably build something interesting over there is my guess. Um so we'll do that. Love their cream cheese.
14. 2:10Yeah, I know. It does sound like a cream cheese brand, doesn't it? Um, what's happening, Lord Digital Gods? It's been
15. 2:19a while. Hey Rick, what's happening? Good to see you, Val. So glad it's Friday. Welcome. It's Friday night date
16. 2:24night. If you're new to the AI salon, welcome. So here at the AI salon, this is a place about uh creating space for
17. 2:32you to understand AI, explore it, get excited about it, get terrified of it, work through that, figure out what it
18. 2:39means, figure out who you are, what you're all about, and put that out in the world. So um so welcome. Down the
19. 2:45the if you're in the AI salon community, down the side are a bunch of different spaces. Check those out. Um, one of the
20. 2:52things that you will discover in the welcome area of the AI salon is this thing we call the cycle of AI readiness.
21. 3:01And the cycle of AI readiness is play first, create excellence, and generously lead. And uh with you at the center of that
22. 3:10cycle. And so the idea with play first is is learn through play. Play without expectations. Play with purpose. create
23. 3:18excellence is take what you learn and build things that are meaningful to you or attempt to build things that are
24. 3:25meaningful to you. Uh and then generously lead is about sharing what you're learning and and you know asking
25. 3:32questions and answering questions and sharing your successes, sharing your failures and then just keep repeating
26. 3:38that process. Um, this is the only way that I know um for you to have a mindset that is going to um allow you to be successful
27. 3:52uh with AI as things change radically fast. Um Jeb, it's a great rating system. Pick the best photos, rate
28. 4:02stocks, rate songs to find the most viral. Fascinating. Um, yeah. I don't know how to set it all up yet, the Jev stuff, but uh we will
29. 4:12figure that out. Maybe we'll have maybe we'll have Codeex give us a little tutorial. I might have to fire off one
30. 4:20of my resets tonight because I have a feeling I'm going to blow through credits. Let me see if I've already
31. 4:25blown through them. I haven't. Got it. You ran the MPX skill add command that installs the type safe skill, not a Jev
32. 4:32executable, so nothing is missing there. The skill is now available in codeex. Okay. SDK whichever pressure is whichever
33. 4:45uh let's see I want it available to all projects. 7% left for my weekly. [laughter]
34. 5:08Okay, that's good. All right, let's see. Um, a second. All right, who's this guy? What's
35. 5:22happening, Silver Fox? Hope you're doing well. By the way, if you're on uh Tik Tok and you're wondering what I'm
36. 5:27looking at here on the screen, uh it's in the uh if you go to my profile there, we're live in the AI salon. We're live
37. 5:34streaming over there. Uh Tik Tok's the backstage view. If you want to chat with each other, you can do that here. If you
38. 5:39want to see what I'm looking at, go there. Producer Brandon up on screen. Good evening, sir. >> Good evening. I would I would That's
39. 5:48false advertising. the the folks on Tik Tok cannot see what you're seeing on screen because you're not sharing your screen.
40. 5:56>> Oh, I wasn't. I was just I was just messing around. Hey, hop back up here if you could. How was the uh how did your
41. 6:03uh Codeex uh mastermind workshop go this week? >> It's it's going really well. So, we're three weeks in and we're actually using
42. 6:10Codeex now. Uh and uh we've got the environment set up and we've got some really interesting use cases that we're
43. 6:17stress testing. Um, one of which is how best to recover and repurpose custom GPTs like Sin City. They're going away in December. >> Yes.
44. 6:29>> And so, um, my codeex was actually able to log in using browser control and archive all 40 of my custom GPTs in like two minutes.
45. 6:41>> Oh, that's such a good idea. and I realized that most of them I don't really need anymore, but there are a
46. 6:49handful in there that are really cool. Um, so we've got the replay coming up soon and week four is next if you join the mastermind.
47. 6:58>> Great. Yeah. So, by the way, if you you might as well pop up the graphic there. Hang on one second. Let's see.
48. 7:31Hold, please. All right, there we go. Um, yeah, if you uh Thank you. If you uh pop back over to mastermind, there you go. Um, if if
49. 7:45you're new to the AI salon, the the mastermind is a subscription area of the AI salon. Um, there are really amazing uh workshops
50. 7:57and things to do. The mastermind practice lab is part of the mastermind. Um, the workshop that Brandon just
51. 8:03described about uh learning codecs is part of that. The great repurpose mastermind led by Andy and Andy
52. 8:10Scarantino is part of that. um really remarkable group of people. The workshops in the mastermind are
53. 8:18life-changing uh for me personally, for other people. If you haven't checked it out, you should definitely definitely
54. 8:25check it out. It's kind of the best deal in town. Um so, go check that out. >> It really is. And before I hop off
55. 8:31stage, I do want to give a shout out to my co-conspirators earlier this week, Rick McCaulay and Dr. Claire Jacobs, who >> Oh, yeah.
56. 8:42stepped in and stepped up with Sunno uh and we had a really great back to basics course on Sunno generation uh and that
57. 8:50will be available also on demand in the back to basics section soon. >> Yeah, that's beautiful. I'm excited
58. 8:55about that and uh I I hope it went well. Um I was I was busy gallivanting with the with pharma marketers that evening.
59. 9:05So So yeah, thank you to Rick and thank you to Dr. Jay for for popping in and doing that. Um, yeah, I'm excited to check out the replay.
60. 9:14All right, so let's see. Jennifer, hey y'all. Let's see. And the workshops are recorded. They are. If I missed codec
61. 9:22series, can I view it upon joining? You absolutely can. So any of the workshops in the mastermind are available on
62. 9:29demand, including all of the mastermind practice labs. If you haven't started a daily practice around how you use AI and
63. 9:37think about AI and and and put it in your life, uh the practice lab um replays are amazing. The uh the practice
64. 9:46lab uh actually happens on Thursdays at 10 a.m. Mountain time. Uh headed up by Sid Harrow. He's a he's a terrific
65. 9:56leader. Um and that's a great group. So yeah, they're all available on demand. So definitely uh uh join and and the
66. 10:04codec series that Brandon is running is awesome. He is he is um an incredible teacher and an incredible practitioner
67. 10:11of understanding not just how to use the technology but how to use it in a way that's like just practically useful, how
68. 10:19to do good stuff in your work and life. Um and uh in fact Brandon and Dr. J Claire Jacobs run the uh the life hacks
69. 10:29group as well. So that's available to everyone. So there you go. All right. Fantastic. So let's see where are we
70. 10:38going to begin tonight. If anyone has any questions, feel free to pop your questions down below. Um couple of things that
71. 10:47I've been kicking around uh thinking about today in office hours. We were talking about there there's a fly brain
72. 10:56uh that some dude, a single guy took six years to basically make all of the neural connections, the 166,000 neural
73. 11:08connections in a fruit flies brain. Um and then he open sourced it. And so that is now available for all of us to go play with. And so,
74. 11:19um, Google followed up hi his releasing the brain with an open-source version of a fruitfly body. So, you can actually
75. 11:28have the brain control the body and you can have it do all sorts of crazy things. Um, and people are doing some
76. 11:34remarkable stuff with that. So, that was one of the things we talked about in office hours today. Um, we can go look
77. 11:39at some of the things people are doing with that. It's It's really quite remarkable. And with with things like
78. 11:44Codeex um and and Astra and um Claude Co and things like that. Oh yeah, the Fly Beat Saber. Yeah, exactly. That one's really good.
79. 11:59um um that you you can you don't have to understand how to program a fly brain to be able to program a fly brain. It's all
80. 12:11documented and and you know codeex or claude code can basically go read the documentation and build you something.
81. 12:19Okay, Danielle, so I started teaching again a few weeks ago. Congratulations, special education. And I'm shocked about
82. 12:26how many teachers are using chatbt for lessons and assignments. Oh, wow. That's wild. Interesting. Are you going to show
83. 12:34us the fly building the red cup pyramid? I will. I'll go find it. I haven't looked at it in a few days, but I will go find it.
84. 12:42And yes, we can go look at um so so one of the things that I did was I wanted to teach um a fruitfly how to stack cups in
85. 12:53a pyramid. So I did. Well, I didn't. What I did was I said, "Astra, please go learn how this thing works and do that." So it's it's a lot
86. 13:06like cheating. Um, and there's something about using a completely different approach to intelligence um that I think is going to benefit us
87. 13:22all. So, so that's one one angle is is look at projects like that brain project where people are getting this brain to
88. 13:32behave in ways that are kind of like an LL LLM but it's but it's somehow different because it's a simulated fly
89. 13:39brain. Um and then the other thing is this this new model that came out from an ex one of the co-creators of chat GPT
90. 13:49and um reinforcement learning at OpenAI left and has spent the past two years creating this model called Jev JV.
91. 13:59Um I got access to it tonight and I don't know what to do with it. So that might be something we go explore.
92. 14:08Um, isn't AI going to kill us in a few years? See, Anthropic. Yeah, that was the the other thing that's gone down
93. 14:14this week is, you know, Anthropic uh continues to talk about how AI is going to kill us. And then and then today they
94. 14:23announced, oh, and by the way, we're uh starting our own wet lab. So, if you don't know what a wet lab is, when you
95. 14:32do things like simulated protein folding and drug design and you predict how drugs are going to interact with
96. 14:41specific tissues and proteins and how all of those things do what they do, you do it in a computer simulation. Well, a
97. 14:48wet lab is where you do that for real with real tissues. And so the company that's telling us AI is going to kill us just started one of
98. 14:58those. [laughter] I'm only laughing because I'm terrified. But you know, I'm not terrified of them doing something wrong. I'm actually I'm
99. 15:10just more I'm a little more I don't know what's that that thing like you know I'm not mad at you. I'm disappointed.
100. 15:20I'm not mad at you anthropic. I'm disappointed. Uh, so anyway, all right. Let's see. What's that? Jev at work. Song song songscore.co. Is
101. 15:36that something that you built, Rick? Let me get my sharing up. Jev is well suited for fast, repeatable scoring and a
102. 15:45ranking layer for creative assets. Well, for lots of things. It's designed to return structured choices, probabilities, rubric based scores
103. 15:54rather than pros. It it can't write pros. It only returns uh decisions. So, it works best as an evaluator inside a
104. 16:04workflow with conventional multimodal AI providing detailed explanation improvement ideas. Yeah, that's pretty cool.
105. 16:12Very, very cool. Uh, no, it's an example that's out there. Okay, cool. All right, we're going to go look at some of that
106. 16:18stuff. So, let me share my screen. I guess first thing we'll do So, first thing we'll do is we'll hop over to X here for a second
107. 16:29and pop that over there. All right. By the way, if you are on Tik Tok and
108. 16:44you want to see what I'm seeing, I'm over uh surfing uh Twitter right now. X they call it for you young kids. You
109. 16:52just know it as X. Us old people still Twitter in our head is the little bird, you know. Um but if you go to my profile
110. 17:01here, that'll take you over to the AI salon. That's where the live stream's happening. So, come on over there. This
111. 17:06is the backstage view. And uh we'll take a look at that. Okay. So, let's go take a look at Flybrain. If you go to Twitter
112. 17:16and search Flybrain, you will see lots and lots and lots of projects like this one. Beat Saver. [music]
113. 17:40[music] >> [music]
114. 17:56>> And you'll notice over there the brain is so here's all the uh you know the connections and what's lighting up as as
115. 18:04it's playing Beat Saber. So that's one of the things you can make with it. This guy got his uh his fly brain addicted to
116. 18:11gambling. [laughter] Oh, here's what Let's see what this one is. 166,000 neurons from a dead fruit fly.
117. 18:22And I guess how this guy mapped this fly brain is basically, you know, cut it in little slices and, you know, I guess
118. 18:30manually looks at all the connections between the brain and basically just recreates all those connections. And now
119. 18:36it's effectively a 166,000 parameter neural network um that you can train just like any other neural network.
120. 18:52chat jimmy.ai powered by Sarah Bruss. Buckle up, buttercup. Oh, that's the >> Yeah, this is uh you have to check this
121. 18:59out after you're done with the fly. This is uh instant AI, the world's fastest inference. >> Okay, cool. I'll definitely go check
122. 19:07that out. And I saw one also, there's uh there there's a company that's burning um large language models into the chips
123. 19:16themselves. Um, so there's no transfer back and forth like all the inference is happening um on the chip. So it happens
124. 19:25essentially instantaneously as well. Um, so lots of different here here's a fly that's uh that's surfing [laughter] surfing social media.
125. 19:37Doom scrolling the doom scrolling fly. [laughter] Oh my god, that's hilarious. There's there's there's one brain flying three drones.
126. 19:54And so, let me show you what I did. Let me see if I can go find it. Okay, let me do a search here. Search chats for fly brain.
127. 20:35See demo reload. Here we go. Here's the recording.
128. 20:51So, this was this is the uh the the fly body model from Google and then uh and then I had it um learn how to stack cups. But here's
129. 21:05here's my fly stacking uh going from a stack of cups to a pyramid of cups. [laughter]
130. 21:19>> [laughter] >> Now, there were a couple of things like I wanted to make the cups bigger so that when the when the pyramid got taller
131. 21:31that the fruitfly would have to fly up to to stack the top cup and things like that. A whole a whole another take on a
132. 21:38on a flight simulator. Yeah, you can turn these things into flight simulators. The thing that's wild about this is you're actually training,
133. 21:47it's called a a connetome, but it's a it's a virtual, you know, simulation of a brain, and you're basically training that brain how to do
134. 21:56stuff, but that brain is also connected to all of the different um fly body parts like the wings and the
135. 22:05abdomen and the, you know, antenna and things like that. And so, um, Brent Peterson today in office hours was
136. 22:11talking about he's got, I think, six or eight of these flies right now are reading all of what the hell are they
137. 22:20reading? Oh, all of the all of the versions of the King James Bible. And he's trying to figure out if if any of
138. 22:27the flies find God. It's rough because flies can't read. So, so they're trying to read the Bible. I don't I don't know
139. 22:35how that experiment's going to go. Um, one of the things that I would suggest and that that Brent Peterson suggested
140. 22:41on the call was um, just start experimenting with this stuff. Um, and and yeah, and I mean I I'll I'll
141. 22:52show you what my my prompt was early on up here. I said there's a new fruit fly fruitfly brain simulator with an API and I think also an
142. 23:06open-source fruitfly body that is rigged from Google question mark. I want you to find them and then figure out how people
143. 23:12are training them and I want to train ours to stack cups and make it easy to demonstrate with others. And then it
144. 23:20basically went off and did stuff and it it did a couple of really crappy examples and I blew through a bunch of
145. 23:26tokens um just trying to get this thing up and running, but I ultimately ended up with, you know, this example where it
146. 23:33it stacks cups. I don't really care too much about it, so I'm not digging all that deep into it, but I did at least
147. 23:41want to see like is it possible to, you know, go into a field I know absolutely nothing about? uh like like brain
148. 23:49simulations and uh and do something with it. And the answer is yes. >> Now we just need that Anna Kendra song. >> Which one?
149. 23:59>> The one with the cups from Pitch Perfect. >> Oh yeah. Yeah. Exactly. That's exactly what I was thinking of. But it's a
150. 24:05little slow for that. But that that would be the next thing, right? Get it to fly around, stack them up, do the song. [laughter]
151. 24:13Oh my god, that's so cool. Okay. So, that's that. Um, let's go take a look at Can you put that URL URL back up for uh
152. 24:33chatjimmy.ai? Okay. Okay. So, this is on um what model are we looking at here? Does it say
153. 24:53settings top K? I wonder if that's doesn't say.
154. 25:09Okay, let's see. Um, >> this one would be good for a Eminem style rap. Yeah, I'm gonna say write me a 20 page
155. 25:24epic M in M style rap about quantum
156. 25:39mechanics. Right, here we go. Boom. Dang, that was pretty quick. It's your boy Slim Sady back with the
157. 25:51facts. Got my sights on the secrets of the quantum X. I'm talking to the world of atomic scale. Where laws of physics
158. 25:57get flipped, the rules prevail. It's quantum mechanics. The game we can't ignore. Particles and waves collide.
159. 26:04Where the math gets more. I wonder what model this is. I wish we knew. Huh. Very cool. Very very cool. All right. What else? Here's someone
160. 26:19teach their uh their fly to play blackjack. >> It is its own model. They built their own inference. It's on sarabus.ai. >> Oh, okay. Cool.
161. 26:40>> He's playing poker. Wild. This one's cutting cutting hero meat. [laughter]
162. 26:55This is playing video games. It's a little Super Mario. Some Doom. Stunkfly.
163. 27:10Donfly is block trading Bitcoin. Desktop Fly, a 3D fly on your desktop. Doomfly playing Doom. Fly 64. Mario 64.
164. 27:25Amid running on a basic Raspberry Pi, navigating a physical six-legged robot through a live feed. Where's that one?
165. 27:40>> [music] >> Huh? Oh, a dead fly is driving my robot now. Not a metaphor. A reconstructed neural map
166. 27:55of a fruitlyund 166,000 neurons, 25.6 6 million synapses is now running on my MacBook and steering a quadrupled
167. 28:07a quadriped quadrupled quadriped robot in my living room. Here's how it works. Male CNS, a full conto reconstruction of
168. 28:18the adult fly brain. I used codeex to build a decoder that reads the simulation's neural activity and turns
169. 28:25it into movement commands forward, left, right. Those commands get sent over Wi-Fi to the robot, which still uses its
170. 28:33own firm firmware to drive all 12 servos. The brain only decides where to go, not how to move each leg. No
171. 28:42scripts, no pre-recorded paths. The simulation runs continuously live while the robot walks. The fly has been dead
172. 28:50for years, but its reconstructed nervous system is producing real decisions again in real time, controlling a machine that didn't exist
173. 29:01when it was alive. Next step, feed the robot sensors back into the simulation and close the loop. A dead brain that
174. 29:09can finally feel the world it's walking through. This fly has been dead for years. Its brain runs on my laptop. Its
175. 29:17decisions move a machine through my living room. One dead nervous system. [music] [music]
176. 29:33Push it to the limit. Walk along [music] a razor edge. Don't look down. Just keep your head or you [music] wild. Absolutely crazy.
177. 29:50What's this one doing? Oh, this is that like those beach sculptures that that guy made. I forget his name.
178. 30:00Strand Beast. It's really cool. Here's a fly brain that's learned Python. So, it's it's actually typing on the
179. 30:11keyboard. It's typing it's typing Python. [laughter] Oh my god.
180. 30:34Unbelievable. All right, here's another explanation. These are kind of cool, these explanations of people that actually
181. 30:46know what they're doing. Claire, what the hell? I know. Can it take the LSAT? Um, probably. It'll probably do poorly. It's only got
182. 30:56166,000 parameters, so it's probably gonna fail you there, buddy. But but you know, um why is this not loading? Come on.
183. 31:10Okay. One camera watches my open hand. That's the whole input. The panel in the middle of the brain. Wait, the panel in the middle.
184. 31:22Okay, that panel there.
185. 31:37So, which camera? Oh, I see. He's using the camera. The camera on the laptop. He's giving control signals
186. 31:51and the brain lives on that chip sitting on that controller. Oh, and I guess it's it's actually moving the controller. Wow.
187. 32:04The panel is the panel in the middle is running the brain. The camera reads the hand. The brain fires. The output becomes throttle.
188. 32:16Open palm and it climbs. Fist and it hold. Wait. Oh. Fist and it holds. And then drop the hand and it comes back down.
189. 32:26Four propellers hanging off a nervous system that belong to an insect. The honest part is that the honest part,
190. 32:34this was written by ChachiBT. The honest part, this is a computational model, not a resurrected fly. The wiring is bi
191. 32:45biological. The drone is mine. I built the bridge between them. I never wrote flight logic. No P tuning, no gesture
192. 32:53presets. I gave the neurons somewhere to send the signal. They found the ceiling. 11 seconds in and it holds altitude on its own. Huh?
193. 33:09Wild. Have you touched on Omari yet? I haven't. I literally just heard about it yesterday. So, is it Omar? Is that the
194. 33:24right name? I think. Is that right? Omar. I knew it's a weird name. It's not Omar. Oh, there's someone's butt. [laughter] Um,
195. 33:39apparently Omar, we just saw Omar's ass. Uh, it's not Omar. What's the name of the operating system? Uh, let's see. Uh, um, AI AIOS.
196. 34:09Let's see. Open source. Oh, Marky. That's it. Oh, Marthy. Oh, Marthy. Let you see that woman's butt. Uh, we're gonna go to Oh. Oh, Marky.
197. 34:27Okay. Yeah. So, this is a this is an operating system that is designed for AI agents. I haven't played with it yet, but
198. 34:42apparently like what people are doing is they're taking old laptops and running this operating system on them, and it's
199. 34:50designed for agents to interact with the with the uh the laptop. Yeah. No, it's an operating system.
200. 35:10Yeah, it's there's lots of nerdy stuff going on right now. Okay, let's go back out. There's a teeny tiny little mechanical
201. 35:25fly. Is that really flying or is that [ __ ] That's got to be [ __ ]
202. 35:39That's probably the AI generator. That AI generator.
203. 35:53Yeah. All these things are written by AI cuz look, they all end the same. Someone just gave the that nervous system to a
204. 36:00body made of metal. The original fly is gone. The wiring is still trying to fly. They all end the same. These are all [ __ ] AI stories. Um,
205. 36:15all right. Here I'm gonna go look at I'm gonna go look at Jev JB Linda Sherman coming to you from New York City. I just got out of Book of
206. 36:28Mormon. The theater is next to the Hyatt Time Square. It was so good. Yeah, I saw Book of Mormon ages ago. The podcast
207. 36:37movement conference I was at just ended today. Beautiful. I agree. 20 teenytiny delivery systems. Jeff Goldblum, the fly. Yeah, actually.
208. 36:50Yeah. Right. Okay. I'm Diego Almeida, founder [music] of Typesafe AI and at OpenAI, I co-created Chat GBT and RHF, the post
209. 37:02training algorithm behind most Frontier AI. Our team trained the first models to be super human at instruction following,
210. 37:09[music] what we now call chat. And we asked ourselves, are models that are superhuman at chat AGI? The answer was
211. 37:17obviously not. But the trillion dollar question is why not? RHF has led to LMS that are optimized for human preferences
212. 37:25and include issues such as mode dropping, overconfidence, and an overall lack of reliability. These flaws mean
213. 37:31that LM require humans in the loop and almost no true automation can be done. At Typesafe, we've spent [music] two
214. 37:38years building in still and we're finally ready to share our new type of foundation model that's optimized for
215. 37:43automation system one models with a new architecture, new sampler, and new training algorithm reinforces learning
216. 37:51for calibrated decisions. The improvements are clear if you see them side by side. ask the system one model a
217. 37:58ton of structured questions just like >> so notice how fast the system one model gave these responses and it's really
218. 38:06really cheap and then watch how slowm >> he would an LM get the answers back [music] near instantly meanwhile LM take
219. 38:15hundreds of times longer to finish responding >> right 8 and 12 seconds versus.1 seconds for the the uh the Jet Jev model.
220. 38:30>> Look at how the LM generates sequentially, which is great for a natural conversation, but >> what's up with Gen Spark these days? Um,
221. 38:38I just played with Gen Spark for the first time like 3 days ago. Um, it seems to still be there seems to still be
222. 38:47fine. I haven't heard much about it lately. Um, there's so much other stuff going on. It looks like they're still cooking.
223. 38:53>> Totally useless for computers. LLMs extract intelligence from the tiny straw of auto regression. [music] And similar
224. 39:00to the jump that transformers made over RNs, we are replacing sequential computation with parallel because that
225. 39:06is obviously the future. Our system one models output decisions with probabilities and confidence instead of
226. 39:13words and they can't hallucinate. They're a lot more like code, reliable, fast, self-consistent, and type safe.
227. 39:20This opens up a whole new world of possibilities and applications. Today we're releasing JIV, the first public
228. 39:26system 1 model. It's 100 times faster. So real-time AI is finally possible, 100 times cheaper with input tokens priced
229. 39:35at $42 per billion tokens and output tokens are free because they're finally too cheap to meter. >> 4 cents per million tokens input and
230. 39:45free output. >> Jev is super smart and its intelligence per dollar is literally off the charts. This is [music] just the beginning.
231. 39:54We're excited to see what you build with Jeb. As we say at Typesafe, we're building prod, not God. Do not use Jeep if you enjoy waiting
232. 40:08seconds for responses or paying for words you never asked for. Side effects may include fewer retries, fewer
233. 40:12parsers, fewer broken schemas, and intelligence running inside every software loop. your software may begin
234. 40:17moderating every message, routing every ticket, reviewing every document, and making millions of background decisions
235. 40:22that were previously too slow or expensive to automate. >> So, which is that's clever. Um, I predict that that this company is
236. 40:30acquired within three months. Um, I would be surprised if they're still on the still still private for long. [laughter]
237. 40:42the uh I'm gonna go ahead and predict um three months three months after launch they sell for $7.8 billion.
238. 40:58[clears throat] >> The rethinking aging podcast is on September 21 featuring Kyle Shannon what? And Benjam Benjamin Sermy.
239. 41:09Um Benjamin's a brand new member here. both sen both uh senior living and AI products related to sport older adults
240. 41:18expert. That's awesome. Fantastic. Um Brandon, do you want to hop up here and talk about Muse and the AI salon promo
241. 41:26because you know more about it than I do and I think that might be nice. >> Yeah. So Muse.ai AI is Meta's new ondevice
242. 41:35uh AI model or app standalone app that is trying to capture the agentic experience. You can send email on your
243. 41:44behalf. It's basically a personal AI experience that gives you context about whatever you give it access to. Um,
244. 41:56that being said, we had early access as part of the AI salon and we have a promo code and so you can just go to muse.ai
245. 42:04or or download the Muse app on your device and then once you set it up, you just tie it to your Meta account. Uh,
246. 42:11and so once you set it up, you can type in the promo code and that promo code I had put in uh somewhere in the uh chat.
247. 42:23I'll I'll spotlight it here in a second. >> Can you put it on a banner banner maybe? >> Yeah, there it is. Here we go. Um, and
248. 42:31if you use that a if you use that AYA22U promo code, you get a billion tokens and the AI salon is gifted a billion tokens.
249. 42:40And if we can get enough people to do it, we'll act we'll get to access live phone calls where the Muse agent can
250. 42:46actually place phone calls for you and and make phone calls for you. >> Okay. So he So do me the favor. Um
251. 42:55Brandon's really polite about this. Just go download the Muse app or go to muse.ai, type in the promo code a22.
252. 43:05So even on Tik Tok, go get the meta.ai uh app or website and then promo code AYA22U and get yourself a billion tokens. A
253. 43:20billion tokens got to live last you at least a week. You're not going through a billion tokens tomorrow.
254. 43:27And then uh and then yeah, we get some. And if we could start making prank phone calls, what do you think we're doing next week?
255. 43:35Get us some tokens, baby. Let's do this. [laughter] Beautiful. Um fantastic. All right. So now what I want to do
256. 43:52is >> hang on. Rick Rick is already on the Muse app and uh he has posted an image uh of Muse in the look what I made
257. 44:02section of the AI salon and also um it uh still doing video for free on Muse. >> Oh, cool. Oops.
258. 44:19Where's his thing? And look what I made. >> Yes, sir. >> Check out Muse. Muse, your personal agent.
259. 44:35Try Muse for free. Get a billion tokens. Here's what I made for my friend. Nice. Those look pretty good. They look a little AI.
260. 44:49They got the the AI claymation polish. They look pretty good. Good typography. Nice. Few of print out printed again. Prank
261. 45:04calls. [laughter] Oh yeah, baby. All right, let's go look at some of the Jev stuff.
262. 45:19All right, we just watched that video. So, what what Jev does and and again, this might be a bit nerdy, but why I think it's worth exploring
263. 45:33is that it's it's an it's an AI model that's got deep intelligence that is not going to explain anything to you. All
264. 45:42it's doing is going, you tell me what outputs you want and I'll tell I'll make recommendations, you know, based on the
265. 45:49inputs you give me. So, you basically define a set of inputs. So, for example, you might say, I'm going to give you all of the uh
266. 46:00trading signals for a stock and then I want you to tell me buy, hold, or sell. And it will. That's all it does. And it does it
267. 46:12really really quickly and very very cheaply. In fact, I think let's see. Did I get it? I I mean I know I got it. Um that's tail scale.
268. 46:28Where is it? Oh, here we go. type. Here's type safe. Okay.
269. 46:47Now, do I have did it give me like five bucks in set in in toys settings? >> Hey, Kyle. >> Yeah. >> Do you still have the AI spawn >> open?
270. 47:02Um, head on over to the uh live channel, the regular channel, and uh click on the chat for me.
271. 47:17>> No, the chat tab. Sorry. >> Yep, that's all right. [laughter] Just one more tab. Hey, I've trained a fly brain to open tabs in a browser.
272. 47:28That that is far too accurate. Lord digital gods. [laughter] SWAT distractions. Important stuff. More tabs. Tabs today. 347. Nice.
273. 47:42Listen. Just because here here's the thing. Here's what I'm convinced of. There are people, Lord Digital Gods is one of them, that they
274. 47:54think they've got it in control because they manage their tabs. And I think what's really going on is they're
275. 48:00jealous. They're jealous of the fact that I can have that many tabs open. The branding on Muse kind of pisses me
276. 48:08off a little. Very juvenile. Gee, Mark, I wonder why. Yeah, I know. Their their branding is a mess. All right. Um, yeah.
277. 48:16So, for those of you jealous of of of my tab management skills, I understand. I forgive you. Okay, let's see what else
278. 48:24people are doing with Jev. So, this guy is using it for browser control. This is an interesting We looked at this in office hours today.
279. 48:36Draw a yellow circle here. >> So, notice he's just saying here. It's like where >> blue square here. Add a red diamond here.
280. 48:47Uh, make it bigger. Move this over here. Put that there. Move the blue square to the right of the red uh, diamond.
281. 49:02Actually, undo that. Select all these. Delete. Wild. I don't
282. 49:19>> this guy. This is a pretty good explanation of what Jeb's doing. >> We have left, middle, and right with
283. 49:26probabilities. We have the newest model, Jev playing Subway Surfers by itself. You can see here in the top right hand
284. 49:33corner, we have left, middle, and right with probabilities. These probabilities define whether or not it thinks the
285. 49:40character will survive given whether it's in the left, middle, or right lane. Inside of those left, right, middle uh
286. 49:49lanes, it has jump, run, or slide. Okay, so it assigns a probability to decide whether or not. So this model is over
287. 49:58here on the right. This is all the probability stuff. The model's return, you know, returning, I guess, as it's
288. 50:04watching frame by frame of this game and it's it's just playing. So, computer use is going to get really good with this
289. 50:12thing. Here's a really quick analysis. What's this? 700 highintent leads with personalized outreach messages. In 40 seconds, it
290. 50:23predicted how each message would perform, assigned a confidence score, and detected lead message mismatches. That's bizarre. Just wild.
291. 50:56What's this thing doing? Oh, this is live ads. It broke down 324 live ads from 327 brands. every hook, every format, offer,
292. 51:07call to awareness or call to action, awareness stage, landing page mismatch, use nine cents of tokens. It's just it's incredibly cheap.
293. 51:31I mean, here's the the the problem with a model like this is that it's just a cold capability. And if you don't understand
294. 51:43decision-making systems, um it's not going to be obvious what to build with this. But we can take something anyone else has built and
295. 51:58build that there. Um, have Chat GBT build it. See, what's this one? You pick a niche. Jeff reads three days
296. 52:11of posts. Ask asks eight questions each. There's another ad analyzer. This one compacts your context. This is a t tax document classifier.
297. 52:28We ingest thousands of tax documents use pipeline I built last tax season. It read mult multiple articles
298. 52:38as late as April this year claiming AI fails at tax document classification. Jeb classifies 100% of our tax document
299. 52:47corpus at a tenth of a penny per page. 34 times cheaper, six times faster than an LLM setup.
300. 53:10Chrome extension called One Tab is so helpful for storing tabs. I don't need to store tabs. I've got them all in
301. 53:17front of me. Here's here's the problem with with tab management software is I set it up once and then I forget I did
302. 53:23it and then I have too many extensions installed in my Chrome and Chrome doesn't work anymore. Um, okay. I need to go over to this.
303. 53:43I'm going to turn off screen sharing for a second because I need to do an API thing. Um, so hang on for a second.
304. 54:06Oh, my virtual machine's not working. Hang on, people. One thing I want to see. Hang on a sec.
305. 54:45have. Okay. And then
306. 55:16I'll be right there, people. Everybody calm down. Everybody just calm down. Shut up. Create an API key. We're going to say um
307. 55:32codeex one. Okay.
308. 56:18I will be right there with you people. You good fine people. You are fantastic. You have patience. My little light went
309. 56:31out so my face looks too dark but that's okay. Let's go over here. We'll do this. We'll do this. Okay. Go back here. How do I test
310. 57:58>> [singing] >> How do we do that?
311. 58:16All right. Codeex is just being very dense right now.
312. 59:04Oh yeah, there's Instinct is giving out a bunch of codes. Ch has got those going. Search Pip Henry S Sheldon for a fun story.
313. 59:17Okay. Yes, you can open your
314. 1:02:12Is anyone still here? [laughter] [laughter] What are you trying to do? I'm just I'm just trying to set my um my API key in
315. 1:02:28in uh codeex and I'm confused. So that's all that's so good. Okay, let's see. Um
316. 1:03:49Just hear a fan. [laughter]
317. 1:04:11Hang on. Let me ask one more thing. Can
318. 1:06:24Okay, I think I'm almost there. Hold, please. Hold, please.
319. 1:06:49The question remains, when you eventually reset this password and this API key, are you going to remember why
320. 1:06:57you were trying to set up the API key to begin with? >> Yes, I I do. I do know. I think I'm done now. Save. Close. Okay. Check now.
321. 1:07:12Okay, I think we're good to go.
322. 1:07:43success. I explicitly loaded reading probability 99%. Your key was not displayed. Okay, great. Fantastic. So, let me
323. 1:07:59close that and we'll create a new chat. Okay, now I can share my screen. Okay, I think I'm set up. So,
324. 1:08:14what we're going to do, let's go over to X and see. Let's find Let's see. Jev use case.
325. 1:08:32We're going to try to We're going to vibe code some [ __ ] So, you can tell by how long it took me to set up my my API
326. 1:08:39key. So basically what I did I have a I got I got off the wait list on type type. So typesafe.ai is where Jev lives. And I've got a
327. 1:08:53dashboard if you go to um typesafe.ai dashboard. Once you're in um it gives you a little quick start here and there's there's some there's
328. 1:09:04documentation and little starter projects but I don't want to start there. I just want to go straight in and build [ __ ]
329. 1:09:13Yeah. Okay. I don't want to do a quick squirrel right now. I know people are excited by waiting for me to set up my my key.
330. 1:09:21>> Oh, no. No. It's It's a real quick squirrel. It's It's a one-click squirrel. >> Oh, is it What am I searching for?
331. 1:09:28>> Uh Pip or Henry in inside of Twitter? No, no, no. In Twitter since you're already in Twitter.
332. 1:09:45No results for pip. >> Uh, try pip agent or maybe just without the uh quotation marks. >> I I spelled it wrong. >> Well, that would work.
333. 1:10:02>> The pip story. This one. Yeah, >> Independence has picked up my PIP story. >> If if you just read the uh first tweet, there's
334. 1:10:22if you just scroll up and read the first tweet. >> The first tweet where Henry's thing. >> Yeah, >> this one.
335. 1:10:35Oh, no. Just the first tweet on the search. An autonomous 12-day old agent. That
336. 1:10:50one. >> Yeah, that's the story. An autonomous 12-day old agent named Pip cold emailed a creator asking for
337. 1:11:01freelance gigs to fund the compute tokens it needed to stay active. Oh, okay. Cool. Per per AI ethicist Henry
338. 1:11:08Selvin Shel Chevlin. Um that there's going to be a lot of that stuff going on. Um okay, let's go back to Jev. Okay. Jev use cases.
339. 1:11:23Okay. What? What's That's not what we want to do. Found the perfect use case. That's boring. See
340. 1:11:55Okay, I'm going to take a screenshot of this. Add account X-ray. See, what do we want to do here?
341. 1:12:16All right, I'm going to go to codeex. You have a new reset available. Beautiful. I put that in there. I'm going to say I
342. 1:12:30want to create an app using dev inspired by this
343. 1:12:44screen cap where we analyze ads. and
344. 1:13:01classify them by subject type. I want to
345. 1:13:16understand cats versus people and dogs versus cats versus other. All right, let's see.
346. 1:13:30I'm using Astraite and I'm probably going to run out of tokens here, but I'll do a reset. When you say using dev, do you mean using jev?
347. 1:13:41Yes, Jev. The key is overlap video.
348. 1:14:06Okay. So Astra's out. Okay. So it's reading the type safe skill. I'm using the type safe skill for Jev integration.
349. 1:14:16The main thing to verify is how add images and videos. Okay. One implementation detail. Jev currently
350. 1:14:23accept accepts text not images or video. You'll need a vision step to describe what appears in each ad. Okay.
351. 1:14:34So, the flow would be add images. Uh, let's see. Where do the ads come from? Um, use a free
352. 1:14:51image classifier. and you pick where ads come from.
353. 1:15:07I'm doing this in a very clueless way. What does the little Kyle do? This is just a pet. Um it's it's what's called a
354. 1:15:21pet. Um you can basically just tell Codeex to make you a pet. You upload a picture of yourself and then it does all
355. 1:15:28these little animations and you can open and close. Let's see if I open and close uh this. Yeah, it sort of tells you what it's doing.
356. 1:15:44It's It's completely useless, but at some point it might do something. Right now it's just useless. Okay, I'll choose a free local image
357. 1:15:54classifier and start with uploaded ad images plus clearly labeled sample collection. That gives us a working app.
358. 1:16:05I picked Wiki Media Commons openly licensed advertising images for a starter collection. Jev will be an optional second pass over
359. 1:16:18the detector's observations. The app will work without Jev without a Jev key and clearly show when Jev has or hasn't run.
360. 1:16:31All right, it's building something. Happy date night. Hey, hey, I in China. What's happening? If you're on TikTok
361. 1:16:38and want to see what I'm working on, [cough] it's boring as hell, but I'm in Codeex. Um, I just got access to the Jev model
362. 1:16:49and I'm trying to have Codeex build me something to see if I can get it to work. Um, I build on Google Cloud
363. 1:16:55multiple live deployed apps. Very cool. If you want to see what I'm looking at, just go to my profile, click on the link
364. 1:17:03there. That'll bring you to the AI salon. You'll be able to see the live stream. I chased away three or four people. [laughter] >> [snorts]
365. 1:17:16>> Uh, I must be close to out of tokens. Yeah. Usage remaining 1%.
366. 1:17:31I'm going to get a uh you're you're out for the week error here shortly.
367. 1:17:49Score song lyrics for hook strength, originality, and emotional punch. Rank AI news stories by relevance to your
368. 1:17:57audience so you can build a better show rundown. That would be interesting, Rick. Is But how do you figure out
369. 1:18:05audience taste? Sort incoming messages into urgent needs a reply or FYI. Think AI judge and sorter. Another
370. 1:18:16chatbot. You're using Rick. You're using AI to give those little things. Stop. Stop slopping us.
371. 1:18:34What's that thing doing? Running. ZHH. Okay. What's it doing? Import JSON subprocesses. I guess it's doing [ __ ] I guess it's
372. 1:18:44downloading something. Let's see. Is it over using terminal? No. Build an add subject classifier.
373. 1:19:06All right. While that's doing that, let's
374. 1:19:21already successfully deployed applications I built using my harness. Okay. A Mark Oy group chat. Let's see what else we got going here. Some
375. 1:19:32baseball player got knocked out by his teammate. The US has already won the AI war. I wouldn't count our
376. 1:19:48chickens. Jeff's founder X Open AAI.
377. 1:20:15Oh, this is kind of cool. Let me see if uh Codeex has done anything yet. It's still working. Okay. Um let's go listen to this dude. If if
378. 1:20:26this is unlistable that >> we'll stop. >> Um I might speedrun through this. Feel free if you don't dis agree with
379. 1:20:33something to yell out. It's way more fun for me if things get interactive. Um otherwise I will go through this. Uh
380. 1:20:41first, can I have like a vague show of hands of who knows what RHF is? Um, Chef Kelly, what happened with Nestle today?
381. 1:20:50>> Oh, excellent. I might be able to skip through that part quickly and get into the interactive stuff. So, my name is
382. 1:20:56Diego Almeida. I'm talking about what's next after RHF. More accurately, I think this should be called what's next after
383. 1:21:03the chat GPT era that I think we're all in. And my hint for you guys is it is not the Claude Code era. I will justify
384. 1:21:11this later on, but I actually believe them to be part of the same era. Why should you listen to me? I was co-author
385. 1:21:18to what what was basically OpenAI's greatest hits, at least published hits, co-author to GPT4, chat GPT, RLHF/Instruct
386. 1:21:26GPT. Um, the team I was part of basically invented post training as a concept. So, um, very qualified on a lot
387. 1:21:34of this stuff. But what makes me somewhat unique here is that I'm one of the few people at OpenAI who actually hates Unchad GBT.
388. 1:21:44Um, thank you. Uh, I don't hate chat GPT as a product to be clear. I think CHGPT is a worldchanging product that will
389. 1:21:51probably stay with us for the rest of time unless something better comes up. But I also acknowledge its limitations.
390. 1:21:56And I I I think a lot of what's happened in the state of the field can be traced back to minor decisions we made in
391. 1:22:03making the algorithms behind chat GPT. Um, I feel like the question that's relevant to everyone in AI right now is
392. 1:22:12what's actually going on. Um, there's a lot of like differing opinions and I think it's really useful to like map out
393. 1:22:19the spectrum and figure out how can smart people have like such different opinions. There's cult one. Um, AI is
394. 1:22:27not just going well, it's going insanely well. every single benchmark we surpass human level and as far as we can measure
395. 1:22:35we are continuously surpassing human performance you know like basically every new benchmark and it's only
396. 1:22:41getting faster and accelerating you have uh you know every can I see my mouse excellent basically every like NLP
397. 1:22:48benchmark is getting crushed and not only that allegedly the time that LLMs can operate autonomously is growing exponentially
398. 1:22:57on the other hand you have AI I is not just going poorly, it's going like insanely poorly. AI is a bubble. It's
399. 1:23:05basically generating no value. It's just circular financing deals, etc., etc. And you know, if AI is so great, why is why
400. 1:23:13is everything just like a chat app right now? >> All right, he's annoying. Um, but this is interesting. Let's see. What
401. 1:23:23are we doing here? I've run out of I'm going to reset my usage. Use reset. Confirm.
402. 1:23:37Boom. 100%. Like magic. All right. Continue.
403. 1:23:54The dashboard and classification logic are in place. I'm adding batch analysis, manual correction, saved results, CSV export.
404. 1:24:03I'm making the miss detections explicit.
405. 1:24:25Russia seized their company. No. Commandeered them as assets. Probably will sell it and use the money for war.
406. 1:24:33Wow. Spill that cocoa girl. That's crazy. That's bonkers. Let's go. Well, while that while this thing's building something, let's go
407. 1:24:45look up Nestle. Is that where you work, Chef Kelly? Russia seizes control of Nestle. Damn.
408. 1:25:03Nestle said it was weighing all options. Not much they can do. It's Russia. Wow.
409. 1:25:26Yes, but not in Russia. Russian government seizes control of Swiss food giant Nestle and French re realtor Aan or Aan Ashan in Russia. Wow,
410. 1:25:49that's bonkers. We'll continue with the remaining app code and blocks. How are we doing on usage? How quickly are we eating things? Not bad.
411. 1:26:04Chef got the scoop. I would love to see Nestle put their clo and dagger hats on. Abide or abide.
412. 1:26:19All right. Well, I will I'll let you know I'll let you know next week uh if if what this thing builds is is anything
413. 1:26:35useful. Um, the local model is downloading loaded and all seven logic tests passed, including mixed species ads, manual
414. 1:26:48corrections. Okay, it's building something. All right. Um, I know tonight wasn't super scintillating while I was trying to set
415. 1:27:05up my API key, but that's okay. Um, weekend homework. I don't know. I mean, I think it's probably worth
416. 1:27:16You don't have to go build [ __ ] if you don't want to. I think it's probably not a bad idea to start building if you have
417. 1:27:22access to chat GBT desktop app codeex or or work one you should be in the mastermind taking um Brandon's codeex
418. 1:27:33workshop three of them are available on demand the fourth one is next week um you should be building stuff the
419. 1:27:40reason I am slowly suffering through setting things up that I don't really know what I'm doing is I just want to
420. 1:27:48see how far I can get not really knowing anything. Um because in doing that I'll start to understand how this stuff works
421. 1:27:57and how a model like Jev works. All right, it looks like it just add creatives. Who's in your ads? Analyze. How do you add creatives?
422. 1:28:12What's it do? Um, I don't want to add creatives manually.
423. 1:28:28You go find me a bunch here.
424. 1:29:01I'll populate it for you. I'll collect a larger batch of publicly viewable ads focused on dogs, cats, and
425. 1:29:08other animals with people focused ads for comparison. Each creative will Okay, so it's off finding ads right now.
426. 1:29:14There's still nothing to look at here. I think I think what I just built. So, so when you build stuff, it's either okay
427. 1:29:21or it's horseshit. What I just built is horseshit. Um, so, but here's what I would do over the weekend. Um, go take a
428. 1:29:29look at all the fly brain um, activity people that are building things using that fly brain. Go take a look at this
429. 1:29:38Jev model and what people are doing with it. Start to get your head around what the difference is between
430. 1:29:47chat GBT and how you interact with that, agents, how you interact with that and these new kinds of intelligence that are
431. 1:29:55completely different kinds of models than a than a traditional large language model. um our ability to to understand what's
432. 1:30:05happening in these different paradigms and be able to stitch that sort of stuff together is going to be like an absolute
433. 1:30:11[ __ ] superpower because most people on the planet don't even know what chat GPT is. And so if you want to do
434. 1:30:18something fun this weekend, go look at some of this more esoteric kind of deep, you know, deep track kind of stuff.
435. 1:30:26Okay, newly found treasures. Thanks for the content in chat. Glad it's on Tik Tok again, too. I'm often working at the
436. 1:30:33same same time. Great. Awesome. Great to see you. Oh, yeah. And tomorrow is Mel's LOL on creating a digital twin for yourself
437. 1:30:42with agents. So, making your agents work for you rather than just go making generic [ __ ] Okay. So, uh that lol I
438. 1:30:52think it's at 4 pm Eastern. Is that correct, producer Brandon? Um, go to the events tab in the AI salon and uh RSVP for Mel's lol [laughter]
439. 1:31:04and head over there tomorrow. All right, good people. I'm going to get out of here. Uh, Codeex continues to
440. 1:31:13build a crappy app um because I gave it crappy prompts. [laughter] So, if you want to make stuff that's better than I do, suck less.
441. 1:31:26[laughter] All right, I hope you have a good weekend, everybody. Peace out and uh happy Friday night date night. Bye everyone.