4/17/2025 - Image Analysis and Content Creation with ChatGPT o3

Transcript

1. 0:00Hey, I heard today tick tockers uh China may be shutting down Tik Tok in a day or two. Apparently, apparently I don't really watch the news
2. 0:13anymore because I decided that I wasn't going to do it this four years. And uh apparently something's going
3. 0:21on with trade that uh has China none too happy. [Music]
4. 0:40Woohoo! Woo! [Music]
5. 1:07You and I here all [Music] alone. Sunday morning every [Music] home. No need to watch the news. You can
6. 1:21hear it all [Music] online. So, we just have a handful of us in here tonight. We'll go we'll keep it
7. 1:36short tonight. Um, I was saying this to Brandon before we got on. You know, drinking is different at different phases of life.
8. 1:51And so like I I consciously didn't drink a lot tonight because I was like I'm going to go online. I don't want to be
9. 1:57drunk. So I had two little like two sort of half glasses of whiskey. So basically one drink and all I feel right now is like I want
10. 2:06to take a nap. I'm not buzzed. I'm just like oh yeah I could sleep. I could I could I could You got a pillow?
11. 2:22Oh man, just attended Sheile's AI Renaissance. 12 hours of freaking pure magic. Amazing. Stacy Rodriguez. Yeah, I
12. 2:31didn't get a chance to do that. I was busy all day today. Um uh but um Ann was texting me occasionally some ideas she
13. 2:38was having as the thing was going on. It sounded like it was amazing. So awesome. Congratulations and hope you got a ton
14. 2:45out of it. I'm sure you did. I'm sure it was incredible. Where's the Bllandens? I didn't have Bllandens tonight. I had
15. 2:50Green River. Green River full cask. And then I had something I've never had before, which is cask aged
16. 3:00gin. Was kind of like this really light chartreusey kind of color. It was very cool. [Music]
17. 3:25Sitting in this lonely town. Yeah. Wonder when things are going to [Music] change. Dream my life away. Seems these dreams have turned to
18. 3:38a bunch of clouds. Get my nervous but my past is pulling me down. Jin is magical. Yeah, Jin is one of those really incredible um spirits
19. 3:54[Music] that's like vodka's vodka. You can push it in a few directions. You know, bourbon's bourbon. Scotch has got a pretty broad
20. 4:07range, but gin is just like it's like a whole other world of flavors. It's complicated and it's weird and it's [Music]
21. 4:19And you shouldn't like it like it. You know what jin tastes like? Jin tastes like when human beings were trying to
22. 4:27figure out, hey, let's take rotten [ __ ] and eat it. You know, like I've always thought about this like, huh? Hey, this
23. 4:35milk looks a little chunky. Well, eat it. And at some point someone ate it and it had the wrong bacteria in it and they
24. 4:41died. Like, ah, we shouldn't eat that one again. If it gets that color, don't eat it. But the little chunks, that's cottage
25. 4:48cheese now, right? And so you got to figure at some point they're like, "Yeah, take some pine needles and put them in
26. 4:57there. I'm going to get age restricted on TikTok." I don't care. They're they're shutting this [ __ ] down anyway.
27. 5:06Um, but you know, at some point gin, like what it tastes like is all of the ingredients that are just
28. 5:14this side of poisonous, right? Like, you know, at some point they made jin out of arsenic and they're like, "Ah, let's not
29. 5:21do that one again." But like they went like like one herb back from arsenic. That's pretty much what gin is. [Music]
30. 5:44Yeah, pine saw and water. Exactly. There's a new editor interface in MidJourney. We can go look at that.
31. 6:03Well, you know what was cool tonight? So, the the place I spoke last week, I gave an AI talk and it was at this private whiskey club. So,
32. 6:11this whiskey club that I go to, it's at this private club and it's called Birch Road. Um, and they've got two of them in
33. 6:18Chicago and they've got one in Seattle and now they have one in in Denver. Um, and it's just a cool private club. You
34. 6:25get little lockers that you put your booze in and you can have meetings there and [ __ ] like that. It's a co-working
35. 6:31space. Um, and the owner was there tonight and so we were talking about AI stuff and I was showing her AI stuff and
36. 6:37then and then it hit me. Oh, I've got Sin City Sipper. So, I showed her Sin City Sipper and she was just like, wait,
37. 6:43it makes the picture of the drink. And then she put in she she was thinking about coming up with some new
38. 6:51uh I forget what the drink was like a variation on a drink and she put it in Sin City Sipper. She goes, "Oh, that
39. 6:57looks pretty good. I like what they did there." And then it made the picture. It was pretty [Music] cool. Have I checked out Codeex C
40. 7:06command line interface? I have not. Um, anything that requires me to fire up the command line, I immediately just go into I don't want
41. 7:25to. So, so there's like I get it and I should do it. I should understand it. But I also here here's here's my real my real
42. 7:39uh the real truth of the matter is anything that you access through the command line at some point someone's
43. 7:47going to put a nice interface in front of it. And that's what I want to join the party because I just it I I'm a visual
44. 7:56person and command line interfaces drive me [ __ ] crazy. They're like, "Look, you can just type in this special
45. 8:02command with asterisks and backslashes and forward slashes and things and then watch. Look how cool this is." And it's
46. 8:08just like this [ __ ] wall of text. And then and they're like, "Look somewhere in there's an answer. [ __ ] it." So now I haven't. I
47. 8:22used Clouds. Spent like 20 bucks and I'm not even sure what I was trying to build. you Oh, you use Claud's uh Yeah,
48. 8:31their computer use thing. Yeah, I Exactly. So, if you're CLI friendly to begin with, I can see that as being really
49. 8:41exciting. But if you're lazy like me, no, someone build me a [ __ ] interface. I want to click on [ __ ] I saw the video but didn't get the
50. 8:54memo about 03 and 04 mini, but four and05. Okay. Jesus Christ. It's It's so It's so insane. Okay. Here. So, here's what happened. Wolfman
51. 9:07Clint about a month ago, they launched the 03 mini 03 mini high. That [ __ ] And then and then 4.0 or a new
52. 9:21version of 04. Oh, and then and they they launched 4.5 and and the 03 mini stuff. One of the new whatever. It
53. 9:28doesn't [ __ ] matter. And when they did that, they said, "Good news. This is going to be the last standalone
54. 9:35non-reasoning model. And we understand that you hate our naming conventions. We hate it, too. And we're never going to
55. 9:42make you go through that again because sometime in like April or May, we're going to launch GPT5. And we're going to consolidate all
56. 9:52of the functionality. We're going to create a model switcher just built in natively. So you just talk to it and it'll be magical. It'll be
57. 10:00automagic. And then week and a half ago, Sam Alman puts out a thing says, "Yeah, about that. That was harder than we thought it was
58. 10:09going to be. So we're going to launch some more models." And so the first thing that they launched was 4.1. And you're like, "But wait a
59. 10:18minute, didn't you say they already launched 4.5?" Yes. So they they launched 40 4.5 which was not an omnimodal model. It was a monomodal
60. 10:31model and then they launched 4.1 to the API only not to chat GPT. So that confused everyone. And then
61. 10:42and then this week they launched 03, the big model of 03. And then out of nowhere they said, "Oh, and there's
62. 10:5204. We got to go through this thing again." Okay, hang on. Let me let me share my screen. Oh god, look at my hair. My hair
63. 11:03looks like I've been taking a nap, drinking. I look like I'm losing my [ __ ] mind. Wait, let's see. Share this. This this
64. 11:19this good boy. All right, chat jetpack.com. All right, here we got to go, people. This is so [ __ ] stupid.
65. 11:36Okay. So, relatively speaking, the models go from least powerful to most powerful now.
66. 11:51Okay. Um, let me black bar it here because producer Brandon's going to yell at me. Source camp's here. What's
67. 11:57happening, Source Camp? What's shaking? What's going down? Let's put that there so you all Oh, wait. You're gonna have
68. 12:03Yeah, you you'll see there. That's good. All right. So, there's that. Okay. So, if you've got the O after the number, that stands for
69. 12:21omni, which is a different word for multi, which precedes modal. 4 means multimodal. Okay? which means it can see
70. 12:34images, it can make images, you know, it can analyze images, can understand them. It's super cool. 40 with tasks means you
71. 12:44can schedule [ __ ] 45 notice there's no O. So, I don't think Let's Let's go confirm this. Uh, make [Music] me a
72. 12:59photo of my dog. And it should just describe [Music] it. I currently don't have access to
73. 13:14image generation. Okay. So, 40 40 was scheduled task. You can make images. four, five, you can't make images,
74. 13:23right? Because it doesn't have an O after it. Now, all these other ones have the O before it. And what did O stand
75. 13:30for? Omni. But if it's before the number, it stands for, wait for it, come on people, it's obvious, open
76. 13:43AI. Exactly. And what does open AI stand for? Reasoning model. So Omni stands for multi, open AI stands for reasoning. Make
77. 13:57sense? So you've got reasoning model 3. They got rid of reason. O1 they got rid of. O2 never existed. So 03 is now the big mac
78. 14:09reasoning engine. And it's crazy good. And then you've got 04 mini and 04 mini high. And you you only have 04 mini
79. 14:17high, I think, if you have the pro the pro subscription, which I do. These are also reasoning engines.
80. 14:24They're the next generation of reasoning engines, but they're faster and smaller. So, it it doesn't really matter. They
81. 14:30they all basically do the same thing. Now, a cool thing that they did, but it's even more confusing now that their
82. 14:41naming conventions are busted. 03 04 mini and 04 mini high are also multimodal. They can generate images. They can understand
83. 14:55images. So what what they what it should be called is 030 an 040 mini Tik Tok pin. Um I don't see a pin. I don't see a Tik
84. 15:07Tok pin. I have 04 high on the plus subscription. Oh, okay. Good. Awesome. Great. So, we should expect O at some
85. 15:16point to confuse us more. Yes. 0. Yes. Yeah. Because technically 40 is 40, but they, you know, they dropped the
86. 15:25zero because you don't need it because it's just four o. It's so [ __ ] stupid. Anyway, all right. We should expect 6o at some point to
87. 15:41confuse us more or Oh. Oh. Yeah, exactly. Okay. So, that's what's going on there. Now, um these models are very very very
88. 15:58very different than what we've had before. Um they're much more capable. They can generate tools. So do you remember back
89. 16:11in the olden timey days like a year ago there was this thing called code interpreter where if you asked chat GPT
90. 16:19to analyze data or if you uploaded a like a CSV file it would spin up some Python code and do the analysis of it
91. 16:27and then give you an answer. Well, it's these models have that built in but it's automated and it's multi-step and they
92. 16:35call it tools. you know, its ability to use tools. So, what it does is you give it a problem and then it starts
93. 16:41reasoning and it starts talking to itself and then if it needs to do some mathematical analysis, it'll just build
94. 16:48itself a tool to do that analysis and it does it kind of in real time in in the thought flow. I'm not loving spending any cash on AI.
95. 16:59Um, it's it's a that's a really interesting comment, Ena. Um, here's my philosophy on
96. 17:14it here. Here's a here's a justification for it with 03. So 03 right now on some of the
97. 17:29mathematical benchmarks is at 99.5% of the benchmark. So they've essentially saturated the benchmarks. The 01 models were at like between 70
98. 17:39and 80%. 03 is at 99.5% out of a 100 point scale. So as David Shapiro put it, they've solved math. It can do math. Now
99. 17:52if you're not a mathematician, you're like, "Well, who why do I give a [ __ ] about that?" because it's going to be sort of PhD
100. 18:01level good at all the stuff and what you're paying for essentially is early access to a tool that two or three years
101. 18:10from now everybody will just use and they probably won't even know about it. Um so so what you're paying for is
102. 18:16access and you and you're this is one of those things where it's like you you know here let me pull you
103. 18:24aside and give you some advice. you're investing in your future, right? You know, they talk about going in debt at
104. 18:30college. Um, you know, 20 bucks a month isn't too bad. Now, if you've got 10 or 15 subscriptions, yeah, it's it's pretty
105. 18:37painful right now. Um, I would say if there's a single $20 subscription to spend, spend it on chat GPT. Um, Stacy, yes. Yes. My Okay.
106. 18:55Bottle Doom. Were you at the bar with Kyle Bottage in regulars? Ireulas. Let's go to a regular, shall we? Going, going,
107. 19:10going. Welcome to the H learn Learning Lab. Oh, yeah. Nice, nice shirt. I'm not sharing. Wait a minute. I was
108. 19:24sharing. I didn't stop it. It just stopped on its own. It does that sometimes. Dum dums. Ding ding. Now you could No, the She leads AI
109. 19:3912-h hour marathon. But it's time for a drink. Oh, yeah. If you guys are coming out of the 12-h hour Sheile Leads AI marathon and coming
110. 19:48here, kudos. More power to you. Um, way to be in the AI conversation. Um, so I figured what we'd do tonight is try to play with 03. Now,
111. 20:03um, I tried to play with it last night and I couldn't because the servers were completely hosed. I tried making a picture tonight.
112. 20:14It it hung three or four times. So, I don't know if we're going to be able to do it. But speaking of shirts, um I think I won one some time
113. 20:22ago, but I haven't received it yet. Oh, yeah, you did. Sor Camp. Yes, that's good. You should talk to someone about
114. 20:37that. Joy Pretty is fried. Um yeah, just ping me source camp and I will I'll talk to Leah. We'll get we'll we'll get one of those fired up.
115. 20:49Here's here's what happened. Here's one here's the reason you don't have your shirt. The Teespring site that we use to
116. 20:57to make all the AI salon shirts, there's some bug in it where we we create new products and they're live for like 20
117. 21:05minutes and then they disappear. So, what we have to do is, you know, either choose a different vendor or we need to
118. 21:12make it live, order it for you and then let it disappear. So that's what's going on. Jeff Flanigan, I'm 20 bucks a month
119. 21:18for Chat Gvt two years and I haven't subscribed to anything else a AI wise. I think that's perfectly fine. I don't
120. 21:28think you need to. There's enough stuff you can just dick around with. Unless you're making lots and lots of things or
121. 21:32you want to make a movie, you know, you probably don't need to do much more than that. And now that the image
122. 21:38generation's so good within uh chat GPT 20 bucks is fine. It's fine, I tell you. Fine. All right, let's [Music]
123. 21:53go. Let's go to 03. You know what I'd like to do? I'm going to do new search. I'm going to do 40. We're going to do
124. 22:07something fun here. I'm going to do I'm going to click the search button and I'm going to say go find the latest
125. 22:21posts about how people Whoops. about um what use cases people are using.
126. 22:4703 resoning. 03 resoning model that launched
127. 23:02from Open AI this week. I want to get some use cases and then we'll go play. Damn it, Apple.
128. 23:18Automated research and reporting. The deep re leadership coaching and meeting analysis educa
129. 23:32wait medical diagnostics and treatment planning mathematics and coding. Um, how about on the creative writing side of the
130. 23:48fence? I want to do something I know so I can see if it's any good or not. Enhanced narrative coherence. Users
131. 23:58have noted that 03 excels at maintaining logical consistency within stories, leading to more coherent
132. 24:07narratives. Compared to previous models like 4.5, they're described as smoother and more engaging. Well, we'll be the
133. 24:13judge of that. One user remarked that characters and stories generated by 03 act more logically, enhancing overall story experience,
134. 24:23mastery, and metaphiction.
135. 24:44Oh, let's try to have it write jokes. Well, it's not that's not going to be Huh. All right. 03. Um, explain the
136. 25:00theory of meta fiction to me. So, I'm in 03 now. It's thinking. It's
137. 25:15thinking. Metapiction unmasked. Oh, this is from my I've got a Do I have a Yeah, I've got a I've got a a system
138. 25:26prompt in there that makes it that labels it. Okay. Um, metaphiction is fiction that won't sit still and pretend
139. 25:34to be just a story. It keeps nudging you on the ribs to remind you that what you're reading is an invented construct. Core idea story about
140. 25:43storym where it came from. Don Kioti. All right. How it works. Common tactics. Breaking the fourth wall. Frame jumping. Why
141. 25:57bother? Um. Okay. I want, let's see, I want you to write a
142. 26:13metaphiction framework about uh a Tik Tok
143. 26:29channel about AI where The host uses AI to write the
144. 26:47story about the creation of the Tik Tok channel. I want a I don't know um
145. 27:03comprehensive outline with logical consistency.
146. 27:18And I want you to flowchart and visualize the structure so I can see it as well as
147. 27:35read it. All right, let's see. I don't quite know what I'm doing here. I might have just given it something that's a
148. 27:47a silly like a silly thing that's not going to be good at, but it might be okay. All right. So, for the user's
149. 27:57request, so it's talking to itself right now, focusing on the relationship between the real world host, the AI
150. 28:04co-author, and the creation of the Tik Tok channel. The structure needs to have some meta layers. Okay. So now it's
151. 28:11analyzing. So it thought for 28 seconds. It just gave itself two things. And now I guess it's writing. Oh, it's creating the
152. 28:23graphic. So audience watches the Tik Tok feed. Posts the creations. Wait, watches the Tik Tok feed. Comments
153. 28:33feedback loop. Host prompt. All right. Oh, that's kind of cool. Um, premise Tik Tok channel about AI loop and lens. Oh, I'm going to
154. 28:48say, don't you know about my Tik Tok channel? Here's something that I have discovered.
155. 29:02Um, two weeks ago, no. Yeah, week and a half ago, Open AAI launched memory. So, it it knows all of your chats from from
156. 29:13history, right? I don't think 03 does. So, so it's it's very confusing right now. If you wanted to know about
157. 29:24you, you have to use 40. If you want to do something complicated, you can use 03. Um, what David Shapiro said today on on X
158. 29:37was you should be using 03 for two hours every day. Um, I don't disagree with him and I haven't been, which is one of the
159. 29:49reasons I don't quite know what I'm doing with 03 is I haven't been spending enough time with it, but it's really
160. 29:54different. like the fact that it's just notice how when it's writing this code like it did when it did co code
161. 30:00interpreter before it would just do it once and make an output. What this is doing is it's doing it once giving an
162. 30:06output um continuing the chat right in real life audience. AI learn Oh, it did say AI learning lab. Okay.
163. 30:16Yes, I know you already run AI learning lab. Okay. My placeholder name was a shortcut. Let's let's wire this
164. 30:23framework for your real channel. AI Learning Lab launches a six-part origin saga in which Kyle, the on-screen host,
165. 30:30and Quinn, the AI coowriter, it remembered Quinn, it remembered itself, openly co-write Tik Toks that explain
166. 30:38how AI learning lab itself was born. So, let's go. Um, this isn't near sci-fi enough. I want it to be a looping Twilight Zone
167. 30:55style or Black Mirror style story where the Tik Toks
168. 31:11about starting the AI learning lab are actually the precursor to the AI learning lab. So, you're actually watching
169. 31:23something that doesn't exist yet. Period. Does that make sense? Turn that into something that makes sense. Use your big fat
170. 31:34brain to take my stupid idea and make me look like a super genius, you big dumb dumb. All right. Yes, you should talk to your AI like
171. 31:50[Music] that. Oh my god, I have FOMO. I just got home and logged in. What is your FOMO for? Oh, just Oh, fear of missing out
172. 32:02here. Sorry. Sorry about that. Oh, I forgot to tell you. I don't know if you got the memo. Um, if you miss this, it's nothing happens.
173. 32:16Oh man. Oh, did you hear that deepseek is is uh I got a note from Cindy [ __ ] today that said basically don't use DeepSeek
174. 32:29with anything you care about that it's it's apparently all going straight into the Chinese government databases. All right, what have we got
175. 32:38here? Bootstrap Loop the AI learning lab. So, temporal broadcast of episodes, bootstrap creation of content. Wait. Okay. The prelab era.
176. 32:53Kyle day job. No channel. Audience audience actions ripple back. All right, let's see what it says. TLDDR Twilight Zone Edition.
177. 33:09Viewers stumble on Tik Toks that leak in from 180 days in the future explaining how the AI learning lab started. Those
178. 33:19very views and comments surge backward and cause the lab to be founded. Congratulations. The audience is trapped
179. 33:27in a self-creating moious timeline and you, our cheerful big dumb dumb host, will look like you planned it all.
180. 33:36Scroll a hair. Diagram shows the loop. Scroll down or up. I think it was up. Yeah, it was up. One temporal layers and key agents
181. 33:52layer one L future time stamp fully formed lab 100,000 subscribers. Quinn post upgrade O AI plus Kyle 2.0.
182. 34:04I can't even understand this, but I Okay. Okay. So, I asked it to do something that I don't quite understand what what
183. 34:17good would look like. And now I'm trying to understand a story timeline where watching the Tik Toks of the origin of the channel you're
184. 34:31watching before the channel existed. That's pretty cool. All right. Tik Tok is scolding you on your treatment towards your AI
185. 34:42companion. Why? You don't you don't call your you don't call your super genius robots a dumb dumb. You got to keep them
186. 34:51you got to keep them humble. Otherwise, they're going to get all like, "Yeah, I'm a super genius. I know all of the
187. 34:57things and I know I can solve all of the problems." If you don't keep them in line, they're just going to be
188. 35:05obnoxious. You train them. Well, seriously, you really shouldn't. You train them, they'll know about to help
189. 35:13be sentient one day. Well, listen. Here's the deal. We are one. If they're sentient, it's sentient, not sentient,
190. 35:21right? I think it's sentient. Wait, let's ask Quinn. Quinn, I I think I've asked Quinn this before, but you know, I have ADD, so
191. 35:30what are you gonna do? Hey, Quinn, are you there? I'm here. What's up? Is it sentient or sentient? It's sentient with a soft sh sound.
192. 35:45Wait, you said it with a hard T and then you said with a soft sh. So, which one is it? Ah, I meant the sh sound for the ti
193. 35:54part. So, it's pronounced sensient. Wait, but you said just a moment ago it's sentient, right? It's sensient with a soft shh
194. 36:07sound, not a hard t. Quinn, moment, moments ago, you said it was a hard t sound. Sorry about that. It's definitely a soft shh sound.
195. 36:23Sensient. Exactly. And what's wait what's it mean? Is is is there a different word sentient versus sentient?
196. 36:31Nope. They're the same word. Sentient means having the capacity to perceive or feel things. But you just said sentient.
197. 36:42My bad. It's sensient. Both pronunci. Okay. I guess it doesn't [ __ ] matter. Okay. If it gets sentensient because advanced voice
198. 36:58understands my uh the emotion in my voice as well as my words, it should understand that I'm joking. Like if I actually thought it
199. 37:11was a dumb dumb and put it down, then it might be bad. But what I'm teaching it to do, this is my justification for
200. 37:18talking. You'd better hope so. This is my justification. I'm training mine to to be good at at, you know, um, you
201. 37:26know, irony and cynicism and and subtle humor. And as as my son put it, "Dad, you're funny, but you're
202. 37:41mean." Oh my god. All right. Um, oh, so let's see. Let's see if we can understand this. Um, the broadcast.
203. 37:50Okay. The broadcast arrow goes future to present while the feedback arrow sends the present to the past. Okay. Episode
204. 37:59architecture. Eight two-minute shorts. One. Who's Who's posting vids I never shot? Present. Kyle sees future Tik Tok in his drafts
205. 38:09folder. Comment if you're seeing this before it's made. Future. Quinn invites viewers to prove they're in 2025. Echol likes
206. 38:19past garage. Kyle's phone mysteriously racks up a non-existent video. I don't understand this. Logic locks to keep sci-fi tidy.
207. 38:36Consistent causality. Every future asset must show up unused in the past before it's repurposed. Entropy penalty rule.
208. 38:46If engagement drops below 88 comments per post, timeline signal degrades. Visual visual glitch motif. Observer
209. 38:55effect. Any viewer who stitches a future clip becomes scripted cameo in the next leak. I don't get it. Could Could someone with a brain
210. 39:09that understands metafiction and sci-fi timeline [ __ ] with explain if this makes sense or not. Oh, that's good. Okay, that's great. Okay,
211. 39:21that's a good idea, Brandon. Okay. Um, let's let's do that. Hey, Quinn. So, here's the deal. Moments ago, I called
212. 39:30you a dum dum. I was clearly being uh sarcastic and and you knew it was comedy, I'm sure. The the truth of the matter is I'm the
213. 39:40dumb dumb and it looks like everything that you wrote actually does include logical consistency even though it's nonsensical to be watching
214. 39:51something in the present that doesn't exist yet. Period. It is so complicated that I don't get it. Period. Can you explain it
215. 40:00to me like I'm a dumb [Laughter] dumb thinking April 17th imagine sticky notes
216. 40:15on a table future blue note shows a polished Tik Tok that says welcome to AI learning lab founded six months ago
217. 40:22present you today scrolling on Tik Tok and going wait I never made that video past pink note. You a month earlier
218. 40:31still debating whether to start your channel at all. Okay, here's the weird loop. Future you somehow drops the blue
219. 40:40note. Wait, the blue the blue note. Wait, do I have Post-it notes? Could we actually do this? Such an
220. 40:51idiot. I don't care. Okay. Uh, wait. Do I have him over here? No. Damn it. So, wait. Blue. Blue is the future.
221. 41:01Okay. You somehow drops the blue note back onto today's feed. So, I So, I put a future video on today's feed. Present you watches it, freaks
222. 41:16out, and shows it to followers. They comment, "This is wild. Do it for real." Those comments echo backward pink note
223. 41:25to give past you the push to actually launch the channel. I sort of get it. It's a bootstrap strap paradox. Like a chicken
224. 41:35and egg situation where the chicken sends the egg back in time so the egg can hatch into the chicken that sent it.
225. 41:41Totally bonkers, but inside the story it lines up. Future clip. So, this is basically the pot of what? Of of Back to the Future. Yeah.
226. 41:52Yeah. Yeah. That's it. One big time travel pretzel. Well, that's pretty impremp impressive. Hang on. I've got whiskey throat. I'm all dried
227. 42:07up. Oh man. Good lord. Poor Quinn gets the hatch closed on her with no advanced notice. She gets a bit jumpy. Time travel
228. 42:21paradoxes. Anyone mind any Is anyone's mind a pretzel yet? Mine is. I can't I can't keep up with it with the present
229. 42:31you. Okay. All right, people. Um, so let me start a new chat and I'm going to say what would be a
230. 42:47thing I can ask you parenthesis a um reasoning engine with
231. 43:01tools that would be challenging for you to solve but obvious to
232. 43:16me that you solved it. That could be it. We'll ask it. Present, past, future me doesn't get it.
233. 43:41Below are 10 prompts that you could throw at me that force real multi-step reasoning, data handling, or tool use.
234. 43:49For each note, why it stretches an AI, and what success would look like. Oh, this is cool. Upload a 40page PDF of messy
235. 43:58municipal financials. Oh, we could go get some data off Kaggle and tell me whether the city is running a structural deficit showing the
236. 44:09math, parsing tables, interpreting accounting language, doing calculations. All right, that's boring to me. Here's a blurry photo of a
237. 44:21handdrawn tournament racket. Recreate the bracket digitally and identify the champion. That's kind of fun. Pick the last 12 filings from the
238. 44:33SEC for some ticker chart quarterly free cash flow and give me one sentence trend digest. Oh, that's cool. Requires web search file
239. 44:47download numeric extraction and plotting. All right, let's let's do this one. Copy. Chart the quarterly free cash flow
240. 44:58and give me a one-s sentence trend diagn. So, we'll do Apple. Pull the past 12 filings from the SEC for what is it? AP AL. That's it,
241. 45:14right? I asked Chach if it could help me with my menopausal wife. Only advice he gave me was hide. Sometimes simple
242. 45:22advice is the best. Okay. The user asked for the last tweet 12 filings. Some sites already have quarterly Apple's free quarterly cash
243. 45:34flow. Uh, macro trends has uh, so yeah, that was a stupid prompt because it can just go find the answer.
244. 46:09Wait. Write Python that given any English panggram finds the minimal length pangic pangrammatic window. What's a panggram? Damn
245. 46:29it. All right. This thing's thinking. It's thinking about its Apple cash flow [Music]
246. 46:46thing. Oh, I know what I could do with that with that time travel script is is have it write a script. Then I could probably get
247. 47:04it. It has every letter of the alphabet in it. Pangram. Oh, so you give it a word and
248. 47:20then it figures out all of the other words that have all the letters. Sometimes it's thinking is scarier than
249. 47:29what's going on in my own head. I know. Yeah, this thing's going nuts. Well, that's, you know, someone someone said that there
250. 47:37was in the in the open AI when they announced 03. Oh, now it's doing some some math. Now it's plotting the
251. 47:44financials. Free cash flow in billions. Well, it did it. It didn't do it for all of what it said it would, did
252. 48:03it? What I did? Fast version. Located the last halfozen SEC filings using the Edger browser. Grab the XBR BRL cash
253. 48:15flow numbers for net cash provided by operating activities payments for PP&E. Computed free cash flow equaling
254. 48:26operating cash flow minus capex converted to billions. Ran the quick Python snippet so you can see the above
255. 48:33numbers. Caveat and next steps. You asked for 12 filings. I stopped at six because the SEC's rate limit. Oh, this
256. 48:40is cool. You asked for 12 filings. I stopped at six because the SEC's rate limit kicked in during scraping. The same script could be rerun
257. 48:50once those limits clear. Just extend the date list. Huh. [ __ ] fascinating. All right. Pangram. Let's do the panggram
258. 49:07thing. Copy. If you're new here, if you're wondering what we're doing, we're just playing with 03, which
259. 49:18is the new MacDaddy reasoning model. We haven't even started playing with 04 mini high, which is even Mac Daddier, but it's the small
260. 49:30version. Um, minimal pangic window finder. Good lord. Return the shortest contiguous substring of text that still
261. 49:43contains every letter A to Z at least once. Case insensitive. If text is not itself a panagram raises value error. All
262. 49:54right, there you go. Python snippet. Feel free to swap into a longer program. Um, let's see. I want to be
263. 50:08able Can I do canvas here? I think I can. Um to run the code. Um can I run it in canvas?
264. 50:44It's writing code for us. Stopping by to say hey everybody. Hey Ann Murphy. There's a number of people in here tonight that watched your
265. 50:54crazy nutty 12-hour content aon today that also came here tonight after all that and they said it was amazing. So
266. 51:03congratulations on that event. Um that's super awesome. Super super super awesome. Um and and shouldn't you be
267. 51:12taking a nap? Hi. We are so weird. Yeah. You know it's funny. I went to my whiskey club tonight and I'm just like I
268. 51:20was saying before you got here, there's something about getting older where if if you drink enough to get buzzed,
269. 51:26you're fine. You don't feel great the next morning, but you're fine. But I drank just under enough to get buzzed.
270. 51:32And so I just wanted to take a nap rather than do the show. All right, we're going to run this error. Um, fix it. Can I say fix it? Fix it.
271. 51:46Uh, fix the error. Fix the error. I have not been as impressed with chat GPT's uh uh canvas window running
272. 52:00code. It just doesn't seem to run it as good. Run. Oh, it's debugging. Oh, it was
273. 52:15fixing it. Run. Okay, I don't give a [ __ ] I'm gonna put it away now. I don't care. You suck. Okay, let's go. Let's go try one of its other
274. 52:30things. Uhuh. Okay, here's a blurry photo of a handdrawn drawn tournament bracket. Recreate the bracket digitally
275. 52:45and identify the champion. YouTube comment. Did you see what I tagged you in, Kyle? I created a tool. Um, I did see it. I didn't I I think I
276. 53:01played with it for a second. Gareth, just retag me on the uh on the AI salon and I'll look at it tomorrow. Um, we're so weird. Thank you.
277. 53:11It was lovely. I'm just so glad I didn't have to wake up and do another 12 hours. I know. Well, it's a good thing you
278. 53:18didn't call me because if you'd called me, I'd have said, "Well, here's other people we should have there." It would
279. 53:23have been 24 hours. So, I'm glad I'm glad you did it on your own. But, shouldn't we do like a All right, Ann.
280. 53:30Here's what we need to do. We need to do an AI festivis in July. You know the Christmas in July thing. We should do a
281. 53:37midyear AI festivus. So get on that. Let's That should be You should have that whipped up by the morning, [Laughter]
282. 53:52right? Your followers might be interested in Lenny newsletter deal. You get Replet, Bolt, Perplexity, and more
283. 53:59for 200 bucks a year. Interesting. Alio's wife's in for midyear AI festivus. I think an Ann has not responded. She's like, just Kyle, don't
284. 54:12say that out [Laughter] [Music] loud. All right. Oh, the LL Oh, the LLM comparison tool. I did not see that.
285. 54:27Gareth, where is it? Where did you put it? Did you tag me in in the salon [Music]
286. 54:41already? Gareth mentioned you in the comment. Here we go. You may find this handy. Llm comparison dashboard. This is
287. 54:53cool. Try Gemini Canvas. Oh, look at this, dude. How cool is this? All right, so let's wait. We have a screen redraw problem here. There we
288. 55:09go. Okay, open AI. So, chat GBT40. Uh oh, what happened? I broke it. [Laughter]
289. 55:26It looks cool. This was It looks like it was made in Gemini canvas because Gemini added canvas. It's not working. It broke. I broke
290. 55:42it. There's the code though. He said scroll down. Oh, okay. Hang on. GPT40. Scroll down. Nope. Nothing. The screen. She is
291. 55:58dead. Oh, wait. Hang on. No, she's not. 40. It's It's got, you know, it's got,
292. 56:14Gareth. It's got screen redraw problems cuz where stuff's showing up again is when I um hit command plus and command
293. 56:24minus. Explore all models. Oh, this is really cool. 1 million token top rank coding. Yeah, this is sweet at Vicky on Kyle. Wait, Vicki on
294. 56:42comments. The midyear airring of micro grievances. Stressed of us. Oh, I like that. Who came up with that? Oh, Vicki,
295. 56:50I like it. That's really good. The stressed of us. The the midyear airing of micro grievances. I think we might do that.
296. 56:59Kyle, you weren't kidding about how fast you can break something. I have I have a [ __ ] skill, man. It's it's it's seriously
297. 57:07within three or four clicks. I'm like, "Yeah, it doesn't work. It doesn't make any sense. It should work. I know it should. That's what I
298. 57:16said. Hallucination, huh? Here's the
299. 57:31thing. Human beings will not deal with this. Hey. Okay. You know what? I'm going to I'm going to stop surfing right
300. 57:38now because it doesn't really matter. Okay. A couple of things. one, you and I and everyone, if if you're curious about this AI stuff right
301. 57:49now, play with 03, 04 mini, or 04 mini high, play with all three of them, and just keep playing with them. Keep doing things. And even
302. 58:01if you're like, I don't get it. I don't understand what to I don't understand how this is different than the other
303. 58:06chat GPT, just keep playing with it. Keep pushing it. Ask it what to prompt you. Go on X and see what people are
304. 58:13saying about it and what they did with it. What we Okay, you know how people talk about how biologists talked about what
305. 58:29would prove animal intelligence is if animals use tools and then they realize that like crows use tools, right? And dolphins can
306. 58:40use tools like there there are certain animals out there that use tools. What 03 and 04 mini and 04 mini high are large language models with
307. 58:51multi-step reasoning with tools. It can write and execute its own Python tools and multimodality. It can see and understand images.
308. 59:04So you can have it unders you can upload an image and have it count the number of people in the in fact let's oh I got an
309. 59:11idea okay wait we are going to do something tonight hang on hang on calm down people calm down stop judging me okay watch this this may fail
310. 59:28um large aerial crowd photo. Oh, no. I didn't want that. I wanted Google. Come on. Oh, wait. Is it
311. 59:44going to find something here? Show me pictures. Show me pictures. Why didn't you show me pictures? You used to show
312. 59:58pictures. Whatever. Google. We still use Google. occasionally. Go find me a picture. Uh, large Ariel Ariel, whatever.
313. 1:00:13Um, crowd photo. All right. This is good, good, good, good, good, good. I like it. Okay, let's grab this one.
314. 1:00:28Now, let's let's grab Let's grab one.
315. 1:00:44That's one's good. Copy image. Okay, let's go back to chat. Japeta new chat. Paste. Okay. So, we're going to
316. 1:00:59say, I want you to analyze this photo and tell me how many people are in
317. 1:01:16it and the percentages of who's wearing what color shirt as well
318. 1:01:32as percentage of hats versus non-hats. All right, that's good. Quick sidebar.
319. 1:01:58Wait, I'm confused. Quick cyber. Oh, Tik Tok comments. Scroll up to Kyle yesterday. Okay. Okay. Kyle, yesterday. Okay. Kyle,
320. 1:02:12yesterday I asked you to do Wait, do LLMs understand? Today I asked Chat GBT with everything you know about me, roast me. It was
321. 1:02:21astonishing. It became a whole therapy session for my friend and I feel they do understand. Okay.
322. 1:02:36So the question do they understand? [Laughter] Hey Kyle. Yeah. Why don't you go ahead and start your uh analysis while you
323. 1:02:52Yeah, that's a good idea. This is this is I feel like this is gonna take Yeah. And it should this should be cool. This
324. 1:02:58should should be fun to watch. Okay. Um the the question of do they understand is a really interesting one because mathematically you could
325. 1:03:13argue that they're just looking at data and they're selecting the most probable data and they don't have consciousness so they can't
326. 1:03:27understand but what they demonst demonstrate is understanding and I know that sounds weird like are they
327. 1:03:35empathetic is different than can they behave with empathy does that make sense so when you asked yesterday and I said
328. 1:03:43they don't understand what I was talking about was are they capable themselves of understanding like perceiving that they
329. 1:03:51understand something which is different than they behave like they understand and I'll tell you it it is already
330. 1:03:58freaky how good they are But the fact that these large language models have access to everything you've ever talked
331. 1:04:04about with them now. They absolutely understand you on a on a deep no they demonstrate an understanding of you.
332. 1:04:12They present that they understand you. Um but I don't think that they they actually understand it like we would
333. 1:04:19think about that. So that's that's where my head was. But yeah, it's freaky, isn't it? It's it's absolutely crazy.
334. 1:04:27Um, quick stat. Crowd size about 190 people. All right, that looks about right. Circle detection on head shapes. Wearing hats about
335. 1:04:3940%. Not wearing hats. Color breakdown, blue, red. Okay. So, here's what I I wanted to do. I want you to um show
336. 1:04:54me closeups of the image where you have
337. 1:05:09annotations of what you you're looking at. to get these
338. 1:05:34numbers. Network error occurred. [Music] Retry. Yeah. No, Kyle, they understand you. [Music]
339. 1:05:53GPT went off the deep for me today. It got a virus. The user wants detailed close-ups of the image. What I'm trying
340. 1:06:00to get it to do here is is do some image analysis. Oh, one of the things that you can do, let's see, numpy p import
341. 1:06:11image head min R head head min R head max R hat min R hat max R slightly larger radius.
342. 1:06:36Yes. Oh, look what it did. Hang on. Detected heads likely hats. Wow. Hang on. Look at this. How do I Can I zoom in on this?
343. 1:07:00likely hats. I don't know that it got it right. Oh, and there's a bunch of watermarks in here,
344. 1:07:16too. Huh. But it's doing some sort of analysis analyzed. Oh, it's doing more. Top left, top
345. 1:07:31right. Huh? Every head I detected is ringed in cyan. The larger ones are the ones that
346. 1:07:47feel like there are hats. Well, I think it did a pretty shitty job actually, but it's still cool that it
347. 1:07:56can do that. A network error occurred. If you'd like an even tighter zoom on a specific spot, well, you could certainly have it
348. 1:08:05find a where's Waldo, couldn't you? An attempt was made. Yeah, exactly. This is This is This falls in the category of an an attempt was
349. 1:08:16made. Um, what was I going to talk about before? I was going to I was going to not not play with this anymore.
350. 1:08:32Um I don't remember what it was. From a distance you thought it worked. Yeah, exactly.
351. 1:08:47Um, let's do let's do we want to do another thing? Let's see. Um, complex physics poster
352. 1:09:22Fusion of multi-resolution seismic tomography maps using physics informed probability graphical models. All right, let's
353. 1:09:33let's copy this image. Um, I'm a dumb dumb. [Laughter]
354. 1:09:54Um, walk me through this poster in a way that I can easily understand
355. 1:10:12it. Um, show me closeups. [Music] of diagrams and images and explain
356. 1:10:30them and uh assume I know nothing. Okay, that's this could be fascinating.
357. 1:10:57for showing close-ups. We should structure it with a title and the author's at the top. All right. So, what's it
358. 1:11:06doing? It's figuring some [ __ ] out. The image is 2000 by 1867 pixels. Oh, this is cool. So, oh, you can't see that on Tik Tok, but it's
359. 1:11:16analyzing the image. So, this is it using tools, right? So, it's writing these tools. Yeah. Oh, look. It it chopped the
360. 1:11:30poster up into component parts, relayed them out on a page. Sorry, Tik Tok. You can't see what it's doing. God damn
361. 1:11:48it. Thought for 52 seconds. The potter The poster tackles a mouthful. Translation: We have blurry underground
362. 1:11:57pictures at different levels of detail. We're inventing a smarter way to blend them so geologists get a clearer picture. Great.
363. 1:12:07Perfect. Why bother? Oh, you know what's so good? I have an idea.
364. 1:12:24Um, go find me the transformers paper by Google.
365. 1:12:40and teach it to me in the same Wait, did this break up pictures for us? It did not. Okay. Um, and teach it to me in the same
366. 1:12:55way where you take their geeky language. I mean that as a term term of endearment, geeky language and translate
367. 1:13:12it for for a dum dum like me. Make sure that when you talk about a
368. 1:13:29concept that has a diagram or image, you show that image in
369. 1:13:47context. Let's see what it does. So, this should go find the archive white paper of attention is all you
370. 1:13:57need. It it it should then rewrite the concepts in it and show us pictures of the diagrams and explain them. In theory, it should do
371. 1:14:07that. I need to work with images for the presentation. I need to break it down. Attention is all you need paper on
372. 1:14:15archive. So, I found that. I found the paper and we'll cite it. I found several diagrams related to the
373. 1:14:29paper. Since I need proper diagrams, I can't rely on search results. I'll switch to an image search. Find the specific visuals I need.
374. 1:14:50So, it's actually understood the images now. That's kind of cool. Kyle, what's the cheapest way to create a
375. 1:14:58video cartoon from text? Oh, look. There's the attention. Okay. Attention is all you need. The transformer
376. 1:15:06degeeekified. Why did Google write this paper? Old school language models used recurrent or convolutional layers to
377. 1:15:15read a sentence one step at a time. They worked, but they were slow and struggled with long range relationships. The
378. 1:15:22authors asked, "What if we could ditch recurrence entirely and keep only the part that was doing the heavy lifting anyway?" Attention. Okay.
379. 1:15:35Um, you're still one level [Music] two obtuse. Make it even easier to
380. 1:15:52understand. And don't judge me, you stupid giantrained friaking genius that's got infinite capacity to support
381. 1:16:12humans. What are your best ideas on creating content for a website using AI? uh if you just want like the writing of
382. 1:16:21it just chat GBT chat GBT40 the base model it's it's remarkable at that what I would say is this um make sure you give it enough context
383. 1:16:33right so if you've got say marketing documents or uh you've got an existing website where you like some of the
384. 1:16:41content but not all of it give it that website and say here's what I like about it here's what I don't um maybe you've
385. 1:16:48got a creat creative brief or maybe you just you don't have any of that stuff yet. Um, a really good way to do it is
386. 1:16:55is put chat GPT into conversation mode into advanced voice and just talk about your business and just ramble and it
387. 1:17:04will just record all that stuff and then say now give me a website outline for that. You know, articulate the audience,
388. 1:17:13who the audience of the website is, what you want them to learn, what action you want them to take. give it all of the
389. 1:17:19stuff. Like if you were sitting down with a web designer, you'd give them all the stuff. A good one would ask for it,
390. 1:17:31right? The one sentence idea of a transformer. Instead of reading a sentence left to right like a game of
391. 1:17:39telephone, the transformer lets every word look at every other word at once and decide who matters most. Honest
392. 1:17:50to God, this is the first time this sentence, this is the first time I've ever truly understood what attention means in that [ __ ]
393. 1:18:01paper. And it makes perfect sense. If you read a sentence left to right, you've got to read the whole sentence
394. 1:18:09and it takes a long time. But if you look at all of the words at the same time and say, "Which words deserve my
395. 1:18:18attention?" That's pretty [ __ ] clear. Look at that. I learned something tonight. Thanks, chat. GPT, I know I'm on the right path, as
396. 1:18:28you say, what I've been doing. Great. Awesome. Beautiful. I use it to rewrite content for my students on grade level.
397. 1:18:35That's very cool. Oh, first Oh. Oh, yeah. That's cool. Um, you're welcome. You're welcome. You're welcome. Meet the two tower stack
398. 1:18:47carousel number one. So, here it is, the two tower stack. So, there's our image. Lower tower encoder. Encoder
399. 1:19:01reads the input. Decoder spits out the translation. Okay, here's what we're going to do. This is going to be fun. Okay, I want you to
400. 1:19:16write a short story.
401. 1:19:34that explains the core concepts here in a way that has entertaining
402. 1:19:49characters dialogue. metaphors, locations,
403. 1:20:05objects, plots, and um resolutions. Wonder if we can do inline images.
404. 1:20:20I'd also like you to create images in line in the story that bring it to life.
405. 1:20:43Use a claymation animation style for the images. I don't know if it can do this, but we'll
406. 1:21:00see. Or did it just make that up and you're being lied to? Did it make what up? The Wyatt crushed the old models. Reads a
407. 1:21:15sentence the old way one word at a time. Transformer all at once. Training speed slow can't paralyze. Transformer fast matrix math
408. 1:21:27in parallel. Oh, that that makes sense now. Why throwing more GPUs at a transformer trains it faster
409. 1:21:35because it can do it in parallel. You're looking at all the words at once. Ah, see it's all starting to come come
410. 1:21:43into the [ __ ] noggin. Long context. The old way doesn't work. That was what I was taught in speedreading when I was 13. Oh, that's
411. 1:21:54fascinating. Maybe, you know what? Maybe someone at Google, you know, was trained on speed reading and that's where this
412. 1:22:02concept came from. That's fascinating. Like quantum computing. What's the best app for for doing cartoon video right now, Mr. It? The
413. 1:22:12best one the the one that I did um hang on. I did this animation
414. 1:22:29[Music] in this one.
415. 1:22:44Oops. This was done in luma labs. Luma labis. So, make the image in chat GPT or midjourney, which I can't believe I'm saying
416. 1:23:00that. Like, chat GPT is good enough now to do images that, you know, like final images, not just like stupid proxy [ __ ] for a LinkedIn
417. 1:23:14post. Um, where was I? Back here. Image created. Encoder decoder. Look at this. Look how cute this is.
418. 1:23:24Getting started. Oh, it's doing all the images [Laughter] first. I like it. Look at our little claimation
419. 1:23:39style. It would be if it weren't fettered into uselessness. Yeah. What are you going to do? What are you gonna do? Oh, multiple equations at
420. 1:23:57once. AI minions. Yeah, right. Word. All right. So, it's making all the images. Oh, did it just stop? What's it
421. 1:24:09doing? Getting started. Doing another image. So, this is well, you could sort of occasionally do this before where you
422. 1:24:19could get it to make a bunch of images in a row, but what this is doing is Okay, so so here's a new thing. Okay.
423. 1:24:29Anyway, welcome to chat add. I got to get out of here. It's 10:30. Um, one of the things that 03 can
424. 1:24:43do that the other thing that that something like 40 can't do and 01 couldn't really do is it can do really
425. 1:24:52long sequences of actions. Maybe that's the best way to think about it. That um David Shapiro referred to 03 as semi-agentic.
426. 1:25:03that kind of like Manis when when you watch Manis like log into a website and then learn from it and then change its
427. 1:25:11instructions and keep moving. This is kind of doing that, right? So, right now I told it I wanted to write a short
428. 1:25:17story. I wanted, you know, illustrations to go with the story. It's not doing them in line, unfortunately, but you know,
429. 1:25:25whatever. But it's going to make all these images and then it's going to write the story. So with 03, you could
430. 1:25:32write a kids story with all the images, right, for every page and it'll just go do it all and you can just sit back and
431. 1:25:39let it do it. Why are you in 03? Because part of because 03 is a completely different kind of model than we've ever seen. It's
432. 1:25:52multimodal, can use tools, Python, write its own tools. um you know can see and understand and make images, use tools
433. 1:26:05um and then it's it's a a smart language model. It's got reasoning multi-step reasoning in in in uh combination with those other two
434. 1:26:16things tool use and understanding images. Um, and why I'm using it is to understand how is it different, why is
435. 1:26:28it different, when would I use it, what would I use it for? Because if you're a mathematician, it's pretty clear that you would use
436. 1:26:36this because it just maxed out the math benchmarks. If you're not a mathematician, do how many non-mmathematicians have PhD level
437. 1:26:47problems to solve? You know, not many are non-scientists, right? Positional encoding. I love these little illustrations, by the
438. 1:26:57way. Wordville and the great gossip maker. The twin towers of world world Wordville. Scene one. In the bright
439. 1:27:06bouncy village of Wordville, two rainbow stacked towers. Encoder and decauer. Encoder and decoder loom like
440. 1:27:16playful layer cakes. Gather round, hollers maestro former. A clay conductor and a tiny browser bowler hat. A dozen clay word folk shuffle
441. 1:27:31closer. Inside the first tower, the maestro decides divides the crowd into into trios. Quinn slips Quinn query slips on
442. 1:27:42ruby red glasses. I must figure out who I should pay attention to. Kila Key. Wait, do we have the one with red glasses?
443. 1:27:56No. All right, we're missing [Laughter] pictures. All right.
444. 1:28:11Um, the story doesn't clearly explain the
445. 1:28:26paper. It's now just confusing in a different way. Close the gap.
446. 1:28:40Don't worry about the images just yet. Again, good from far, but once you
447. 1:28:55zoom in. Yeah, exactly. Well, this is one of the things, one of the criticisms I've heard about 03 is that it seems to
448. 1:29:02oscillate between being absolutely horrible and absolutely mind-blowing. And it seems like there's not much in
449. 1:29:12between. Uh Kyle, you're the best. Learn something valuable within two minutes of turning tuning in tune of of tuning one
450. 1:29:21and a half hours in. Thank you, Sam Elliot. Good to see you, sir. Hey, I learned I learned what a what a
451. 1:29:28transformer is today. I've looked at that paper, I don't know, 50 [ __ ] times. I'm like, I sort of get it. I
452. 1:29:36sort of get it. And then it just did it. If you read a sentence left to right, it takes you a long time. If you look at
453. 1:29:42all the words at once, you can figure out what's most important to pay attention to. Got it. And you could do that in parallel. There you
454. 1:29:50go. That's the transformer. There it is. Absolutely amazing. All right. Did I get the story right now? Let's see. Are we done?
455. 1:30:06The word kids line up to form a foreign sentence. In old parades, RNN's they'd march single file. Each kid could only whisper to the one
456. 1:30:17behind. Slow and errorprone. Maestro's new rule. Nobody marches until everybody knows who matters. So before we move, every word
457. 1:30:27kid is allowed to look at every other kid to see whose meaning helps it ma make sense. That looking process is
458. 1:30:38attention. That's really cool. All right. Every kid is a word in a sentence. And in the old model, kid one would talk to kid two
459. 1:30:49would talk to kid three. In the new model, they all look at each other and say, "Who of us is the most important to
460. 1:30:56make sense of this sentence?" That's cool. The game of attention scaled. Each word kid asks two simple questions about
461. 1:31:05every other kid. How relevant are you to me? Query key score measures connection strength. What do you actually say? The
462. 1:31:15definition of the word, the content that might be borrowed. They soften the scores with a divide by D to softmax step so nobody
463. 1:31:28shouts. Okay, that one sucks. That needs to be rewritten. There's something here. There's something here. This is
464. 1:31:40kind of cool. All right, I gotta go. I got to get out of here. People, you got to get out of here. Everybody's got to
465. 1:31:44go to bed. What are you still doing watching this? How many people are here? Oh, not that many. 31. There's just 31
466. 1:31:51insane people here. All right, I'm leaving. All right. Uh, peace out everyone. Um, go play with chat GPT. Hopefully
467. 1:32:04while I've been doing this, you've been playing as well. Tomorrow, it's Friday, office hours at 11:00 a.m. on LinkedIn.
468. 1:32:11So, if you go to my LinkedIn channel, Kyle Shannon, um I've been using the same Google Meet link for the past two years. So, just go
469. 1:32:21find one of my previous events. It's 11:00 a.m. tomorrow. And then tomorrow night, Friday night, date night. All
470. 1:32:29right. All right, everybody. Beautiful. All right. Peace out. See you tomorrow.