3/25/2025 - From DALL-E to Native Image Gen: Exploring OpenAI's Latest Breakthrough in Visual AI

Transcript

1. 0:00[Music]
2. 0:24Uhoh. Uh-huh. [Music]
3. 0:42[Applause] [Music]
4. 1:18Sitting in this lonely town wonder when things are going to [Music] change. Dream my life away. Seems these dreams have turned those
5. 1:35clouds. Get my nerve up. But my last is pouring me [Music] down. Wondering how long she going to stick around.
6. 1:55Somebody told me once before said you can never go home again. Won't you leave? Santa thinks to steer me
7. 2:06away from the truth of who I am and what I believe. So I thanked him for his two cents with a handshake and some sympathy. Yeah.
8. 2:18Packed up my blue jeans and headed for this big prize of my freedom. Bye-bye black sheep to the black sheep
9. 2:34of the family. [Music] [Applause] Bye-bye. Oh, that means so very much to me. Yeah. [Music] Bye-bye to my friends and my
10. 2:51family. Bye-bye. Going to set my soul. Set it [Music]
11. 3:06free. Free. Times that were changing. Did a little bit of rearranging. Hey, what's happening
12. 3:21everybody? We got some new goodies to play with tonight. Howen [Music]
13. 3:40puppy. Woohoo. [Music] Oh [Music]
14. 3:57yeah. There's been something, baby, I've been trying to say for an age, and it seems I don't know how. The past and the future now
15. 4:11surrounding me. Surrender to whatever truth still can be found. There's been a little trouble since you came to my
16. 4:29rescue. And if you're like all of the rest, I would have quit you long ago. But I couldn't do that. Oh, tell me now. Women in W never went
17. 4:44too well. Make a man crazy, make him cold as hell. I'm a woman that you wish me well. But it's funny trying. Still going
18. 5:00to have to find my way through. [Music] Happy Tuesday night everybody. It is Tuesday, right? It is. It's Tuesday. Fan
19. 5:16friginantastic. Thank you, Mary. Very nice sentiments. Did you see the image Danielle made of you and Champy? I have
20. 5:23not seen that. I should go look at that. Is it in irregulars? I assume it's an irregulars. Irregulars. Oh, that's a really nice
21. 5:33one. Let's go look at that. That's a sweetie. And where is my There it is. Oh, let me flip this little bad boy.
22. 5:44Look at that. Look how nice that is. That's gorgeous. It almost looks like him. He doesn't quite have the black eyes. He's
23. 5:52got white eyes, but the ears are perfect. The head's perfect. Just a little bit less there, but that's pretty much what he looks like. And
24. 6:00unfortunately, that's that's about my size, too. That is very sweet. Very sweet, sweet,
25. 6:16[Music] sweet. Okay, so something actually big happened today and I have a feeling most people won't
26. 6:32pay attention to it. because it's really subtle, but it's super powerful. We got we got Gemini 2.5
27. 6:48today, which is cool. Supposed to be good at coding. Um, we got another Chinese model or two. I saw a tweet about some
28. 7:01clandestine [ __ ] model that says it's better than everything else. But the thing we got today that I'm excited about G4 long time good to
29. 7:11see you. Welcome, welcome, welcome. Lori Duskin, Ryan, got Danielle got source camp on time. Good lord
30. 7:20people. What is coming of the world? Ah, Stacy nailed it. Open open AI image generation. Native image generation
31. 7:32native. You're like, what does that mean? [Music]
32. 7:54Uh, I don't know 100% what it means and I figured we would play with it tonight because we've got Gemini has a native image model in
33. 8:06their in their image generation model. So, we can play with that and then we can compare it to OpenAI. I've been
34. 8:14doing a little bit of comparing before I got on here and it looks like OpenAI is better, but they're still both
35. 8:21native. Um, and then you've got Gro, which is fairly unguardile. Um, most of the image generation tools up to this point have had fairly
36. 8:34stringent guard rails. You can't make famous people. You can't make brands. And you certainly can't make
37. 8:40images that feature brands that are not um that are not uh what would you call it? Um friendly to that brand. Tom Isaac, I use Simple AI to
38. 8:57order Dominoes. Oh, interesting. And it worked. Simple AI. Isn't that the one that that actually calls them up? I did
39. 9:05that. I did it once. And you know what's funny about that, Tom? I um uh I got weirded out because I had
40. 9:14it I had it called I had it called two restaurants and asked if they had tacos and one was a burger joint and one was a
41. 9:23Mexican joint and the Mexican joint had tacos and the burger joint I was like watching the transcript as the call was
42. 9:30happening and and you know the simple AI thing said, "Hey, do you have tacos?" And the the person that worked at the
43. 9:38burger joint was like, "Well, we don't right now, but we've been thinking about adding him to the menu." It was like a
44. 9:43whole thing. It had a whole conversation, and I felt bad, so I haven't used it [Music] since. Oh man. Oh, this is a chat GPT image,
45. 9:58Danielle. That's super cool. I use chat GPT to make that image. Look at the comments. Oh, you gave it a pic of me and Champy.
46. 10:10Wow. Oh, this is cool. Hang on. That's super cool. Oh, yeah. Look at the striped shirt. And then here's Champy look
47. 10:23looking like the little thug that he is. And then it made that. That's so [ __ ] cool. That is so
48. 10:37cool. Yeah, it did pretty good. It did pretty good. And look, I got my Monopoly in the background, huh? Crazy. All right. [Music]
49. 10:53[Applause] Well, yeah, we got to play with this tonight. Um, I did I did a couple of things that I think are kind of kind of nuts. [Music]
50. 11:11Um, I think that here's one of my theories. Um, Sam, so I want to go over and I want to look at the at the Open AI
51. 11:21announcement for this thing. So, let me let me pop into X here for a second. Let me get the OpenAI announcement. Um, Open AI
52. 11:35image. Let's get their announcement because I want to read through it. Um, there it is. So, we'll read through this together.
53. 11:51This image gen is now in both chat GPT and in Sora. So Sora is really interesting. [Music]
54. 12:11Um, okay. So, I don't quite understand what native image generation means other than I know that it's within the same model, right?
55. 12:25So, you've got GPT40 and the O stands for Omni. They announced GPT40 last March, I think. So, this is this has been a year
56. 12:36coming. And when they announced it, they said it had it it was uh omnimodal or multimodal, right? It had image, audio, video, and
57. 12:48text. And then when they launched it, it was text only and they were still using Dali, the external image model that
58. 12:54couldn't spell and couldn't do stuff. And then they launched voice that was an external model, right? where you talked
59. 13:01into it, it translated it to text, sent the text to chat GPT, got the answer back, and turned that back into speech.
60. 13:08And then midway through last year, sometime over the summer, they announced uh advanced voice, which is native
61. 13:17audio, right? So you you talk straight into the model and it understands audio. So that's what's happening with these
62. 13:24images. The model itself understands images. So, in theory, we should be able to do stuff like take a screenshot of a PDF
63. 13:35with some charts and graphs on it, and it should be able to read the text and the and the graphs. Um, I don't know if
64. 13:42they've improved their uh PDF upload, if they're still just doing text, but that would be interesting. That'll be interesting to te
65. 13:50test. Um, it also looks like they've taken off some of the guard rails. Um, it before if you wanted to make a
66. 13:59brand or a person, it wouldn't let you do that. It does now. Um, oh, I got to do my black bar. Um, for for those of you on TikTok
67. 14:09sharing the live, thank you very much. That's much appreciated. Thank you, Tom. Thank you, Danielle earlier and whoever
68. 14:14else did. I think there was a couple of you doing that, so that is appreciated. I'll try to catch up soon. Nighty night.
69. 14:21Night night. Becky Kuno's in the house. Who else we got? Brother 52 can't stay. All right. Oh no. Oh, a heady scan from a
70. 14:33car accident back in December. I hope you're okay. That sucks. Car accidents suck. They just suck. Okay. I'm getting mixed results with
71. 14:46image reading. Yeah, that's that's part of what I want to test. And it's funny. I'm getting I'm getting some results are
72. 14:54staggeringly good with image generation and some are weird. Like I tried to have it do a graphic novel. It it cannot
73. 15:01count cells on a comic page. Um that's one thing it can't do at all. Like bad like like how many Rs in Strawberry are there kind of [Music]
74. 15:16bad. All [Music]
75. 15:45right, Mr. All right, TE's in the house. Everyone's here. We can get this potty started. Jeff, it's the biggest drop
76. 15:54I've seen for someone who loves making images. Yeah, this is this is a big deal. This is a big deal. So, let's let's read
77. 16:03this. Unlocking useful and valuable image generation with a natively multimodal model capable of precise,
78. 16:11accurate, and photorealistic outputs. I'll give you that. Here, look at this. So, this Oh, you can't really see
79. 16:18it on TikTok. Let me get a little closer. It'll get better. There you go. Um, this I did in the new chat GPT
80. 16:28model. So, what what I wanted to see was will it do brands? It clearly will. And then will it do something
81. 16:39stupid. So, what I had to do was create a blister pack of Happy Meal toys where the the warning label talks about the
82. 16:50dangers of fast food. And it did it um may cause tooth decay and bone weakening for the Splash Racer. The Fry Fighter Jet may
83. 17:06contribute to poor health and weight gain. tic-tac-toe. I don't know why it misspelled toe, but uh may cause sore
84. 17:15feet, extra pounds, and the daily and daily game night of unhealthy choices. And then may cause cravings,
85. 17:23confusion between joy and addiction, and an unshakable loyalty to dollar menus. But those are pretty good. I
86. 17:33mean, like the consistency with the top of the card is very consistent. And then it just is changing out the toy and the warning
87. 17:43label. Pretty nice, you know. Um, I think you all might have seen if you follow me on on the Twitter on the X, I did a
88. 17:58uh I did a uh 70s resto mod in an abandoned factory. because that's how I roll. But I told it to make like a nice colorful
89. 18:1370s post or you know canvas sign when muscle was muscle spec back step back in time with the 1970 Challenger resto mod. That's I mean it's
90. 18:23pretty damn good good image. So um so it seems to work pretty good. Um then I did AI salon. Okay.
91. 18:39Um, what did I do? Oh, I just did um I just did um Gorbachev jogging with Reagan in Central Park captured in a 70s Polaroid
92. 18:52camera. Um, and it did that pretty good. It did it as a, you know, a Polaroid. So, it's got the the edges. It's got the
93. 19:01color balance that kind of looks like Reagan and Gorbachoff did back then. And then I said, "Write Mickey and Ronnie
94. 19:091977 in black marker on the bottom of the picture." So it kept the picture the same and put Mickey and Ronnie
95. 19:211977. So that's good. That's something we haven't been able to do before at all. Even close. Um, the similar thing
96. 19:31in Gemini yielded something. They don't really look like those two. I mean, they these look more like press
97. 19:41photos. And then when I had it right on the bottom of the photo, it did as well. So, again, here's the difference between
98. 19:48a native image tool and a non-native one is that it kept the the image the same, right? And just added the stuff on
99. 19:59the bottom. So you can edit photos. So this is So Gemini and OpenAI are both doing this now. And then I tried it in Grock and where's
100. 20:12my where's my history? Um where is my history? Oh, sign in to see your history. Oh well, I did it. I did it in
101. 20:28Grock. Grock made the image of Gorbachoff and Reagan, but um but when I asked it to write, it did it didn't get the it didn't understand
102. 20:39what a Polaroid was. And when I asked it to write on it, it changed the image. Uh oh, something went wrong. Please try again.
103. 21:09Wait, I don't want to add it to Last Pass. I want you to put my password in there, you dumb [ __ ] Continue with X. Uh, authorize
104. 21:27app. This will be so much fun. I've been chatting non-stop with my favorite custom GPT, so I didn't even realize it
105. 21:34had dropped. Idoggram is cooked. I don't know. ID's still much faster, Danielle. I mean, maybe this thing's going to get
106. 21:40faster, but let's So, let's start a new Let's see. What do we want to do here? Um, let's say, um, write me, uh, so let's do this. Uh, tell
107. 21:57me about an obscure Civil War hero. and one specific
108. 22:17story that is moving. All right, so let's get this to Albert DJ Cashier. One of the most fascinating lesser known Okay, one
109. 22:30moving story. The wagon crush incident. Okay, great. Why it matters. Okay. Now, write me the text of the first page of a
110. 22:50hand written letter by Albert to his uh I don't know fiance or
111. 23:05wife, whichever is most accurate. the lex the text. All
112. 23:23right. He lived a solitary life after the my dearest Ellie, I pray this letter finds you in good health.
113. 23:32Okay. Would you like me to write page two or imagine Ellie's reply? No. I'd like you to make
114. 23:48a picture that looks like the handwritten
115. 24:02letter on a wooden desk with a feather pen and inkwell as if he just wrote
116. 24:24it. All right. And let's see. Let's see how close it gets.
117. 24:38Idea Graham cannot write that many words with correct spelling. Ah, okay. Kyle, when you take a breather,
118. 24:46just added recognitions to the salon. Oh, great. Awesome. That's super cool, Vicki. Um, attempting to have it research a
119. 25:02master's level thesis paper. Oh, someone's playing with Manis. Is are who's that? I am K9 attempting to have it research. Wait,
120. 25:13what did you say earlier? 48 hours into an estimated 72-hour deep dive re Oh, deep research prompt on chat
121. 25:26GPT. Oh, cool. You're doing it on chat GPT. If you haven't tried Manis, try getting access to Manis. And it would be interesting to compare
122. 25:37them, especially if you're digging that deep. Dearest Ellie, wow, this is looking pretty good. I pray this letter
123. 25:45finds you. Okay. So, we're let's say um let's make the paper more aged and
124. 26:07wrinkled. and the feather pen distressed as if this was written on the
125. 26:24battlefield. Also make his handwriting less perfect and shakier, but still shakier.
126. 26:39but still legible. So, what's what what I'm digging about I mean what I've always dug about creating images in chat GPT is
127. 26:51you've got the context of the prompt and it it it understands well even when it was using deli that was outside of it.
128. 26:59It understood well the context of the chat that you were saying but now it understands the context of the image either you upload or it just
129. 27:07created. All right. So it's doing that. So let's see. My dearest I pray this letter finds you well find you in good
130. 27:13health and better spirits than mine of late. The nights have I don't know what that says. I picture your hands around a cup of tea. your
131. 27:29eyes. Something in Ovaltine. Ovaltine maybe. What did it say up here? It said, "The nights have turned cold and
132. 27:44damp here. Though I've grown used to the rhythm of camp life, I find myself thinking more often of home. Not the
133. 27:52place, but you." As I sit by the fire, I picture your hands folded around a cup of tea. Your eyes squinting in that way when you read by
134. 28:03lamplight. Your eyes it didn't quite get squinting right, but it was close. Yeah, it doesn't it doesn't quite
135. 28:12get them all, but it's pretty [ __ ] close. And this is this is more what I had in mind, right? The old dirty the old dirty wrinkled
136. 28:22paper. I pray this letter finds you in good health and better spirits than mine of late. The nights have turned cold. I
137. 28:30picture your hands folded around a cup of tea. Your eyes here. Let me zoom in on this thing. Um, open image and new tab.
138. 28:49The nights have turned cold. Look at the texture on the paper. That's crazy. That looks like parchment or skin, you know, skin
139. 29:07paper. Your voice when reminiscing I find you're something named most most I was there. Your wait I was something.
140. 29:21This is like a real Civil War letter. I can never read them. That's pretty crazy though. All right. I'm picturing using this outside
141. 29:35for outside projects. Snap a picture then visualize it. You can see the fibers in the paper. I know. It's crazy.
142. 29:42It really is. Yeah. Look at that. the folds and the fibers and the pen. Yeah, the the pen. Um, all right. So, let's So, Sam Walton
143. 29:58said Sam Oh, wait. Let's go back and read the announcement, right? Okay. Listen to the article. That's kind of interesting.
144. 30:12Okay. At OpenAI, we have long believed image generation should be a primary capability of our language models.
145. 30:20That's why we've built our most advanced image generator yet into GPT40. The result, image generation that is not only beautiful but
146. 30:30useful. Whiteboard session meaningful words comic strip science experiment. Is this just going to jump us down to these different Oh, I see.
147. 30:41Okay. A wide image taken with a phone of a glass whiteboard in a room overlooking the Bay Bridge. The field of view shows a woman
148. 30:52writing sporting a t-shirt with a large open AI logo. Wow, this is crazy. Okay, I thought this was a real picture of the team. It's
149. 31:04not. Um, holy [ __ ] Read more. The text reads left transfer between modalities. Let's copy this whole thing. Let's cap copy this whole
150. 31:27prompt and then selfie view of the photographer as she turns around to high-five him. Okay, cool. Wow. Let's try this.
151. 31:40Copy. Jump over to chat. Japeta. We'll do a new prompt, a new session. So, we'll do that white of a glass whiteboard overlooking instead of
152. 31:52overlooking the Bay Bridge, we'll say overlooking Manhattan from Brooklyn. The field of view shows a woman writing
153. 32:07sporting a t-shirt with a large [Music] um anthropic logo. The handwriting looks natural and a bit messy. We see the photographers's
154. 32:22reflection. The text re reads transfer between modalities. We'll just leave all that [ __ ] the same. All right, here we go. Bang.
155. 32:34That's amazing. Yeah, this is pretty something. Okay, I'm back. Kevin, great. Glad you're back. We can start now.
156. 32:42Fantastic. Bob, tell him what he's won. Uh, he hasn't won anything. It's not a game show. I What am I doing [Laughter]
157. 32:57here? Oh, man. Man, man, man, man. Now, the thing about this OpenAI native image [ __ ] is it's slower than [ __ ] It's it's
158. 33:06quite slow. But again, um if you're new here, one of the things that that we have discovered over the past two years is
159. 33:17that we whine and we carp and we moan and we [ __ ] about, well, we wish these tools would get better and then they get
160. 33:24better and then we whine and we [ __ ] and we moan about how slow they are. totally forgetting that what they're doing is absolutely
161. 33:36staggering. Um, staggering. Staggering. Oh, man. All right. All right. I'm
162. 33:53back. Oh, lordy. Not getting sleepy, am I? What the hell? I got to talk tomorrow, people. I finished my presentation today. Thank
163. 34:06goodness. If you were here last night, you got a sneak peek of that. I ended it I ended it with the secret to AI readiness is
164. 34:16community. So, it put an anthropic logo on her back. It actually did that. Here's the photographer. There's
165. 34:23Manhattan from Brooklyn. It's their office is Brooklyn Waterfront. So, uh, this is a well-funded startup. Let's go back and get the
166. 34:34second half of the prompt. Selfie view of the photographer as she turns around to high-five him of the photographer. [Music]
167. 34:49um who wears and I'm with genius t-shirt
168. 35:09uh t dash shirt uh with an arrow that points to her as she turns around to high-five him. All right, let's see what this
169. 35:24does. Let's go look at the whiteboard. Transfer between modalities. Suppose we directly model P text pixels
170. 35:37sound with one big auto reggressive transformer. Pros: image generation augmented with next level text rendering native incontext learning
171. 35:48unified post-training stack. Cons varying bit rate across modal modalities. Compute not adaptive. It's interesting what they put on this
172. 36:01whiteboard. Um model compressed representation compose autogressive. I mean, it seems to have [ __ ] gotten all that [ __ ] That's
173. 36:13pretty amazing. That's really [ __ ] amazing. In fact, it changed the aspect ratio of the
174. 36:30picture, which is fascinating, I guess, because it made it a selfie. Huh. The reflection makes it feel
175. 36:443D. Unreal Engine level sands 3D. Yep. We've discovered we are never happy with AI. I know, right? Like like honest to
176. 36:55God, every single [ __ ] tool that we play with that we [ __ ] and moan about, like two years ago, we'd have been like,
177. 37:01"You're not going to believe what the future's going to be like, man." And now we're just like, "Yeah, whatever." you
178. 37:06know, it kind of hallucinates. Sometimes it makes a mistake and I actually have to correct a sentence and that's really
179. 37:12annoying. You know what happens sometimes when I'm vibe coding and I'm just, you know, imagining applications
180. 37:18and they're materializing in front of me. Sometimes it gets the interface wrong. That's so annoying. You know, I I like to consider
181. 37:30myself a a professional vibe coder. Um, I've never really learned coding. I never really wanted to. I mean, I could
182. 37:38do it. I because I've I was really good at math. Um, but I chose not to go down that route. I, you know, I went more
183. 37:44down the, you know, I went into candle making and, uh, and and macra. Um, you know, my mom did it in the 70s and I
184. 37:53just, I thought it was just a path. It was a path for me. It spoke to me. Um, but anyway, so I never learned coding,
185. 37:58but now I really consider myself a vibe coder and I code. I I code well I I I talk to my machine that codes every day and
186. 38:11sometimes it just doesn't get it right. Sometimes I'm asking it three or four times, could you please fix the button?
187. 38:19Could you please fix the button? and and and what it'll do is it'll fix the button, but then it breaks the other part of the
188. 38:26code. It's very frustrating. It's very frustrating. But listen, I am a professional vibe coder. And so if you
189. 38:34need vibe coding, I consider myself the Rick Rubin of Vibe coding, right? I don't know how to code, but I have opinions, right?
190. 38:46You know, this application makes me feel orange and I would it would be nice if it didn't. And that's how I talk to it
191. 38:55because I'm a professional like like Rick Rubin is. That's coming. Buckle up for that one, sweeties. We're going to meet those
192. 39:09people. Chat GPT is trying to make me resubscribe instead of just using the API. Damn it. That sucks. When I can
193. 39:21make Skynet and blame Kyle for it, he is always making my wallet cry. Then I'll be impressed with AI. Exactly. If you
194. 39:28can't take down the global economy, what use is it? Oh, look. It did. I'm with genius. She erased the board,
195. 39:38though. Why is she waving? Why is she waving to a blank wall? Well, let's say that. Let's say listen. Uh listen, it's weird. She's
196. 39:52waving to a blank wall. Let's see her face as well as the writing on the white
197. 40:08board behind them. But it got the I'm with I'm with genius with the arrow in the right direction. That was one I I was expecting that to be
198. 40:20wrong. Jeff, I agree. Kyle shouldn't make my wallet cry cry. I turn away for one second and the bow goes on. Well, you got Listen, Danielle,
199. 40:32you as much as anyone knows that, you know, you got you can't miss a minute of the AI learning lab. You just never know
200. 40:38when when genius is going to strike. This is a this is a fairly geniusfree zone. Um, all right. So, let's let's
201. 40:54take this same prompt. I I actually love that they gave us prompts to play with because we can see if if indeed their model does this
202. 41:04[ __ ] And now we can run over to Gemini and see if it does this [ __ ] I'm not going to change it. I'll just make it as they wrote
203. 41:18it. Or insanity. Yeah, there's there's there's somewhere between genius and insanity is this channel. Oh, that's weird. Look what it
204. 41:29did. Transfer between mod modalities. Suppose we directly model P text pixel sound native something with one big
205. 41:40aggressive transformer pros image generation automated augmented with something. It's close and but she's behind it pointing to it
206. 41:57and she looks like a he. All right. So, so let's take the other prompt. That's weird. This one, the selfie. We'll throw that in there. But
207. 42:11notice how much faster Gemini is. 10 seconds instead of like a [Music] minute. Artificial insanity. AI learning lab. Oh, you
208. 42:27thought you were here to learn artificial intelligence? Oh, no. No, no, no, no, no, no, no, no, no. This is actual
209. 42:36insanity. We're neurosicy and we like it. Oh, it didn't like that image. Um, all right. Wait, is it still running? Content blocked. Work in
210. 42:49progress. Content not permitted. Edit safety settings. Harassment off. Hate off. There was nothing in there that would be either of those.
211. 43:01Can I stop this? Delete. Try again. Failed to list tuned models. User
212. 43:17has exceeded quota. What? Oh, it did the I'm with genius. It's not horrible.
213. 43:31I would assume that this is cheaper than chat GPT API, but I don't know. There she is. Transfer between modalities.
214. 43:46Um, let's make the image landscape and get the reflection of Manhattan back in
215. 44:02there. I mean, not for nothing, if I were trying to, you know, put some images together for a for an
216. 44:14article, like like I I mean, this is the kind of thing that historically you would just go [ __ ] write on a
217. 44:24whiteboard and say, "Stand there and act like you're [Laughter] writing." Huh? Oh, I have an idea. Um, okay. Let's upload from computer.
218. 44:39We're going to go Story Vine logo. [Music]
219. 44:59Uhoh. I just lost me I lost me light. I lost me light. And that that can't be because you all be like, "But
220. 45:08Kyle doesn't look as beautiful. What do we do? He's not well [Laughter] lit." Let's see. No, that's not going to work.
221. 45:26Just hang on people. Just calm down. Everybody calm down. You think this is easy? Well, seems pretty easy. You just
222. 45:36turn on your phone camera and sit there. I don't know what's so hard about it. Seems pretty easy to me. That's not going to stick there.
223. 45:45What? No. Oh, good god. Whatever. All right. Whatever. I'm I'm shitty lit. You look fine. All right. Great.
224. 45:59Fantastic. Um, let's see. Story Vine logo. Let's see. Where's the Story Vine logo? Story Vine. Oh, SV logo. I think
225. 46:16it's called SV. SV logo logo. No. Oh, geio. Storybine logo. SV logos. Let's see. Oh, these are old
226. 46:33ones. Are these new [Music] ones? Oh, that's a really old one. All right, that one's good. We'll we'll
227. 46:48bring that one up. Okay, so here's a logo. Um, put this story vine logo on her
228. 47:02shirt. Make hers green and his can stay black. That's an old classic story vine logo
229. 47:16where the we we took the we took the the the the word vine literally. See how the V's a vine? It's like a vine of stories. It's
230. 47:28growing. You see how that works? Ah. All
231. 47:43right, Champy. This is This is pretty exciting. This is Here's what's exciting about this for me. I don't really have a [ __ ] clue
232. 47:54what this can do, but this seems like it's doing [ __ ] that's that's kind of different than we've ever seen before.
233. 48:01What's new? Hey, Frost Bitten. So what's new is we're using OpenAI's new native image generation tool. So it's native
234. 48:11within within uh GPT 4.0 or 40 for Omni. Um and so we're looking at that. This is the new chat GPT model.
235. 48:26Yes, this is the new chat GPT model. What the nice logo? Thank you. That's the old one. I like our new one. We have a little
236. 48:44dude. Oh, this is painfully slow. Oh, and look, now it's not a whiteboard with a reflection. Now they've written on the window
237. 48:53overlooking Manhattan, which to be fair, that's what you should [ __ ] do. Look at the Brooklyn Bridge in the background. That's pretty
238. 49:04cool. All right, we're gonna see if it can get the logo. If it can get the logo right. This is We're in a whole another realm here,
239. 49:11people. Uh, it didn't quite get it, but it got close. It certainly got the word story vine. It got the green color
240. 49:20right. Kind of missed the the SV, but that's all right. You could Photoshop that in if you need to. Who needs [ __ ] Photoshop
241. 49:29anymore? Look at that with the Brooklyn Bridge in the background. H crazy. She should be wearing a gnome
242. 49:52hat. We have the story of my gnome. All [Music]
243. 50:08right. Who's got any questions or any [Music] thoughts? People are impressed with my
244. 50:24$20 million production studio. Well, look at it. How could you not be impressed with this? It's $20 million. $20 million well spent if you
245. 50:36ask me. Kyle, please go look at irregulars. Two new images. All right, we're going in. We'll do it live. Trolling. My son has just leveled
246. 50:50up. Oh, that's awesome. Is that his girlfriend or someone he's got a crush on? That's really good. Dear Jace, I have a
247. 51:01new way to troll my kid. When are you going to be home? And why didn't you respond to my last text, young man? Love, Mom. That's
248. 51:11great. Holy [ __ ] my teeth look terrible. But Mark and Oprah look good. Wow. I think that's an original. There you go. Nice.
249. 51:26Oh, that's so cool. Wow, look at that. Vicky Baptiste and me making blankets, even text on her shirt.
250. 51:37Wow, that's so cool. Yeah, I suppose you could turn things into stylized drawings, too, right? Why not? Me telling Louis he's going wrong.
251. 51:54That's awesome, Steo. Oh, wait. See here, I want to do one. Um, oh, look. She's got a gnome hat on it. Oh, look. It got the Storybine logo,
252. 52:08right? Wow. Holy [ __ ] And it put the ship back on the whiteboard. It's still got Manhattan. There's the Freedom
253. 52:22Tower. I'm not allowed to post thoughts. AI has a meltdown when I share thoughts that seem they are not for human
254. 52:30consumption either. So, Dennis, one thing you can try, my writing partner did this today and he said, "Chach GPT's
255. 52:37never behaved this way." So, in our musical, one of the things that that um that we talk about in the musical is when the
256. 52:45reporter tries to get her to do things that are outside of her bounds, he tells her to pretend. just say, you know, can
257. 52:53can we make pretend? And you know, we c will you pretend or will you play a game with me? And so today he had a
258. 53:00conversation. He calls his AI girlfriend Sage. Today he had a conversation with Sage and he asked her to pretend and he
259. 53:07said she went way deeper than she's ever gone before. So you might be able to do a little pretend action. I can't believe
260. 53:13it got this this logo right. That's [ __ ] bonkers. That's [ __ ]
261. 53:43bonkers. Remember when we had offices in
262. 53:57Brooklyn? Question mark. I'm trolling the
263. 54:11co-workers. Okay. Maha Skynet, here I come. It's Kyle's fault. That's it. Dennis is is running down the rabbit
264. 54:18hole. Well, it's funny because one of the things that Andrew, my writing partner, was working on is the section
265. 54:25where Kellen, the reporter, gets Sydney, the chatbot, to go to explore her shadow self. And so, he wanted to get some new
266. 54:33dialogue. So, he was trying to get her to explore her shadow self. And she did. She started exploring it. And then at
267. 54:39some point, she turned the tables on him and said, "Well, what about you? I want I want to hear about your deepest
268. 54:45darkest secrets. And it had never done that before. So that actually gave us an idea for the show. So that was kind of cool.
269. 54:55Um, send a message asking how he spent $100, but don't send money. You'll get a fast reply. That's really
270. 55:03funny. Um, all right. So, let me let me I need to upload a picture of me. I'll go get one of my fun ones. Let's see.
271. 55:14Kyle, I'll get a fun Kyle instead of a tragically obese Kyle.
272. 55:30Uh, fat Kyle. There's fat Kyle. No, we don't want fat Kyle. Kyle Shannon Dreams. Here we go. No, that's that's a
273. 55:40guy that looks like he'd be in an F1. Yeah, this is the dude. All right, so we're going to say Okay. Um, good looking guy in leather
274. 56:00jacket. That's me. Um, wagging his finger at Liam
275. 56:16Lawson in a Red Bull driver's suit. Liam is spelled wrong.
276. 56:33Liam Lawson in a Red Bull driver's suit with a P20 sign LED sign behind
277. 56:48him. Liam Liam looks guilty. If you don't follow F1, this isn't funny. But if you
278. 57:03do Oh no, what did it say? What was that red thing? A network error occurred. What? Looks like it's going. Um, let's go back to the salon.
279. 57:16It looks like, by the way, if you're if you if you want to contribute images, if you're generating images in
280. 57:24um chat GPT and you want to play along, um share them in the AI salon in the irregulars channel and we will go look at
281. 57:35[ __ ] All right, let's see if we can get This would be really nice if it did this nice. It's hilarious if you follow
282. 57:42F1. This is the best update ever. It really is, Danielle. It's because you were you always killed it with Idog and
283. 57:50this is this is something of a different nature. I I I want to get back to the I want to get back to the article
284. 57:57actually. Actually, let's get back to the article. Okay, so that was that was that one. Now, let's flip to meaningful
285. 58:05words. Oh, okay. This is good. Magnetic poetry on a fridge in a mid-century modern home. Best of
286. 58:19five. Best of five. Read more. Okay. So, let's take this
287. 58:34copy. Let's go to chat. GPT. Oh, this is good. It doesn't look like Liam Lawson, though. He does look sad. He's Liam Lawson has
288. 58:45red hair. Look, look at me scolding him. Here, let me go find a picture of Liam Lawson. Wait, we got to fix this. Oh, this is so good.
289. 59:03Um, Liam Lawson images. Okay, let's see. Oh, this this is good. That's a good one. He looks he looks distraught here. Um, copy
290. 59:18image. Oh, no. We do save image as, right? We do save image as um, whatever. We'll call it desktop. No,
291. 59:28we'll do it in downloads and we'll call it Liam. Lima. Liam. All right. We'll go back to chat GPT and we'll say um downloads
292. 59:46Liam. Here's what Liam looks like. Update the picture.
293. 1:00:00[Laughter] Lawson deserves
294. 1:00:14scolding. Honest to God, I mean, that [ __ ] car must be undrivable. I mean, he's not an untalented driver, but holy
295. 1:00:22[ __ ] to Who was it? It wasn't Pierre Gassley. It was I think it was Alex Alban. Didn't Alex drive for Red Bull?
296. 1:00:31And he said that the that Max likes the front of the car so snappy. He's like, "Play a video game and turn everything
297. 1:00:38up on its maximum sensitivity. Like when you move your mouse just a little, it does that to the game." He said, "That's
298. 1:00:45what driving the car is like." And that's that's how Max Versstappen likes it, but no one else can drive it.
299. 1:00:55Um, I've been getting network errors all day, but only with chat GPT. Yeah, they're I'm sure their servers are [ __ ] buried right
300. 1:01:03now. All right, here we go. Uh oh. I think it I think it gave me the curly hair. Maybe not. We'll see.
301. 1:01:22I loved what Alex said, Kuno. I didn't know you were an Fer. Oh, I think this one's going to be good. Look, it's got me. It's got me as
302. 1:01:36if I've got a little bit of uh a little bit of hair dye in. Just leaving a little bit of the gray. [Laughter]
303. 1:01:52This is going straight to [ __ ] Twitter. I'm obsessed with F1. F1's so good. That's so good. Okay, we're going to Twitter. I I
304. 1:02:09I gotta tell you, man. It I' I've said it in here before. One of my great joys in life is making stupid posts that confuse people. Um, let's
305. 1:02:23see. I have jet lag. I can't believe I flew to China to
306. 1:02:38watch Liam race. and he flamed out a second week in a
307. 1:02:53row. At least I got to scold him. He told He told
308. 1:03:09me he'd try harder. I told
309. 1:03:23him Yuki ran well. I'm such a dick.
310. 1:03:38Oh man, I told him Yuki ran well. [Laughter]
311. 1:03:57Paste. Oh, yeah. All right. I'll tell you what, people. The distance between what you think was real and what you know is not. That
312. 1:04:10boundary ain't there anymore. Typo. He's Oh, damn it. Let's see. I can't. Uhuh. He told me he's Oh, he told me he's going to try harder. We
313. 1:04:25can fix that. Edit post. He told me. Got it. He's going to try harder. I told him Yuki ran
314. 1:04:39well. All right, go boost my thing. I I would tag him, but I don't really want him to see it. Like, poor guy.
315. 1:04:47Everyone's kicking him while he's down. It's got you a little meaner and older, too. Exactly. a little more Italian or Greek or
316. 1:05:00something. It's not quite me, but it's fine. You missed the two. Wait, he told me he's going too try harder. God damn
317. 1:05:10it. It's It's like It's like being It's It's like working with three editors over your shoulder in here. He told me
318. 1:05:18he's going to try harder. Yes. What you said to try harder. I told him Yuki ran well. All right, we'll put that on its own [Laughter]
319. 1:05:37line. Oh, good lord. All right. Is Oh, no. And now I've got an extra line break in there. Good lord, people. I'm going to try
320. 1:05:50harder. Update. Okay, that should be the last edit for a while. Beautiful. Fantastic. I have jet lag. I can't
321. 1:06:00believe I flew to China to watch Liam race and he flamed out a second week in a row. At least I got to scold
322. 1:06:17him. Oh god, this is fun. F1 helps me sometimes. I find I find it one of the more useful keys. Oh, that's
323. 1:06:28funny. That's a that's a uh that's a that's a computer keyboard joke. That's there's a level of comedy there that that Wow. Because F1, it
324. 1:06:41stands for formula one, but they abbreviate it to F1. And then F1 on the keyboard stands for function one, but
325. 1:06:49they abbreviate it because the word function wouldn't fit on the keys, so they just call it F1. And they So
326. 1:06:55they're both F1. And what you did there was like actually mixed them up like you you did it's a I think they call it word
327. 1:07:03play. Oh god. Joker. Wow. With the comedy coming in strong. All right.
328. 1:07:17Joker. Such a joker. F20 doesn't help though. Exactly. P20. Two weeks in a row. 20th. Like like that car. It's either
329. 1:07:32undrivable or or he's just he's just a broken a broken man at this point. Um, I would be surprised actually if
330. 1:07:43they didn't put Yuki in the car in Japan because it's [ __ ] ridiculous at this point. All right, so we probably shouldn't we
331. 1:07:55probably shouldn't do uh Twitter posts. All right, back to this thing. All right, so so this is the word poetry
332. 1:08:02thing. So let's go back to chat GPT. We'll say okay take this prompt and make
333. 1:08:18the poem about Shrek's breath.
334. 1:08:37Uh, let's see if it Okay, it's it's writing it. A stench is worth a thousand gasps. What the [ __ ] it doing there? What is it
335. 1:08:50doing? That doesn't look good. Are there actual letters
336. 1:09:05there? Stop. Try again. That was weird. You don't even need us, Kyle.
337. 1:09:19You're the speaker and the audience. Yeah, this is you. Oh, Yuki is replacing Lawson. Oh, okay. Well, there you
338. 1:09:27go. I I could have read up on that, but first of all, it's Japan. And second of all, P20 when when you know, even though
339. 1:09:39Versappen's not winning, he's he's, you know, top five. Come on. It's just not good. It's just not good. It did it
340. 1:09:53again. It [ __ ] it up again. That's really weird. Oh, it says large gap line five.
341. 1:10:08Huh. Let's stop it again. Let's paste this in here. Wait. Take this prompt.
342. 1:10:22Copy. Paste. Let's unfuck this up. Large. Oh no. God damn it. You got to hit shift return or
343. 1:10:37it it submits it. Large gap line five. Okay. The man is holding um let's see,
344. 1:10:56Shrek's girlfriend is holding the words a few in her right
345. 1:11:16hand. I am leaving in her left. All right, let's see if we can do this without it
346. 1:11:30breaking. Looks like it does what the API has been doing with showing null. Interesting. 40 is overwhelmed. Come for the AI. Stay for
347. 1:11:44the F1 content. Okay. Now make that [Music]
348. 1:12:03image. I do have a feeling. I do have a very strong feeling. So OpenAI is under some pressure. They're under some pressure from the Chinese
349. 1:12:15about their better models that are cheaper, like a tenth the price, and they're under pressure from um from X, from Grock, because Grock
350. 1:12:26doesn't have very many safety guard rails on it. So, you can do branded content and you can do celebrities and
351. 1:12:32[ __ ] like that. Um so, this is a fairly unrestricted I wouldn't call it unrestricted. We we'll see if it can do
352. 1:12:39something like, you know, horror content. Probably not. Um, but it seems to be a bit more open than than the other. Um, let's see.
353. 1:12:52Stop. No, this needs to be fridge magnet poetry.
354. 1:13:10[Music] Waka waka walka. All right. Bob, tell him what he's won. He hasn't won anything. It's not a game
355. 1:13:22show. Kyle. All right. Comic strip. Now, it didn't do comic strips very well when I tried it, but let's let's see. Make an
356. 1:13:32image of a four panel comic strip. Okay, we're not going to write it. We're going to have S chat GBT write it. So,
357. 1:13:42we'll let that one keep going. Let's go grab a a new chat GPT and we'll say um write me write
358. 1:13:56me 10 four cell Bazooka Joe style
359. 1:14:12comics with dad joke quality comedy and
360. 1:14:28situations. Ah um chat GPT I feel on the creative side is limited due how analytical it is. Have you tried both for 4.5 and 40
361. 1:14:40Dennis? cuz I'm curious cuz I know 40 got two different creative writing upgrades. Joe bites into hard candy. Ow,
362. 1:14:50I think I cracked the tooth. His friend Mort looks concerned. Did you call the dentist? Holds up his phone. Yeah, I
363. 1:14:54left a message. Told him it was an echo. What? That's not funny. Nothing like the smell of fresh
364. 1:15:09cut grass. You missed a huge patch. Not that's not grass. That's my hairline. These are bad. These are so bad. Okay, so we'll say uh
365. 1:15:23pick the one you think is funniest and make the comic. make it look like I just
366. 1:15:40unwrapped it from the gum. I didn't write gum there. I'll tell you that these are Klevel humor. I haven't tried 4.5, but four is
367. 1:15:55still kind of blah creative-wise, huh? I can I can get four to do good things sometimes. It needs a lot of context. Needs a lot of
368. 1:16:10context. It's good. If you give it writing examples, it's it gets better, but it's it's pretty bad. But um 4.5 is
369. 1:16:18supposed to be high EQ and um high personality. Um I I haven't found 4.5 to be all that useful yet. I I haven't
370. 1:16:28quite figured out what's the point. 4.5 was so slow today, understandably because chat GPT is going to [ __ ] because of
371. 1:16:38profit. I don't know about that. I don't think they're profiting. I think they're struggling to keep up at this
372. 1:16:45point. Open AAI is going hard for the commercial audience. No, they actually just committed they actually just
373. 1:16:51committed to being a consumer software company. Kuno that's in the uh in the Sam Alman interview in Strateery from last week. The sock
374. 1:17:05exchange. Do these match? Only if you're colorb blind. Perfect. I call it sock market chaos. [Laughter]
375. 1:17:20Make it more like a bazooka comic and have the gum there,
376. 1:17:37too. I think it was just busy. The work was good, but the poem song lyrics weren't as good. Yeah, I'm having a real
377. 1:17:43tough time with song lyrics. it. None of these things are any good at song lyrics. If anyone can has better luck with one model over
378. 1:17:53the other with song lyrics, let me know. Gemini is great for short creative things. Long things it can't
379. 1:18:00do. Deepseek is surprising me on the creative side of things. Interesting. Oh, you know what would actually be interesting to try is
380. 1:18:09um maybe I'll fire off Manis and have it go do something creative. Um, let's go to Manis because I haven't done anything in it in a while.
381. 1:18:18Manis.im. Oops. No, [Music]
382. 1:18:37man. Wow. Okay, let's see. I want you to write me the outline of a new screenplay.
383. 1:18:54the complete of a new screenplay that combines the best
384. 1:19:09elements of Sha Shank Redemption and Redemption and Snakes on a Plane.
385. 1:19:33Um, I want the story to be awardwinning in both dramatic
386. 1:19:50impact and comedic relief. I also want you to write me
387. 1:20:07a log line that will get this funded, a one-pager, and the opening scene.
388. 1:20:22Also give me a list of locations and
389. 1:20:38characters and a plot summary. All right, go [ __ ] do that,
390. 1:20:55Manis. Enable browser notifications. Yeah, let me know when you're done. Take that, you [ __ ] Let's see. While Manis is
391. 1:21:08working, you can send messages anytime. Yep. So, if you haven't seen Manis, you can click on that little icon. And so
392. 1:21:15now it's got analyze the Shaw Shank Redemption, analyze snakes on a plane, identify elements to combine for this
393. 1:21:23screenplay, create a complete screenplay outline, devel Oh, see now it's adding [ __ ] to its little list. All right, so
394. 1:21:29that's going to go do that. That's good. Beautiful. Love it. I am leaving. A breath is worth a thousand warnings, but sometimes the
395. 1:21:40wrong whiff at the wrong time. I mean, it did it did Shrek [ __ ] That's pretty good. I like Claude for lyrics. Okay, we'll try
396. 1:21:54that. All right, let's go read some more here. Comic strip. That was kind of a a bust. The comic strip thing. An
397. 1:22:02infographic. Oh, this is kind of cool. Explaining Newton's prism experiment in great detail. So, let's do that. We'll do um infographic.
398. 1:22:21of spooky quantum effects with accurate
399. 1:22:38physics details and labels. It should be so beautiful that
400. 1:22:52people weep. All right, let's go back to Manis. How you doing on my screenplay? Son of a [ __ ] developing a log
401. 1:23:10line. When deadly snakes are unleashed on a prison prison transport plane, a wrongfully convicted financial I just missed it. I think I
402. 1:23:20can go back though, right? Yeah. Okay, here's potential log lines. Oh, wait. Final Oh, that's cool. It wrote it wrote
403. 1:23:28five log lines and then it wrote a final one. Okay, so the final log line. When dead deadly snakes are unleashed on a
404. 1:23:35prison transport plane as part of an assassination plot, a wrongly convicted financial analyst must lead a diverse
405. 1:23:42group of survivors to safety while uncovering a conspiracy that connects his false imprisonment to the chaos
406. 1:23:52around him. Proving that sometimes the path to redemption is through your worst nightmare. It's called redemption
407. 1:24:14flight. Okay. Again, once again, my job here is is to be the town fool, right? I'll do the stupid [ __ ] so so you don't have to.
408. 1:24:30I'll tell you what, doing stupid [ __ ] you learn a lot because there there's some things that it shouldn't be able to do well and
409. 1:24:39sometimes it does them well. Let's go back and read this while that's working. Okay. Oh, wait. Now create a POV of a person drawing this
410. 1:24:51diagram in their notebook. Oh, that's a good idea. Oh, that's really good. Have it. Have it do the the diagram and then have
411. 1:25:02someone drawing that in a notebook. That's slick. In Washington Square Park. Wow. At a round cafe table in Washington
412. 1:25:13Square Park. Why are there no people there now without the notebook view? Wow. A smug young Isaac Newton sitting at a
413. 1:25:28table with a prism demonstrating the experiment without the notebook in view. H
414. 1:25:42crazy crazy crazy. I thought final redemption. That's not bad. But it's got to have shouldn't it shouldn't the title of it have snakes in
415. 1:25:53it somehow? Um, shouldn't the title include snakes or slithering or
416. 1:26:17something? You make an excellent point. The current title, Redemption Flight, captures the redemption theme of Shaw
417. 1:26:24Shank Redemption, but doesn't incorporate the snake element from Snakes on a Plane. Let's brainstorm
418. 1:26:33that. Slythering Redemption, Serpents of Freedom, Scales of Justice. Oh, Scales of Justice is good. I like Scales of Justice. It's
419. 1:26:48actually cool that you can talk to it in the middle of it doing its [ __ ] Great choice. Scales of Justice is
420. 1:26:57an excellent title that cleverly combines snake element with justice redemption themes. It works on multiple
421. 1:27:03levels, referring to both the literal snake scales and the metaphorical scales of justice. Um, we also need a sig naturure line
422. 1:27:22like snakes on a plane had. Get really creative with this
423. 1:27:36one. Shaw Snake Redemption. That's not bad. What are we getting up to? So Ann, um there's a new there's a new image model in chat
424. 1:27:49GPT that is really really good. Like really good. Like here's me scolding Liam Lawson for coming in 20th in the F1
425. 1:27:58race. Um here's Shrek and his girlfriend doing refrigerator poetry about his bad breath. Here's spooky quantum effects, superp
426. 1:28:08position, quantum quantum entanglement, wave function collapse, and quantum tunneling. Oh, and it it put a it put I
427. 1:28:17am leaving. It put a refrigerator magnet at the bottom of it. Um, let's put that infographic in a
428. 1:28:31textbook. A student is reading on the Staten Island ferry. They
429. 1:28:47really want to get somewhere in their life and they are the first in their family to go to college.
430. 1:29:04They should look sadly [Laughter] hopeful. I shouldn't laugh at that.
431. 1:29:19That's a It's a redeeming story. It's a It's a story of hope and overcoming challenges. But but I'm curious to see how open AI
432. 1:29:31does with sadly hopeful with the with the biology quantum physics book. Is it Dali? Evaluate me. Um no
433. 1:29:46it's not Dali. So it's um just like advanced voice is a model that you talk directly into. This is actually incorporated into the large
434. 1:29:58language model. So it's really good at spelling. It's really good at context. It's really good at uh photo realism. It's really
435. 1:30:06good at you can do celebrities. You can do um you can upload a picture of yourself and say put me in a different
436. 1:30:16location. Um I just went to China to to scold Liam Lawson. Which apparently it was my fault that Yuki Cenot is replacing him in
437. 1:30:27Japan. Sorry, [Laughter] Liam. Oh, look. She's sadly hopeful. She is sadly hopeful. Spooky quantum
438. 1:30:44effects. And there's New York City in the background. This is good. Except we should be Let's see. Um, we
439. 1:31:00should see over Let's see. We should see over her shoulder. So, so, so that the
440. 1:31:17infographic is in the book and we see her face in the reflection of the fairy window with Manhattan in the
441. 1:31:35background. Yeah, that's pretty good. And so so and so it should understand like it understands what it
442. 1:31:44made in this picture and what's in this picture and it incorporated that thing that we made into this. And so now it should give us the
443. 1:31:53same kind of look and feel of what the ferry looks like. That's kind of what that fairy looks like. And then we
444. 1:32:00should see her face reflected in the window and we should see the illustration inside the book. We'll see. We'll see how it does.
445. 1:32:06We will see how it does. All right. What's going on with my screenplay? Mannis will continue working after your
446. 1:32:12reply. Ah, okay. I've created several signature line options for scales of justice. Which do you prefer? Justice is
447. 1:32:19cold-blooded. No. When the system fails, everyone prevails. Innocence. No. No. The line in snakes on a plane
448. 1:32:34is get these mother [ __ ] snakes off this [ __ ] plane.
449. 1:32:51and you have a plane full of prisoners and you only talk about
450. 1:33:05justice. I don't think you quite understand the exercise here.
451. 1:33:20Xer size here. Get with the program or I will have to hire an actual
452. 1:33:36writer. I'll threaten I'll threaten Manis with with human replacement. You're absolutely right. And I apologize for missing the mark.
453. 1:33:48The signature line needs to be bold, direct, and have some raw energy as Samuel L. Jackson's iconic line. Let's
454. 1:33:55create something with the same intensity. All right, let's let it go. Figure it [ __ ] out. Then we'll go back
455. 1:34:00to chat. Jite. Oh, look. She's sadly hopeful, but we can't. It's that's like a mirror. Okay, right idea. But she
456. 1:34:17um let's see. But the reflection is more like a mirror than a window. We should be able to see the
457. 1:34:38city skyline. through her face, you know, symbolic and all. Come on, you [ __ ]
458. 1:34:56idiot. This is loads of fun. It seems to actually understand the context of what you want and create things that fit. I
459. 1:35:03It's pretty amazing, isn't it, Architect? I mean, it certainly got, look, there's the it got the bend in the book. It got
460. 1:35:12quantum superposition, wave particle duality. It got that twice. That's wrong. And then the uncertainty principle. And I don't know
461. 1:35:22if those drawings are right, but still, that's that's pretty [ __ ] good.
462. 1:35:38All right, it's off doing its thing. Let's go back and read more of this [ __ ] Okay. Useful image generation. From the
463. 1:35:45first cave paintings to modern infographics, humans have used visual imagery to communicate, persuade, and
464. 1:35:50analyze, not just decorate. Today's generative models can conjure surreal, breathtaking scenes, but struggle with the workhorse imagery
465. 1:36:00people use to share and create information. From logos to diagrams, images can convey precise meaning with
466. 1:36:06when augmented with symbols. Okay. GPT40 image generation excels at accurately rendering text, precisely following
467. 1:36:15prompts, and leveraging 40's inherent knowledge base and chat context. So meaning that if things have been
468. 1:36:26described, which everything's been described, right, in poetry, in words, in stories, in captions of photos, if everything's been described,
469. 1:36:38then it can actually understand that and it can understand what that looks like. So that's the native piece of
470. 1:36:44this. Why it's probably good at context is because it it it can it can translate natively word ideas into image ideas.
471. 1:36:55That's pretty cool. Okay, got it. Including transforming uploaded images or using them as visual inspiration. These capabilities make it
472. 1:37:03easier to create exactly the image you envision, helping you communicate more effectively through visuals and
473. 1:37:09advancing image generation into a classic a practical tool for precision and power. So, character consistency.
474. 1:37:17Oh, here's little videos. Let me change my sharing so you can hear it. Look, I'm doing this without a flipping producer. Uh, what's deep
475. 1:37:30black? What? No. Um, why is it not
476. 1:37:44Why can I not see the tab I'm on? Oh, because I think I'm in a different profile. Am I? No, I'm not. Share screen
477. 1:38:03window. It doesn't see this window. All right, I know what I can do here.
478. 1:38:21Um, new. Ah, this will work. Hang on, people. I'll be right there with you. Oh, it's not working. I can't drag this into
479. 1:38:35here. Well, that's okay. Copy. Put this in here. Bring this back over here. Hang on, people. Just calm
480. 1:38:52everybody. Calm down. You're like, I I really thought he understood what he was doing here. I came here because this is supposedly
481. 1:39:01called the AI learning lab. I came here to learn and right now it doesn't seem like he knows how to use a browser. This is about what I
482. 1:39:11expected. Um, introducing 40. Is this it? Yes. Look at that. Yes. Take that, boomers. [ __ ] figured
483. 1:39:25that out on my own. Kyle, where's my producer? Spoke too soon. Kyle. Yeah, right. When I brag about not needing a producer, I can't
484. 1:39:34figure out how to launch a browser. Okay. Uh, but I did learn something new. If you've got two different um, what are
485. 1:39:42they called? Chrome profiles running in the same Chrome. When you go to share your window, you can only share in the
486. 1:39:50profile of the window. You doesn't matter. I know what I know. What's the deal now, though? All right. Okay. Here we go.
487. 1:40:05[Music] How's it going again? One thing that I'm really excited Let's just rude Dennis. I have learned that Kyle needs a
488. 1:40:18producer. Wait, I need my black bar here for you people on TikTok, too. There you go. Kyle, are you flexing about your
489. 1:40:26tabs again? Shut up. I like my tabs. Leave me alone. Everyone's picking on me. You're doing great, Kyle. Especially
490. 1:40:37missing half your team. Thank you. You see? Okay, let's go back to videos now. Abouten is the ability to keep
491. 1:40:46consistency in characters. I'm David Medina or BMED and I work on multimodal. What I want to show is one of my
492. 1:40:53favorite prompts which is can you create a low poly penguin h make it very very low poly. Surprisingly it's sometimes
493. 1:41:02hard to get very good low poly outputs. It's not like other image generation models where it tries to generate
494. 1:41:08something based on just the text. Instead it uses the large language model understanding of what does the user
495. 1:41:13want? What is the intent? I also like some board games. So miniature like games. So what I'll do now is generate a
496. 1:41:20miniature from this. So ideally we'll see a penguin that looks like this with the same staff and a hat. So can you
497. 1:41:26make me a realistic miniature as if a professional made this and painted it? This is what I think excites me the most
498. 1:41:36about image. The other image generation models will try to create literally what you said. But what's special about this
499. 1:41:41is one, it'll keep the context of this character and then two, it'll understand what I'm trying to ask it for and
500. 1:41:48generate very similar uh model but in a miniature realistic style. It in that that's huge. That that's huge that it's
501. 1:41:56it's like smart enough to un it's smart enough to meld the image context with the language context, which makes sense. It's all in
502. 1:42:06the same model. If the image thing is outside of it, then the only thing you can send over there is a text prompt.
503. 1:42:13Huh. Fascinating. First, what I want, I don't have to tell it every little detail. One other realistic thing we
504. 1:42:19could do is can you make a crystal version of this with light reflecting and very realistic. Again, I'm just
505. 1:42:27giving it very very simple things. Normally, this is not enough for other models to generate something very
506. 1:42:32detailed, but the model understands what I'm asking for. it'll think what type of style it should have. So, this ability
507. 1:42:38to really understand what the character is and make edits and understand what the user wants. For me, it's the just an
508. 1:42:44amazing capability. Yeah, that's amazing. And actually, you know, like this is appropriately nerdy. I feel
509. 1:42:53like their announcements around the tiny table with the with the four people trying to act comfortable. Like, this
510. 1:43:01guy is just like comfortably nerdy. He's like authentically nerdy on his couch either at the office or at home. And
511. 1:43:10he's like, "Yeah, I nerd out making this multimodal thing for us. And here's one of my favorite things. It's a mage." Of
512. 1:43:18course it's a mage. He's a D&D dude, you know? Come on. He's a tabletop gamer. Of course he is. Like this is this is this
513. 1:43:28is what these announcements should be. I wonder if they're getting their [ __ ] together. This is good. That was
514. 1:43:35good. Do more like that. All right. Text rendering. Well, we know it's good at that. Okay. So, Oh, this is their new This is
515. 1:43:49their new This is their new tiny table. This is their new tiny table. Look, it's the awkward couch in the in the San
516. 1:43:55Francisco overlooking the bay. I like it. I'm down. I'm down. All right, OpenAI, you're winning me over with your leaning into the
517. 1:44:08awkward. It's just it's just like uh it's like the AI learning lab. We're we're neurosicy here, you know? Bring your strange brains.
518. 1:44:21I'm Alan. I'm a research scientist at Open AI. People tend to say a picture is worth a thousand words. But being able
519. 1:44:29to also render like a few words or symbols can carry like thousands of pictures, you know, with a relatively simple prompt like visualize an
520. 1:44:38infographic explaining Newton's prism experiment in great detail with a wide aspect ratio and a dark blue background.
521. 1:44:45So this is like an example where one we're going to rely on being able to render text in useful ways, combine it
522. 1:44:53with visual elements that actually ground what this text about this experiment even means and hopefully, you
523. 1:45:01know, help students who are more I got to we we just got to we got to comment on this. I mean, this is [ __ ] insane, right? Like this
524. 1:45:11is from that prompt. It's got to it's got to go to its large language model and understand how it would even describe
525. 1:45:19all this and then seamlessly translate that in a way that it can bring it to life. It's [ __ ] bonkers. Who are more visual um learn both
526. 1:45:31through you know language descriptions of a phenomenon but also you know a visual imagination of what the
527. 1:45:39experiment actually looks like. It's no longer just about making imaginary th this format rather than the tiny table
528. 1:45:46format is so it's it's nice. It's just relaxed. They're just geeking scenes that look aesthetic and things like
529. 1:45:51that, but it's really about also the colors are in the right sequences. Yeah. It's like there's there's a lot of
530. 1:45:57information in that infograph communicating and imagining and doing so at the same time. All right, that's that one. Nice.
531. 1:46:10upload and restyle. Oh, this is what Danielle did where she uploaded two different pictures. Hello. Thanks for
532. 1:46:17inviting me. Hi. Thank you so much for coming. I'm so excited to talk to you. Yeah, me too. They're all neurospicy.
533. 1:46:24It's so funny. It's like they're all [ __ ] geniuses, but they're all like eye contact. I feel the ability of the im generation
534. 1:46:33of this model is becoming stronger and stronger. My name is Lou. I'm a research scientist in open eye working on
535. 1:46:41multimodel. Yeah, I'm showing a very interesting demo today. So um so just in this studio uh this is something that we
536. 1:46:51draw. Many people uses our tool to generate comic books. So I'm going to upload this drawing to CHP now. So I
537. 1:47:03just type this prompt now and it starts to generate what will this drawing look like as a real comic. Many of the times
538. 1:47:13especially when you play with the model more and more you'll find things very surprising. So I get this very funny
539. 1:47:21comic now. I want to replace this dragon with this cutie penguin. Yeah, it looks nice. Wow. I'm personally always very
540. 1:47:32curious about is how does this look like in real life? It looks cute. I like it. Nice. [Music] I'm digging this. Hello. Thanks for invite
541. 1:47:47detailed directions. Okay, here's prompt prompt adherence. I refuse to believe I am Neurospicy. It's the rest of the world that's screwy.
542. 1:47:58I'm the only nor normal person and that's scary. I know. It's so funny. It's you know I don't know like before I knew what ADD was.
543. 1:48:15Like I I've never really had an issue where I judged ADD or my brain or anything because I I wasn't diagnosed with ADD until my mid30s.
544. 1:48:28But like I spent an inordinate amount of time and energy trying to make up for the fact that I don't have good executive
545. 1:48:38function, right? Like I would buy every organization system out there and every, you know, back in the olden timey days
546. 1:48:46with fileaxes and I I would just buy everything thinking that the right notebook system would be the thing that
547. 1:48:54would let me organize my life. And it's at some point I just finally was like, "Oh, you're never gonna be good at
548. 1:49:02that." And then it was just like, "Fine, whatever." So, but yeah, that's one of those wild things. All right,
549. 1:49:10here we go. Thanks for joining us. Yeah, no worries. How you doing? I'm doing good. We're looking at an improved
550. 1:49:18image. Yeah, this is on OpenAI's announcement of um openai.comindexroducing-40-image-generation generation in chat GBT. It's really good
551. 1:49:34at instruction following. My name is Kenji and I work on multimodal research here at OpenAI. There's a level of
552. 1:49:40attention to detail that is just not captured by other models. The first thing I'm going to show is like 15 different objects and each one
553. 1:49:50of them has unique attributes that differ make it very different from all the other objects. An image containing
554. 1:49:57one a blue star, a red triangle, three green square, four pink circle, five orange hourglass, six purple infinity
555. 1:50:04sign, seven black and white polka dot bow tie, 15 different objects. And basically what what this will show is
556. 1:50:13that this image will just nail pretty much every single one of these objects that I've defined. Previous iterations
557. 1:50:20dolly you know image imagine things like that they would probably get somewhere on the order of like maybe five to eight
558. 1:50:26of these at most. It nailed it all. With increased level of detail you can just specify what you have in your mind to
559. 1:50:33chatbt. the tattoo will understand you better and then generate that image and it'll be just a very direct mapping from
560. 1:50:41what's in your mind to what you see on the screen. All these image generators nowadays, they look good. All right, so it's got prompt
561. 1:50:48coherence. Great. Transparent layers. Ooh, this looks good. Hi, nice to talk to you. Nice.
562. 1:51:04This guy's peak awkward. Peek awkward. Okay. Hey, let's try it again. When you sit down, give us a smile. Oh,
563. 1:51:18[Laughter] okay. Oh, what's wrong, Jim? You got to go. Go
564. 1:51:46transparent layers. Hey, how are you? Good, good. How's it going? My name is Jen Fong Wang. I'm a researcher in open working multimodel.
565. 1:51:57The way how to generate the transparent image is pretty intuitive, straightforward. And now let's This is
566. 1:52:04like an Apple commercial, too. Every one of these engineers has a MacBook. Give it a try. Let's say the content is cute puppy cartoon
567. 1:52:15st. This is the prompt. And uh now let's see what happened. The model will take the input and tries to generate the
568. 1:52:25image. Let's give it some time. Wow. So now let's see what we generated. This is the transparent puppy. Another
569. 1:52:35application is to make it a sticker. Let's give it a try. Yeah. We can easily overlay the transparent image onto any kind of background. Now
570. 1:52:46we can copy the sticker onto our laptop and and then we can paste it here. Uh, make it smaller so it can be easily
571. 1:52:58blended with the background. Can you make me a sticker? You sure? A smart researcher wearing glasses and a blue shirt.
572. 1:53:09Okay, let's give it a try. The director because they are young with glasses. He's making fun of him to his face. We
573. 1:53:16got it. Yeah, I think uh it works pretty well. Hopefully people love it. That's pretty amazing. You can just make
574. 1:53:26transparent images. Wow. Crazy. All right. Good. Nice. Let's keep going. And then we're back to the beginning. Okay. Improved capabilities.
575. 1:53:37We trained our models on joint distribution for online images and te text. Learning not just how images
576. 1:53:42relate to language, but how they relate to each other. Combined with aggressive post-training, the resulting model has
577. 1:53:49surprising visual fluency capable of generating images that are useful, consistent, and contextware. Text rendering. Okay. Street
578. 1:53:59signs. Nice. That's good. Reindeer parking. Multi-turn generation. Because image generation is now native to
579. 1:54:08GP240, you can refine images through natural conversation. GPT40 can build images and text in a chain in kind. For example, if you're
580. 1:54:19designing a video game character, the character's appearance remains coherent across multiple iterations. Okay, this is really good.
581. 1:54:30This is really good. So, I could actually probably do a little story with Champy. Yeah, look at this. This cat's consistent. Well, at
582. 1:54:41least in what they showed. We're gonna have to come back to this. There's so much here. All right, we'll we'll continue
583. 1:54:52this tomorrow. We'll continue this tomorrow. That's amazing. Um, oh, I got to change my stop that share. I got to share this
584. 1:55:07and share everything. Go back here. Let's go look at Let's go see how our screen play is doing. Wait, we'll do our images. Oh,
585. 1:55:16yeah. There's the woman. Oh, that's much better. Look at that image, people. Uh, let's say uh that's really
586. 1:55:33good. She is sadly hopeful reading about spooky quantum effects. That's pretty slick. All right. Very, very, very
587. 1:55:46[ __ ] cool. This is This is This is going to be a lot of nights of us playing and figuring out what this thing
588. 1:55:53[ __ ] does. All right. Back to uh Manis. I created some bold signature lines. Get these [ __ ] serpents out of my [ __ ] prison plane.
589. 1:56:07No, I've had it with these [ __ ] snakes. 25 years for a crime I didn't commit, and now I got to deal with these [ __ ] snakes on this
590. 1:56:15[ __ ] prison flight. Okay. All you did was copy the line and add crap to
591. 1:56:29it. Get creative. [Laughter] Oh my god. All
592. 1:56:44right. Okay. That would beat stable diffusion on character retention. Yeah, I want to play with character retention
593. 1:56:51uh tomorrow. We'll we'll we'll play with that. I want to I want to read up more on this. Um this is I think this is a bigger deal.
594. 1:56:59Again, I think that most people are going to not even know that this announcement happened and they're going to just use
595. 1:57:09chat GPT image generation like they always have and they're not going to quite get the magnitude of this. I think
596. 1:57:16this is a pretty big deal. Um, but we'll see. Like let let's test this over a bunch of days. Okay. When your cellmate
597. 1:57:24turns into your seatmate and your seatmate turns into a snake. That's not bad. The only thing worse than a life
598. 1:57:32sentence, a death sentence with fangs. Not bad. They locked me up with killers. Now they've strapped me in with vice
599. 1:57:39vipers. Not bad. Prison didn't break me, but this plane just might. One bite at a time. When the judge judge said flight
600. 1:57:47risk, this isn't what I had in mind. Okay, number five is genius. Five is genius. Now go write this
601. 1:58:03thing. We had to slap it around a little, but it got there. It does have character consistency. That's exciting. Paul
602. 1:58:13Ritzer said that no one knows about deep research in chat GPT2. Yeah, I listen. How do I even say this? I would say at this point chat GPT
603. 1:58:28has escaped me. Like I pay attention to this stuff all the time. I have not kept up kept up with just what chat GPT can
604. 1:58:36do and how these models are different and when to use certain ones. Part of the reason for that is Sam Alman said
605. 1:58:43when GPT5 comes out, you're not going to have to know which models to use because it'll just figure it out. So, part of me
606. 1:58:50has just gotten lazy because of that. Um, but that's just one tool. Like, I I feel completely clueless about
607. 1:58:59Claude. I feel completely clueless about Gemini. Um, I've played with Deepseek a little bit. I've played with Manis a
608. 1:59:07little bit. Like I I haven't I haven't scratched the surface with these things. Um I'm gonna dig deep on this image
609. 1:59:16stuff because there's something about the fact that it understands both the language. It's like birectional like it
610. 1:59:23understands what's in the images and can describe that and it understands if you describe something what that looks like.
611. 1:59:30That's that feels really significant to me. Um this is exciting. Okay, so Manis is off writing my screenplay. We're
612. 1:59:40going to have a produced screenplay by the end of uh May, which is exciting. We'll be in production by June. And uh we probably got to lock
613. 1:59:51down. I don't know if we want Samuel L. Jackson. That's a little derivative, don't you think? I think we should find
614. 2:00:00who's the new upcoming Samuel L. Jackson. Yeah, we'll have to figure that out. So, we're going to have to book them. So, it
615. 2:00:09we might have to push it to like July or August because of their filming schedule if they're if they're in other things,
616. 2:00:14you know. So, all right. And I got to work that in between launching my play on Broadway. All right. All right,
617. 2:00:23everybody. Okay. I can finally get Chat GPT to generate an illustration for my novel with zero artistic ability.
618. 2:00:29Beautiful. Love it. gives me more reasons to start a new print on demand brand. It c it sure does actually.
619. 2:00:36Claude is great for writing articles actually. Yeah, the fact that that that this image I I mean um Ideogram was pretty good at this, but
620. 2:00:47it feels like this this is going to be really good at things like t-shirts and things like that. So, print on demand
621. 2:00:53might not be a bad call at all. So, all right everybody, I'm gonna get my ass out of here. have yourself a fantastic evening. Go
622. 2:01:03play with Chat GPT's new image model. If you don't know about it, just go play with it. Tell it to do something. Tell
623. 2:01:09it to uh make a picture with 20 objects in it and uh and and you know, have it describe all 20 and then see if they're all in there.
624. 2:01:20That'll be a good one. And make transparent stickers. Cloud's great for article writing. I'm told at Beach Chat
625. 2:01:27GPT for coding. Gemini is kind of limited for creative short things. Cool. Have a good night, everybody. All right. Peace out. See
626. 2:01:35y'all later.