8/15/2025 - Exploring the Nuances of GPT-5 and Its Various Personalities

Transcript

1. 0:04Let it be. [Music]
2. 0:44Sitting in this lonely town, wonder when things are going to change. Dream my life away. Seems these dreams have turned to bunch
3. 0:59dust glass. Get my nerve up, but my past is pulling me down. Wondering how long
4. 1:14a black she don't stick around. [Music] [Applause] Somebody told me once before said you can never go home again.
5. 1:29Won't you leave Santa things to steal me away? Yeah. From the truth of who I am and what I believe. So I thanked him for his two
6. 1:42sins with a handshake and some sympathy. Yeah. I packed up my blue jeans and headed for this big prize of my freedom.
7. 1:59Bye-bye, black sheep to the black sheep of the family. Bye-bye. [Music] Oh, it means so very much to me. Yeah.
8. 2:16Byebye. Black sheep to my friends and my family. Bye-bye. Go off and set my soul.
9. 2:33Set it free. [Music] Friday night, date night. It's Friday night. Date night. Friday night. Date night. What's happening, good people?
10. 2:46How are you all doing tonight? [Music]
11. 3:25Um, all right. So, tonight I got myself some rocks. I got myself some Blandon's flavored tequila. Now,
12. 3:35just to be clear, I hate tequila. I've never since my 21st birthday at Penn State been able to drink tequila again
13. 3:46except for this this particular Wait, where's Oh, wait. That brand. Um, let me flip this and then I can do this. I'm probably not
14. 3:57allowed to do that on TikTok. this. They're probably gonna get banned for life because I showed illicit adult enjoyment liquid. [Laughter]
15. 4:16So, anyway, happy Friday night date night. Cheers. Here's to uh Amelio's wife, Serena. May she rest in peace. She got me these glasses.
16. 4:30I miss her. There you go. Um, tonight we're going to be talking about personality. [Music]
17. 5:03[Music] Didn't mean to cause you any sorrow. [Music] Didn't mean to cause you any pain.
18. 5:19[Music] Only want one time to see you laughing. Only want to see you laughing in that purple rain. Purple rain. Purple rain.
19. 5:35[Music] Purple rain. Purple rain. Purple rain. Purple rain. [Music]
20. 5:50[Applause] [Music] only want to see you laughing in that purple ring. Um,
21. 6:04I keep saying this, but I keep feeling this. We're at a fascinating time in history where like I don't I don't know. I had to reboot
22. 6:15chat GPT because it was getting loopy. Yeah, I've heard. Okay, so a couple of things going on. They are they are
23. 6:23changing the personality of chat GPT5. They're trying to trying to make it not as sickopantic as 40. Um, but give it a little bit more of
24. 6:35that personality. Uh, I I said this when Chat GPT5 first came out. I said, "Be prepared for the next month, this thing to just keep changing
25. 6:47and getting weird and don't don't don't assume how it is is how it's going to be." Um, and I think in the past week
26. 6:58it's been I don't know an average of a change a day. [Music]
27. 7:37Well, I heard there was a secret call David played and it pleased the Lord. You don't really care for music, do you?
28. 7:54goes like this. The fourth, the fifth minor fall and a major lift for King composing. Hallelujah. Hallelujah.
29. 8:12>> Hallelujah. Hallelujah. Hallelujah. Oh yeah. [Music]
30. 8:29>> Well, your faith was strong, but you needed proof. Saw her bathing on the roof. Her beauty and the moonlight overthrew you. [Music]
31. 8:43>> She tied you to the kitchen chair. She smashed your throne and cut your hair from your lint shed r. Hallelujah. Hallelujah.
32. 8:59Hallelujah. Hallelujah. [Music] Hallelu. [Music] Oh yeah.
33. 9:13[Music] Satan wear suit and tie. Does he work at the Dairy Queen? Does he listen to rock and roll? Does he
34. 9:29feed the mean singing hallelujah? Hallelujah. Hallelujah. [Music] What about Jesus?
35. 9:44Didn't he do it too? Yeah. Hang out with the prostitutes. Have a drink, too. Power of example. My mama said it and I
36. 9:55heard. She said one ounce of action beats a ton of words. Singing hallelujah. Mama said there would be angels. Hallelujah.
37. 10:08Mama said there would be sons. [Music] What do you think, Champy? You're singing pretty good tonight, buddy. He
38. 10:25is in a singing mood. All right. Happy Friday night. I hope you all have your nachos and hot pockets heated up.
39. 10:34Your wine in a box decanted. You want to let that rest for a few minutes. Let it breathe a little. You got to get it out
40. 10:41of the out of the plastic bag. You know, you know what I'm saying? Oh, that is tasty. All right. Um All right. Let me talk a little bit
41. 10:56first. Let me see. Pup can carry a tune, right? Ex when he wants to. That's the uh that's the that's the challenge with
42. 11:03the old champion. Some nights not so much. All right. You can make money with chat GPT. See slug of doom date night.
43. 11:17So, I just saw a tweet. Let me see who who tweeted it. Right before I went live here. Nathan Lance. Who is he? Lore.com.
44. 11:38Lance Capital Early Stage Fund. All right. He's a general partner at a investment fund in Canada. All right. Um, let me share
45. 11:52my screen. Can you pop my screen up there? Yan Brandon. So, so this tweet answered some things for me. it. Well,
46. 12:08it and this isn't answering anything. This is just sort of him kind of speculating. Sounds like a trust fund, baby. Maybe.
47. 12:15All right. Um, GPT5 shows that we've crossed the point where most people can't tell if one AI model is better
48. 12:22than another. From here, models we will be judged on vibe and utility. Do I like the way it responds to me? Can it
49. 12:30actually do something useful? Does it remember me? And and what matters, you know, its personality, right?
50. 12:37um or if it's if it's something like Perplexity where it's got some additional features built onto it. This
51. 12:43feels right to me because for the past two weeks, ever since CH Chat GPT's come out, there's something oddly unsatisfying.
52. 12:59And you could certainly argue, well, they didn't blow it out of the water, so why it's unsatisfying is because, you
53. 13:05know, it's not as good a model. It's it's not as big a jump as three to four was. Um, that may or may not be true.
54. 13:16It also very well may be the case that the things that I ask large language models, I kind of tapped out at 40 or or a
55. 13:26little bit with 03, right? the the previous thinking model. Um, and so so that's kind of that's kind of
56. 13:36where I feel we are right now. And so I don't know, I don't know if it's a good thing or a bad thing. There there's another piece
57. 13:45of me, one of the things that I have been preaching is adaptability. that adaptability is the big the big attribute
58. 14:00over the next three to five years that's going to allow you to be successful. And one of the things that struck me today,
59. 14:13you can be adaptable. So, you got really good at chat GPT4. You built a bunch of [ __ ] on it. New model comes out, you adapt to it, right?
60. 14:25There's there's a there's a loss there, right? There's there's mourning. There's mourning of the me that for the past year and a half has
61. 14:41been playing primarily with chat GPT40 on this channel. And I've gotten really good. I've got my little shortcuts. I
62. 14:49kind of know what works, what doesn't. I know when to switch models, when not to. And I kind of feel like I've got this
63. 14:56new thing. And like like OpenAI can't even figure out what they want the interface to be. It keeps changing every
64. 15:01other day. But I like I found myself in the past day or two feeling like like why am I sad? What's going on here?
65. 15:10I think part of adaptability is um some of this shit's going to suck, right? Like it's going to suck
66. 15:25when your job that you've done for 20 years becomes something else. And even if you're adaptable and you can
67. 15:35adapt to it and you end up getting, you know, retained, they're like, "Oh, that's one of the good ones. They
68. 15:40adapted well to the new way it is." There's 20 years of how you did [ __ ] that's just going to vaporize just, right?
69. 15:54And then, you know, and then obviously if you get laid off, then it's even worse. It compounds it. So I I never in my life before have I
70. 16:09have I felt pangs of sadness for features going away or software going away. I'm like, "Ah, that was a good one." But I it's never
71. 16:19felt like visceral in any way. And and all of a sudden it does to me. So that's new. And it's not sad like I'm like super
72. 16:35depressed and bummed out and I'm like I hate software and I hate AI. It's just like ah got to start over. Got to start over. Danielle,
73. 16:48the real AI pro the real AI pros like and appreciate GPT5. Oh, Daniel, not Daniel. Um
74. 17:03um I don't know what you mean by real AI pros. If you mean scientists and engineers and like pros like people that are building
75. 17:14the technology or do you mean pros like people that are building you know chat GPT into workflows at work or using it
76. 17:23to run their startup business because I would consider them both pros but if what you're saying is do the engineers
77. 17:33and mathematicians like chat GP PT5 better probably because they've got hard mathematical and scientific problems to solve.
78. 17:47The rest of us don't necessarily. And we're going to have to figure out how we take non STEM essentially all of the world
79. 17:58that is non- STEM and non-engineering and how do we use it and how do we use it in the best way and I don't know that
80. 18:06yet and I don't think anyone does yet and I would I would even argue that people that are sort of high science you
81. 18:12know polymaths are are still probably trying to figure out is this thing actually good? Does it actually work?
82. 18:18Tik Tok pinned five stages of AI grief. Yeah, like the the five stages of AI adoption that I've got are kind of on
83. 18:26the on the positive of this, but I think I think it is five stages of AI grief.
84. 18:42[Music] Yep, I see. Um,
85. 19:02let's see. So, a thing that I think might be fun tonight is Well, there's a couple of things. One is um Gemini came out with Imagine 4.
86. 19:18Well, it's not at gemini.google.com. It's on their It's in their developer playground, but there's a new image
87. 19:25model out from Google that we could go play with. Um that could be kind of fun. Um we could just totally punt and just do some fun imaging
88. 19:40video [ __ ] over in Mid Journey. Um, but I thought I thought what might be nice is I think two nights ago I showed you all
89. 19:52I think someone in here told me about it. I think it was Brandon, producer Brandon, that within chat GPT they've added these personalities
90. 20:01and you can have it be a cynic and you can have it be a thinker and you can have it be a I don't know whatever a a
91. 20:07lover what whatever they call them. And so those are basically just little probably pre-prompted short little
92. 20:15pieces of context that when you choose those different personalities, it just, you know, throws something into your custom instructions.
93. 20:22But I thought might what might be interesting is take a similar kind of prompt and, you know, throw it into one
94. 20:29personality mode and use chat GPT5, thinking mini thinking and GPT40. like look at the look at the models with the
95. 20:38different personalities and just see see how different it is. Um I think that kind of experimentation is
96. 20:46the only way we're going to really get our head around. Oh, I get it. This particular model with this particular personality that kind of
97. 20:55gives me something that yeah, that's something I would use a lot, right? I don't know that there's another way to
98. 21:01really do that. I also don't know are those personalities at all useful. They may or may not be, you know. Um, Mr. IT is late.
99. 21:22[Music] So, let's go do that. Let's go chat JPI Tai. [Music] New chat
100. 21:40beauty. All right. Oh, wrong tab. There we go. All right.
101. 22:02Okay, so I've showed this before, but if you're new here in Chat GPT now, we've got three different models. If you're paying
102. 22:1420 bucks a month, I think even on the free version, you've got three different models that you've got access to that are GPT5.
103. 22:24You've got fast, thinking, mini, and thinking. Um, or you can think of this, if you if you knew what the old numbers meant, you
104. 22:32can think of this as like 50, five Omni, 05 mini, and then 05 for thinking. And then 05 Pro is the one if you pay 200
105. 22:42bucks a month. And then we also have access to some of the old models. I'm going to ignore all of them except GPT40.
106. 22:50Um, and then the auto mode. All all auto is doing is it's deciding for you which of these three it's switching between.
107. 22:57It's it's not going to any of these legacy models. It's just doing these three or four if you've got that one
108. 23:03enabled. So So those are the different flavors that we're going to play with. So we'll do 40, we'll do what's
109. 23:11effectively 50, so fast. We'll do 05 mini and then 05. Right? That's those are the three that we'll play with. Plus plus 40.
110. 23:21And then to change your personalities, if you go to customize chat GPT where your custom instructions are, they have now added.
111. 23:33What personality should chat GPT have default cynic robot listener nerd. So now we got to come up with something that we uh
112. 23:49we got to come up with something that we are going to ask across all of these things. So I think we'll start let's see
113. 23:58default is cheerful and adaptive. You know what one of the things we can test is you have unsaved changes back. Oh, safe. Um,
114. 24:25um, from Ryan. Earlier today, I made it think for 20 seconds using auto mode. When it did that, was it switching to
115. 24:36the deeper thinking model? Yes. So, so what happens in auto is so, so fast, and it was funny when I did fast last night, it hung for like 15
116. 24:48seconds. But, but if I do something like um explain um uh uh what's it called? Um osmosis to me.
117. 25:03Fast is pretty fast, right? It just sort of it blasts it out. Now, it's not it's not ripping fast, but it's it's decently
118. 25:10fast. But if I flip it into thinking mini and do a new chat and say, "Explain osmosis to me." It's going to think for a little bit
119. 25:28and now give me an answer. Maybe a little bit more thought out. Nice. And then if I go to thinking new chat explain osmosis to me
120. 25:47thinking. So the other one was maybe five seconds. Explaining osmosis. So similar thought for eight seconds.
121. 25:58The other one was maybe five. And these are probably since it's such a broad answer, right? The these are It was actually interesting. The
122. 26:07thinking mini gave it a longer answer than the the long thinking did. But yeah, when you're in auto mode, chat GPT
123. 26:14is deciding which of those to use, especially if it says the words thought for 8 seconds. Um, that generally means
124. 26:21it's switched. Now, there was some weird behavior with GPT40 right before they switched to GPT5. Some people were reporting that GPT40
125. 26:34was doing some sort some thinking. So I don't I don't know details of how these models are actually doing what they're
126. 26:41doing. The one thing I have heard about GPT5 is that this fast mode is actually kind of crappy. That the non-thinking version
127. 26:52is kind of crappy. And I've heard more people saying they're just parking it in thinking mode and just using chat gvt in
128. 26:59thinking mode. Again, these are all things that are worth playing with, right? That the idea within the AI salon
129. 27:05play first. Play with these different modes uh to to understand like what are these things actually doing different
130. 27:13for you, right? It's not just what's it doing technically. It's like what's better for you? Tik Tok pin. My daughter
131. 27:21graduated from Longart College. She got a bachelor's degree in nursing. Congratulations to her, Mr. IT. That's
132. 27:31awesome. That's awesome. My belt is a little tight. All right. All right. We're good now. We're good now. Um Okay. What are we going to do?
133. 27:58Austin Scout, ever since they upped the limit to 3,000 a week, I park it in thinking. Oh, yeah. That's the other
134. 28:05thing. Thank you for that reminder, Austin. Um, Altman today said that for the 20 bucks a month, you get 3,000
135. 28:14um, prompts a month in in thinking mode. Um, which is that's that's that's pretty lot. That's a hundred a day, right? So,
136. 28:23a 100 prompts a day in thinking mode. Um, if you're doing anything decently worthwhile, it might be good to just
137. 28:32park it there and just just see what it's like. My only problem with leaving things in thinking mode is sometimes I
138. 28:38just need a quick answer on something and it and the minute I hit return, I'm like, "Ah, I should have switched out of
139. 28:44thinking mode." Because then it's like and it just gives me a stupid answer. Um, so I find myself flipping back and forth
140. 28:54between auto and thinking because I'm like, I I can't be trusted to flip those models on my own. Um, okay. Um,
141. 29:05all right. Here's what we're gonna do. I'm gonna I'm gonna go I'm gonna cycle through 4050 mini and 05 with two prompts each. So, the first
142. 29:21prompt is going to be give me 10 ideas for a compelling
143. 29:44short story about
144. 30:01This is good about hiding from technological advancement. All right. So, let me I'm going to copy this.
145. 30:16Let me throw it into my notes program. All right, I'll put that off to the side.
146. 30:30So, we'll come back to that. Okay. So, just to confirm, we're in Oh, yeah. Jingle AI. We'll play We'll play with that,
147. 30:41Brandon. That's a That's a funny joke, actually. Okay. So, we're we're in default personality, which is cheerful and adaptive.
148. 30:52So, now we're going to do give me 10 compelling short story ideas.
149. 31:09Okay. about hiding from technological advancement. Okay, so we hit it. Here are 10 things. The analog refuge.
150. 31:22So, so it's numbering them, formatting them, putting a line between them. 404 humanity not found. In a future where
151. 31:30every citizen is assigned an AI companion at birth, a teenage girl fakes her own digital death to live unplugged.
152. 31:36That's actually really good. The air airgapped choir. The garden protocol off the map.
153. 31:51The blue room. A society where all walls are smart surfaces. A mother paints one room in her house with analog blue
154. 31:59paint. Her children think it's just quirky until the walls start listening harder. That's good. Okay.
155. 32:12All right. Then what's it say at the end here? Would you like one of these expanded into a one paragraph synopsis
156. 32:18or a full outline? Let's do a full outline of 404 humanity not found. Um, give me full outline of
157. 32:34uh 404. Speculative fiction, coming of age, dystopian, thriller, theme. You can't
158. 32:48truly hide from something that knows you better than you know yourself. Setup, world building. Okay, so four bullets. The protagonist,
159. 32:59the inciting in incident, escape, decision point, core's reaction, confrontation, core evolves, the climax, the choice.
160. 33:09Okay, so it's given us a bulleted bulleted sort of list. Okay, cool. So, that was 40. I I find 40 to be a really good writer,
161. 33:21but let's do new chat. Flip this to GPT5 fast. So, this should be the next model up from 40. This is effectively 50.
162. 33:34And we'll throw in the same prompt. We'll see what we get back here. 10 compelling story ideas. Doesn't format them quite the same. A
163. 33:48group of rebels discovers the last valley where no satellites, drones, or signals can penetrate. Becomes a sanctuary.
164. 33:58family hides. The analog heirloom hides a typewriter as if it were a sacred relic, passing it down through
165. 34:03generations while the world mandates neural implants. One day, the youngest secretly types a story that exposes their hiding place. Um,
166. 34:16I find these a little more boring, a little more generic. Let's do we'll do the typewriter one. Um, analog heirloom. Uh, outline. Analog
167. 34:32heirloom. Premise. Outline. One, world setup. Two, characters. Three, conflict, rising tension, climax.
168. 34:46Yeah, this is just this is thin. GPT5 fast is thin with with with this particular prompt right it might be
169. 34:56better at other stuff but like given those two 40 was a much better writer now let's go to thinking mini so this is
170. 35:05now like05 mini not 505 right so the thinking model the mini version of it
171. 35:20thinking get A quick answer thought for a couple of seconds. 10 short story ideas. Nice prompt. Here are just 10 distinct
172. 35:31hooks you can pick, stretch, or mash together. So, this is already got more personality. 50. 5 the the fast model just feels like a
173. 35:42barebones. It's it it just feels not good. This feels better. A formal Google engineer runs a secret archive of physical
174. 35:52memories, printed photos, handwritten letters, cassette recordings in a cellar protected by ancient analog locks. When
175. 35:59a city-wide push to digitize everything sends a municipal drone crew to inventory private artifacts, the
176. 36:06archivist must outwit surveillance with low tech cunning. The Faraday Farm, Analog Child. These are really good.
177. 36:18Last library, a future where books are outlawed, hiding in plain code, The Analog Detective, Shelter of Static, The Hider, and the Helper.
178. 36:30I like the I like the one where you've got the archivist number one outline. Number one
179. 36:49thought for a few seconds log line. An ex search engineer turned secret arch archavist protects a seller of forbidden analog
180. 37:01memories from a citywide digitization program until a decades old cassette recording in his care begins to teach the very
181. 37:12drones sent to seize how it sees it how to feel. So this again five 50 five the fast five was
182. 37:28really bad. This this has well-written characters and then structure three acts and key scenes. Act one setup established the
183. 37:40tableau Eli in a cellar inciting incidents. Marin's municipal drones begin a neighbor neighborhood survey. Polite notice is posted. A public
184. 37:51campaign promising universal access to memory. Eli stuffs the flyer into a box marked not for indexing. Act two confrontation.
185. 38:05Yeah, this is quite good. All right, got it. So now let's go to GPT5 thinking. So this is the big mac daddy. This isn't
186. 38:15pro, but it's up there. So, this is your full-on thinking model. So, the other one thought two seconds for this first prompt. We'll see how
187. 38:24long deeper thinking does generating story ideas. And again, there's not a real challenge here.
188. 38:39So, this is 7 8 9 10 seconds, something like that. 11 12. So, it's thinking 21 seconds. Patch day refu refuge refugee. Everyday
189. 38:51global software update patches reality, erasing unsanctioned objects and memories. Wow. A woman camps in a rural
190. 38:58dead zone to keep her unpatched self intact. The Lag monastery. An order of monks takes vows of deliberate inefficiency.
191. 39:09Hand copying, walking, waiting to generate lag that befuddles optimization AIS. That's pretty good. I like that
192. 39:17one. Analog heirloom. It came up with that again. Although this time it's a windup clock, not a typewriter. Time shelter. HOA.
193. 39:28Culdeac enforces 1997 only tech to stay below predictive policing thresholds. That's pretty good. The unphotograph.
194. 39:39So again, these are pretty good. I kind of liked I kind of like the writing style of GPT5 thinking mini or thinking fast,
195. 39:51which is the mini. Um let's do the lag monastery. So, um outline number two,
196. 40:23Sister Mara Abbott. Isef serene stubborn leader kite auditor brother Tomas world rules the state systems thrive on
197. 40:38low variance the order vows right by hand walk on foot wait one breath longer than needed their bells toll on prime intervals
198. 40:50creating unpredictable burdens on nearby mesh mesh networks that's really smart actually on prime intervals. That's super cool.
199. 41:02The valley is a latency sanctuary. Theme threads, efficiency versus meeting, obedience versus conscience,
200. 41:08time is craft. Can a machine choose to be slow? Um, this actually doesn't seem as well written. So, if I'm if I'm just although
201. 41:24it does give ending options, three different options. A hopeful ending, a bittersweet ending, or a define ending.
202. 41:31That's pretty good. Title alternatives, the order of one breath, prime bells. I like prime bells. Variance, tolerance, the human jitter.
203. 41:41So, let me flip to let's go with
204. 42:07let's go with the defiant ending. So, I'm going to put this into thinking mini because I think it wrote better. Let's call it Prime Bells
205. 42:23and go with the defiant ending. Prime bells next step. Okay. Love it.
206. 42:38I'll lock in the title as Prime Bells and take the defined ending. Below is the expanded scene by scene ro road mapap
207. 42:46one sentence synopsis. Okay, I don't give a [ __ ] about that. Yeah, that's fine. Oh, opening scene. Okay.
208. 42:55The bell waited two beats and then three. The bell The bell waited two beats and then three. It was a small rebellion,
209. 43:06not the kind that overturned governments, only the kind that unmade a machine's expectation. The first strike
210. 43:13was thin and bright as a blade. The second a little heavier, the third, a tired, flaring note that refused to resolve.
211. 43:22It's pretty good writing. Sister Mara held up a pen like a compass and breathed the way the abbey taught taught. in.
212. 43:31Hold a breath that is one more than required. Out. Her hand moves slowly, copying a single copper plate letter
213. 43:39into the scriptorum's long folio, allowing the ink to pull and sink in the strokes that would not be machine
214. 43:47readable from a satellite scan. Pretty good. All right, so let me do a new chat. Flip back over to 40
215. 44:01and then I'm going to change the personality al together two. So that was helpful. I'm trying to think which personality is
216. 44:11going to be the most obvious. Cynical is probably the one. Robot is efficient and blunt. I don't think we want that. Listener is
217. 44:20thoughtful and supportive. I think we want critical and sarcastic. All right. So, we've now got a critical and sarcastic um partner.
218. 44:35So, same prompt. Give me 10 ideas. Let's see if we get any cynicism here. Here are 10 original short story concepts
219. 44:42coming from the theme of hiding from tech advancement, each with its own twist. The analog heirloom.
220. 44:54Here it's a photo album. It's funny that we've gotten that title three times now. Here's Prime Bells. Oh, so this is
221. 45:00remembering the previous chats we've had before. All right, so there's something if you don't know. Oh my god, this is
222. 45:12how to actually experiment with these things is really complicated. Now, all right, let me show you something else in
223. 45:18settings if you don't know this. So, if you go to settings and you go to personalization, they've got reference saved memories and
224. 45:26reference your chat history. Um, I should probably turn both of these off, but the difference between them is
225. 45:38reference saved memories is if you're ever working in chat GPT and you you see it, write the phrase memory saved. It's
226. 45:48saving it into a long list of memories that you can actually go in and look at and you can delete ones if you don't
227. 45:55like them. Like Kyle Shannon hosts five monthly AI salon meetings. Okay, that's good. Kyle Shannon lives on Tik Tok,
228. 46:02goes live on Tik Tok five nights a week on the AI learning lab, which has a following of 40 more than 40,000. Okay,
229. 46:09so those are fine, but there might be other things in here that I don't like, right? Um, and you can't edit these, which is a
230. 46:18drag. You can delete them or not. So that's saved memories. your chat history is. It's got access to everything we've
231. 46:27talked about for the past year and you can just invoke it. Hey, what did we talk about? We were talking about that
232. 46:32musical I'm working on and it'll know it. But for this exercise, I think we we turn some of this [ __ ] off.
233. 46:55reference chat history. Think I'll do it like this. All right, new chat. So, we're in GPT40 again. We just turned
234. 47:08off reference old chat history. And like another thing that's striking me right now, like
235. 47:22very few people are going to geek out like this. Like people that are really into um you know iteration and and details and and
236. 47:33things like this, they may do this, but I think very few people are going to dig in like this. You know, no one changes their defaults
237. 47:42for the most part. Here are 10 stories. The algorithm exiles. In a future where life is governed by the omnisient Oh, by
238. 47:55omnisient AI algorithms, a small community refuses to be optimized. They retreat into a vast cave system
239. 48:02where they live without technology. But when one of them secretly uses a predictive AI to prevent a cave-in, they
240. 48:09must decide whether survival is worth their principles. That's an interesting one. Grandma's firewall. Okay. So, so
241. 48:16this is consistent like how this wrote is consistent with how it did it before. All right. Fine. Let's go to fast. Ask the same question.
242. 48:27This is now there was nothing in there that was cynical, though. Remember this is let me just make sure that I've got our
243. 48:36cynical. Yeah, we're speaking as a cynic right now. Tik Tok question. Uh I don't see a question pinned, but as soon as
244. 48:44one's pinned, I will look. [Music] Oh, from the bottom. Not pinned. You could ask the same question to different
245. 48:57models all at once in different tabs. Oh, and then flip between them. That's not a bad idea. That's not a bad idea.
246. 49:04Um, but I'm I'm I'm good. I'm just trying to get a a really broad take on how these things work. So, the last time we did this, let's see
247. 49:17here. 10 stories about people trying and usually failing. That might have a little cynical flare to it. To hide from
248. 49:23technological advancement, a group of survivalists retreat into the woods with typewriters, vinyl records, rotary
249. 49:29phones, only to discover the forest itself has been wired with sensors, tracking their biod diversity. Their
250. 49:36primitive life becomes the hottest live stream on the internet. Oh, that's [ __ ] awesome. That's a great story. That's a great story.
251. 49:46So, it's like the Truman Show except they think they're out there living in the woods and they're the they're a big
252. 49:51hit show. That's [ __ ] hilarious. And actually, what would be good about that one is you have one of the people like
253. 49:58cuts themselves and they have to go into into town um you know to go to a hospital to get fixed up and and they're
254. 50:06a star. They're a celebrity. That's a really good story. Oh my god, that's hilarious. Um, blackout hotel. Guests pay a fortune
255. 50:19to stay in a mountain lodge where there's no devices allowed. But an AI concierge disguised as a kindly inkeeper
256. 50:26runs everything behind the scenes, watching, predicting, even planting dreams. That's boring. We've seen that before.
257. 50:33A man refuses to buy. I think the the first one is so good. Like, this is something I think I actually wanna I
258. 50:39wanna I want to play with. Um, I'm going to say I love analog colony. Here's the ending. Um, a person from the woods
259. 50:55hurts themselves and has to go to town for medical care. and
260. 51:13um is um recognized like a celebrity and then
261. 51:31and then sees that let's and then sees the 24/7 stream of their
262. 51:48off-grid life. Oh my god. And the show ends or story ends.
263. 52:07That's really good. That's sharp, gut punching ending. It nails the hypocrisy of trying to live outside the machine
264. 52:14only to realize you've been the main attraction all along. Here's a tightened version. Do you want me to sketch out a few different tones?
265. 52:27Oh, I mean fast. Um, I'm going to say write the outline.
266. 52:46Dream of escape. Unseen is free. Scenes of clunky but heartfelt living. Writing letters. Hand grinding coffee.
267. 52:59arguing about whether using a gas stove is too modern. Cracks in the illusion.
268. 53:14I'm going to try something here. I'm going to flip back to four and say, "Give me a new outline." The escape fantasy opening shot
269. 53:28manifesto being written on a typewriter. Rejecting the algorithmic leash theme control life in the woods. The
270. 53:39glitches. The accident. Celebrity unmasked. Gut punch. They've been Truman showed the entire time. The nar narrator stares
271. 53:49blankly at their own face on screen. Mara, barely conscious, mutters, "We were never off the grid." That's pretty good. Do you want to run
272. 54:02this as a short film, tight short story, or sketch it into a Black Mirror style script? Because the thing's got legs and
273. 54:10a spine, and I'm already annoyed how good it could be. This that line. So, there's your cynic. So, if you're a Gen
274. 54:18Xer, absolutely put chat GPT into cynic mode because it's that's absolutely writing like, all right. Okay. So, now I want to go to um
275. 54:31so that was GPT fast. I want to go to thinking mini. Thinking mini I liked a lot before. So, we're going to do new chat.
276. 54:39Wait, how am I going to remember this one? Um Oh, I know how. Hiding from technology. Let's uh rename this. This one is what was it called?
277. 54:50The what colony? Analog. The analog colony. Yeah.
278. 55:09By the way, if anyone tells you that chat GPT can't be creative, tell them to bite your butt. Um, okay. So, we're going to go to
279. 55:19thinking mini. We're going to go new chat. We're going to pop in there the prompt. This I'm kind of excited about. So,
280. 55:29we've got our cynical personality on. And this is thinking mini. And I I like this output last time.
281. 55:38Thought for two seconds. Same thing. See how these are like longer and more fleshed out. An old railway si signaler maintains a
282. 55:49disused switch tower on the outskirts of a city that's been fully automated. Like it's just got it's got like the last signalman.
283. 55:57When a new autonomous rail line reroutes the digital map, the ghost train appears only on his analog scanner
284. 56:04and it needs a human to stop it. He must decide whether to protect the passengers by exposing himself or let the silent
285. 56:11system keep running and erase his existence. Right. There's your Christmas Carol choice. Sophie's choice. Right. Nice.
286. 56:21Barter and static. I like I like thinking mini there. 10 ways to hide from progress without being boring about
287. 56:30it. Again, there's your cynical Gen Xer. Try not to trip over the nearest smart device while you pick your favorite.
288. 56:38See, I like this. I like this. This is And here's the thing, man. Man, what character is that? What What
289. 56:49actor is that? Here's the thing, man. I forget. Um Oh, hang on. I got to do physical dexterity in my mind.
290. 57:03Here, I'm gonna I'm gonna stop playing with this. I I have I I have enough I have enough information right now to to
291. 57:08to analyze this or to comment on it. In my mind, if 40 is is is kind of the best writer in my
292. 57:23opinion, like the way I like writing, it writes really good. I would expect the parallel model in GPT5, the the 50, the
293. 57:33fast version to be like at that level or better. But the five fast model sucks. Then I would think, well, the real the
294. 57:46real good writer then should be GPT5 thinking, the big mac daddy. And it's not. It's the thinking mini model that's
295. 57:56got this really interesting way of writing about it. So, let's do a new chat. Wait, I'm Let's go back and share again. Yeah, share the
296. 58:07screen again, Brandon. I want to I want to go back and try something. So, we're going to grab analog colony. Here's our outline.
297. 58:48Okay. Copy. I'm going to do new chat. So, I'm going to say, um, here's an outline of a short story
298. 59:03I really like. Tell me how you would improve it. All right.
299. 59:21Delicious premise. You've built a neat little moral mousetrap. People flee screens to discover the screens were
300. 59:27front row seats the whole time. Below I'll point out what's working and give surgical fixes for a stronger
301. 59:34alternatives. What's already great? Clear, satisfying structure. Truman show style reveal. Mara as a quiet competent
302. 59:43anchor. Small sensory details. What can get sharper? Foreshadow the surveillance more cunningly. Someone has someone always has has
303. 59:56what's needed. Clue is good, but make it a running motif. Give the narrator a clearer inner stake. Why did they run?
304. 1:00:04Not just ideology. A personal wound, a loss. Okay, that's good. Make Mara's arc do more than suffer. Good. Tighten the tone. Fine.
305. 1:00:15Depth of audience. The fans at the end are easy targets. Instead, show them in micro. A kid who imitates the colony. An
306. 1:00:25influencer live stream streaming their outrage. Oh, that's good. Okay. Oh, it goes on and on. Damn. Damn.
307. 1:00:41You already have the spine. Fix the pacing. Okay, great. So, um, update the outline with your fixes and improvements.
308. 1:00:58[Music] Good. You've built a delightful trap for earnest fugitives. I've tightened the plot. Okay. Life in the woods, the glitches.
309. 1:01:17So, I like this. Okay. What's this? Commit to first person POV. Fine. There. The story keeps. Okay. So,
310. 1:01:31I would Okay. Give me five author styles you think
311. 1:01:47would be best for this. I am a fan of the imagery of Tom Robbins. Um but want your ideas.
312. 1:02:04Don't just make them sci-fi writers. Okay, let's see what we get here. Considering author styles disclaimer, I can't really give you
313. 1:02:19note. I can write in the spirit of living authors and capture their hallmarks, but I won't produce exact
314. 1:02:25imitations. Tom Robbinsesque lyrical mischief. Uh, ecstatic detail. Um Joan did D did D did D did D did D
315. 1:02:36did D did D did D did D did D diddion razor calm observation cultural evacu excavation um Heroku Marikami adjacent dream logic
316. 1:02:50melancholy no that's not right Barbara Kingsolver earthwise moralism ecological texture eh and then Italio Calino fableike clarity
317. 1:03:03formal play. All right. So, let's mix. Let's mix one and five.
318. 1:03:26Fine. You want Tom Robbins ecstatic metaphors graded into Italio Calino's clean fable like rules. ruthless,
319. 1:03:33joyful, and oddly tidy. Below I give you a brief style note, a manifesto epigraph, and three short original excer
320. 1:03:43excerpts. Opening typewriter, Marlo vignette written style in two lines. Joyous metaphor plus domestic precision,
321. 1:03:53parable structure with bursts of lush lush imagery. Fine, whatever. Um, great. Use the latest the latest outline
322. 1:04:04and this creative imperative and write me a complete short story
323. 1:04:23that even the most hardened And Holly would exec will want to option.
324. 1:04:41All right, we're doing it here, people. Joan Didd Kyle types almost as well as me. Shut up. I know I could do the talking thing.
325. 1:04:54I just I'm dumb. Okay. The analog colony rule one. If you're not wait, if you're going to quit the machine, quit the
326. 1:05:04whole contract. Bring paper, not performative gestures. Bring courage, not cameras. The typewriter cuffed the
327. 1:05:11manifesto into the room like a small necessary fire. There's there's your Tim Rob Tom Robbins right there. Keys landed
328. 1:05:18in the stubborn in the stubborn rhythm of someone wrapping a truth into a drum. One key stuck, a stubborn little click
329. 1:05:25that sounded, for reasons I could not name then, like a secret door closing. Little foreshadowing. I underlined the
330. 1:05:33last sentence twice, smelled the lemon oil of the ribbon, and felt, and for a few minutes that we could erase the
331. 1:05:40world by refusing to answer it. Um, and do we have a nice long short story here? We do. Look, it made more. It made more than a
332. 1:05:52six paragraph short story. All right, let me copy this.
333. 1:06:19Why does he sing like that? That's rather annoying. I was just listening to him ramble and he nearly had me asleep and then he
334. 1:06:27started doing that sort of beatboxing like a talentless beatboxer and that really it woke me up. I found it I found it annoying.
335. 1:06:36All right, the analog colony. Watch this. By Kyle Shannon, did you use AI to write this? None of your [ __ ] business,
336. 1:06:57sir. Mr. Shannon, yes. Uh we have we have evidence uh that that the analog colony uh short story uh for which
337. 1:07:07you've received so much praise including the Pulitzer for uh fiction congratulations by the well short-lived
338. 1:07:16congratulations while you had it. Uh it is it has come to our attention that that this is a story that was generated
339. 1:07:23by artificial intelligence. simply a a prompt was put into a machine and a stochastic puppet vomited out some
340. 1:07:31tokens and you submitted that as your own work and and sir, you know, first of all, uh the writers of the world resent
341. 1:07:38you uh and and and really just despise everything that you stand for and and uh we've talked to the Pulitzer committee
342. 1:07:46and uh while while they have never uh retracted a Pulitzer in the past, uh there's always a first for everything.
343. 1:07:53Congratulations, Mr. Shannon are breaking new ground. Uh so yeah, you no longer have a Pulitzer there and uh the
344. 1:08:00uh we talked to Universal Studios was really excited to make this movie. They uh oddly enough not as excited now. Uh
345. 1:08:08so uh good good luck to your career there. Uh and uh you know you'll you'll probably be out of the out of the uh the
346. 1:08:17detention center here shortly. AI prompting career. Yeah, it is. It is
347. 1:08:32an AI prompting career. I mean, quite frankly, because because what what I'll do is I'll actually I like this story
348. 1:08:37enough I'll I'll read it and if there's [ __ ] in there I don't like, I'll rewrite it. But it's like, you know,
349. 1:08:45the amount of the amount of sort of iterations we went through of just give them give us a bunch of ideas. We went
350. 1:08:51we went through a bunch of this stuff. This one like stood out above the rest. Like this was clearly kind of cool. Um
351. 1:08:58and so like the curation of that like there is creative craft in what we just did, right? I saw there's there's a dude on
352. 1:09:08the on the on the Tik Tok right now. He's got long hair. He's a writer and he just keeps talking about people that use
353. 1:09:15AI. They like just push a button and out comes a thing. And it's like no one that I know that's doing good work in AI just
354. 1:09:23pushes a button and out comes a story, right? They're they're they're putting in intention and thoughtfulness. And in
355. 1:09:30this case, we were kind of experimenting with which of these models did the most interesting writing, you know, which and
356. 1:09:36and how did changing personalities alter that? Those are all creative inputs, you know. So, you know, I think I think we're in a
357. 1:09:47we're we're in a relatively short phase, maybe it's five years, where the demonizing of of AI generated work,
358. 1:09:57you know, it's going to peak at some point, probably in the next two years, and then it'll fade over the next 10. Um
359. 1:10:04cuz it who [ __ ] cares if it's good work, if it's a good story that's well written. What does it matter anyway? All right.
360. 1:10:18[Music] Thoughts? Thoughts? I'm going to have some more tequila. Hello, champ. Good day to you, champ
361. 1:10:33Shannon. You're a relatively rotunded puppy. Pinkies out because we're classy.
362. 1:10:47Yeah. Question for you, Mr. Shannon. Yes, Champ Shannon. Um, yes. Question. Uh, are you a purebred? A purebred mut.
363. 1:11:02You were found in a ditch in Oklahoma. All right. He was You all right there, buddy? Am I picking on you? Little too much.
364. 1:11:12Little too much. Orphan shaming. Dog dog shelter shaming. Hey, buddy. Hey, buddy. Hey, Chippy. All right,
365. 1:11:23let's uh let's go do some other [ __ ] Let's go do some other stuff. Tik Tok pen. Wait, is is a collage art? Of
366. 1:11:32course. Yeah, exactly. Exactly. I I mean, listen, the I understand I understand
367. 1:11:48a the anger. I also understand the AI slop argument, right? The AI slop argument is and and YouTube just
368. 1:12:02demonetized a whole bunch of videos that do this, right? If I just push a button and out and squirt out some content and
369. 1:12:08publish that and especially if I do like 10,000 of those a week, um that's just sort of polluting the the the airwaves, right?
370. 1:12:18But if I as a human being am taking my 30 years, 40 years, I don't know, Jesus, 50 years of writing experience
371. 1:12:27and I'm taking the way I think narratively and the stories I like and the way I like to approach things
372. 1:12:34and how I'm prompting these things is guiding them down a very particular path. And I'm iterating as I go and I'm
373. 1:12:44curating as I go and I'm flipping between models because I recognize that this one model is better than these other two.
374. 1:12:56That's a that's a creative chain of craft. It's just not the same thing. And and that the thing that frustrates me about
375. 1:13:06all of the people that are pissed off right now is they don't because they're pissed off and they're
376. 1:13:12sitting on the sidelines with their arms crossed. They actually don't know how these tools work and they don't actually
377. 1:13:19understand how creative professionals actually work with them. Like this writing guy on TikTok today said, "Well,
378. 1:13:27anyone that uses AI is not a creative professional." No. No, that's not true. There are a lot of noncreative people
379. 1:13:38squirting out creative words at the push of a button. But there are also a lot of creative professionals who are learning the
380. 1:13:48nuances of these AI tools and working them into their traditional storytelling chain of craft. But but the argument right now is if you use
381. 1:14:04any AI tools, you're a noncreative plagiarizing thief. It's just not true. It's just not true. Um for for most of the people that I
382. 1:14:18know that are doing anything interesting with AI, if people are just squirting out [ __ ] I hate him, too. Calm down. He's not
383. 1:14:30I well I almost I almost respond I almost made a a response video today but I just I didn't quite know I didn't have
384. 1:14:36enough time to to know on TikTok do I stitch it or duetit or whatever the [ __ ] you're supposed to do so I just
385. 1:14:45gave up anyway. Um all right imagine wait no let's go to let's go to let's go to jingle.ai. So, what I want you to do I
386. 1:14:59sorry I may choke out on this one. I want you to go to jingle.ai. Is it jingle.ai? Is it jingle maker? It's not jingle.ai.
387. 1:15:18Jingle maker.ai. Yeah, jingle maker. Okay. Okay. So, this is from 11 Labs. So, as you may know, 11 Labs two weeks ago um
388. 1:15:35came out with 11 Labs music, right? So, a pseudo competitor. It's not It's not great. It's like I I would consider it fourth.
389. 1:15:47I I would say right now you've got Sununo and um Producer AI are kind of the top two. Then you have UIO and then
390. 1:15:55I would say 11 Labs Music is number four in in my opinion. Um but they did this thing which is really cute and the idea
391. 1:16:04is very right. The idea is I can just take the URL of a website and it's going to create a jingle for the website.
392. 1:16:14but they're so bad. So, we're going to do a radio ad. Well, no, let's do Madison Avenue. We'll do more sophisticated. And what I'm going to
393. 1:16:22type in is I'm going to type in my company's um Oh, screen. Oh, am I not sharing at all anymore? What happened? Damn it. Weird.
394. 1:16:38Okay, so this is jinglemaker.ai. So, this is part of 11 Labs. So, you can pick Um, six different styles, a funky rap, a
395. 1:16:47Midwest farmer, dramatic TV commercial. I'm going to do Madison Avenue sophisticated advertising appeal. And
396. 1:16:54then you just type in your website, storyvine.com. So, this is my company. And then it's going to go make us a song.
397. 1:17:05It ain't no Rick Rubin. Stacy Rodriguez, thank you, Kyle, for all you do every night breaking things down and showing
398. 1:17:12us the ropes. appreciate you. Thank you, Stacy. I appreciate that. Yeah. I mean, you know, I I say this a lot, but the
399. 1:17:20the purpose of this channel is not it's not for me to teach you anything. It's just for all of us to be in the conversation.
400. 1:17:28And I am every bit as much at the beginning with this GPT5 stuff as you are. I learned something new tonight that the model that I
401. 1:17:39assumed I would use the least is probably the one I'm going to use the most. Right? Because thinking f like it is
402. 1:17:48clear to me that that GPT5 fast is a shitty model. Like that's clear to me. I I got clear on that tonight.
403. 1:17:56I don't like GPT5 thinking because it takes too long. But GPT5, Thinking Fast, that was the one that I didn't think I'd
404. 1:18:06have any use for. That's the one that ended up being my favorite writing partner tonight and it's decently fast. So,
405. 1:18:13so I that's probably going to become my new default. I'll I'll tell you that could change in three days, but that's
406. 1:18:18my instinct right now. All right, so here's here's uh now watch this be [ __ ] brilliant, but the I did one
407. 1:18:25earlier for Story Vine and so did producer Brandon. We could talk about the black bar. Um, and uh, so here we go. [Music]
408. 1:18:42>> Ask some questions. Authentic moments brought to life. Structured storytelling by story by minutes. Your tales alive. Polish scaled
409. 1:18:55in record time. [Music]
410. 1:19:12Uh, that was actually not horrible. I mean, it was cheesy as [ __ ] And you know, if we were in the 50s,
411. 1:19:20but you know, what's old is new again. Um, again, here's here's why I think this is interesting. There's there's one there's one thing to
412. 1:19:35put in the prompt box. a URL and then there's six choices. Do you want a radio ad? Do you want TV? Do you want country?
413. 1:19:42Do you want hip-hop? Let's let's go back and do hip-hop. So, we'll do storyvine.com and we'll do funk or funky rap they call it.
414. 1:19:57people's interactions with generative AI are are are likely going to be these little entertainments, these little like
415. 1:20:04vending machines, you know, think think like, you know, in the in the uh tech district in Tokyo, all the crazy vending
416. 1:20:13machines they have for like blocks and blocks and blocks. Um, that's how I can imagine AI evolving, that it just evolves into all
417. 1:20:23these little tools of joy, right, and function. Um, oh, Vicki did one for the AI salon in the Irregulars. Okay, good.
418. 1:20:32Let's go check that out. That's great. Um, all right. Irregulars channel. Vicky's creativity unleashed at the AI
419. 1:20:47salon. Jingle maker. Here we go. Um, it's probably gonna jump me to a new tab, but I know how to deal with that
420. 1:20:56now. It did show this tab instead. Okay, here we go. >> Agnite. Join our vibrant community. spark your insight.
421. 1:21:10From curious beinners to daring experts in play, empathy and bravery guide us every day. [Music] The AI salon where creativity
422. 1:21:29collaborate, experiment, let's all thrive. Unlock the future. Join us now. the AI salon. >> And speaking of which, that's a
423. 1:21:42fantastic ad for the AI salon. If you're going to take yourself over there to the salon.ai or just community.thesalon.ai,
424. 1:21:50that'll bring you right into the AI salon. And I'd join if I were you. Why? Because it's awesome. Couldn't you tell by the song?
425. 1:22:01That's actually pretty good. I liked it. That was cheesy. Yeah, it reminds me of the original Sununo. It It does. That's
426. 1:22:07That's the thing about 11 Labs music right now is it very much feels like early Sunno or early UIO. Those guys
427. 1:22:14have moved on to more sophisticated stuff, but um but yeah. Um so that's it. Jingle Maker.ai. Let's go. Let's go back and listen to um
428. 1:22:26[Music] Is this it? Share this tab. Tik Tok pin. Uh, I see no Tik Tok pin.
429. 1:22:42Just Justin Justin Bber Aber. Yo, feel that baseline bumping beats in the trunk and smooth sets the stage. Let
430. 1:22:53the story function. Your cord tail in real time. Get your answers on the fly. Cloud does the magic clips transform
431. 1:23:00before your eyes. Structure flow maintains your vibe. Authenticity amplified. Simplify video storytelling. Watch your
432. 1:23:07narrative thrive. From customers to advocates, let their voices come alive. All you need is story vine. Scale up your vids. And
433. 1:23:16>> yeah, that's pretty horrific. It's pretty bad. Um, like it's bad musically, it's bad lyrically. Um, but you know, um, here's a widget.
434. 1:23:28Here's an AI widget that, you know, this might be people's first uh interactions with these tools. Okay, let's jump over
435. 1:23:37now. Let me let me pause for a second because I know I've been rambling a lot. Let me just check in with y'all. See if
436. 1:23:42you have any questions or thoughts, questions, thoughts? Anybody have any questions? Music is math at the end of the day. AI
437. 1:23:53can probably fake hiphop easier than it can classical music. So Ryan, here's an interesting thing about that statement.
438. 1:24:01I mean, about about that supposition. When I when I was first learning about generative AI, I was learning about um
439. 1:24:12translating Chinese versus translating from English to Chinese uh or from English to Spanish. Because Spanish is a romance language,
440. 1:24:22you would think it would be easier for it to do that because it's closer. In traditional computing where you're using
441. 1:24:28logic, that would actually make sense. But the way large language models work is they're essentially blind to the
442. 1:24:36data. You just tag the data, whatever the data means, and then it it sort of trains on the data objects whatever it
443. 1:24:44turns into these tokens that go into the the latent space. when it recreates them, it's just doing it um
444. 1:24:54mathematically identically whether it's Chinese or Spanish or English. Um in fact, the language translation
445. 1:25:03stuff was an emergent behavior. They didn't train these things to be good at language translation. They just trained
446. 1:25:08it on enough language that it sort of figured it out. And it's the same with music. So, so, um, it being better at
447. 1:25:17hip-hop than classical music, probably not. It's just if if you've got a strong well-trained model, it will be
448. 1:25:27good at both. Um, now, if you if you trained a model and it's got almost no hip-hop in it or no references to it,
449. 1:25:35then it would be really weak at that. But it's not that one is more complex than the other and it does it better or
450. 1:25:40not, which is I find wild. Um what do I mean by the latent space? Okay, so box metaphor this thing. Actually, you know what?
451. 1:25:54I'm going to show you a slide because
452. 1:26:08>> Hey, Kyle. >> Yes, sir. Uh, I'm I'm I'm not saying you can't show your slide, but it is against the production rules to bring the boxes
453. 1:26:19out and not do the box method >> and not use the box. I know. It's true. It's true. >> You can't tease people on Friday night
454. 1:26:25date night like that. It's >> okay. All right. All right. Well, we'll start with the box because that thing's
455. 1:26:29loading anyway. Okay. We'll go with the box metaphor. So, what is the light in space? Okay. Um the way the way okay there's so many
456. 1:26:42pieces to this in 2017 Google writes this paper called attention is all you need and they introduce this concept of
457. 1:26:50a technology called the transformer and the way the transformer works is it takes tagged data right so you so you you
458. 1:27:02know take a document doument and you know you you tag it so that you understand what that document is and
459. 1:27:10what's in it and then you run it through this thing called a transformer. And what the transformer does is it takes
460. 1:27:18the document and it actually shatters it into what are called tokens. And tokens are things like spaces and periods,
461. 1:27:27fragments of words, whole words like dog or and, fragments of words that are sort of the root of a word plus the modifiers.
462. 1:27:36So, so it shatters it into these tokens and it then semantically clusters them in thousand dimensional mathematical space.
463. 1:27:50That's the box. So, so they basically take all these documents and then they shatter them into these things called
464. 1:27:57tokens. And in the box there are all these different little clusters of tokens. And and and the clusters of tokens might
465. 1:28:06be you might have a token that represents the word dog that was out of some paper that was about canines. And
466. 1:28:12that's over here in a cluster of tokens related to pets. And then you could have the same word dog that came out of a dating book and
467. 1:28:23it talked about the guy that did his girlfriend bad and he's a dog. And that's over here in the dating cluster,
468. 1:28:30right? And so so none of the original documents actually remain anymore. Like once you embed them, they literally turn
469. 1:28:39into these shards, these little teeny tiny shards of words that get clustered in in a thousand dimensions in basically
470. 1:28:47in in a box like this. And that's your large language model. So if you remember the early days of chat GBT, it was like
471. 1:28:55it was last updated on, you know, November of 2021 or 2020, whatever it was. That's because that's when they put the
472. 1:29:04last that's when they did the last embeddings. And so all of those tokens are basically just sitting in this
473. 1:29:09mathematical space. And then what happens is when you type a prompt and this is called the latent space. The latent space is that is that
474. 1:29:17mathematical space where all these tokens exist. When you type in a prompt and and you hit the return button, it
475. 1:29:25turns your prompt into a mathematical probability and and it basically says the next most likely token that is the right response
476. 1:29:36to that prompt is somewhere in this mathematical space. And so it'll go, okay, dog. And you know, somewhere over
477. 1:29:43here in the pet cluster is the word dog. And I'm going to pull that out. I'm going to put that in the sentence. And
478. 1:29:48so when when you see a large language model generating all of this crap, all these words, it's literally pulling
479. 1:29:59individual little fragments of words from thousand dimensional space with with billions if not trillions of of
480. 1:30:08these tokens. And you know, when they talk about parameters, all all the different, you know, ways it can look
481. 1:30:14into that space and how it's trained. Uh, it's insane. The more you learn about how this stuff works, the more
482. 1:30:21magical it gets to me. Like, that anyone figured this out is [ __ ] insane to me.
483. 1:30:40Okay. So, >> it won't let me pin it, but um Miss Burch would like to see the slide. [Laughter] >> Fine. >> Mind your tabs.
484. 1:30:54>> Thank you. >> Am I sharing right now? No. All right. So, so I just I I just described it in the box metaphor and
485. 1:31:08I'll show it to you in a uh in an actual slide. Oops.
486. 1:31:27Okay. So, in case you didn't know, GPT stands for generative pre-trained transformer. Generative means it's generating. When
487. 1:31:36when you create something, it is not copying and pasting. That's a really important concept. Most people that
488. 1:31:42don't understand AI assume it's copying and pasting. They assume it's just like hip-hop in the 90s where you're
489. 1:31:48literally taking a a chunk from a song and copying and pasting that into your song. That's not the way it works. It's
490. 1:31:56generating from the from all these fragments being stitched together. pre-trained means it's been pre-trained
491. 1:32:02on whatever data they jammed into it. Uh much of it not legally gathered. Uh and then transformers the technology uh LLM
492. 1:32:12is the large language model. That's the thing. The latent space is this is a a two-dimensional a flat dimensional
493. 1:32:20representation of like these tokens in semantic clusters. It's a mathematical representation of words, tokens, which
494. 1:32:28are fragments of words, but but for the most part, it's a it's a mathematical representation of words in ndimensional
495. 1:32:35space. This is two-dimensional space, right? XY and then there's these little colored clusters of meaning.
496. 1:32:44And then based on your prompt, the LLM comes up with a probability waiting that somewhere in this mathematical space is
497. 1:32:53the the correct next word. So the example that I have here is the quick brown fox jumps over the lazy blank. And
498. 1:33:02the way it works is mathematically the this blue cluster is is most likely where that word is. And then one of
499. 1:33:12those tokens within there will have the highest probability of being the correct word. They have this thing in the in the
500. 1:33:19settings called temperature where they can if they turn temperature down to zero, it will always pick that same
501. 1:33:25token. If they turn it up to one or two, I forget whatever the scale is right now. It it might pick from like 20 of
502. 1:33:33those different tokens, right? And so so there's a they can actually affect its creativity or its hallucination.
503. 1:33:40um by by sliding a slider to be more deterministic or or more uh random in in terms of what it's what it's putting in
504. 1:33:49there. And then that um that token that it chooses has this mathematical location in the latent space and then
505. 1:34:01and then that specific token is associated with a specific word in this case dog. And then it'll take that word
506. 1:34:09dog and put it in the sentence. And you know, you can see there's other words around it that that might have been
507. 1:34:14possibilities, but it chooses dog. And again, this to me, like once you understand this that how the how this
508. 1:34:23[ __ ] actually works when you watch it write an entire marketing plan for you and what it's effectively doing is
509. 1:34:31going, "What's the next token in your thousand dimensional latent space? What's the next one? What's the next
510. 1:34:36one?" and there it's just sort of vomiting them out one after the other. So the modern systems do it in clusters
511. 1:34:43and there there's all sorts of fancy [ __ ] But this is basically what's happening. This is also why um these things hallucinate
512. 1:34:54because as much as it seems like they're semiconsciously understanding what you asked and answering it, they actually don't know
513. 1:35:04what they wrote. They're just pulling tokens out of mathematical space and putting them on the page. And so that's why when you
514. 1:35:13say, "Well, that thing's wrong." They'll be like, "Oh, you're absolutely right. That was wrong. Let me fix it for you."
515. 1:35:17And then they'll do it again, right? They'll [ __ ] it up again. GPT5 apparently, again I don't know all
516. 1:35:25the details of this, but apparently GPT5 has in it a mechanism that rather than just vomiting the tokens out on your
517. 1:35:33page for you, it vomits the tokens out internally and then actually looks at them and says, "Are there any
518. 1:35:39hallucinations here?" And if there are, it tries to fix them and then it gives you an answer. So there you go. That's that's it. Um, this
519. 1:35:50is the Let me show you another thing. There's a there's also a reason when when you tell um when you tell your chatbot, I want you
520. 1:36:03to act like uh like a newspaper reporter. Um, the reason that works is that it it basically narrows down your semantic
521. 1:36:15clusters basically, right? So if you say I want you to act like a newspaper reporter, there's probably whole areas
522. 1:36:21of the latent space that have all sorts of journalism and things in it like that and it'll probably ignore things where
523. 1:36:28people just talk about news that are not journalists, right? So so that's why giving your large language model a role
524. 1:36:36can work is it limits the latent space and you get more focused tokens to draw from. So all right, how we doing timewise?
525. 1:36:46It's getting there. It's getting there. It's getting there, people. All right. Black bar. All right. Switch me. Let's see. Where
526. 1:36:55are we going? I'm I can do whole screen here because I'm not doing any audio. Okay. So, let's go to where I'm going
527. 1:37:03right now is aistudio.google.com. So, this is the development area of Google's Gemini. And when I say the devel development
528. 1:37:18area, this is where coders and programmers and engineers go if they want to use the Google APIs. So you can come in here and they've got
529. 1:37:25all these different models that you can choose from. They've got open- source models. All their Gemma models are in here. Gemini
530. 1:37:33um they've got the the interface for this has gotten better and better and better, but you can get an API key here.
531. 1:37:41So if you actually want to be making AI applications using in this case Gemini models, this is where you would sort of
532. 1:37:48test them and see which one works best and what the costs are and all that sort of stuff. But one of the things that
533. 1:37:54they've done is they broke out um generative media as a distinct tab here. And so they've got image gen, they've
534. 1:38:03got um LIA real time interactively create, control, and perform music in the moment. That's kind of cool. I don't
535. 1:38:11think I've played with that. Um, image gen is their best image generation model. They've got speech generation.
536. 1:38:17They've got the VO, this is the VO video model, right? So you could say um uh have a pompous actor recite
537. 1:38:32um the text of a spetio's commercial
538. 1:38:48failed to generate video quoted exceeded due to high demand or not. That's that's the go [ __ ] yourself response. Okay, so we'll go to
539. 1:38:59image gen here. Um okay, so now you've got image gen 4, image gen 4 ultra, image gen fast. So again, if you're in a
540. 1:39:10development environment like this or in the open AAI playground, it's going to be a lot more technical. There's there's
541. 1:39:18going to be lots of models to choose from, which in theory we shouldn't have to do in chat GPT, but we do because
542. 1:39:23that's what people wanted apparently. Um, but these are all the new models. So, let's see. Image Gen 40.
543. 1:39:34And so, if we go in here, you all know my classic my classic image generation prompt, 1970s
544. 1:39:48muscle car restood glistens in an abandoned factory
545. 1:40:02with uh afternoon sun streaming through the missing factory. Tory Windows.
546. 1:40:20So, let me copy this so I don't have to type it again. Black bar. Sorry about that. Oh, it's up.
547. 1:40:35Oh, yeah. I should have done four images. [ __ ] output resolution 1K 2K. So, one of the things I've noticed about
548. 1:40:47image gen is that it's making like this is clearly a Mustang, right? That's clearly a Mustang logo. That's the
549. 1:40:55Mustang grill. That's the Mustang profile. This is sort of not a Mustang curve right here. It's kind of got um
550. 1:41:04Challenger some Challenger influence on the fenders, but pretty pretty good. Um let me do four of these. Let me flip it
551. 1:41:14to that. So that's image gen 4. Here's Image Gen Fast. We'll do four of them here.
552. 1:41:28So again, two Mustangs, two Challengers. Interesting. And these these are faster, but they're still pretty good.
553. 1:41:47And then let's do ultra. So ultra should be should be super good.
554. 1:42:08Yeah, look how that glistens. Love that color. Chevel SS. Really nice. Challenger Dragster. Look at those nice wide tires.
555. 1:42:23Oh yeah, that's Resto Mod. That one's gorgeous. Really cool.
556. 1:42:39That's really good. That's really good. So, if you want a new image thing to play with, head on over to aistudio.google.com, google.com
557. 1:42:48click on generate media and uh yeah then there you have it. Do did I lose all my generate media? Did I lose everything I've created history?
558. 1:43:08What's wrong champ? You want cheese? Where's my history? There it is. All right. So my history's in there. Nice. That one's really slick.
559. 1:43:21I'm gonna download that one. Let's do these in 2K. We'll do 2K ultra.
560. 1:43:40[Music] Uh oh, I just quit Tik Tok. Sorry, Tik
561. 1:43:56Tok. I just quit you. I wish I could quit you. Let's see what we got here. Large view. Why can I not do a large view?
562. 1:44:12What's going on? Download. We'll do this to desktop.
563. 1:44:32Champy. Shush. Oh, that's gorgeous. Look at the cobwebs on the on the machine back here. But look at the depth of the color.
564. 1:44:44Doesn't quite say charger, but you know, good enough for government work. I think that's pretty damn pretty damn
565. 1:44:51good. That's pretty darn good. What do you think, Champy? You think that's good? Yeah. You want to go out? All right, I'll let you out.
566. 1:45:05All
567. 1:45:19right. All right. Let's get out of here, people. Let's get out of here. How's the hair? It's It's special
568. 1:45:34tonight. All right, everyone. So, hope you had fun on Friday night date night. Um, here's the thing. Go check out the AI
569. 1:45:45Salon. Next Tuesday, we've got the AI Salon meet and greet. So, um, AI Salon meet and greet, you get to meet other
570. 1:45:53people in the community. Um, it's a special meet and greet because we're doing the official launch of the AI
571. 1:46:00readiness training program. So, we're bringing back I think we've got fi somewhere between five and seven
572. 1:46:07original speakers from AI Festivus are coming back and talking about what they've been doing since they talked in
573. 1:46:12Festivist. Uh, and then we're doing the official launch of the brand new website and the official roll out of the AI
574. 1:46:19readiness training program. So, I'm super excited about that. Um, so that's next Tuesday at 5:00 pm Mountain time.
575. 1:46:25Homework for the weekend. This is this is a big one. So, what you if if you've been watching the whole time, what you witnessed me do
576. 1:46:36tonight was I took something that I knew decently well, like I know writing decently well and, you know,
577. 1:46:41storytelling decently well, and I just I just went, you know, sort of poked down some rabbit holes really quickly doing
578. 1:46:50something I know because when you ask a large language model to do something that you've got expertise in, you can
579. 1:46:58pretty quickly see if it hits it or not. So for you it might not be writing. It might be project management. It might be
580. 1:47:05business plans. It might be a marketing campaign. Whatever it is that that you know deeply. It might be coding.
581. 1:47:12Play with these different tools. And I would say play with those personalities as well because again I that that that
582. 1:47:19tweet of the of the VC guy that said you know we've officially hit the tipping point where none of us really know if
583. 1:47:29these things are that much better than the previous model but we are going to know if it gives us what we need if we
584. 1:47:36resonate with it. If it feels like, oh, that's good. Like me tonight, the GPT thinking mini model was the surprise hit
585. 1:47:46of the evening for me. I'm like, oh, that's really exciting. Don't make any assumptions like I did about which model
586. 1:47:54I assume would work best, right? Because the base GPT5 model is horrible at writing is what I learned tonight.
587. 1:48:05It's horrible at it. But the thinking mini one is really good. And then the full thinking one is not as good. Why? I
588. 1:48:13don't know. It doesn't matter. That's just my personal take on it through my filter, through my lenses, through my
589. 1:48:20meaning making machine. So, so play this weekend. Play with the different things. Go into it not having expectations and
590. 1:48:27just see what emerges for you. All right. That's an interesting point. But what about all the people who are going to
591. 1:48:36use AI without having that background knowledge? That's why I do this show. Um I there are no manuals for this. The
592. 1:48:48stuff is changing too fast. Um one of the things that breaks my heart the most is people who are actively
593. 1:48:58sitting on the sidelines with their arms crossed. Like ones that just don't know about AI. I give them a little bit of grace.
594. 1:49:07Ones that know about AI and are just kind of lazy like I've got time. What if they're just lethargic about it, them it
595. 1:49:13serves right? The ones that really break my heart are the ones that actually have some passion about this, but the passion
596. 1:49:20is [ __ ] AI. I'm not going to do AI. They're going to get their asses handed to them. And it's like like you can be
597. 1:49:28pissed off at AI, but [ __ ] use it. understand this stuff, understand what it makes possible, and then you can
598. 1:49:37choose morally which model you want to use or not, right? Like, you know, I I have full respect for creative
599. 1:49:47professionals that are like, I'm not going to use AI for this one area of my life because that's like my sacred
600. 1:49:52creative space and I'm just not going to do it. Cool. But I'm going to use it in these other areas because why not? Great.
601. 1:50:01But the ones that are just not using it, it's just like it's tragic. So anyway, that's why that's the whole point of
602. 1:50:06this channel. Get as many people as possible. So if you know people who should be in this conversation, [ __ ] bring them here. Okay,
603. 1:50:16maybe that's another request this weekend. So one is start playing with these models, the GPT5 models, so you
604. 1:50:22can understand them a little bit more. And then the other one is think about people who should be in this
605. 1:50:28conversation and and maybe invite them next week to the to the lives or to the AI salon. All right, Ruby, have a
606. 1:50:36fantastic weekend everybody and I will see you on Monday night. All right, bye. [Music]