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

Description

Chapters

Transcript

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