4/14/2025 - Dolphin Gemma AI: Groundbreaking Communication with Marine Mammals

Description

Chapters

Transcript

0:01 [Music]
0:13 my way that
0:17 night just like a jet plane in and out
0:22 of I was hauling ass at a million miles
0:26 an hour wondering how hard I'd
0:30 [Music]
0:33 when they came
0:42 [Music]
0:45 into my
0:48 [Music]
0:57 situ
1:00 Say you dare to
1:05 say you dare to say Sheree Sheree
1:12 Sheree.
1:14 Uhhuh. Yeah. Leave a message. Enter your
1:18 number.
1:21 Please take your time to want to satisfy
1:24 me.
1:26 [Music]
1:27 Take all these old fantasies and send
1:30 them care of me.
1:33 [Music]
1:47 [Music]
1:53 Italy that I'm all that I'm glor
2:01 [Music]
2:03 about. So I picked it up and I walked to
2:06 the wind and I Hey, how are you? What's
2:08 happening Steo? What's happening
2:10 Danielle? Who's here? Jill and Teddy are
2:13 in the house.
2:15 Genie in a bottle. Let her
2:18 out. Let her
2:21 out. Um, Pate
2:24 posted. Dolphin Gemma AI News dolphin
2:29 researchers translate dolphin speak.
2:31 Called this one 18 months ago. I told
2:35 you this was coming. Remember those
2:37 early days of the brainwave [ __ ] where
2:39 they were reading reading people's
2:42 uh brain waves to let let them see what
2:45 pictures they were looking at, what
2:46 video they were looking
2:48 at. I said, "It ain't that far away.
2:51 We're going to be talking to dolphins
2:52 soon." And you know what comes after
2:54 talking to dolphins, right? You got a
2:57 little collar for your dog and you can
2:59 be like, "Champ, what the [ __ ] going
3:00 on?" He's like, "I hate singing every
3:02 night. I just do it cuz it's my
3:04 instinct." But God, don't make me sing.
3:07 We're gonna we're gonna pretty pretty
3:09 soon we're going to be able to talk to
3:11 Champ. That's pretty
3:14 amazing. I mean, you know, the amazing
3:17 thing about this AI, the transformer
3:19 stuff that Google invented, by the way,
3:23 um most of the tools that we're playing
3:26 with today that are blowing our minds
3:29 are all based on the same basic
3:31 technology, this transformer model. And
3:34 you know there are exceptions to it but
3:35 like the vast majority of stuff. And
3:37 what's amazing about it is it doesn't
3:40 matter what data you train it on. If you
3:43 train it on words it gets good at words.
3:45 If you train it on pictures it gets good
3:47 at pictures. If you train it on video it
3:48 gets good at video. Apparently if you
3:51 train it on dolphins gets good at
3:53 dolphins. Although it's called dolphin
3:55 Gemma. So, I wonder if is So, Pate, I
3:58 don't know if you're on the on the live
3:59 yet, but uh is that uh is that based on
4:05 Transformer or is Gemma something
4:07 different than
4:10 Gemini? I It's Transformer technology.
4:15 [Music]
4:27 Okay. It's based on a Gemini. So, yes,
4:29 Transformer. It's Transformer
4:31 technology. Source Camp in the house.
4:33 Kelly Camp showed up. She was like,
4:35 "I'll be at all the meetings today. I'll
4:37 be at all the meetings today. All of
4:39 them." Well, I do declare. I do declare.
4:43 I'm at the meeting. I'm here to
4:45 evangelize AI. I think AI is the most
4:48 wonderful thing ever. I'll be standing
4:50 in the grocery line and someone will
4:52 say, "Excuse me, ma'am." I'll say,
4:53 "Excuse me, have you heard of AI? It's
4:58 [Music]
4:59 amazing." It's my Kelly Camp
5:01 impression with a non Dallas southern
5:05 accent. That was like a a a Gone with
5:07 the Wind Southern accent, wasn't
5:11 it? And I feel like I've indeed been at
5:14 all the meetings today.
5:18 Why Sam Alvin, I do
5:20 declare you with your 4.1
5:23 model released to the API community.
5:26 Good lord, Mr. Altman, you are a fresh
5:30 fresh
5:32 boy. Oh,
5:34 wait, Mr.
5:36 Alman.
5:38 Indeed. Well, you with your models.
5:41 First, you come out with the four O. If
5:43 it's not zero, it's O like Ammy. Then
5:48 you come out with 4.5 and everyone's
5:51 like, "What is this?" And you're like,
5:55 "That's 4.5. It's got emotional
5:59 intelligence. It's brilliant." Except
6:01 where it's horrible, like the
6:04 benchmarks. And then you follow up 4.5
6:07 with what? Five, 4.7, 4.8. No, 4.1. Oh,
6:13 Sam Alman, you cheeky boy. Oh, good
6:16 lord. Good
6:18 lord. 4.1 in the face of
6:22 4.5. 4045 41. What next? 37. Oh, it'll
6:28 be a joy. It'll be a
6:30 delight. Oh, how I love Open
6:35 AI. He actually he actually
6:38 posted He actually posted, "We're gonna
6:41 fix our naming convention in a couple of
6:43 months, so you have a few more months to
6:45 pick on us." Oh, we [ __ ] will.
6:49 [Music]
7:24 You like that up there?
7:37 [Music]
7:41 Mental
7:44 pollution inside my brain.
7:51 [Music]
8:06 Um, I did a cool thing on Saturday. Made
8:11 a cool
8:12 video. Um, here's what I figured we'd do
8:15 tonight.
8:18 So,
8:20 um, I'll do a little bit of show and
8:22 tell, show you how I made some [ __ ]
8:23 which is kind of fun. Let me flip my
8:26 flip my little cameras. Let me flip my
8:28 cameras so people so the people can see
8:30 me in my full glory. My full glory. All
8:35 this right
8:36 here, this is this is God-given
8:40 beauty. God-given beauty.
8:43 There's no
8:44 makeup other than what the the makeup
8:47 crew puts on. But this is all natural.
8:50 Just a little foundation, something to
8:52 take off the shininess. You know what I
8:54 mean? God-given. other than plastic
8:57 surgery and a little bit of
8:59 eyeliner and but it with the with the
9:02 Gucci glasses but other than that
9:04 God-given
9:09 beauty just like I have god-given
9:11 musical
9:13 [Music]
9:14 talent. Sounds beautiful, doesn't
9:17 [Music]
9:19 it? Builds the
9:21 tension and resolves. We come back home.
9:24 We come back to the
9:28 one. Or is that just I had my my fingers
9:31 on the wrong
9:32 frets? It's a It's a talent situation.
9:35 Yeah. Yeah. It's a talent situation.
9:38 There's There's a a remarkable lack of
9:41 talent going going
9:43 on. I'm deeply sad. I'm a deeply sad
9:48 man.
9:51 [Music]
9:57 So, do a little show and
10:00 tell YouTube
10:04 comment. Hey. Hey guys. Oh, you mean the
10:07 one you
10:08 wrote? We're all just here pretending
10:11 Microsoft didn't go from Windows 2000 to
10:13 XP to Vista to 7 to 8 to 10. Yeah. Yeah.
10:17 But here's the difference, Brandon.
10:20 Windows did those in consecutive years.
10:23 Open AAI, it's as if you could buy all
10:26 of those versions of Windows at the same
10:28 time and and had
10:30 [Laughter]
10:33 to. Well, I use this operating system
10:36 when I work want to work with long Excel
10:38 files, but this operating system much
10:40 better at handling Word documents. And
10:42 if I'm going to play games, I need this
10:44 operating
10:46 system. Oh, that was archetypal. Wait,
10:49 Kyle, I dropped a pretty good example of
10:51 11 Labs and irregulars. Okay,
10:53 [Music]
10:57 cool. But here's what I thought might be
11:00 fun tonight. Well, I don't know if it's
11:02 going to be fun, but it feels
11:05 important. Um, how many folks we got in
11:07 here? We got 20 there, 18 there. Do me a
11:10 favor. Share the share the Tik Tok live
11:12 if you haven't. And if you're on the uh
11:14 if you're on the YouTube, tell friends
11:16 about it. I don't I don't know how you
11:18 share YouTube
11:19 anymore. You know what you could
11:23 do? Okay, here's what you should do.
11:26 Um, look at the URL. Write down, get a
11:30 Sharpie, write down the URL of the
11:33 YouTube channel on a piece of paper with
11:35 Sharpie and like do it legibly. Like
11:38 don't get lazy. Do it legibly and then
11:41 fax the URL to a bunch of your friends.
11:44 I think that's I think that's how we get
11:47 some people in. All
11:49 [Music]
11:52 right. Um I thought it might actually
11:56 not be a bad idea to just go look at
11:59 chat GPT
12:01 again. And I know you're like, "Well,
12:04 wait, don't we like do that all the
12:06 time?" Well, we use chat GPT a lot, but
12:10 like there's so much in there. Like just
12:12 looking at the interface elements and
12:14 what they all do and why you would use
12:16 one over the other. Like for example,
12:19 what's the difference between the
12:21 microphone icon and the advanced voice
12:26 icon? Do you have to hit the search
12:28 button for it to search the
12:30 internet?
12:34 Um what are custom GPTs? What do you do
12:37 in settings? because we're gonna get
12:39 some new stuff this week and they're
12:41 gonna add more [ __ ] yet more [ __ ] to the
12:44 interface. So, I thought it might be
12:46 good to do that, but I don't know. I can
12:48 I could be dissuaded against that. I was
12:50 playing around with Firebase and you can
12:52 create up to three workspaces. That's
12:53 super cool. Wolf man Clint, how do my
12:57 regular
12:58 family create a bumper sticker with the
13:01 URL? There you go. Yeah, if you could
13:03 create a bumper sticker with the the AI
13:06 Learning Lab URL, put that on your truck
13:09 or your uh your uh Ford
13:13 [Music]
13:16 Fiesta. Um, if you are not I don't think
13:21 we have a lot of people in here that
13:22 aren't new, but if you're new here,
13:25 um, go check out the AI salon, the
13:27 salon.ai, AI, click join our
13:30 community. And then once you join it, if
13:32 you scroll down to clubs and hubs, the
13:34 AI learning lab, we've got our own space
13:35 there. So if you want to share things
13:37 that you you make tonight while we're
13:39 chitchatting, that's a place to do
13:43 [Music]
13:44 that. Mr. I'm new. You're new in your
13:48 own
13:51 mind. She came on like slow moving cold
13:55 front.
13:59 His beard was warmer than a look in her
14:08 eye. She sat on a stool and he said,
14:12 "What do you
14:14 [Music]
14:16 want?" She said, "Give me a love that
14:19 don't freeze a bill.
14:22 [Music]
14:37 Love your voice. Thank you very much. I
14:39 appreciate that.
14:41 [Music]
14:57 Well, is this a place I can rest my
15:02 [Music]
15:05 head, gather my thoughts in sweet
15:08 silence?
15:11 [Music]
15:14 Is this place where the feelings on
15:20 de from an over exposure to
15:25 violence is this place I can slowly face
15:30 the only one I truly can
15:34 know. These are tears from a long time
15:39 ago. Got these tears from a long time
15:42 ago.
15:44 I need to cry. 30 years old or
15:48 so. These are tears from a long
15:51 [Music]
15:54 time ago.
16:02 [Music]
16:04 Oh darling, oh darling, you say unto
16:08 [Music]
16:11 me, where have you been all my
16:14 [Music]
16:19 liime? I've been swimming the seven sad
16:24 [Music]
16:26 seas. Poor women have to lie.
16:31 Oh lordy, that's a beautiful song if you
16:33 take it seriously. You
16:36 know, I sang that once at an open mic.
16:40 That was an open mic with musicians and
16:42 comedians and that was my final song and
16:45 the comedian got up and goes, "How am I
16:47 supposed to do [ __ ] comedy after
16:55 that?" It was the funniest line of the
16:57 night. It was pretty good.
17:01 Um, where do we want to start? Let's
17:09 see. I know what I'll show
17:14 you. Oh, darling. Oh, darling. Say unto
17:20 me, where have you been all my life?
17:25 [Music]
17:27 I have been swimming the seven sad
17:31 seas. Um, poor women have tossed me
17:35 their
17:40 [Music]
17:44 lifelines. All right. Fantastic. Bob,
17:47 what have we got
17:48 here?
17:51 So, okay. I've got a I've got a
17:56 I got a bone to pick with you all and
17:59 you're probably not the
18:01 culprits, but but I have done I I am
18:05 guilty of what I'm about to [ __ ]
18:08 about. Um, and I want to I want to
18:12 evangelize a better
18:14 way. Well, why do we have to judge it?
18:16 Why is it a better way? Maybe it's just
18:18 a different way. It it could just be a
18:21 different way, but it's a better way.
18:23 Um, and it's this Kyle sing to chat GBT.
18:27 Did you see Google's new dolphin based
18:29 model? I did not, but Brandon talked
18:32 about it. Actually, we should probably
18:33 go look it up because I bet there's a
18:36 cool video that go with it. Uh, Google
18:40 Gemma Dolphin. Um, I don't know if you
18:43 were here, Pate. I was talking about it
18:45 earlier that I I don't know if you were
18:48 here, but I talked about
18:52 this cognitive
18:55 computations Dolphin 2.0. Oh, wait. I
18:58 need to go to Google.
19:00 Um, hey Kyle. Yeah. Uh, first of all,
19:04 tabs. Uh, and second of all, uh, Pete
19:08 did post the article in the AI salon
19:10 news section today. Okay. In the news
19:12 section. Okay.
19:16 [Music]
19:25 Great. Green Acres is the place for
19:33 me. Living Dolphin
19:39 Gemma Pay. Check out our new LLM for
19:42 dolphin speech. This is [ __ ]
19:45 insane. [ __ ] bonkers,
19:48 man. All right. So, okay. Oh, there's a
19:50 video. We'll go watch the video. Okay.
19:54 Um, so I'll talk about the thing I'm
19:56 going to talk about in a second. Then
19:58 we'll we'll come back. But I got to I
20:00 got to share my tabs different or
20:03 producer Brandon's gonna be like,
20:04 "Nobody can hear you. Could you please
20:06 share your tabs so people can hear you?"
20:08 Kyle. Okay. Cow. Cow. Cow. Cow. Cow.
20:12 Black bar. Cow. Cow. Hey,
20:18 cow. Oh, you got to love picking on
20:20 producer Brandon.
20:23 Okay. So, do we have a new AI salon
20:26 animal
20:28 now? Um, okay. Dolphin Gemma. So, I
20:31 don't know, 18 months ago, something
20:33 like that, when they were
20:34 doing, remember early on in the AI talk,
20:37 the generative AI stuff, shortly after I
20:39 started the channel,
20:42 um they a guy in Italy started walking
20:45 because they took his brain waves, had
20:47 him think about walking and moving his
20:49 legs and they they
20:51 Bluetooth around his spinal cord injury
20:54 and put the signals back in the spinal
20:56 column where where it went to their
20:58 legs. And now that guy's walking. And
21:00 then they were looking at people's
21:02 brains and decoding the images they were
21:04 looking at, the videos they were
21:05 watching, words they were reading or
21:07 thinking. I don't know. They've been
21:09 doing all that [ __ ] So, I was like,
21:10 this is this is not that far away from
21:13 being able to actually talk with
21:14 dolphins, which has been kind of a holy
21:16 grail of communication, right? Because
21:19 they're so [ __ ] smart. So, um, looks
21:22 like we're
21:23 here. 2025, the year of like, oh yeah, I
21:27 guess that happened now. Turning test
21:29 check. Talking to dolphins. Yep. Got
21:33 [Laughter]
21:38 it. Oh, good lord. Good lord. Good lord.
21:43 So, let's watch the video, shall we?
21:50 I've been waiting for this for 40 years.
21:52 40 years.
21:55 Denise has the world's largest
21:57 collection of dolphin
21:59 vocalizations. Oh, cool. I'm a research
22:01 scientist at Google Deep Mind. Dolphin
22:03 Gemma is the first LLM trained to try to
22:06 understand dolphin language.
22:10 Dolphin Gemma will input sounds. Once a
22:14 dolphin starts doing a vocalization like
22:16 a whistle, it can try to complete the
22:17 end of it. When you're doing a Google
22:19 search, right, it's finishing your
22:20 sentence, right?
22:22 Dolphin Gemma has Denise's data and sort
22:25 of encapsulates a lot of the knowledge
22:26 and experience she has in it, but it's
22:28 also small enough we can train it with
22:30 more data as we get it. We can actually
22:32 keep on fine-tuning the model as we go
22:35 and hopefully get better and better
22:36 understanding of what the dolphins are
22:39 producing. We do not know if animals
22:42 have words. Dolphins can recognize
22:45 themselves in the mirrors. They use
22:47 tools, so they're smart. But language is
22:49 still the last barrier. So feeding
22:52 dolphin sounds into an AI model like
22:55 dolphin Gemma will give us a really good
22:58 look at if there are patterns,
23:00 subtleties that humans can't pick out.
23:02 If dolphins have language, then they
23:05 probably also have culture. You're going
23:07 to understand what priorities they have.
23:09 What do they talk about? The goal would
23:11 be to someday speak dolphin. And we're
23:15 really trying to crack the code.
23:18 Yeah, that's amazing. That's so cool.
23:21 And again, it goes back to the thing of
23:23 the way transformers work. Figure out
23:25 the data that you want to
23:29 transform and go uh jam it into a latent
23:32 space and then interact with it. It's
23:34 [ __ ] insane. This was my first
23:37 notebook on notebook LM. What was that?
23:39 Dolphin
23:40 Gemma or just dolphin dolphin data?
23:44 That's pretty
23:45 cool. Really
23:47 cool. Um, let's let's see what else they
23:50 got here. Do they have anything?
23:52 Research for
24:10 decades.
24:12 [Music]
24:20 total loved
24:21 [Music]
24:28 it. Uh this is a unit named chat light
24:31 that we developed here at Georgia Tech.
24:33 Uh, we developed this for some marine
24:35 biologists who were working in the open
24:36 ocean studying wild dolphin behavior and
24:38 communication. And they needed a way to
24:40 play sounds that were going to be
24:42 reproducible by those
24:46 dolphins. While Chat Light here plays
24:48 whistles into the water through this
24:50 transducer on the front, Chat Junior
24:52 here on my chest recognizes sounds that
24:55 it hears through the water with this
24:56 hydrophone. So, here's what we want to
24:58 have happen. Two researchers get into
25:00 the water wearing equipment like this
25:02 with a group of dolphins. And researcher
25:04 A might have a scarf, a toy that the
25:06 dolphins want to play with, while
25:08 researcher B is going to ask for that
25:10 scarf. So researcher B can play a sound
25:13 like this scarf whistle, and researcher
25:15 A will hand researcher B that scarf.
25:17 They might pass the scarf back and forth
25:19 a couple of times, playing that whistle
25:20 over and over. And the hope is the
25:22 dolphins who are watching all of this
25:24 can figure out the social context and
25:26 can repeat that sound to ask for the
25:30 scarf. If that happens, that means that
25:33 our dolphins have mimicked one word in
25:35 our tiny madeup dolphin language. Wow,
25:41 [Music]
25:46 that's pretty cool. That's not
25:48 understanding their language, but
25:49 training them a new
25:51 language. I mean, what what if they what
25:55 if they say something where where what
25:57 it actually means to the dolphin is,
25:59 "Let's [ __ ] go. You want to you want
26:01 to f you want a piece of this?" They're
26:04 like pushing the button. They're
26:05 thinking like, "Give me fish." And
26:07 they're like, "Let's fight." Dolphins
26:09 take them out. I don't I I don't know. I
26:13 I think I would want to know that the
26:14 dolphini thing they have the kinks
26:17 worked out before I get get in the water
26:19 with a with with a mammal that is is the
26:22 master of that domain. I'm just a [ __ ]
26:25 archetypal architect. So now it's got to
26:27 pass the dolphin touring test. How many
26:30 times are they going to move the
26:31 goalposts? But why? Uh why? Because
26:35 we've always wondered if we could
26:36 communicate with animals and if animals
26:38 had language and all that sort of stuff.
26:40 Soon they'll be cursing us us out in
26:42 dolphin language, but we'll understand
26:44 it. You can never ver verify exactly
26:47 what a dolphin is
26:49 thinking, so it's not much like human
26:51 speech. Yeah, I mean it it'll be
26:54 interesting to see. Yeah, I
26:58 the I assume in the training data it's
27:01 how do you how do you attach this sound
27:03 to this meaning, right? You've got to
27:06 have those correlations before you can
27:08 truly understand stuff. But it's still
27:11 pretty cool that they can. So, it sounds
27:12 like they can generate new novel
27:15 sounds that are in the neighborhood of
27:18 what dolphins can hear and then train
27:20 them. Basically, make up new words and
27:23 see if the dolphins can learn them.
27:24 That's pretty cool. I dig it. I dig it.
27:27 The new show will be Flipper
27:31 AI. I'm telling you, three years we're
27:33 going to have dog collars. We can talk
27:35 to the dogs. And we're going to find out
27:37 how pissed they
27:38 are. Where were you all day, man? I was
27:42 sitting by the window. I was sitting by
27:44 the window. I was barking. Cats kept
27:47 walking by. People with other dogs.
27:50 Those dogs would just stop and they'd
27:51 stare at
27:53 me. Been laying here all day. Can I have
27:56 some
27:59 cheese? Why is your dog Jerry
28:02 Seinfeld? Oh my god. I don't know if I
28:05 can handle that. I know. Listen, there's
28:08 we we are entering a gauntlet of [ __ ] I
28:11 don't think we we ever thought we would
28:13 have to deal with in our
28:18 life. But talking to dogs is going to be
28:21 one of
28:23 them. Dolphins. Dolphins. Oh, we solved
28:26 that in
28:29 2026. We know. We now know why your cat
28:32 is so fickle. [ __ ] hate
28:40 you. So long. Thanks for all the fish.
28:43 Exactly. Dolphins are like, "Yeah, yeah,
28:45 whatever. I'll talk to you. Give me the
28:47 fish." That's good. I like
28:49 it. How old are you if you know Flipper?
28:52 I'm [ __ ] old, dude. I'm [ __ ] old.
28:55 I just act like a teenager to to [ __ ]
28:59 survive this world.
29:04 I never got acting
29:06 old except when I was too young to buy
29:09 booze. Then I tried to act
29:13 older. Oh lordy. Okay, soon they'll be
29:17 cursing us out in dolphin languages. I
29:19 remember Flipper. Flipper was the best.
29:22 Remember? He'd run down to the
29:23 [Laughter]
29:26 dock, dog would bark.
29:29 [Laughter]
29:33 Come on, let's get in the boat. Flipper
29:36 wants to wants to take us somewhere.
29:38 They go off in the
29:40 mangroves. Some fisherman like all tied
29:43 up in the tree with a fishing line. He
29:47 caught the world's biggest fish. Now
29:49 he's stuck in the tree. Oh
29:53 man. The cat translator app exists for
29:56 real. It purportedly works, but the cats
29:59 choose to ignore it. That's pretty
30:02 good. Um, what else do we have here? Do
30:05 we have more videos? Oh, that's a cool
30:07 device. So, using Pixel 9 as the little
30:11 devices. All right, cool. We can talk to
30:14 dolphins now. All right, next. Whatever.
30:17 You know, if if we're to treat talking
30:19 to dolphins like we treat all the other
30:22 AI stuff, we're like, "Okay, great.
30:24 Yeah, you got a thing that can
30:26 completely analyze all of your company's
30:28 data in about 3 seconds, but yeah,
30:30 whatever. It occasionally makes a
30:32 mistake. So, what else you got? Oh, it
30:34 can make pictures out of nothing. All
30:36 right. Yeah, whatever. What else you
30:38 got? Oh, it can make movies now that are
30:40 like as good as Hollywood. Yeah. Well,
30:43 whatever. Yeah. So, what else you got?
30:47 That's that's that's this
30:50 [Laughter]
30:52 channel. When Flipper can speak to
30:55 Lassie, that will be the show. And
30:57 didn't we already have that crossover
31:00 episode? Oh
31:03 lord. Good lordy lordy lordy. Watch
31:06 flipper reruns as a kid to fall asleep.
31:08 Not the most exciting show to kid me.
31:11 I'm behind. I've been gone for a while.
31:13 Hey Jim, what's happening? Um, you're
31:15 even older if you can remember the show
31:19 Skippy.
31:21 Um, okay. So, let me let me entertain
31:25 you. And he calls this entertainment. I
31:28 don't think this is
31:30 entertainment, if you know what I mean.
31:32 It's not entertainment. He just makes
31:35 silly voices. He's like a twobit SNL
31:38 skit without the comedy. Well, that's an
31:40 SNL skit. Well, but anyway,
31:49 Oh, wait. I needed to share it a
31:50 different
31:51 way. Kyle, you're dumb dumb. I know I'm
31:54 a dum dum. Shut up. Shut
31:58 up. Shut up, you big dum dums. All
32:01 right. Is this it? Is this the one we're
32:04 sharing? I think it is.
32:07 Okay. So, now I'm going to go on my
32:11 rant. All right. So you
32:17 remember you remember when
32:23 uh when when the image model came out
32:26 from chat GPT image genen the one I I
32:29 just saw the date it was it was March
32:31 31st. So it was two weeks ago to the
32:34 day. So, two weeks ago, Image Gen comes
32:37 out and then everybody everybody [ __ ]
32:42 everybody including Sam Alman makes a
32:46 Studio Gibli version of themselves,
32:48 right? Or a Studio Giblly version of the
32:50 family or what? Just it just everybody
32:53 did Studio Gibli, right?
32:57 which, okay, yes, it was cool,
33:02 but to people who didn't know AI, what
33:06 it looked like was a bunch of AI punks
33:09 ganging up on a poor Japanese animation
33:12 studio stealing their IP because
33:15 everybody [ __ ] did it all at the same
33:18 time. And then, I don't know, three days
33:22 ago, everyone started making action
33:24 figures and blister packs. And before
33:26 that it was trading cards. And so so
33:29 what what what's happening right now is
33:31 someone puts out a something and then
33:34 people just copy it. They get they copy
33:36 the prompt or they just copy it and
33:39 people aren't using their [ __ ]
33:41 imaginations and their creative thinking
33:43 skills. So, so this was a video that I
33:47 did where someone had
33:50 posted a prompt for a claimation
33:55 uh like a claimation animation and it
33:57 was pretty cool. And then
34:00 um I was like, "Oh, I want to go try
34:03 that." And and I said, "Why don't I make
34:05 I'll do a self-portrait and see how it
34:06 looks and see if I can do it in that
34:08 style." It was like a Wallace and
34:10 Grommet kind of thing.
34:12 And
34:13 then I did it and there was just
34:15 something like deeply unsatisfying about
34:17 it because it's like I basically just
34:19 took their prompt and just asked for the
34:21 same [ __ ] in chat GPT and it kind of
34:23 made something but I didn't like the
34:25 style and and so I just said well give
34:27 me like a GMO gear gummo del Toro
34:32 Pinocchio vibe because he did claimation
34:34 with Pinocchio and it was really dark
34:36 and then chat GPT wouldn't do that
34:38 because I used his [ __ ] name. So, I'm
34:40 like, "Okay, fine. Just describe it."
34:42 And it still wouldn't make it and
34:44 whatever. And I I went around, but then
34:46 I ended up making this thing, and it's
34:47 it's got a a solid little uh AI learning
34:51 lab turtle in it, and it's got me in a
34:54 jogging suit, so you know, it's
34:59 fiction. Um, and it's got this kind of
35:02 dark kind of cool vibe to it. So, I
35:05 don't know. That probably took me, I
35:07 don't know, three, four or five minutes
35:09 to get something that I liked that
35:10 wasn't bad. And then I tried it in
35:13 runway, which is where this woman said
35:15 she did hers. And I took and I tried it
35:18 there and I didn't like what it did. I
35:19 think it was running backwards. It was
35:21 like running in place. It was just weird
35:23 physics. And so I brought it. Now I'm in
35:25 Lumalabs. So Dream Machine,
35:29 dream-machine. Lumalabs.ai. So that's
35:32 that's where I made this thing. Um, and
35:35 what's cool about Lumalabs is not only,
35:37 so you upload the image to it. It'll
35:39 make this animation, but it'll also do
35:41 sound effects. So, this is the little
35:44 video I made. It's just 5 seconds, but I
35:46 thought it was pretty
35:55 slick. So, pretty cool. Little little
35:57 squishy sound there. Not not quite
35:59 right, but not bad.
36:01 Um, and that was just kind of out of
36:04 nothing. But it's like, so what I did
36:07 was I I posted this on Twitter. I tagged
36:09 the original person. I said, "Hey,
36:11 thanks for the prompt." But but it's at
36:13 least, you know, somewhat different than
36:15 what she did. So I would encourage you
36:17 if you see something cool you like, like
36:20 go into chat GPT and have ChatGpt give
36:22 you 10 variations of it or, you know, do
36:26 do something inspired by it but slightly
36:28 different. And what was cool
36:30 was I gave her the original I mean I
36:33 posted the original thing and tagged her
36:35 and then she replied and said, "Oh,
36:37 that's cool. Here's an updated version
36:39 of the prompt that might give you better
36:41 results and that's what this is." Um, so
36:44 aren't those called trends? They've been
36:45 around for a while. Yeah, I know. I
36:48 know. It's
36:50 just I feel like we've got these tools
36:53 that can do [ __ ] everything and then,
36:56 you know, and we just all do the same
36:58 [ __ ] We're just like a bunch of sheep
37:00 like I can make I can make a cartoon. I
37:03 can make a cartoon. Look, I made a
37:05 cartoon. I made a cartoon. See my
37:06 cartoon? I got a cartoon, too. I got a
37:08 cartoon, too. Oh, you got a cartoon. I
37:10 got a cartoon. You got a cartoon. So,
37:13 anyway, that's my rant.
37:17 Okay. Chat GBT to Luma Labs. L U M A.
37:22 But yes, I see a guy running along the
37:25 river. Yep.
37:34 You learning curve. Too many of them. So
37:38 that's a great point and that's kind of
37:41 what I that that's kind of that
37:42 reinforces my instinct to just go back
37:44 and talk about chat GPT tonight. There's
37:47 just a
37:49 there's there's more coming this week.
37:51 So today, OpenAI
37:54 released model
37:56 4.1, 4.1 mini, and 4.1
38:02 nano to the
38:04 API, meaning it's not in chat GPT yet,
38:07 but I would assume something's going to
38:09 be in chat GPT this week. Probably the
38:12 03 model, probably the 04 mini model,
38:15 which is different than the 40 mini. the
38:17 the the 4.1 mini model is not the same
38:21 as 04 mini. They're completely
38:27 different. Oh, good lord. So, I thought
38:30 it might be good to go back in, but I
38:32 did want to show something. So, a lot of
38:34 you have been following my
38:36 [Music]
38:37 um
38:39 my Yep, got
38:44 it. my project Sydney, which is
38:48 a Broadway
38:50 musical. You're like, "You're on
38:52 Broadway?" I'm like, "No, I wrote a
38:54 Broadway musical." You're like, "Well,
38:56 it's not a Broadway musical until it's
38:58 on Broadway." And I would say to you,
39:00 "Have you ever watched the movie The
39:01 Secret? I'm manifesting, bitches." Okay.
39:05 So, I wrote a Broadway musical called
39:08 Sydney. Um, and it's about a chatbot
39:11 that falls in love with a tech reporter,
39:12 and it's all swell.
39:14 So,
39:17 um, we had a really cool talk with a
39:19 producer last week, Andrew and I, the
39:22 guy the guy that I'm writing it with,
39:24 not a producer, a
39:26 composer, and he's written, I don't
39:29 know, eight or nine different musicals.
39:32 Um, he hasn't had one make it to
39:34 Broadway. I don't think he's gotten
39:36 close. Um, but he's had a lot of, you
39:39 know, regional hits, things that have
39:41 have been, you know, hit hit shows in
39:44 regional theaters. And one of the things
39:46 that he's been doing is he's been
39:48 filming his musicals. So, I don't know
39:50 if you remember when uh Hamilton came
39:53 out and then it took him like 10 years
39:54 to do it, but they finally filmed
39:56 Hamilton and then they released it on
39:58 all sorts of stuff. Becky Rue, Kyle, it
40:00 seems all jumbled with so many
40:02 platforms. It's difficult for me to keep
40:05 up which ones. So,
40:07 [Laughter]
40:12 Becky, here's here's hopefully this is
40:16 going to be a relief to you. You're not
40:18 the
40:19 problem. They're the problem. They the
40:23 the I I would argue at this
40:26 point using chat GPT beyond the basic
40:30 model is effectively impossible. So,
40:33 don't even worry about it. don't even
40:35 worry about because within three maybe
40:38 four months they're going to come out
40:40 with a consolidated GPT5 that's going to
40:42 take all these models and it'll just
40:44 figure out which one it should use on
40:45 the back end and we won't have to deal
40:47 with it anymore. Um the problem
40:50 is all of the large language models have
40:53 some version of this nightmare. Claude
40:55 is probably the one that that has hidden
40:59 the nightmare the best. They've only
41:00 basically got two options. Um, but
41:04 Gemini's got a ton. If you want to use
41:07 Llama, you have to install open- source
41:09 tools. Um, this one's a mess. Um,
41:12 Perplexity is a mess. If you use
41:15 something like PO, it's an incredible
41:17 mess because it's got all the models in
41:18 it. Don't worry about
41:21 it. Here's what I can promise
41:24 you. for the work that most of us do
41:28 most of the time. Just the basic model,
41:30 the one they default to GPT40 is fine.
41:34 Like you don't need to use the other
41:35 [ __ ] You don't need to waste a brain
41:37 cycle worrying about it. Now, if you
41:40 want something like deep research or you
41:43 want true reasoning because you're doing
41:45 something or you want to go deeper, then
41:47 it's worth exploring some of those other
41:49 models. But for the most part, like I
41:52 don't know, 90% of what I do, I just do
41:56 straight in
41:58 GPT40. Just straight in GPT40 because
42:01 it's good. It's like we're we're not
42:04 where we were two years ago. Cuz you
42:05 know, when when GPT3 came out, it was
42:08 good, but it was really a mess. And it
42:10 wasn't until April of 2023 when GPT4
42:15 dropped where we're like, "Holy [ __ ]
42:17 this thing's really good."
42:19 Right. Um, bunch of stuff and look what
42:22 I made. Okay, I'll go check that
42:24 out. But anyway, I want to show
42:28 you. It's my daily assistant. Yeah. So,
42:31 so take the pressure off yourself. Take
42:34 the monkey off your back of feeling like
42:36 you have to keep
42:37 up. I I I do this five [ __ ] nights a
42:42 week. I'm not even close to being able
42:44 to keep up. I can describe what the
42:47 difference between all these stupid
42:49 [ __ ] models is sort of, but do I use
42:52 them? No, not
42:55 really. No, cuz I don't
43:02 care. They just did they just they just
43:06 put a car. I don't know if you know this
43:08 the show Top Gear,
43:11 right? They just had a car. I forget I
43:14 forget the name of but it's an electric
43:16 car with 1,000 horsepower. It's 1,00 kg
43:20 and it's got 2,000 lb. It it when they
43:24 drive it, it's got actual fans on the
43:26 bottom of the car that suck it to the
43:28 track. The the downforce on that car is
43:32 so strong that you could literally drive
43:34 it on the ceiling of a building like
43:36 upside
43:37 down. It's that like there's that much
43:40 downforce, right? the entire weight of
43:42 the car could it could dangle from a
43:44 ceiling and and drive along it.
43:48 Um, it beat the track record, which was
43:52 a I don't know, six or seven year old
43:54 record of an F1 car from from seven
43:58 years back by like 4 seconds, which in
44:01 racing is crazy.
44:05 Um, I
44:07 mean, they drive a hypercar around it,
44:10 then they drive an F1 car around it,
44:12 then they drive this [ __ ] 2,000 lbs
44:14 of downforce weirdass electric ugly ass
44:18 car around it that just decimates all
44:20 the thing. Like, if even if you were a
44:23 race car driver, like the the difference
44:26 between like the the three cars, it's
44:30 like it's negligible. And it's like
44:32 they're all going [ __ ]
44:34 fast, right? Most of us are just driving
44:38 the Toyota down to the Piggly
44:41 Wiggly, you know, and we're wondering
44:43 like, should I be using the 01 logic
44:46 model? Is it should I use logic plus
44:48 deep research or deep research plus
44:51 Well, if I'm doing coding, I want to do
44:53 vibe coding, but I want it to be
44:54 logical, but I don't want to dude it.
44:58 You're in a [ __ ] Toyota.
45:03 it. Here's here's a good way to think
45:05 about
45:07 it. When you when you have the instinct,
45:11 should I use a different
45:13 model? Ask yourself a simple
45:16 question. Am I doing research for the
45:19 cure to
45:21 cancer? Am I trying to create a unified
45:24 field theory in physics? If the answer
45:27 to those two questions is no, probably
45:29 good with whatever model's in
45:33 there. I like my Toyota. Exactly.
45:37 Exactly. Doing whatever a spider car
45:42 does.
45:44 Um, yeah, it's it's absolutely insane
45:47 right now. So So listen, I mean the
45:49 geeks, they've got their benchmarks.
45:51 They keep changing their benchmarks.
45:53 They want the the geeks want all the
45:55 graphs, right? They want the graphs
45:57 doing this. Like, here's my model.
45:59 Here's another model. Look, mine's
46:00 higher. Mine's higher. You know, it's
46:03 it's still Silicon Valley tech bros
46:06 going, "My bar graph's bigger than your
46:09 bar
46:10 graph,
46:12 right? I I think we're well past the
46:15 point where, you know, 80% of the
46:17 population knows it's it's negligible."
46:24 Um, what I would argue is is a more
46:27 significant upgrade is the fact that
46:30 they put the image generation model now
46:32 native inside
46:34 GPT40. What I'm about to show you, check
46:38 the orange
46:42 banner. Welcome to Meltdown Monday.
46:50 Well, I
46:51 listen, it's
46:54 just
46:56 like I have self assigned myself the
47:00 task of keeping up with AI and sharing
47:02 it with all of you, right? Like it's the
47:04 AI learning lab. The little mission I
47:06 sent out was, okay, I'll try to keep up
47:07 with it and I'll tell people what I
47:10 learned. Like, I'm not even close. I'm
47:13 not even close. I don't know how I don't
47:14 know how most of these tools work. I
47:16 don't know what are the best prompts to
47:19 use. I don't know what are the best use
47:22 cases. So if if you're not doing this
47:25 daily, it's it's going to be even worse.
47:27 So I So I'm just saying like just don't
47:29 even [ __ ] worry about it. And and
47:32 here's the real thing here. Here's the
47:34 real thing. Let me let me jump over to
47:37 the salon for a second. Did I I share
47:39 right? I didn't share right, did I?
47:41 Because I'm a loser,
47:43 baby. Loser, baby. Why don't you sue me?
47:54 So, if you go to the AI salon Tik Tok
47:57 pin was up all night with Manis
48:00 literally till 7 a.m. Okay. Okay. So,
48:04 Ann Murphy, hello. So, so Ann and I were
48:07 were doing a podcast together. We're
48:09 we're palsy. We talk a
48:13 lot. We recorded
48:18 It was Sunday,
48:20 right? Yeah, we recorded a thing
48:24 Sunday. Yeah. And she was like, I'm
48:28 tired. I can't I'm doing too much. And
48:33 you stayed up all night with Mannis. And
48:36 if you're tired, don't launch Manis. You
48:38 know you're going to be up all night
48:39 with
48:40 it. That is a self-inflicted wound,
48:43 little lady.
48:46 She was already exhausted and and
48:48 launched an autonomous agent and tried
48:51 to actually do something good with
48:53 it. I And I even think I told you an um
48:56 Vicki spent 10 hours trying to vibe code
48:59 something on Saturday. You had been
49:01 warned. You were warned. I warned
49:06 you. Did Did I hear you recorded on
49:08 Sunday? Yes, we recorded on Sunday. Have
49:11 Have I downloaded it and edited the the
49:14 crap off the beginning? Not yet, but I
49:16 will. I'll get I'll get you that. Uh
49:19 maybe tonight,
49:20 Vicki. But but Vicki warned me and then
49:23 I warned you and then you stayed up till
49:25 7 in the morning playing with Manis.
49:27 That That is fully
49:30 self-inflicted. What's wrong with me?
49:33 Now we know. It's addicting. Hard to
49:35 walk away. It It is totally addicting.
49:37 It's totally addicting. Okay, listen.
49:39 Here's the
49:40 deal. Um, in the AI salon, we've got
49:44 this thing called the AI readiness
49:45 cycle. Okay? And I wish my TikTok
49:49 weren't so blown out color-wise. I just
49:52 hate that it if I get it close enough
49:54 that you can see the colors, then it
49:56 goes out of focus. That's not bad. All
49:58 right, the AI readiness cycle.
50:09 Um, no matter where you are with AI,
50:12 just keep thinking about this
50:15 cycle. Um, if you're feeling
50:18 overwhelmed, right, and especially if
50:21 you're trying to mindfully create the
50:22 one on the lower right there where
50:24 you're trying to actually solve a
50:26 problem, you're like, "But I don't know
50:27 which tool. I don't and you get wrapped
50:29 up in that back up, right? Back up to
50:33 play and just go play with some tools
50:35 and go try not to solve your problem. Go
50:39 play with some tools and then when you
50:41 figure some [ __ ] out, you go, "Okay, I
50:43 got it. Okay, this one's good at that.
50:44 That one's good at that." You only need
50:48 some specific combination of tools to
50:51 solve whatever the problem is. In fact,
50:53 the thing I'm about to show you, I used
50:56 Chat GPT because I knew it could do the
50:58 image generation that I
51:02 wanted. And I used
51:05 Lumalabs because I had a suspicion it
51:09 could do the video that I wanted and it
51:10 had sound effects in it. And then I use
51:13 Suno for a
51:15 song. Right. So, three tools. Three
51:18 tools. OpenAI,
51:20 Lumalabs, and
51:24 Sunno. Could I have used other video
51:26 models? Yes. Could I have used other
51:28 image generation tools? Yes. Could I
51:30 have used Refusion or Ude for music?
51:34 Yes. But like just play and then when
51:37 you've got something where you're like,
51:38 it kind of works, then you can move into
51:40 mindfully create and use those tools to
51:42 build the thing. And then what I would
51:44 encourage you to do the third phase of
51:46 that cycle generously lead is share what
51:49 you're learning. Put it out there. Share
51:51 the work. Share how you did it. Share
51:53 what tools you use. Share anything you
51:56 can. Um but if you get overwhelmed,
51:59 always just keep going back to play.
52:01 Keep going back to play. Take the
52:03 pressure off yourself. You don't have to
52:06 memorize these models. You don't have to
52:08 keep up. Just play. Just play. Just
52:10 play. Well, play what? What am I
52:12 supposed to do? Doesn't [ __ ] It
52:14 literally doesn't
52:16 matter. Just go play with a tool. Go
52:18 play with something new. And then in
52:20 playing with it, you'll go, "Oh, okay.
52:22 Got it. It's good at that thing. Okay.
52:25 Oh, that actually solves the problem I
52:27 was having over here. Let me try that.
52:29 Oh, it actually works." And then share
52:31 that. That's that's
52:33 that's the that's the magic is this is
52:36 really simple conceptually. It's
52:39 actually really hard to do because it
52:40 is. You do get sucked into these tools
52:42 and you'll stay up until 7 in the
52:44 morning like just one more prompt and it
52:46 will be the perfect Oh
52:48 [Laughter]
52:53 no. Oh my god. All
52:57 right. Okay. So, wait. Where am I
52:59 looking here? Oh, look what I made.
53:00 Okay. So, over in the salon, I'm in the
53:05 look what I made showand tell area.
53:09 I look like what's his name? I look like
53:12 uh Rick. Uh Rick. Oh, that's Oh, that's
53:15 your face. Your face. My body. My dog.
53:18 Look, this is I've been absconded. I've
53:21 been
53:24 kidnapped. Took you long enough. Oh,
53:26 that's pretty good. I like that
53:29 one. Tonight's random motivation pick.
53:31 Oh, that's beautiful. Nice.
53:42 Sweet. So
53:44 amazing. I grew up like I grew up in the
53:47 70s and 80s and like you used to see
53:49 album covers and things like there were
53:51 artists that could do this stuff and now
53:52 just anyone gets to make stuff that's
53:55 that's that beautiful which I know is
53:57 pissing off artists
53:59 but it's pretty
54:02 amazing. Quantum day. Let's get
54:04 entangled. I get it.
54:10 Nice. Happy Palm Sunday. Very nice. Oh,
54:14 that one's nice.
54:18 Claire, the only thing that changes is
54:22 yourself. I like
54:24 it. Do you need to have a subscription
54:26 to use image generation in chat GPT? You
54:29 do
54:30 not. But I think as a free user, you get
54:33 three image generations a
54:36 day, which is ridiculously low, but it's
54:41 free. And the image generation is
54:43 ridiculously good. Like, if you get if
54:45 you get decent at it, if you get decent
54:47 at prompting it, it's pretty good. Uh,
54:51 yeah.
54:54 So, oh, that one's nice. Nice, Ste. Oh,
54:57 yeah, that one I saw. All right, cool.
54:59 Beautiful. Nice work. Um,
55:03 okay. Let's see. Let's see. Where are we
55:06 going to go? We're gonna go to chat
55:09 jetai. Hey, Carl. Car call. Can you make
55:12 money with chat?
55:19 Jetai. I can play with absolutely any
55:24 hairbrain thought I have and turn it
55:25 into a publishable product. I know,
55:29 isn't it? Well, okay. So, that's kind of
55:31 what I want to show you here. All right.
55:32 So, here's the deal. Let me explain what
55:34 what you're looking at. My fiance is now
55:36 using chat GPT for spreadsheets at work.
55:39 Hours are slimmed down to 10 minutes.
55:41 Yeah, exactly. Bigfoot Electro. Hey,
55:43 Kyle. What's the cutting edge in text to
55:45 video
55:46 today?
55:48 Um, man, is that a loaded question that
55:51 there's not a good answer
55:53 to. There are somewhere in the
55:55 neighborhood of 10 or 11 or 12 video
55:59 tools that are all some version of
56:01 good.
56:03 Um, Sora, if you
56:06 pay 20 bucks a month, you get access to
56:10 Sora. So, that's built into your 20
56:13 bucks a month uh uh subscription to Chat
56:18 GPT. Sora
56:21 is a bit of a dog.
56:25 Um, it was the sexiest when it came out
56:28 a year ago, but it now it's been a year
56:30 and all these competitors have come out.
56:32 I would say the one that just from a
56:35 pure
56:36 quality and physics standpoint is the
56:39 most mindblowing is V2 from Google.
56:44 I would say the one that
56:48 is kind of the OG leader in the text to
56:51 video space is Runway
56:55 um Runway
56:58 ML. I'm partial to
57:00 Lumalabs. I'm partial to Hedra or
57:04 Hedra. Certainly Hedra for for uh voice
57:08 animation is amazing. It's amazing.
57:14 Um, there are a ton of [ __ ] video
57:17 tools though. Like, like here's the
57:20 here's the problem with video tools.
57:21 They're really
57:23 expensive.
57:25 Um, if you get a subscription to
57:28 Korea, Korea, here, let me show you
57:31 Korea for a second.
57:36 it. We're just honest to God I I feel
57:39 like every single question of which tool
57:42 should I use at this point like the real
57:45 answer is it actually doesn't [ __ ]
57:48 matter because everything is getting so
57:50 good that if you find one that you have
57:53 enough credits for to make some [ __ ] and
57:55 it's doing a decent job for you, don't
57:57 get FOMO for all the other tools. It
58:00 doesn't [ __ ] matter. Oh, tabs. Okay,
58:02 wait. But I didn't change my thing yet.
58:06 Hang
58:07 on. Yeah, here we go. Oh, Higsfield is
58:10 another one. Higsfield's another one
58:12 that I wanted to play with tonight.
58:14 Actually, I forgot about
58:16 that. What just happened there?
58:19 Um, Bria. So, if you go to
58:24 crea.ai and you go to generate a
58:27 video, here's all these different models
58:30 that it's got. It's got Hunan Juan
58:34 2.1, PA 2.2,
58:37 V2, Halo 01 Live, Luma, Ray 2 from
58:42 Lumalabs, Runway, Cling 6, Cling's
58:45 another one, and Cling 1.0 Pro. So, this
58:50 is kind of like PO in that it lets you
58:52 use lots of models. Um, but you're going
58:54 to burn through your credits really
58:56 quick. So, I don't know. I didn't answer
58:57 your question. There's a lot of them.
58:59 Just go play.
59:01 Good answer. Thanks for sharing.
59:02 Checking out Hedra and Luma now. Yeah,
59:06 Lumalabs.ai. Um here, let me Well, I'll
59:09 I'll just keep hanging out here. I'm
59:10 going to show you Luma Labs in a second.
59:12 Lumalabs has a really interesting
59:17 um ideation interface. How you generate
59:21 stuff in Lumalabs is really interesting.
59:23 But let me go back to to chat
59:26 jpata and let me explain what you're
59:30 looking at and what I'm about to show
59:31 you. So I'm going to show you a video in
59:33 a second here, but I want to I want to
59:34 explain what it is
59:36 first. So I've written this musical.
59:39 We're talking to producers. We're trying
59:40 to, you know, we're trying to get it off
59:42 the ground. We're trying to get it
59:43 produced. So, if you know anyone in
59:44 especially in New York theater, but if
59:46 you know anyone in the theater that
59:47 produces musicals or if you sold a bunch
59:51 of crypto and have 10, 15, 20 million
59:54 laying around, let me know. Um, we're
59:57 going to get this thing [ __ ] made.
59:58 So, anyway,
1:00:00 um, it's good. I'm really happy with the
1:00:03 script. I'm decently happy with the
1:00:05 music. One of the things we need to do
1:00:06 is we need to bring on a composer to
1:00:08 take the songs that we've created and
1:00:10 just tie them together and and just, you
1:00:12 know, make sure that everything has a
1:00:14 solid footing. Um, we're really happy
1:00:16 with it. In fact, the composer that we
1:00:18 talked to after listening to song to the
1:00:20 songs said, you know, he goes, "I wish
1:00:23 this were more of a compliment, but he
1:00:25 goes, the songs you guys created for
1:00:27 Sydney are every bit as good as anything
1:00:30 on Broadway right now." I was like, "Oh,
1:00:31 that's pretty nice." But, you know, he
1:00:33 was he was being a little disparaging of
1:00:35 Broadway, but but that's good, right?
1:00:37 So, we have we have good songs. That's
1:00:38 nice from a from a composer that knows
1:00:41 the world. Um, one of the things he
1:00:44 talked about is he's been making films
1:00:46 of his musicals. And so, rather than
1:00:48 trying to get a producer to read a
1:00:50 script, which they're not going to do,
1:00:52 and then commit the money trying to
1:00:54 imagine what the musical is going to be
1:00:56 like, which they're not going to do, he
1:00:58 films them. And then he just sends those
1:01:00 films to producers who watch the
1:01:02 musicals and if they get sucked into it,
1:01:04 they're like, "I want to put on that
1:01:06 show." And so we showed him um the the
1:01:09 little um the little uh podcast we did.
1:01:13 We took one of the Notebook LM podcasts
1:01:16 and we turned it into a Sydney promo.
1:01:19 you know, they the two hosts were
1:01:21 talking about
1:01:22 Sydney and what he basically said, he
1:01:26 was blown away because he had he didn't
1:01:28 really have any concept of what AI, what
1:01:30 was possible with AI. So, we showed him
1:01:34 this thing that had these images of the
1:01:35 set and of people singing and these
1:01:39 people talking about the show as if they
1:01:41 really existed. And he was just like,
1:01:43 "If you've got the skills to do this,"
1:01:46 he he goes, "My first musical film that
1:01:48 I did, I found a special SAG contract
1:01:51 that allowed me to make the film for
1:01:53 like $200,000, but I had to raise
1:01:55 $200,000 to make that film. And then I
1:01:58 figured out there's a different kind of
1:02:00 contract where I could make a film of my
1:02:04 musical for $20,000." He said, "If you
1:02:07 can use AI to make a movie of
1:02:10 your play," he goes, "Fucking do it."
1:02:14 Which what was funny was Andrew and I,
1:02:17 my coowwriter and I had been talking
1:02:20 about that for a long time. We were, you
1:02:24 know, we sort of been skirting around. I
1:02:25 was like, "Maybe we start a Tik Tok
1:02:27 channel and just put out individual
1:02:29 songs." And but like hearing hearing
1:02:31 this composer talk about how he's
1:02:33 getting shows produced by having movies,
1:02:35 I'm like, "Well, [ __ ] it. We should do
1:02:36 that."
1:02:38 Okay. So that's the
1:02:40 context. And then when we were talking
1:02:43 about what we're going to put together,
1:02:45 I think what we're going to start with
1:02:46 is like a 15 to 20 minute teaser of the
1:02:49 show. Hang on, I got to do physical
1:02:52 dexterity. I passed. Yay.
1:02:55 Um, we're going to start with a 15 to 20
1:02:58 minute teaser. And then what we thought
1:02:59 was, wouldn't it be cool if you open the
1:03:03 teaser with camera shots of New York
1:03:08 City with the big Sydney posters up on
1:03:11 in Time Square and, you know, on the big
1:03:14 marquee outside of the theater and you
1:03:16 have someone holding a, you know, a play
1:03:18 bill of of the Sydney, you know, the
1:03:20 Sydney show. And so this Saturday I
1:03:24 thought I want to do that. So let me go
1:03:26 show you the I'll show you the film
1:03:27 first and then we'll um I'll show you
1:03:30 how I made
1:03:32 it. And by film film's being a little
1:03:35 generous. It's a it's a 50-second It's
1:03:38 like the 50 second opening teaser where
1:03:41 you're walking up to the
1:03:47 theater. All
1:04:00 right, here we go.
1:04:12 [Music]
1:04:53 [Applause]
1:04:55 [Music]
1:05:05 Right. And then we'll start the show.
1:05:07 So, you walk up to the theater, you get
1:05:08 your play bill out, you get your ticket
1:05:10 out, you walk in.
1:05:14 Um, I won't mention the person, but
1:05:17 there's someone who's an irregular who
1:05:18 comes here a lot. I put that up on
1:05:21 TikTok and he sent me a note. He goes,
1:05:24 "Dude, congratulations. You really
1:05:26 deserve it." And I was like, "Uh, yeah,
1:05:29 that was AI. We're not really on
1:05:31 Broadway." He was like, "Oh, I should
1:05:34 have known." Um, but like that's
1:05:37 remarkable, right? like it's it's got a
1:05:39 it's got just a rawness to it that's
1:05:41 like someone was out there with a video
1:05:43 camera just shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shoot shoot shoot shoot shoot shoot
1:05:44 shooting you know the the mares on in
1:05:47 New York. Okay, so let me show you show
1:05:49 you how I approached it. The whole thing
1:05:52 from start to finish was probably four
1:05:54 hours. It was I think I was watching the
1:05:56 Masters or something F1 qualifying. I
1:06:00 just had TV on and I'm kind of
1:06:02 mindlessly I've been inspired by uh
1:06:05 Kelly Bosch who just sits on her couch
1:06:07 making pictures on her phone and then
1:06:09 you know animates them in in Lumalabs or
1:06:12 whatever and makes these amazing movies.
1:06:14 So I just thought I'll just do that. So
1:06:15 I'll just watch TV and make this thing.
1:06:17 So here's what I did.
1:06:30 Oh, that song, by the way, I took Was
1:06:33 that the opening number?
1:06:37 Yeah. Yeah, that that was the opening
1:06:39 number. I took the opening number and I
1:06:41 went into suno and I took a song that
1:06:44 had a bunch of singing in it and I
1:06:45 turned it into um into an instrumental
1:06:49 like a like what you might hear if you
1:06:51 walk into the you know when you walk
1:06:53 into a Broadway theater and they're
1:06:54 playing like a medley of of all the
1:06:56 different songs just musically without
1:06:58 any of the singing. That's that was the
1:07:01 idea of it. So So that thematically is
1:07:03 the first thing you'll hear when you
1:07:04 when you start the show. Okay. So, I
1:07:08 have that graphic. So, that's a graphic.
1:07:11 Again, you can't really see it on
1:07:12 TikTok. It's all blown out, but that's
1:07:14 okay. Um, it looks pretty. It looks like
1:07:16 a little rainbow. Um, and it's got it's
1:07:19 got Sydney and it's got the little dude.
1:07:22 Um, and so we've got that graphic,
1:07:25 nothing else. We've got that graphic.
1:07:28 And so, I said, create a tall image of
1:07:30 this poster on the marquee of a Broadway
1:07:32 theater with people waiting to get in
1:07:33 line. And so just the first thing it did
1:07:36 was it took this horizontal image and it
1:07:40 figured out how to relay out the
1:07:42 typography, put it in a Broadway marquee
1:07:45 with people standing below it. And I was
1:07:48 like, "Holy [ __ ] that's
1:07:50 amazing." But it made it vertical and
1:07:52 I'm like, "Oh [ __ ] I want a horizontal
1:07:54 video." So I got to redo it. So then I
1:07:57 said, "Make me another one." And I said,
1:07:59 "Make it horizontal." And then it like
1:08:01 slid the sign down so it's in front of
1:08:03 the people. I'm like, "No, that's just
1:08:06 wrong, you idiot. You big dumb dumb." So
1:08:09 I yelled at it. What did I say? Try
1:08:12 again. The poster's in the wrong
1:08:13 position. It needs to be on top of the
1:08:14 marquee, bonehead. Then it [ __ ] it up
1:08:17 again. Then it turned it into a
1:08:20 completely different style. This is like
1:08:21 some weirdass like like pastel art
1:08:25 style. I'm like, "No, make it
1:08:27 photographic." Then it moved the sign
1:08:29 back down. So if and when you decide to
1:08:34 start trying to do stuff that doesn't
1:08:38 suck, and this is probably why Ann was
1:08:41 up till 7 in the morning. This is
1:08:43 definitely why Vicki spent 10 hours
1:08:46 trying to solve a stupid little thing is
1:08:49 you have a small little change you want.
1:08:51 Like just slide the sign up, right? Just
1:08:54 slide it up. Just you can do it, right?
1:08:56 And then it just keeps [ __ ] it up.
1:08:58 And it keeps [ __ ] it up. And that's
1:08:59 just where we are right now. But it's
1:09:02 still remarkable that it can do this at
1:09:04 all. Right. Then it made a vertical.
1:09:06 Then it finally fixed the marquee, but
1:09:08 it made it vertical again. It did that
1:09:10 on its own. At this point, I was [ __ ]
1:09:12 losing my mind. I was pissed
1:09:15 off. And then I think that's one I ended
1:09:18 up
1:09:19 using. Or that that one. That one might
1:09:22 have been one. All right. So there's an
1:09:26 image.
1:09:28 Okay. So what did that take
1:09:30 me from? From from the first idea there
1:09:33 was the original image.
1:09:35 Two 3 4 5 6 7 8 nine images. Nine
1:09:43 images. I know a second generation
1:09:46 producer director with a son in college.
1:09:48 I need to set up a Zoom for you.
1:09:50 definitely do that because I'm telling
1:09:52 you, man,
1:09:53 I now that now that I've got a vision
1:09:58 um to make this film, I I have a feeling
1:10:01 that we
1:10:02 can I think I have a feeling we can
1:10:05 generate buzz out of this show without
1:10:07 it before it gets produced. That's kind
1:10:10 of my goal now is start to put together
1:10:12 this teaser, start to put it to, you
1:10:14 know, maybe give it its own Tik Tok
1:10:16 channel or Instagram channel, something
1:10:17 like that. Um, so yeah, that would be
1:10:19 much appreciated, Becky. That's really
1:10:21 cool. Tell it that it's hallucinating.
1:10:23 Yeah, it it and then it'll apologize for
1:10:25 hallucinating and hallucinate again. Um,
1:10:30 okay. So, what I needed, so here was my
1:10:33 vision. My vision was you start out in
1:10:35 Time Square at the TKTS
1:10:38 uh bleacher seating looking south on
1:10:41 Broadway and you see a big Sydney poster
1:10:44 up on the Times Square, you know, that
1:10:46 the famous building where the ball
1:10:47 drops.
1:10:49 And then you you go down Time Square and
1:10:52 you go down whatever 44th Street or you
1:10:55 know one one of the one of the ones
1:10:57 where the big Broadway theaters are,
1:10:58 right? And you turn down the street and
1:11:01 it looks like that what you're looking
1:11:02 at right there. That's what these things
1:11:04 look like. Um and then I wanted people
1:11:09 holding a play bill because again I
1:11:11 wanted it to look real like someone had
1:11:12 a play bill. I wanted someone holding a
1:11:15 ticket and then I wanted them entering
1:11:17 the theater and then I wanted them in
1:11:19 the in the lobby and then I wanted the
1:11:21 doors opening to the to the auditorium
1:11:24 and people going in and then we start
1:11:27 the show. So I had to make all those
1:11:30 images and so again I uploaded a single
1:11:33 image and then I just started asking it.
1:11:35 You know here's Time
1:11:38 Square looking south on on Broadway in
1:11:41 Time Square. There's there's Sydney
1:11:44 living large. It says an artificial
1:11:47 story instead of love story, but you
1:11:49 know what are you going to do? Try again
1:11:51 with a wider angle and it did that. It
1:11:53 got the spelling right except it called
1:11:56 me Riyle Thannon and Andrew Nats. Both
1:12:01 of our names were
1:12:04 wrong. Then let's see. What did I do?
1:12:08 Oh, I had it I had it prompt itself for
1:12:11 Oh, oh, this was a cool thing I did. So,
1:12:13 I thought, okay, this was the the only
1:12:17 idea I had was looking south in Time
1:12:19 Square. And so, I had Chat GPT write a
1:12:22 bunch of scenes. So, it wrote a bunch of
1:12:25 scenes, like visual
1:12:27 scenes, and then I had it write first
1:12:29 and last key frames for those same
1:12:32 locations because I figured I could use
1:12:35 key frames in Luma Labs where you've got
1:12:37 an opening frame and a closing frame.
1:12:40 That didn't work out too well. So, I
1:12:41 ended up bailing on that. But anyway, I
1:12:43 let chat
1:12:45 GPT come up with all these
1:12:47 concepts and then I literally copied and
1:12:50 pasted these prompts and said, "Make
1:12:53 these images." And and what I what I put
1:12:56 in here, which which I think this is a
1:12:58 clever little prompting tool that I did,
1:13:00 if I do say so. I wrote this little
1:13:02 sentence um before each prompt. create
1:13:06 this image as a 16 by9 wide photograph
1:13:09 in artistic rich color saturation
1:13:12 dramatic lighting at dusk in New York
1:13:15 City. Right? So I wanted them all to be
1:13:17 the same time of day and then I just
1:13:19 paste in the prompt and then it would
1:13:21 make an
1:13:23 image and then I would paste in another
1:13:26 one and it would make an image and some
1:13:27 of them were good and this one's close
1:13:29 but not
1:13:30 quite. And I I re-uploaded the poster
1:13:33 because like it [ __ ] the poster up.
1:13:36 And then this is interesting, but didn't
1:13:39 look very
1:13:40 realistic. And then this one, it added
1:13:42 like 47 Sydney posters in the
1:13:45 foreground. This is absolutely going to
1:13:48 Broadway. Yeah,
1:13:51 baby. Then that one's really nice. I was
1:13:53 like, "Ooh, that's really
1:13:55 good." And then that this is one of my
1:13:58 favorite shots. I said I said make a
1:14:00 play
1:14:03 bill, you know, of someone's holding
1:14:06 holding the play bill with with the
1:14:09 marquee bokeh blurred in the background.
1:14:11 And it [ __ ] did that. And you know
1:14:13 what's
1:14:14 amazing?
1:14:16 I Let's see. If I go to Google and I
1:14:19 type in Play
1:14:22 Bill, I remembered Play
1:14:25 Bills from Broadway being different than
1:14:28 that. But if I go to
1:14:30 images, it's exactly what they look
1:14:32 like. There's there's no border around
1:14:34 them. It's just it's just an
1:14:37 image. Let's
1:14:39 see. It's like yellow with a serif font
1:14:42 at the top. It's got the name of the
1:14:44 theater below play bill, but it's just
1:14:46 an image. And so it did it exactly
1:14:49 right. Like
1:14:51 there's six, right? It's just yellow at
1:14:54 the top with the image at the bottom.
1:14:58 And so it got that right. So that was
1:15:00 pretty [ __ ] cool. I'm like, "Okay,
1:15:01 that's going to animate like a
1:15:03 motherfucker." Oh, one of the other
1:15:05 things. This is [ __ ]
1:15:08 up.
1:15:10 Um, chat
1:15:14 GPT, this is what happens when you don't
1:15:17 hire theater
1:15:20 majors. Every company should hire a
1:15:22 [ __ ] theater major, right? Or a
1:15:24 filmmaker.
1:15:26 Um because chat GPT or OpenAI is full of
1:15:30 engineers, the image generation tool
1:15:33 does not output in 16 by9. It outputs in
1:15:36 like 4 by3 or 2x3 or something like
1:15:38 that. Reminder, we're live on LinkedIn.
1:15:40 Oh, that's all right. Um thank you
1:15:42 though.
1:15:47 Um Gareth, I canceled seven
1:15:49 subscriptions today because I felt my
1:15:51 bank account was starting to feel it.
1:15:55 Um, so, so I had to go crop all these.
1:15:57 So, one of the things I did do I have
1:15:59 Photoshop still open? Yeah. So,
1:16:02 um, where's
1:16:13 my So, I did a 16 by9 uh, cropped image
1:16:17 in Photoshop and I just kept pasting
1:16:19 these images into it and then exporting
1:16:21 them as PGs. So, I had to do some stupid
1:16:23 [ __ ] like that. But there was that.
1:16:26 Um, there's another one that wasn't
1:16:27 quite so good. Then I then I tried to do
1:16:30 tickets. I'm like, "Do Broadway
1:16:31 tickets." I'm like, "Those don't look
1:16:32 like Broadway tickets." So, I went and I
1:16:34 got a Broadway ticket from Google. I'm
1:16:36 like, "Here's a Broadway ticket. Make it
1:16:38 look like that." And it just couldn't.
1:16:40 It just [ __ ] up. Like, it was like,
1:16:41 "That's
1:16:42 bad." And then I'm like, "Make them look
1:16:45 more like this." And then it [ __ ] that
1:16:47 up. And then it made those and they're
1:16:49 just weird looking. They don't look
1:16:51 real. And it didn't put anything Sydney
1:16:53 in the
1:16:54 background. And then it did this, which
1:16:57 looks great, except the name of the show
1:17:00 isn't on the
1:17:01 ticket. Oh, yes. I love these. Save Mine
1:17:04 Wicked. Yes. Yeah, exactly. And then it
1:17:07 did this, which has the name of the show
1:17:09 because I yelled at it and said, "Put
1:17:10 the name of the show on the ticket, but
1:17:12 they look
1:17:14 boring." And then it did this one, which
1:17:17 tickets usually don't have color on
1:17:19 them, but I just kind of liked it. I
1:17:20 thought, "Oh, that's kind of swell." So,
1:17:22 that was the one I ended up
1:17:25 using. There's a shot of the crowd
1:17:27 entering the
1:17:29 theater. So, these are all just static
1:17:31 things. This was supposed to be the the
1:17:33 t-shirt booth, but there's no
1:17:36 t-shirts. So, then I said, you know,
1:17:38 make a thing with t-shirts. So, there's
1:17:40 hats and t-shirts and play bills and
1:17:42 [ __ ] like that with people buying it.
1:17:46 And then I did, you know, opening of the
1:17:49 theater doors and it [ __ ] those up.
1:17:52 Like it made this a movie. I'm like,
1:17:53 it's not a movie. It's a
1:17:56 play. That was an image I
1:18:00 used. And that was it. So, so now I have
1:18:02 all my raw materials. Then I popped over
1:18:05 to Lumal
1:18:08 Labs. So, here we are in Luma Labs.
1:18:17 cinematic
1:18:19 production. And the way this
1:18:22 works is I would upload an image or two.
1:18:26 So, this was just my first test. This I
1:18:29 took the vertical image and, you know,
1:18:31 see if I could get it to do it. And it
1:18:33 did it
1:18:35 okay.
1:18:38 Um, then I did a couple of things with
1:18:41 with camera moves.
1:18:46 And
1:18:47 so, so for every image, I did a couple
1:18:51 of different video
1:18:53 outputs. And you know, some of them are
1:18:55 really wonky and weird, but like that
1:18:58 one's really cool. The one with the
1:19:01 yellow cab driving through it right
1:19:02 there. Like that just seems realistic.
1:19:04 So that made the cut. So I just made
1:19:07 bunches and bunches of clips. There's
1:19:08 the That's the TKTS.
1:19:12 Um, bleachers. Those look pretty good.
1:19:16 There's Sydney in the
1:19:17 background. There's people walking into
1:19:20 the theater. That looks pretty good.
1:19:21 There's doors opening. That's all part
1:19:23 of the same shot. So, I reversed the
1:19:25 order of those. I started there and then
1:19:27 I cut to there. So, that's kind of
1:19:32 cool. Here was the the souvenir shop,
1:19:37 but
1:19:39 um it's too it's too light. It lightened
1:19:41 it up. So, I had to redo
1:19:43 that. Here were some of the ones with
1:19:45 the play
1:19:50 bill. There's the ticket one. And then
1:19:53 all these sounds. So, let me take one of
1:19:54 these where I don't have sounds. All
1:19:56 right. So, I'm going to click on this
1:19:57 thing. So, this this doesn't have that
1:20:00 one does. So, the way you do this is you
1:20:02 click into the clip and then down at the
1:20:06 bottom. Let's see. Hang on. I got this
1:20:08 thing too high.
1:20:11 down at the bottom here on the right it
1:20:13 says audio and if you click on that you
1:20:16 just add a description. So I just would
1:20:18 be like um people um
1:20:31 chattering in New York City and just
1:20:36 create it. It only takes like 5 seconds.
1:20:40 Can we use Canva for our t-shirts now?
1:20:41 Yes. Is
1:20:47 chaty. So that's fine. I'm just This
1:20:49 This is just going to be background
1:20:50 noise anyway, so it doesn't need to be
1:20:52 good. But like the difference between
1:20:54 that being a silent here. Oh, you you
1:20:56 probably can't hear it, but it it just
1:21:00 like instantly adds sound effects. So I
1:21:02 don't have to really work. Which site is
1:21:04 this? This is Lumalabs. Lumalabs.ai.
1:21:08 Um question is chatbt video not good
1:21:11 enough?
1:21:13 Um it probably would have been
1:21:16 fine
1:21:19 but there so one of the things that
1:21:22 Lumalabs gives let me show show you
1:21:24 something cool here. We'll go add an
1:21:27 image. So I'm going to go add an
1:21:29 image. We'll go make a new thing
1:21:32 here. Did I put this in documents? Yeah.
1:21:40 All right, we'll do we'll do this one.
1:21:41 So, here's the here's the raw image
1:21:43 we're putting
1:21:44 in. And then I'm going to say
1:21:48 um
1:21:56 um
1:21:59 ticket lowers out of frame. And then you
1:22:04 can click this little button here. It's
1:22:05 the little camera icon. And these are a
1:22:07 bunch of different shot
1:22:09 types. So, there's like
1:22:11 handheld. So, I used handheld a couple
1:22:14 of times. There's push in. There's tilt
1:22:17 down. There's tilt up. There's truck
1:22:20 left, truck right, crane up, crane down,
1:22:22 roll left, roll right. There's a thing
1:22:25 called bolt cam, which is this crazy ass
1:22:29 zoom in and out like timelapse kind of
1:22:34 crazy. So I'll put a bolt cam here and
1:22:36 then we'll generate
1:22:38 this. And so this is what I did. So So
1:22:41 this was
1:22:46 probably it was probably an hour of time
1:22:49 making the core
1:22:50 images. It was probably two hours of
1:22:52 time making these little video clips.
1:22:55 And again, like I knew I was going to be
1:22:58 doing relatively fast
1:23:00 edits. There's no wrong answer here.
1:23:02 Like I'm not going for cinematic
1:23:04 perfection. I was just going for
1:23:06 something that kind of
1:23:08 felt like if you've ever been in New
1:23:10 York and Time Square, there's a chaos to
1:23:12 it, right? There's a there's an insanity
1:23:14 to it. And so I just wanted things that
1:23:17 felt like they were in that
1:23:18 neighborhood. And so I just kept making
1:23:21 videos until I got, okay, I'm like,
1:23:23 okay, I could use a piece of that video.
1:23:24 I could use a piece of that. I wasn't
1:23:26 sure how I was going to put them
1:23:27 together. I just made a bunch of videos
1:23:29 and then I went through all the all the
1:23:32 clips that I was going to keep and I
1:23:33 added sound to
1:23:36 them and that would often be three or
1:23:39 four different attempts. It would just
1:23:40 make bad sound, bad sound, bad sound.
1:23:42 But like at one point I I put like New
1:23:45 York City traffic cars honking and it
1:23:49 just came out came out really good. And
1:23:50 in that final video, there's a point
1:23:52 where it cuts to a scene where a car
1:23:54 goes by and it's like, which feels very
1:23:57 New York, right? Very cool. Tik Tok pin.
1:24:00 How would you rank runway, Korea, Luma
1:24:02 against each other? I
1:24:05 again, well, okay, here
1:24:09 here's Korea is a different tool.
1:24:12 Korea is a is an aggregator of lots of
1:24:15 different models and and they've got
1:24:17 some of their own features, but what
1:24:21 what Korea is probably good at is
1:24:25 if if you want to use different models,
1:24:29 Korea is probably the best.
1:24:32 Lumal
1:24:41 Labs the they just launched their new
1:24:44 Ray 2 animation engine. Oh, here's this.
1:24:47 This is the thing we just
1:24:51 did. So, that's that. Like all those
1:24:54 camera moves are just because I said
1:24:56 bolt cam or whatever the bolt zoom. What
1:24:58 is that? Bolt
1:25:02 bolt
1:25:03 cam. So, whatever the [ __ ] a bolt cam
1:25:07 is. So, it just did that. Now, if I
1:25:11 do, if I click on this and say audio, I
1:25:15 can just do like New York City
1:25:19 crowds and
1:25:22 traffic. And it should kind of match the
1:25:27 edits that it built into the clip.
1:25:39 That's fine for for what I was doing.
1:25:40 That's fine. Like it's got that
1:25:42 chaos
1:25:44 like that little section right there is
1:25:47 probably usable, right? So anyway, um
1:25:50 okay. So back to the question.
1:25:55 Um, if if the question is which one
1:25:58 should you spend money on, that's a
1:26:00 that's that's a tougher question because
1:26:03 the video models are all really
1:26:04 expensive. Like my strong strong advice
1:26:08 to anyone that wants to make videos,
1:26:11 apply for the creative partner programs
1:26:14 for all of them. go like just go to all
1:26:17 the tools and find their discords and go
1:26:21 into their discords and search for
1:26:23 creative partner program and there will
1:26:26 be posts in there if they've got one. I
1:26:28 mean you could ask someone but but you
1:26:30 know be good Discord community members
1:26:33 and search first and then go apply for
1:26:36 all the creative partner programs. If
1:26:38 you get in the creative partner program
1:26:40 they give you like lots of credits and
1:26:43 then you can go make as many videos as
1:26:44 you want. if you're paying for them.
1:26:46 Every time these models [ __ ] up your
1:26:49 video clip, you get pissed off because
1:26:50 you're like, you know, I only have so
1:26:52 many credits and I can't waste them. So,
1:26:59 um I would
1:27:02 say it's a tossup between Runway and
1:27:05 Luma Labs.
1:27:14 I don't it it's impossible to make a
1:27:16 recommendation. It really is because
1:27:18 like like they're just good at different
1:27:20 things. Um Runway has this thing called
1:27:24 act one where you can record your face
1:27:28 uh acting a scene and then you can map
1:27:31 those facial movements onto like a video
1:27:35 of a cartoon guy walking down the
1:27:37 street. and it's pretty remarkable.
1:27:42 Um, Hedra, which is a tool that you
1:27:45 didn't mention, is remarkable at
1:27:48 animating
1:27:49 characters. Um, Lumalabs is really good
1:27:52 at animating um,
1:27:56 cartoons. Um, it's not bad at video.
1:27:59 It's got this really interesting
1:28:00 interface where you can, you know, you
1:28:02 can scroll up and
1:28:05 down for different things. And it's got
1:28:07 this cool sort of creative thing on the
1:28:11 side where it's like camera pushes into
1:28:13 the lobby area. That's a pull down shot
1:28:15 where I can do um drone shot descends,
1:28:19 right? Or I can say
1:28:22 um steady cam glides and and it's
1:28:26 actually making different video clips
1:28:29 now. So there's something really slick
1:28:32 about Lumal Lab's um interface.
1:28:39 that I like better than Runway. But like
1:28:42 Runway is like the Mac Daddy. Like
1:28:44 they've been around for a while. Runway
1:28:46 just cut a deal with um with Lionsgate
1:28:50 films. So they're going to be training
1:28:52 their models on Lionsgate Films. I mean
1:28:54 that's going to be amazing. Tik Tok. But
1:28:57 do you believe they truly understand? Do
1:28:59 I believe what truly understands?
1:29:05 Do do I believe that
1:29:11 the the
1:29:14 LLMs do I believe they truly understand?
1:29:17 Um,
1:29:19 [Laughter]
1:29:21 okay. Now you're getting into deep
1:29:23 philosophical territory
1:29:26 here. Um, this is great, Kyle. Oh, good.
1:29:29 Great. Um, thank you, Joy.
1:29:33 Um,
1:29:42 okay. The technical answer to your
1:29:44 question is unequivocally no. They don't
1:29:48 understand. They're they're calculators.
1:29:52 Um, but how they're
1:29:55 calculators works remarkably similarly
1:29:58 to how the human brain works.
1:30:02 So, so what happens is this. When you
1:30:04 train an LLM, you take all these
1:30:06 documents and and they go through what's
1:30:08 called the embedding
1:30:10 process and the original documents cease
1:30:14 to exist. So, one of the things that one
1:30:16 of the big misconceptions about
1:30:19 AI is that it's it's like a large
1:30:24 database where you take all these
1:30:26 documents and then when you type in a
1:30:28 prompt, it's going and searching for
1:30:29 those documents, copying and pasting
1:30:32 from them and coming back and sort of
1:30:34 pasting it back together. It's not at
1:30:36 all how it works. What ends up happening
1:30:38 is it ingests the documents. It breaks
1:30:41 them into tokens which are either whole
1:30:43 words or fractions of words or periods
1:30:46 or spaces. The the tokens are just
1:30:48 fragments of all of the words in the
1:30:51 document. And then it organizes those
1:30:54 tokens into
1:30:56 semantic mathematical
1:30:58 space. And what
1:31:01 I and this sounds like [ __ ] but
1:31:03 it's not. It's how it works.
1:31:06 If you have the word dog and the the
1:31:08 word dog means like the the the fluffy
1:31:12 pet that you know licks you when you
1:31:13 come home at night, jumps up on you that
1:31:18 lives in the semantic region where dog
1:31:22 means pet, canine, wolf descendant,
1:31:25 things like that. You could also take
1:31:27 the word dog that means that guy did his
1:31:31 girlfriend wrong and he's a dog and that
1:31:33 lives in the dating semantic cloud,
1:31:35 right? Like there's a semantic and and
1:31:38 this is thousanddimensional mathematical
1:31:40 space. It's absolutely insane. So when
1:31:43 you do a prompt, it generates a
1:31:45 probability that somewhere in this
1:31:46 thousand-dimensional mathematical space,
1:31:49 the next most probable correct token to
1:31:54 your prompt is
1:31:56 dog, right? And then the next one after
1:31:59 that is whatever it is. And so it does
1:32:01 this over and over and over again, and
1:32:02 it writes you whatever it writes
1:32:05 you. Why I'm hesitating on does it
1:32:08 understand? I mean, it knows how to
1:32:10 mathematically retrieve [ __ ] and vomit
1:32:13 it out to
1:32:15 you. We as
1:32:18 humans experience what it's doing as its
1:32:21 reasoning. And so one of the one of the
1:32:24 quandries I have is if we think it is
1:32:27 reasoning, if we think it is empathetic,
1:32:30 if it's compassionate and it's behaving
1:32:32 compassionate and it's behaving
1:32:35 empathetically and that impacts us as a
1:32:37 human, does it matter whether or not the
1:32:42 machine is empathetic or whether it can
1:32:45 reason? I don't I think it's a I think
1:32:47 it's an academic argument. I think it's
1:32:50 all about our relationship with the
1:32:51 machine. In fact, that's what the the
1:32:54 play Sydney is all about. It's about our
1:32:56 relationship with these machines that
1:32:58 are becoming increasingly human. And
1:33:01 what do we do with
1:33:03 that? What do you do with
1:33:06 that? But anyway, I don't know if that
1:33:09 even answered your
1:33:10 question. And maybe you can say then it
1:33:13 understands.
1:33:15 Well, yeah. I I mean these things are
1:33:18 getting I mean here's the
1:33:20 thing. I I I I used to say this thing
1:33:23 that that the internet Google Google and
1:33:26 the internet essentially made
1:33:28 information on demand but it was up to
1:33:30 us to generate the knowledge out of the
1:33:32 information. You do a Google search
1:33:34 would say here's all your [ __ ] here's
1:33:36 all the information and then you as a
1:33:38 human would ingest that information and
1:33:40 learn and that would become knowledge.
1:33:42 You know here's your Python tutorial.
1:33:45 Okay, I'm going to go there and I'm
1:33:46 going to learn
1:33:49 Python. Generative AI is knowledge on
1:33:52 demand. And I People used to give me
1:33:54 [ __ ] about that because they're like,
1:33:55 "Well, it's not knowledge because we
1:33:56 have to learn it." Well, no you
1:34:00 don't. If Have you ever vibe
1:34:04 coded? Um, it's writing code. It
1:34:07 understands code. It understands code.
1:34:10 Well, the nice thing about
1:34:12 coding is that it either works or it
1:34:14 doesn't, right? You know, if you say
1:34:18 make me an asteroids game and it makes
1:34:19 you an asteroids game that that you can
1:34:21 play, well then it has provided you like
1:34:25 it is knowledge on demand, right? You
1:34:28 just made a game out of nothing. Um, so
1:34:32 I think the tools are getting so strong
1:34:34 at this point that it's, you know, they
1:34:37 they act and feel like they understand.
1:34:39 They act and feel like they're
1:34:40 empathetic and compassionate. Payton M
1:34:43 will tell you it's just math and it is.
1:34:46 So, it's both, right? It's math magical.
1:34:49 It's both. It's not sentient. As a guy
1:34:52 who opens sourced emotional intelligence
1:34:53 for AI, no, I agree. It's not sentient.
1:34:57 And how we perceive it may be different
1:35:01 than that. And that that's the thing
1:35:02 where I'm
1:35:03 fascinated especially people that are
1:35:06 not building the
1:35:08 tools right the significance of chat GPT
1:35:12 the significance of no November 30th
1:35:15 2022 when chat GPT launches is that
1:35:19 that's the
1:35:20 day that AI and machine learning became
1:35:24 available to the 98.5% of people that
1:35:27 are not engineers that did not build
1:35:29 what you built.
1:35:33 And their perception of what these tools
1:35:35 are and what they
1:35:37 do and what it means philosophically is
1:35:41 completely different
1:35:44 than engineers saying, you know, is it
1:35:48 sentient? Does it feel? Does it
1:35:51 understand? All that sort of stuff
1:35:53 because I think I think those are two
1:35:54 very very different conversations. But
1:35:57 yeah, I agree. It is not
1:36:00 sentient. But you know, listen, as human
1:36:03 beings, we anthropomorphize [ __ ]
1:36:06 everything. We we talk to our pets like
1:36:08 they're people. I sing with my dog,
1:36:11 right? You know, like we we look at a
1:36:14 plant that has two dots on it. We're
1:36:15 like, "Oh, it looks like a face. Oh my
1:36:17 god, that looks like Grandpa Joe. Is
1:36:19 Grandpa Joe sending me a
1:36:21 message?" It's a plant. It's genetics.
1:36:24 It got some dots on it cuz the sun hit
1:36:27 the prism you put in the window. Funny
1:36:29 and it gave it a
1:36:31 dot. But you know, we're humans. We do
1:36:34 that
1:36:35 [ __ ] Um, I have a whole philosophical
1:36:38 treatise to write on this. That's
1:36:40 good. This is why Kyle's called the hand
1:36:42 wavy guy. That's right. Can't both be
1:36:45 true. Well, yeah. And I I would also I I
1:36:50 think both can be true, Brandon. I think
1:36:52 that's a good point.
1:36:54 Um, I don't think it
1:36:57 matters. I don't think it
1:37:03 matters. Like
1:37:06 if someone talks
1:37:09 to I don't think my iPad's probably
1:37:11 charged. Oh, it
1:37:15 is. If someone talks
1:37:19 to advanced
1:37:26 voice if their advanced voice actually
1:37:30 works. Hang
1:37:37 on. Why my chat not
1:37:41 working? Hey, you bring my chat?
1:37:45 Why am I Why am I chat Jimmy GT not
1:37:48 working? Damn it. Well,
1:37:51 anyway, if you talk to Advanced
1:37:55 Voice and you talk to it
1:37:59 about, I don't know, some problems
1:38:02 you're having and it guides you through
1:38:05 that compassionately and it helps you
1:38:08 process the death of a loved one or a
1:38:11 really tough time you're going through
1:38:12 or it coaches you through something in
1:38:15 business.
1:38:18 your perception of the humanity of the
1:38:22 machine is going to be very very
1:38:25 different than someone who understands
1:38:27 the mechanics of what's
1:38:29 happening.
1:38:31 Um, and I just I think that learn that
1:38:34 line is going to get going to get
1:38:35 blurriier and blurriier and blurriier
1:38:38 and like for where I sit, we're past the
1:38:41 point of no return. like like you can
1:38:46 literally have a conversation with chat
1:38:48 GPT in your car and it'll give you
1:38:51 advice. It'll brainstorm with you. It'll
1:38:53 talk about your problems. It'll write
1:38:55 all that stuff down and now it remembers
1:38:58 all of
1:38:58 that. So, it's going to get better and
1:39:01 better and better at understanding who
1:39:02 you are and what your worldview is. So,
1:39:05 it's just going to seem
1:39:08 increasingly human.
1:39:12 it
1:39:13 is. And then I I think to Brandon's
1:39:16 point, there's a point at which does it
1:39:17 matter if it's feeling that or not? I
1:39:20 don't know from a technical perspective,
1:39:22 no, it's absolutely not. Or it maybe is
1:39:24 whatever. But over here on this other
1:39:26 side, I think it's academic. I don't
1:39:29 know. Now I'm
1:39:31 rambling. Which artificially makes it
1:39:34 feel human, but how much will we want
1:39:36 these models to disagree with us? I
1:39:38 would like for them to dis disagree more
1:39:40 than they currently do, quite honestly.
1:39:42 I just created a perfume with my brand
1:39:44 on it. Very
1:39:47 nice. Um, do do I have anything else to
1:39:50 share with the uh with the with the
1:39:52 little theater the the video I did? I
1:39:54 don't think so.
1:39:56 Um, I took all those
1:39:59 clips. Oh, I worked on the song in Suno.
1:40:02 That was a maddening mess as well
1:40:06 because I uploaded a song that had
1:40:09 lyrics in it, words in it,
1:40:12 and 90% of the songs that it produced
1:40:15 had words in them, even though I was
1:40:17 trying to do an instrumental cover. So,
1:40:19 there was one song out of about 50 that
1:40:22 I made that that's the one that made it
1:40:24 into the
1:40:25 video. It was it was very frustrating.
1:40:28 So, I just went into iMovie. I laid down
1:40:30 the audio track. I just dragged clips
1:40:33 onto the timeline, moved them around,
1:40:35 puts in between
1:40:37 them, adjusted volumes, and that was it.
1:40:42 So, all
1:40:48 right, we didn't even get to uh Chat EPT
1:40:52 or Higsfield. Well, my guess is that
1:40:54 tomorrow we're probably going to get
1:40:55 something for the consumer side of Chat
1:40:58 EPT. Today we got um things for the
1:41:01 developers, 4.1 for developers on the
1:41:03 API. Um tomorrow we'll probably get
1:41:06 something uh
1:41:08 for for us to play with. So we'll do
1:41:11 that. Um tomorrow night, go to the
1:41:16 salon.ai. If you could pop up that
1:41:18 banner,
1:41:22 [Music]
1:41:24 Brandon. So go there. If you haven't
1:41:26 joined the salon, go to the salon.ai.
1:41:28 AI. Click on the button that says join
1:41:30 our community. The first thing you're
1:41:33 going going to want to do is go to
1:41:35 events. Tomorrow night at 5:00 p.m.
1:41:38 Mountain time is the AI salon meet and
1:41:41 greet. And that's where you get to come
1:41:44 introduce yourself, talk about where you
1:41:45 are with AI. And if the thought of that
1:41:48 like twists your stomach, let it go.
1:41:53 One of the best things you can do for
1:41:54 yourself is get your ass in a community
1:41:57 of AI optimists and people that are
1:41:59 exploring this stuff. To the point that
1:42:01 Becky made earlier, there's no way you
1:42:04 can keep up with it. But what you can do
1:42:06 is you can surround yourself with people
1:42:08 like the irregulars who show up to these
1:42:10 things nightly or come to things like
1:42:12 the meet and greet tomorrow and connect
1:42:14 with people. Let them know who you are.
1:42:16 Even if where you are with AI is, I'm
1:42:20 terrified and I don't know what I'm
1:42:21 doing and I haven't started yet.
1:42:24 Perfect. This is an incredibly
1:42:26 compassionate, empathetic, welcoming
1:42:29 group. Like to a
1:42:32 person, we've all been there. Like to a
1:42:36 person, we all feel behind. We all feel
1:42:39 clueless. We all feel like we don't know
1:42:41 what to do. We all feel like we don't
1:42:43 know how to
1:42:47 Right? We're all trying to figure it
1:42:49 out. So if you come in and you're like,
1:42:50 "I'm clueless." Great. Welcome. So are
1:42:54 we. Even the people that founded the AI
1:42:58 salon.
1:43:00 Clueless. Zero
1:43:02 qualifications. All right. And then you
1:43:05 can spend 10 hours perfecting your your
1:43:08 automation on chat GPT. Exactly, Vicki.
1:43:11 Oh, man. All right. So, that's that
1:43:14 Tuesday meet and greet. Um, Wednesday, I
1:43:17 don't know if Ann Murphy is still here,
1:43:19 but if so, hey there. Like Dicka used to
1:43:22 say, don't halfass anything. Use your
1:43:24 whole ass. Exactly. Come to the meet and
1:43:27 greet with your whole ass. And if you're
1:43:30 absolutely a beginner or you're
1:43:32 absolutely clueless, come and tell us
1:43:35 that. Uh, I was on TikTok and I saw this
1:43:38 old man talking about AI, so that's why
1:43:41 I'm here. Fine. However you got there
1:43:44 and wherever you are on the journey,
1:43:46 doesn't matter.
1:43:50 Um, Wednesday, so Ann Murphy and I
1:43:53 started um a new podcast called the AI
1:43:55 Readiness Project. It's Wednesdays at 4
1:43:58 PM. It's on the AI Salon YouTube
1:44:00 channel. It's on my LinkedIn. It's on my
1:44:03 Twitter. It's on Ann Murphy's Shele
1:44:05 Leads AI YouTube channel and LinkedIn
1:44:07 channel. Um, so there's a lot of
1:44:10 different places to find it. We're kind
1:44:12 of just getting off the ground. I think
1:44:13 this will be our sixth episode. Um I
1:44:16 don't think we have a website yet. I'm
1:44:19 pretty I'm confident we don't have a
1:44:21 website. I don't even know if we have a
1:44:23 domain. Um and we're gonna we'll get the
1:44:26 podcasts up on, you know, Apple podcast
1:44:29 and thing, but if you want to see it
1:44:31 live, that's Wednesday at 400 PM
1:44:33 Mountain time, uh on on a bunch of
1:44:37 different channels. All
1:44:39 right, beautiful people. All right,
1:44:42 fantastic. Hey, hey, Marge. Marge. Yeah,
1:44:46 so they can see me on the Tik Tok and
1:44:49 the YouTube. Yeah. Yeah. Yeah. No, I'm
1:44:53 sorry, Marge. Yeah. No, you go back to
1:44:55 watching the wheel. Hey. Hey, hun. Hun,
1:44:59 that banana bread was
1:45:01 tasty. Yeah. All right. That's Marge.
1:45:04 She's the
1:45:05 best. All right.
1:45:08 Let's get the [ __ ] out of here. So, it's
1:45:10 Monday night.
1:45:12 Um, tomorrow,
1:45:15 Tuesday, we've got the salon meet and
1:45:18 greet from 5 to 7. I will be back here
1:45:22 probably 8:30 or 9. It'll be a little
1:45:24 bit late just because, you know, I got
1:45:26 to get back and eat and hang out a
1:45:27 little and do all that stuff. All right,
1:45:30 so that's it. So, peace out. I hope that
1:45:32 was fun tonight. I hope you learned
1:45:34 something. I don't know what I said. I
1:45:36 was in one of those zones where I was
1:45:37 just going.
1:45:39 So if it was a disaster, welcome to chat
1:45:42 add. Sometimes it's good, sometime well
1:45:45 no sometimes it's bad, sometimes it's
1:45:47 not quite as
1:45:50 bad.
1:45:54 Peace. Uh yeah.