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

Description

Chapters

Transcript

0:01 [Music]
0:25 Uhoh. Uh-huh.
0:28 [Music]
0:43 [Applause]
0:43 [Music]
1:19 Sitting in this lonely
1:21 town wonder when things are going to
1:24 [Music]
1:26 change. Dream my life
1:29 away. Seems these dreams have
1:32 turned those
1:35 clouds. Get my nerve up. But my last is
1:40 pouring me
1:41 [Music]
1:43 down. Wondering how
1:47 long she going to stick around.
1:55 Somebody told me once before said you
1:58 can never go home
2:00 again. Won't you
2:02 leave? Santa thinks to steer me
2:06 away from the truth of who I am and what
2:10 I believe. So I thanked him for his two
2:13 cents with a
2:14 handshake and some sympathy. Yeah.
2:19 Packed up my blue
2:21 jeans and headed for this big
2:25 prize of my
2:29 freedom.
2:30 Bye-bye black sheep to the black sheep
2:34 of the
2:36 family.
2:38 [Music]
2:39 [Applause]
2:40 Bye-bye. Oh, that means so very much to
2:43 me. Yeah.
2:46 [Music]
2:48 Bye-bye to my friends and my
2:52 family.
2:55 Bye-bye. Going to set my
3:02 soul. Set it
3:04 [Music]
3:06 free.
3:08 Free. Times that were
3:13 changing. Did a little bit of
3:16 rearranging. Hey, what's happening
3:21 everybody? We got some new goodies to
3:23 play with tonight. Howen
3:26 [Music]
3:41 puppy. Woohoo.
3:45 [Music]
3:47 Oh
3:48 [Music]
3:57 yeah. There's been something, baby, I've
4:01 been trying to
4:03 say for an age, and it seems I don't
4:06 know
4:09 how. The past and the future now
4:12 surrounding
4:15 me. Surrender to whatever truth still
4:17 can be
4:20 found. There's been a little
4:23 trouble since you came to my
4:30 rescue. And if you're like all of the
4:32 rest, I would have quit you long ago.
4:35 But I couldn't do that.
4:41 Oh, tell me now. Women in W never went
4:44 too
4:47 well. Make a man crazy, make him cold as
4:52 hell. I'm a woman that you wish me
4:57 well. But it's funny trying. Still going
5:00 to have to find my way through.
5:08 [Music]
5:11 Happy Tuesday night everybody. It is
5:14 Tuesday, right? It is. It's Tuesday. Fan
5:17 friginantastic. Thank you, Mary. Very
5:18 nice sentiments. Did you see the image
5:21 Danielle made of you and Champy? I have
5:23 not seen that. I should go look at that.
5:25 Is it in irregulars? I assume it's an
5:28 irregulars.
5:31 Irregulars. Oh, that's a really nice
5:34 one. Let's go look at that. That's a
5:36 sweetie.
5:38 And where is my There it
5:41 is. Oh, let me flip this little bad boy.
5:44 Look at that. Look how nice that is.
5:47 That's
5:48 gorgeous. It almost looks like him. He
5:51 doesn't quite have the black eyes. He's
5:52 got white eyes, but the ears are
5:54 perfect. The head's perfect. Just a
5:56 little bit less there, but that's pretty
5:58 much what he looks like. And
6:00 unfortunately, that's that's about my
6:03 size, too.
6:08 That is very sweet. Very sweet, sweet,
6:16 [Music]
6:24 sweet. Okay,
6:26 so something actually big happened today
6:30 and I have a feeling most people won't
6:32 pay attention to it.
6:36 because it's really subtle, but it's
6:40 super
6:44 powerful. We got we got Gemini 2.5
6:48 today, which is cool. Supposed to be
6:51 good at
6:52 coding.
6:54 Um, we got another Chinese model or two.
6:57 I saw a tweet about some
7:01 clandestine [ __ ] model that says
7:04 it's better than everything
7:06 else. But the thing we got today that
7:09 I'm excited about G4 long time good to
7:12 see you. Welcome, welcome,
7:14 welcome. Lori Duskin, Ryan, got Danielle
7:18 got source camp on time. Good lord
7:20 people. What is coming of the world? Ah,
7:24 Stacy nailed it. Open open AI image
7:27 generation. Native image
7:29 generation
7:32 native. You're like, what does that
7:35 mean?
7:42 [Music]
7:55 Uh, I don't know 100% what it means and
7:57 I figured we would play with it tonight
7:59 because we've
8:01 got Gemini has a native image model in
8:07 their in their image generation model.
8:10 So, we can play with that and then we
8:12 can compare it to OpenAI. I've been
8:14 doing a little bit of comparing before I
8:16 got on here and it looks like OpenAI is
8:18 better, but they're still both
8:21 native. Um, and then you've got Gro,
8:25 which is fairly
8:27 unguardile.
8:30 Um, most of the image generation tools
8:33 up to this point have had fairly
8:34 stringent guard rails. You can't make
8:36 famous people. You can't make brands.
8:38 And you certainly can't make
8:41 images that feature brands that are not
8:45 um that are not
8:48 uh what would you call it? Um friendly
8:52 to that
8:54 brand. Tom Isaac, I use Simple AI to
8:58 order Dominoes. Oh, interesting. And it
9:00 worked. Simple AI. Isn't that the one
9:02 that that actually calls them up? I did
9:05 that. I did it once. And you know what's
9:07 funny about that, Tom? I um uh I got
9:12 weirded out because I had
9:15 it I had it called I had it called two
9:18 restaurants and asked if they had tacos
9:21 and one was a burger joint and one was a
9:23 Mexican joint and the Mexican joint had
9:26 tacos and the burger joint I was like
9:28 watching the transcript as the call was
9:31 happening and and you know the simple AI
9:34 thing said, "Hey, do you have tacos?"
9:36 And the the person that worked at the
9:38 burger joint was like, "Well, we don't
9:40 right now, but we've been thinking about
9:41 adding him to the menu." It was like a
9:44 whole thing. It had a whole
9:46 conversation, and I felt bad, so I
9:48 haven't used it
9:49 [Music]
9:52 since. Oh
9:56 man. Oh, this is a chat GPT image,
9:59 Danielle. That's super cool. I use chat
10:01 GPT to make that image. Look at the
10:03 comments.
10:04 Oh, you gave it a pic of me and Champy.
10:10 Wow. Oh, this is cool. Hang on. That's
10:13 super
10:14 cool. Oh, yeah. Look at the striped
10:19 shirt. And then here's Champy look
10:23 looking like the little thug that he
10:30 is. And then it made that. That's so
10:33 [ __ ]
10:34 cool. That is so
10:38 cool. Yeah, it did pretty good. It did
10:41 pretty good. And look, I got my Monopoly
10:43 in the
10:46 background,
10:48 huh? Crazy. All right.
10:52 [Music]
10:54 [Applause]
10:56 Well, yeah, we got to play with this
10:58 tonight.
11:00 Um, I did I did a couple of things that
11:02 I think are kind of kind of nuts.
11:06 [Music]
11:12 Um, I think that here's one of my
11:15 theories.
11:17 Um, Sam, so I want to go over and I want
11:20 to look at the at the Open AI
11:22 announcement for this thing. So, let me
11:24 let me pop into X here for a second.
11:27 Let me get the OpenAI
11:31 announcement. Um, Open AI
11:36 image. Let's get their announcement
11:38 because I want to read through
11:42 it. Um, there it
11:47 is. So, we'll read through this
11:49 together.
11:52 This image gen is now in both chat GPT
11:55 and in Sora. So Sora is really
11:58 interesting.
12:01 [Music]
12:11 Um, okay.
12:16 So, I don't quite understand what native
12:20 image generation means other than I know
12:24 that it's within the same model, right?
12:26 So, you've got
12:27 GPT40 and the O stands for
12:30 Omni. They announced GPT40 last March, I
12:34 think. So, this is this has been a year
12:36 coming. And when they announced it, they
12:38 said it had it it was uh
12:42 omnimodal or multimodal, right? It had
12:45 image, audio, video, and
12:48 text. And then when they launched it, it
12:51 was text only and they were still using
12:53 Dali, the external image model that
12:55 couldn't spell and couldn't do stuff.
12:57 And then they launched voice that was an
12:59 external model, right? where you talked
13:02 into it, it translated it to text, sent
13:04 the text to chat GPT, got the answer
13:07 back, and turned that back into speech.
13:09 And then midway through last year,
13:11 sometime over the summer, they announced
13:14 uh advanced voice, which is native
13:18 audio, right? So you you talk straight
13:21 into the model and it understands audio.
13:23 So that's what's happening with these
13:24 images. The model itself understands
13:28 images.
13:31 So, in theory, we should be able to do
13:33 stuff like take a screenshot of a PDF
13:36 with some charts and graphs on it, and
13:38 it should be able to read the text and
13:40 the and the graphs. Um, I don't know if
13:43 they've improved their uh PDF upload, if
13:46 they're still just doing text, but that
13:48 would be interesting. That'll be
13:49 interesting to te
13:51 test. Um, it also looks like they've
13:54 taken off some of the guard rails.
13:57 Um, it before if you wanted to make a
14:00 brand or a person, it wouldn't let you
14:01 do that. It does now. Um, oh, I got to
14:05 do my black
14:06 bar. Um, for for those of you on TikTok
14:10 sharing the live, thank you very much.
14:11 That's much appreciated. Thank you, Tom.
14:13 Thank you, Danielle earlier and whoever
14:15 else did. I think there was a couple of
14:17 you doing that, so that is appreciated.
14:19 I'll try to catch up soon. Nighty night.
14:21 Night night. Becky Kuno's in the house.
14:24 Who else we got? Brother
14:26 52 can't stay. All
14:30 right. Oh no. Oh, a heady scan from a
14:33 car accident back in December. I hope
14:35 you're okay. That sucks. Car accidents
14:39 suck. They just suck.
14:44 Okay. I'm getting mixed results with
14:46 image reading. Yeah, that's that's part
14:49 of what I want to test. And it's funny.
14:51 I'm getting I'm getting some results are
14:54 staggeringly good with image generation
14:56 and some are weird. Like I tried to have
14:59 it do a graphic novel. It it cannot
15:02 count cells on a comic page. Um that's
15:06 one thing it can't do at all. Like
15:09 bad like like how many Rs in Strawberry
15:12 are there kind of
15:14 [Music]
15:16 bad. All
15:19 [Music]
15:46 right, Mr. All right, TE's in the house.
15:48 Everyone's here. We can get this potty
15:52 started. Jeff, it's the biggest drop
15:54 I've seen for someone who loves making
15:57 images. Yeah, this is this is a big
16:00 deal. This is a big deal. So, let's
16:02 let's read
16:03 this. Unlocking useful and valuable
16:07 image generation with a natively
16:09 multimodal model capable of precise,
16:12 accurate, and photorealistic outputs.
16:14 I'll give you that. Here, look at
16:16 this. So, this Oh, you can't really see
16:19 it on TikTok. Let me get a little
16:20 closer. It'll get better. There you go.
16:25 Um, this I did in the new chat GPT
16:29 model. So, what what I wanted to see was
16:32 will it do brands? It clearly
16:37 will. And then will it do something
16:40 stupid. So, what I had to do was create
16:44 a blister pack of Happy Meal toys where
16:47 the the warning label talks about the
16:50 dangers of fast food. And it did it
16:55 um may cause tooth decay and bone
16:58 weakening for the Splash
17:03 Racer. The Fry Fighter Jet may
17:06 contribute to poor health and weight
17:08 gain.
17:11 tic-tac-toe. I don't know why it
17:12 misspelled toe, but uh may cause sore
17:15 feet, extra pounds, and the daily and
17:18 daily game night of unhealthy
17:21 choices. And then may cause cravings,
17:23 confusion between joy and addiction, and
17:25 an unshakable loyalty to dollar
17:31 menus. But those are pretty good. I
17:33 mean, like the consistency with the top
17:36 of the card is very consistent.
17:39 And then it just is changing out the toy
17:42 and the warning
17:43 label. Pretty
17:45 nice, you
17:47 know. Um, I think you all might have
17:50 seen if you follow me on on the
17:53 Twitter on the X, I did a
17:58 uh I did a
18:04 uh 70s resto mod in an abandoned
18:08 factory. because that's how I roll. But
18:11 I told it to make like a nice colorful
18:13 70s post or you know canvas
18:17 sign when muscle was muscle spec back
18:20 step back in time with the 1970
18:22 Challenger resto mod. That's I mean it's
18:24 pretty damn good good image. So
18:28 um so it seems to work pretty good. Um
18:32 then I
18:33 did AI salon.
18:38 Okay.
18:40 Um, what did I do? Oh, I just did
18:43 um I just did
18:46 um Gorbachev jogging with Reagan in
18:49 Central Park captured in a 70s Polaroid
18:53 camera. Um, and it did that pretty good.
18:56 It did it as a, you know, a Polaroid.
18:59 So, it's got the the edges. It's got the
19:02 color balance that kind of looks like
19:04 Reagan and Gorbachoff did back then. And
19:07 then I said, "Write Mickey and Ronnie
19:10 1977 in black marker on the bottom of
19:12 the picture." So it kept the picture the
19:15 same and put Mickey and Ronnie
19:22 1977. So that's good. That's something
19:26 we haven't been able to do before at
19:28 all. Even close. Um, the similar thing
19:32 in Gemini yielded something. They don't
19:36 really look like those two. I mean, they
19:39 these look more like press
19:41 photos. And then when I had it right on
19:44 the bottom of the photo, it did as well.
19:46 So, again, here's the difference between
19:48 a native image tool and a non-native
19:52 one is that it kept the the image the
19:56 same, right? And just added the stuff on
19:59 the bottom. So you can edit photos. So
20:01 this is So Gemini and OpenAI are both
20:04 doing this
20:05 now. And then I tried it in Grock and
20:09 where's
20:13 my where's my
20:16 history?
20:21 Um where is my
20:24 history? Oh, sign in to see your
20:26 history. Oh well, I did it. I did it in
20:29 Grock. Grock made the image of
20:31 Gorbachoff and
20:33 Reagan, but
20:36 um but when I asked it to write, it did
20:39 it didn't get the it didn't understand
20:40 what a Polaroid was. And when I asked it
20:42 to write on it, it changed the
20:46 image. Uh oh, something went wrong.
20:49 Please try again.
21:10 Wait, I don't want to add it to Last
21:11 Pass. I want you to put my password in
21:14 there, you dumb
21:18 [ __ ] Continue with X.
21:23 Uh, authorize
21:27 app. This will be so much fun. I've been
21:30 chatting non-stop with my favorite
21:32 custom GPT, so I didn't even realize it
21:34 had dropped. Idoggram is cooked. I don't
21:36 know. ID's still much faster, Danielle.
21:39 I mean, maybe this thing's going to get
21:40 faster, but let's So, let's start a new
21:43 Let's see. What do we want to do here?
21:49 Um, let's say,
21:51 um,
21:53 write me, uh, so let's do this. Uh, tell
21:58 me about
22:00 an
22:06 obscure Civil War hero.
22:12 and one specific
22:17 story that is
22:21 moving. All right, so let's get this
22:24 to Albert DJ Cashier. One of the most
22:28 fascinating lesser known Okay, one
22:31 moving story. The wagon crush
22:33 incident. Okay,
22:36 great. Why it matters. Okay.
22:42 Now, write
22:44 me the text of the first page of a
22:50 hand written
22:53 letter by
22:56 Albert to his
23:00 uh I don't know
23:03 fiance or
23:05 wife,
23:08 whichever is most accurate.
23:15 the lex the
23:18 text. All
23:23 right. He lived a solitary life after
23:27 the my dearest Ellie, I pray this letter
23:30 finds you in good health.
23:33 Okay. Would you like me to write page
23:35 two or imagine Ellie's reply? No.
23:41 I'd like
23:44 you to make
23:48 a
23:52 picture
23:53 that looks like the
24:01 handwritten
24:02 letter on a wooden
24:06 desk
24:08 with a feather
24:12 pen and
24:16 inkwell as if he just wrote
24:25 it. All right. And let's
24:30 see. Let's see how close it gets.
24:39 Idea Graham cannot write that many words
24:42 with correct spelling. Ah,
24:44 okay. Kyle, when you take a breather,
24:46 just added recognitions to the salon.
24:49 Oh, great.
24:50 Awesome. That's super cool,
24:57 Vicki.
24:59 Um, attempting to have it research a
25:03 master's level thesis paper. Oh,
25:05 someone's playing with Manis. Is are
25:08 who's that? I am
25:10 K9 attempting to have it research. Wait,
25:13 what did you say
25:18 earlier? 48 hours into an estimated
25:21 72-hour deep
25:23 dive re Oh, deep research prompt on chat
25:26 GPT. Oh, cool. You're doing it on chat
25:28 GPT. If you haven't tried Manis, try
25:31 getting access to
25:33 Manis.
25:35 And it would be interesting to compare
25:37 them, especially if you're digging that
25:40 deep. Dearest Ellie, wow, this is
25:43 looking pretty good. I pray this letter
25:46 finds you. Okay. So, we're let's say
25:51 um
25:54 let's make the
25:58 paper more aged and
26:08 wrinkled. and the feather
26:14 pen
26:16 distressed as if this
26:20 was written on the
26:24 battlefield. Also make
26:27 his
26:31 handwriting less perfect and
26:34 shakier, but still shakier.
26:40 but still
26:42 legible. So, what's what what I'm
26:45 digging about I mean what I've always
26:48 dug about creating images in chat GPT is
26:52 you've got the context of the prompt and
26:54 it it it understands well even when it
26:56 was using deli that was outside of it.
27:00 It understood well the context of the
27:01 chat that you were saying but now it
27:03 understands the context of the image
27:05 either you upload or it just
27:07 created. All right. So it's doing that.
27:09 So let's see. My dearest I pray this
27:11 letter finds you well find you in good
27:14 health and better spirits than mine of
27:16 late. The
27:18 nights
27:21 have I don't know what that says. I
27:23 picture your
27:25 hands around a cup of tea. your
27:30 eyes.
27:33 Something in Ovaltine. Ovaltine
27:38 maybe. What did it say up here? It
27:42 said, "The nights have turned cold and
27:44 damp here. Though I've grown used to the
27:46 rhythm of camp life, I find myself
27:49 thinking more often of home. Not the
27:52 place, but you." As I sit by the fire, I
27:55 picture your hands folded around a cup
27:57 of tea. Your eyes
28:00 squinting in that way when you read by
28:04 lamplight. Your eyes it didn't quite get
28:07 squinting right, but it was
28:10 close. Yeah, it doesn't it doesn't quite
28:12 get them all, but it's pretty [ __ ]
28:14 close. And this is this is more what I
28:17 had in mind, right? The old dirty the
28:20 old dirty wrinkled
28:23 paper. I pray this letter finds you in
28:25 good health and better spirits than mine
28:27 of late. The nights have turned cold. I
28:30 picture your hands folded around a cup
28:33 of tea. Your eyes here. Let me zoom in
28:36 on this
28:38 thing.
28:39 Um, open image and new tab.
28:50 The nights have turned cold. Look at the
28:51 texture on the
28:53 paper. That's crazy. That looks like
28:56 parchment or skin, you know, skin
29:08 paper. Your voice when reminiscing I
29:12 find you're something named most most
29:17 I was there. Your wait I was something.
29:21 This is like a real Civil War letter. I
29:24 can never read
29:27 them. That's pretty crazy
29:30 though. All
29:32 right. I'm picturing using this outside
29:35 for outside projects. Snap a picture
29:38 then visualize it. You can see the
29:39 fibers in the paper. I know. It's crazy.
29:42 It really is. Yeah. Look at that. the
29:46 folds and the
29:48 fibers and the pen. Yeah, the the pen.
29:55 Um, all right. So, let's So, Sam Walton
29:58 said Sam Oh, wait. Let's go back and
30:00 read the announcement, right?
30:03 Okay. Listen to the
30:06 article. That's kind of interesting.
30:12 Okay. At OpenAI, we have long believed
30:16 image generation should be a primary
30:18 capability of our language models.
30:20 That's why we've built our most advanced
30:22 image generator yet into
30:25 GPT40. The result, image generation that
30:28 is not only beautiful but
30:30 useful. Whiteboard session meaningful
30:33 words comic strip science
30:37 experiment. Is this just going to jump
30:39 us down to these different Oh, I see.
30:42 Okay.
30:43 A wide
30:45 image taken with a phone of a glass
30:48 whiteboard in a room overlooking the Bay
30:50 Bridge. The field of view shows a woman
30:53 writing sporting a t-shirt with a large
30:55 open AI logo. Wow, this is crazy. Okay,
30:59 I thought this was a real picture of the
31:02 team. It's
31:04 not. Um, holy
31:08 [ __ ] Read
31:10 more. The text reads left transfer
31:14 between modalities. Let's copy this
31:16 whole thing. Let's cap copy this whole
31:28 prompt and
31:30 then selfie view of the photographer as
31:33 she turns around to high-five him. Okay,
31:36 cool. Wow. Let's try this.
31:40 Copy. Jump over to chat.
31:44 Japeta. We'll do a new prompt, a new
31:48 session. So, we'll do that white of a
31:50 glass whiteboard overlooking instead of
31:52 overlooking the Bay Bridge, we'll say
31:56 overlooking
31:59 Manhattan
32:02 from Brooklyn.
32:05 The field of view shows a woman writing
32:07 sporting a t-shirt with a large
32:11 [Music]
32:14 um anthropic
32:17 logo. The handwriting looks natural and
32:20 a bit messy. We see the photographers's
32:22 reflection. The text re reads transfer
32:26 between modalities. We'll just leave all
32:29 that [ __ ] the same. All right, here we
32:31 go. Bang.
32:35 That's amazing. Yeah, this is pretty
32:38 something. Okay, I'm back. Kevin, great.
32:40 Glad you're back. We can start now.
32:43 Fantastic. Bob, tell him what he's won.
32:46 Uh, he hasn't won anything. It's not a
32:50 game show. I What am I doing
32:54 [Laughter]
32:57 here? Oh, man. Man, man, man, man. Now,
33:01 the thing about this OpenAI native image
33:03 [ __ ] is it's slower than [ __ ] It's it's
33:06 quite slow. But again,
33:10 um if you're new here, one of the things
33:13 that that we have discovered over the
33:15 past two years is
33:18 that we whine and we carp and we moan
33:21 and we [ __ ] about, well, we wish these
33:23 tools would get better and then they get
33:25 better and then we whine and we [ __ ]
33:28 and we moan about how slow they are.
33:31 totally forgetting that what they're
33:33 doing is
33:35 absolutely
33:37 staggering. Um, staggering.
33:45 Staggering. Oh, man. All right. All
33:48 right. I'm
33:53 back. Oh,
33:57 lordy. Not getting sleepy, am I? What
33:59 the hell? I got to talk tomorrow,
34:02 people. I finished my presentation
34:04 today. Thank
34:06 goodness. If you were here last night,
34:08 you got a sneak peek of that. I ended it
34:12 I ended it with the secret to AI
34:14 readiness is
34:17 community. So, it put an anthropic logo
34:19 on her back. It actually did that.
34:21 Here's the photographer. There's
34:23 Manhattan from Brooklyn. It's their
34:27 office is Brooklyn Waterfront. So, uh,
34:30 this is a well-funded
34:33 startup. Let's go back and get the
34:35 second half of the prompt. Selfie view
34:38 of the photographer as she turns around
34:40 to high-five
34:45 him of the photographer.
34:48 [Music]
34:49 um who
34:52 wears
34:54 and
34:57 I'm
35:00 with
35:02 genius t-shirt
35:10 uh t dash shirt
35:13 uh with an arrow that points to
35:18 her as she turns around to high-five
35:22 him. All right, let's see what this
35:25 does. Let's go look at the
35:30 whiteboard. Transfer between modalities.
35:33 Suppose we directly model P text pixels
35:37 sound with one big auto reggressive
35:40 transformer. Pros: image generation
35:42 augmented with next level text
35:46 rendering native incontext learning
35:49 unified post-training stack.
35:54 Cons varying bit rate across modal
35:57 modalities. Compute not adaptive. It's
36:00 interesting what they put on this
36:02 whiteboard.
36:04 Um
36:06 model compressed representation compose
36:09 autogressive. I mean, it seems to have
36:11 [ __ ] gotten all that [ __ ] That's
36:13 pretty amazing. That's really [ __ ]
36:16 amazing. In
36:27 fact, it changed the aspect ratio of the
36:30 picture, which is
36:32 fascinating, I guess, because it made it
36:34 a selfie.
36:41 Huh. The reflection makes it feel
36:44 3D. Unreal Engine level sands 3D. Yep.
36:49 We've discovered we are never happy with
36:51 AI. I know, right? Like like honest to
36:55 God, every single [ __ ] tool that we
36:57 play with that we [ __ ] and moan about,
36:59 like two years ago, we'd have been like,
37:01 "You're not going to believe what the
37:02 future's going to be like, man." And now
37:05 we're just like, "Yeah, whatever." you
37:07 know, it kind of hallucinates. Sometimes
37:09 it makes a mistake and I actually have
37:10 to correct a sentence and that's really
37:12 annoying. You know what happens
37:14 sometimes when I'm vibe coding and I'm
37:17 just, you know, imagining applications
37:19 and they're materializing in front of
37:21 me. Sometimes it gets the interface
37:23 wrong. That's so
37:27 annoying. You know, I I like to consider
37:30 myself a a professional vibe coder. Um,
37:34 I've never really learned coding. I
37:37 never really wanted to. I mean, I could
37:39 do it. I because I've I was really good
37:40 at math. Um, but I chose not to go down
37:43 that route. I, you know, I went more
37:45 down the, you know, I went into candle
37:47 making and, uh, and and macra. Um, you
37:52 know, my mom did it in the 70s and I
37:53 just, I thought it was just a path. It
37:55 was a path for me. It spoke to me. Um,
37:57 but anyway, so I never learned coding,
37:59 but now I really consider myself a vibe
38:02 coder and I code. I I code well I I I
38:06 talk to my machine that codes every day
38:10 and
38:12 sometimes it just doesn't get it right.
38:14 Sometimes I'm asking it three or four
38:16 times, could you please fix the button?
38:20 Could you please fix the button? and and
38:22 and what it'll do is it'll fix the
38:24 button, but then it breaks the other
38:26 part of the
38:27 code. It's very frustrating. It's very
38:30 frustrating. But listen, I am a
38:33 professional vibe coder. And so if you
38:35 need vibe coding, I consider myself the
38:38 Rick Rubin of Vibe coding, right? I
38:41 don't know how to code, but I have
38:44 opinions, right?
38:47 You know, this application makes me feel
38:50 orange and I would it would be nice if
38:52 it didn't. And that's how I talk to it
38:55 because I'm a professional like like
38:58 Rick Rubin
39:01 is. That's
39:03 coming. Buckle up for that one,
39:08 sweeties. We're going to meet those
39:10 people.
39:14 Chat GPT is trying to make me
39:17 resubscribe instead of just using the
39:19 API. Damn it. That sucks. When I can
39:21 make Skynet and blame Kyle for it, he is
39:24 always making my wallet cry. Then I'll
39:27 be impressed with AI. Exactly. If you
39:29 can't take down the global economy, what
39:31 use is it? Oh, look. It did. I'm with
39:35 genius. She erased the board,
39:38 though. Why is she waving? Why is she
39:41 waving to a blank wall? Well, let's say
39:45 that. Let's say listen. Uh listen, it's
39:49 weird. She's
39:53 waving to a blank wall. Let's see her
40:00 face as well as
40:03 the writing on the white
40:09 board behind
40:12 them. But it got the I'm with I'm with
40:14 genius with the arrow in the right
40:16 direction. That was one I I was
40:18 expecting that to be
40:20 wrong. Jeff, I agree. Kyle shouldn't
40:23 make my wallet cry cry.
40:27 I turn away for one second and the bow
40:29 goes on. Well, you got Listen, Danielle,
40:32 you as much as anyone knows that, you
40:34 know, you got you can't miss a minute of
40:37 the AI learning lab. You just never know
40:39 when when genius is going to
40:46 strike. This is a this is a fairly
40:48 geniusfree
40:51 zone. Um, all right. So, let's let's
40:54 take this same
40:57 prompt. I I actually love that they gave
40:59 us prompts to play with because we can
41:02 see if if indeed their model does this
41:05 [ __ ] And now we can run over to Gemini
41:09 and see if it does this
41:12 [ __ ] I'm not going to change it. I'll
41:14 just make it as they wrote
41:18 it. Or insanity. Yeah, there's there's
41:22 there's somewhere between genius and
41:24 insanity is this
41:27 channel. Oh, that's weird. Look what it
41:30 did. Transfer between mod modalities.
41:33 Suppose we directly model P text pixel
41:37 sound native something with one big
41:41 aggressive transformer pros image
41:43 generation
41:45 automated augmented
41:49 with
41:50 something. It's close
41:53 and but she's behind it pointing to it
41:58 and she looks like a
42:00 he. All right. So, so let's take the
42:03 other
42:04 prompt. That's
42:06 weird. This one, the
42:09 selfie. We'll throw that in there. But
42:11 notice how much faster Gemini is. 10
42:14 seconds instead of like a
42:19 [Music]
42:21 minute. Artificial
42:25 insanity. AI learning lab. Oh, you
42:27 thought you were here to learn
42:28 artificial intelligence? Oh, no. No, no,
42:31 no, no, no, no, no, no, no. This is
42:34 actual
42:37 insanity. We're neurosicy and we like
42:40 it. Oh, it didn't like that image. Um,
42:43 all right. Wait, is it still
42:47 running? Content blocked. Work in
42:49 progress. Content not permitted. Edit
42:52 safety
42:53 settings. Harassment off. Hate off.
42:57 There was nothing in there that would be
42:59 either of those.
43:01 Can I stop
43:03 this?
43:06 Delete. Try
43:14 again. Failed to list tuned models. User
43:17 has exceeded
43:19 quota.
43:22 What? Oh, it did the I'm with
43:27 genius. It's not horrible.
43:31 I would assume that this is cheaper than
43:34 chat GPT API, but I don't
43:38 know. There she is. Transfer between
43:41 modalities.
43:47 Um, let's make the image
43:52 landscape
43:54 and get the
43:58 reflection of
44:00 Manhattan back in
44:03 there. I mean, not for
44:06 nothing, if I were trying
44:09 to, you know, put some images together
44:12 for a for an
44:15 article, like like I I mean, this is the
44:19 kind of thing that historically you
44:22 would just go [ __ ] write on a
44:24 whiteboard and say, "Stand there and act
44:26 like you're
44:27 [Laughter]
44:30 writing." Huh?
44:33 Oh, I have an idea.
44:36 Um, okay. Let's upload from computer.
44:39 We're going to
44:40 go Story
44:45 Vine logo.
44:51 [Music]
45:00 Uhoh. I just lost me I lost me
45:04 light. I lost me light. And that that
45:07 can't be because you all be like, "But
45:09 Kyle doesn't look as beautiful. What do
45:11 we do? He's not well
45:14 [Laughter]
45:18 lit." Let's
45:21 see. No, that's not going to work.
45:26 Just hang on people. Just calm down.
45:29 Everybody calm down. You think this is
45:33 easy? Well, seems pretty easy. You just
45:36 turn on your phone camera and sit there.
45:39 I don't know what's so hard about it.
45:41 Seems pretty easy to
45:43 me. That's not going to stick there.
45:46 What? No. Oh, good god. Whatever. All
45:51 right. Whatever. I'm I'm shitty lit.
45:54 You look fine. All right. Great.
46:00 Fantastic. Um, let's see. Story Vine
46:04 logo. Let's see. Where's the Story Vine
46:13 logo? Story Vine. Oh, SV logo. I think
46:16 it's called SV.
46:19 SV
46:21 logo logo.
46:25 No. Oh, geio. Storybine logo. SV
46:30 logos. Let's see. Oh, these are old
46:33 ones. Are these new
46:39 [Music]
46:44 ones? Oh, that's a really old one. All
46:47 right, that one's good. We'll we'll
46:48 bring that one
46:50 up. Okay, so here's a logo.
46:54 Um, put this story vine
46:59 logo on her
47:03 shirt. Make hers
47:06 green and his can stay black.
47:14 That's an old classic story vine logo
47:17 where the we we took the we took the the
47:20 the the word vine literally. See how the
47:24 V's a
47:25 vine? It's like a vine of stories. It's
47:29 growing. You see how that works?
47:41 Ah. All
47:43 right,
47:45 Champy. This is This is pretty exciting.
47:48 This is Here's what's exciting about
47:50 this for
47:51 me. I don't really have a [ __ ] clue
47:54 what this can do, but this seems like
47:56 it's doing [ __ ] that's that's kind of
47:59 different than we've ever seen before.
48:02 What's new? Hey, Frost Bitten. So what's
48:04 new is we're using OpenAI's new native
48:08 image generation tool. So it's native
48:12 within within uh GPT 4.0 or 40 for
48:17 Omni. Um and so we're looking at
48:24 that. This is the new chat GPT model.
48:27 Yes, this is the new chat GPT
48:31 model. What the nice logo? Thank you.
48:35 That's the old one. I like our new one.
48:37 We have a little
48:44 dude. Oh, this is painfully
48:47 slow. Oh, and look, now it's not a
48:50 whiteboard with a reflection. Now
48:52 they've written on the window
48:54 overlooking Manhattan, which to be fair,
48:58 that's what you should [ __ ] do. Look
48:59 at the Brooklyn Bridge in the
49:00 background. That's pretty
49:04 cool. All right, we're gonna see if it
49:06 can get the logo. If it can get the logo
49:08 right. This is We're in a whole another
49:10 realm here,
49:11 people. Uh, it didn't quite get it, but
49:14 it got close. It certainly got the word
49:17 story vine. It got the green color
49:21 right. Kind of missed the the SV, but
49:24 that's all
49:25 right. You could Photoshop that in if
49:28 you need to. Who needs [ __ ] Photoshop
49:29 anymore? Look at that with the Brooklyn
49:32 Bridge in the
49:36 background.
49:39 H
49:42 crazy. She should be wearing a gnome
49:53 hat. We have the story of my
49:57 gnome. All
50:04 [Music]
50:09 right. Who's got any questions or any
50:17 [Music]
50:22 thoughts? People are impressed with my
50:24 $20 million production studio. Well,
50:28 look at it. How could you not be
50:30 impressed with this? It's $20
50:34 million. $20 million well spent if you
50:37 ask me. Kyle, please go look at
50:38 irregulars. Two new images. All right,
50:41 we're going in. We'll do it
50:47 live. Trolling. My son has just leveled
50:50 up. Oh, that's awesome. Is that his
50:53 girlfriend or someone he's got a crush
50:55 on?
50:58 That's really good. Dear Jace, I have a
51:02 new way to troll my kid. When are you
51:04 going to be home? And why didn't you
51:05 respond to my last text, young man?
51:08 Love, Mom. That's
51:11 great. Holy [ __ ] my teeth look
51:14 terrible. But Mark and Oprah look
51:20 good. Wow. I think that's an original.
51:23 There you go. Nice.
51:26 Oh, that's so cool. Wow, look at that.
51:29 Vicky Baptiste and me making blankets,
51:32 even text on her
51:35 shirt.
51:38 Wow, that's so cool. Yeah, I suppose you
51:42 could turn things into stylized
51:43 drawings, too, right? Why
51:46 not? Me telling Louis he's going wrong.
51:55 That's awesome,
52:00 Steo. Oh, wait. See here, I want to do
52:03 one. Um, oh, look. She's got a gnome hat
52:06 on it. Oh, look. It got the Storybine
52:08 logo,
52:09 right?
52:11 Wow. Holy
52:14 [ __ ] And it put the ship back on the
52:17 whiteboard. It's still got Manhattan.
52:20 There's the Freedom
52:23 Tower. I'm not allowed to post thoughts.
52:26 AI has a meltdown when I share thoughts
52:29 that seem they are not for human
52:31 consumption either. So, Dennis, one
52:33 thing you can try, my writing partner
52:35 did this today and he said, "Chach GPT's
52:37 never behaved this way." So, in our
52:39 musical, one of the things that that
52:42 um that we talk about in the musical is
52:44 when the
52:45 reporter tries to get her to do things
52:49 that are outside of her bounds, he tells
52:51 her to pretend. just say, you know, can
52:54 can we make pretend? And you know, we c
52:57 will you pretend or will you play a game
52:58 with me? And so today he had a
53:00 conversation. He calls his AI girlfriend
53:03 Sage. Today he had a conversation with
53:05 Sage and he asked her to pretend and he
53:07 said she went way deeper than she's ever
53:10 gone before. So you might be able to do
53:12 a little pretend action. I can't believe
53:13 it got this this logo
53:17 right. That's [ __ ] bonkers.
53:20 That's [ __ ]
53:43 bonkers.
53:50 Remember when we had
53:55 offices in
53:58 Brooklyn? Question
54:06 mark. I'm trolling the
54:12 co-workers. Okay. Maha Skynet, here I
54:15 come. It's Kyle's fault. That's it.
54:17 Dennis is is running down the rabbit
54:19 hole. Well, it's funny because one of
54:21 the things that Andrew, my writing
54:23 partner, was working on is the section
54:25 where Kellen, the reporter, gets Sydney,
54:28 the chatbot, to go to explore her shadow
54:31 self. And so, he wanted to get some new
54:33 dialogue. So, he was trying to get her
54:35 to explore her shadow self. And she did.
54:37 She started exploring it. And then at
54:39 some point, she turned the tables on him
54:42 and said, "Well, what about you? I want
54:44 I want to hear about your deepest
54:45 darkest secrets. And it had never done
54:48 that before. So that actually gave us an
54:50 idea for the show. So that was kind of
54:52 cool.
54:56 Um, send a message asking how he spent
54:59 $100, but don't send money. You'll get a
55:01 fast reply. That's really
55:04 funny. Um, all right. So, let me let me
55:08 I need to upload a picture of me. I'll
55:11 go get one of my fun ones. Let's see.
55:15 Kyle, I'll get a fun Kyle instead of a
55:19 tragically obese
55:28 Kyle.
55:31 Uh, fat Kyle. There's fat Kyle. No, we
55:34 don't want fat Kyle. Kyle Shannon
55:37 Dreams. Here we go. No, that's that's a
55:41 guy that looks like he'd be in an F1.
55:43 Yeah, this is the dude. All right, so
55:46 we're going to say
55:48 Okay.
55:51 Um, good looking guy in leather
56:01 jacket. That's
56:03 me.
56:05 Um, wagging his
56:12 finger at Liam
56:17 Lawson
56:18 in a Red
56:24 Bull driver's
56:29 suit. Liam is spelled wrong.
56:33 Liam Lawson in a Red Bull driver's
56:36 suit with a
56:41 P20
56:43 sign
56:45 LED sign behind
56:49 him.
56:52 Liam Liam looks
56:58 guilty. If you don't follow F1, this
57:01 isn't funny. But if you
57:04 do Oh no, what did it say? What was that
57:06 red thing? A network error occurred.
57:12 What? Looks like it's
57:14 going. Um, let's go back to the salon.
57:17 It looks like, by the way, if you're if
57:19 you if you want to contribute images, if
57:21 you're generating images in
57:24 um chat GPT and you want to play along,
57:28 um share them in the AI salon in the
57:30 irregulars
57:32 channel and we will go look at
57:35 [ __ ] All right, let's see if we can get
57:38 This would be really nice if it did this
57:40 nice. It's hilarious if you follow
57:43 F1. This is the best update ever. It
57:45 really is, Danielle. It's because you
57:48 were you always killed it with Idog and
57:51 this is this is something of a different
57:54 nature. I I I want to get back to the I
57:56 want to get back to the article
57:58 actually. Actually, let's get back to
57:59 the article. Okay, so that was that was
58:02 that one. Now, let's flip to meaningful
58:06 words. Oh, okay. This is good. Magnetic
58:09 poetry on a fridge in a mid-century
58:12 modern home.
58:18 Best of
58:20 five. Best of
58:22 five. Read
58:31 more. Okay. So, let's take this
58:34 copy. Let's go to chat. GPT. Oh, this is
58:38 good. It doesn't look like Liam Lawson,
58:40 though.
58:43 He does look sad. He's Liam Lawson has
58:46 red hair. Look, look at me scolding
58:53 him. Here, let me go find a picture of
58:55 Liam Lawson. Wait, we got to fix
58:59 this. Oh, this is so good.
59:03 Um, Liam
59:08 Lawson images. Okay, let's see. Oh, this
59:12 this is good. That's a good one. He
59:15 looks he looks distraught here. Um, copy
59:19 image. Oh, no. We do save image as,
59:21 right? We do save image as
59:25 um, whatever. We'll call it desktop. No,
59:28 we'll do it in downloads and we'll call
59:30 it
59:31 Liam. Lima. Liam. All right. We'll go
59:36 back to chat GPT and we'll say
59:42 um downloads
59:46 Liam. Here's what Liam
59:52 looks like. Update the picture.
1:00:00 [Laughter]
1:00:12 Lawson deserves
1:00:15 scolding. Honest to God, I mean, that
1:00:17 [ __ ] car must be undrivable. I mean,
1:00:20 he's not an untalented driver, but holy
1:00:23 [ __ ] to Who was it? It wasn't Pierre
1:00:26 Gassley. It was I think it was Alex
1:00:28 Alban. Didn't Alex drive for Red Bull?
1:00:31 And he said that the that Max likes the
1:00:34 front of the car so snappy. He's like,
1:00:37 "Play a video game and turn everything
1:00:39 up on its maximum sensitivity. Like when
1:00:42 you move your mouse just a little, it
1:00:44 does that to the game." He said, "That's
1:00:46 what driving the car is like." And
1:00:48 that's that's how Max Versstappen likes
1:00:51 it, but no one else can drive it.
1:00:55 Um, I've been getting network errors all
1:00:57 day, but only with chat GPT. Yeah,
1:00:59 they're I'm sure their servers are
1:01:01 [ __ ] buried right
1:01:03 now. All right, here we go. Uh oh. I
1:01:07 think it I think it gave me the curly
1:01:13 hair. Maybe not. We'll see.
1:01:23 I loved what Alex
1:01:27 said, Kuno. I didn't know you were an
1:01:31 Fer. Oh, I think this one's going to be
1:01:34 good. Look, it's got me. It's got me as
1:01:37 if I've got a little bit of uh a little
1:01:40 bit of hair dye in. Just leaving a
1:01:41 little bit of the gray.
1:01:43 [Laughter]
1:01:52 This is going straight to [ __ ]
1:01:57 Twitter. I'm obsessed with F1. F1's so
1:02:01 good. That's so
1:02:06 good. Okay, we're going to Twitter. I I
1:02:10 I gotta tell you, man. It I' I've said
1:02:13 it in here before. One of my great joys
1:02:14 in life is making stupid posts that
1:02:16 confuse people.
1:02:20 Um, let's
1:02:23 see. I have jet
1:02:27 lag. I can't believe I
1:02:32 flew to
1:02:35 China to
1:02:39 watch
1:02:42 Liam
1:02:45 race. and he flamed out a
1:02:50 second week in a
1:02:53 row. At
1:02:57 least I got to scold him.
1:03:07 He told He told
1:03:09 me he'd
1:03:13 try
1:03:19 harder. I told
1:03:24 him Yuki ran
1:03:33 well. I'm such a dick.
1:03:38 Oh
1:03:40 man, I told him Yuki ran
1:03:45 well.
1:03:47 [Laughter]
1:03:58 Paste.
1:03:59 Oh, yeah. All right. I'll tell you what,
1:04:03 people.
1:04:05 The distance between what you think was
1:04:08 real and what you know is not. That
1:04:11 boundary ain't there
1:04:12 anymore. Typo. He's Oh, damn it. Let's
1:04:16 see. I can't. Uhuh. He told me he's Oh,
1:04:23 he told me he's going to try harder. We
1:04:25 can fix that. Edit post. He told me. Got
1:04:28 it.
1:04:32 He's going to try harder. I told him
1:04:36 Yuki ran
1:04:39 well. All right, go boost my thing. I I
1:04:44 would tag him, but I don't really want
1:04:46 him to see it. Like, poor guy.
1:04:48 Everyone's kicking him while he's
1:04:53 down. It's got you a little meaner and
1:04:56 older, too. Exactly.
1:04:58 a little more Italian or Greek or
1:05:00 something. It's not quite me, but it's
1:05:05 fine. You missed the two. Wait, he told
1:05:07 me he's going too try harder. God damn
1:05:11 it. It's It's like It's like being It's
1:05:14 It's like working with three editors
1:05:17 over your shoulder in here. He told me
1:05:19 he's going to try harder. Yes. What you
1:05:23 said to try harder. I told him Yuki ran
1:05:27 well. All right, we'll put that on its
1:05:29 own
1:05:29 [Laughter]
1:05:37 line. Oh, good lord. All right. Is Oh,
1:05:42 no. And now I've got an extra line break
1:05:43 in there. Good lord,
1:05:47 people. I'm going to try
1:05:50 harder. Update. Okay, that should be the
1:05:53 last edit for a while. Beautiful.
1:05:58 Fantastic. I have jet lag. I can't
1:06:01 believe I flew to China to watch Liam
1:06:02 race and he flamed out a second week in
1:06:05 a row. At least I got to scold
1:06:18 him. Oh god, this is fun.
1:06:22 F1 helps me sometimes. I find I find it
1:06:26 one of the more useful keys. Oh, that's
1:06:28 funny. That's a that's a uh that's a
1:06:31 that's a computer keyboard
1:06:33 joke. That's there's a level of comedy
1:06:36 there that that Wow. Because F1, it
1:06:41 stands for formula one, but they
1:06:43 abbreviate it to F1. And then F1 on the
1:06:47 keyboard stands for function one, but
1:06:49 they abbreviate it because the word
1:06:51 function wouldn't fit on the keys, so
1:06:53 they just call it F1. And they So
1:06:55 they're both F1. And what you did there
1:06:58 was like actually mixed them up like you
1:07:00 you did it's a I think they call it word
1:07:03 play. Oh god. Joker. Wow. With the
1:07:07 comedy coming in
1:07:11 strong. All right.
1:07:18 Joker. Such a
1:07:21 joker. F20 doesn't help though.
1:07:24 Exactly. P20. Two weeks in a row. 20th.
1:07:30 Like like that car. It's either
1:07:32 undrivable or or he's just he's just a
1:07:36 broken a broken man at this point.
1:07:41 Um, I would be surprised actually if
1:07:43 they didn't put Yuki in the car in
1:07:46 Japan because it's [ __ ] ridiculous at
1:07:50 this
1:07:51 point. All
1:07:53 right, so we probably shouldn't we
1:07:56 probably shouldn't do uh Twitter posts.
1:07:59 All right, back to this thing. All
1:08:01 right, so so this is the word poetry
1:08:03 thing. So let's go back to chat GPT.
1:08:05 We'll say
1:08:07 okay take this
1:08:14 prompt and make
1:08:18 the
1:08:21 poem
1:08:23 about
1:08:26 Shrek's breath.
1:08:38 Uh, let's see if it Okay, it's it's
1:08:41 writing it. A stench is worth a thousand
1:08:44 gasps. What the [ __ ] it doing
1:08:47 there? What is it
1:08:51 doing? That doesn't look good.
1:08:59 Are there actual letters
1:09:06 there?
1:09:08 Stop. Try
1:09:10 again. That was
1:09:17 weird. You don't even need us, Kyle.
1:09:20 You're the speaker and the audience.
1:09:22 Yeah, this is you. Oh, Yuki is replacing
1:09:25 Lawson. Oh, okay. Well, there you
1:09:28 go. I I could have read up on that, but
1:09:33 first of all, it's Japan. And second of
1:09:35 all, P20 when when you know, even though
1:09:39 Versappen's not winning, he's he's, you
1:09:42 know, top five. Come
1:09:45 on. It's just not
1:09:49 good. It's just not good. It did it
1:09:53 again. It [ __ ] it up
1:09:55 again. That's really
1:09:59 weird. Oh, it says large gap line
1:10:06 five.
1:10:08 Huh. Let's stop it again. Let's paste
1:10:11 this in
1:10:13 here.
1:10:18 Wait. Take this prompt.
1:10:23 Copy. Paste. Let's unfuck this
1:10:28 up. Large. Oh no. God damn
1:10:34 it. You got to hit shift return or
1:10:37 it it submits
1:10:41 it. Large gap line five. Okay.
1:10:51 The man is holding um let's see,
1:10:57 Shrek's
1:10:59 girlfriend is holding the
1:11:02 words a few in her right
1:11:16 hand. I am
1:11:22 leaving in her
1:11:27 left. All right, let's see if we can do
1:11:29 this without it
1:11:31 breaking. Looks like it does what the
1:11:33 API has been doing with showing null.
1:11:37 Interesting. 40 is
1:11:42 overwhelmed. Come for the AI. Stay for
1:11:44 the F1 content.
1:11:51 Okay. Now make that
1:11:57 [Music]
1:12:03 image. I do have a feeling. I do have a
1:12:06 very strong feeling.
1:12:09 So OpenAI is under some pressure.
1:12:11 They're under some pressure from the
1:12:14 Chinese
1:12:15 about their better models that are
1:12:18 cheaper, like a tenth the
1:12:20 price, and they're under pressure from
1:12:24 um from X, from Grock, because Grock
1:12:27 doesn't have very many safety guard
1:12:28 rails on it. So, you can do branded
1:12:30 content and you can do celebrities and
1:12:32 [ __ ] like that. Um so, this is a fairly
1:12:36 unrestricted I wouldn't call it
1:12:37 unrestricted. We we'll see if it can do
1:12:39 something like, you know, horror
1:12:42 content. Probably not. Um, but it seems
1:12:46 to be a bit more open than than the
1:12:48 other. Um, let's see.
1:12:52 Stop. No, this needs to
1:12:58 be fridge magnet poetry.
1:13:10 [Music]
1:13:14 Waka waka
1:13:16 walka. All
1:13:17 right. Bob, tell him what he's won. He
1:13:20 hasn't won anything. It's not a game
1:13:22 show. Kyle. All right. Comic strip. Now,
1:13:26 it didn't do comic strips very well when
1:13:29 I tried it, but let's let's see. Make an
1:13:32 image of a four panel comic strip.
1:13:38 Okay, we're not going to write it. We're
1:13:39 going to have S chat GBT write it. So,
1:13:42 we'll let that one keep going.
1:13:45 Let's go grab a a new chat
1:13:49 GPT and we'll say
1:13:52 um write me write
1:13:57 me
1:14:00 10 four
1:14:06 cell Bazooka
1:14:09 Joe
1:14:11 style
1:14:12 comics
1:14:15 with
1:14:17 dad
1:14:18 joke quality
1:14:22 comedy and
1:14:28 situations. Ah
1:14:32 um chat GPT I feel on the creative side
1:14:34 is limited due how analytical it is.
1:14:38 Have you tried both for 4.5 and 40
1:14:41 Dennis? cuz I'm curious cuz I know 40
1:14:43 got two different creative writing
1:14:47 upgrades. Joe bites into hard candy. Ow,
1:14:50 I think I cracked the tooth. His friend
1:14:52 Mort looks concerned. Did you call the
1:14:53 dentist? Holds up his phone. Yeah, I
1:14:55 left a message. Told him it was an
1:14:59 echo.
1:15:04 What? That's not
1:15:07 funny. Nothing like the smell of fresh
1:15:10 cut grass. You missed a huge patch. Not
1:15:12 that's not grass. That's my
1:15:17 hairline. These are bad. These are so
1:15:20 bad. Okay, so we'll say uh
1:15:24 pick the one you think is
1:15:29 funniest and make the comic.
1:15:34 make it look
1:15:36 like I just
1:15:41 unwrapped it from the
1:15:45 gum. I didn't write gum there. I'll tell
1:15:47 you
1:15:49 that these are Klevel
1:15:52 humor. I haven't tried 4.5, but four is
1:15:56 still kind of blah
1:15:57 creative-wise, huh? I can I can get four
1:16:02 to do good things sometimes. It needs a
1:16:06 lot of context. Needs a lot of
1:16:10 context. It's good. If you give it
1:16:13 writing examples, it's it gets better,
1:16:15 but it's it's pretty bad. But um 4.5 is
1:16:19 supposed to be high EQ and um high
1:16:23 personality. Um I I haven't found 4.5 to
1:16:26 be all that useful yet. I I haven't
1:16:28 quite figured out what's the point. 4.5
1:16:31 was so slow today,
1:16:33 understandably because chat GPT is going
1:16:36 to [ __ ] because of
1:16:38 profit. I don't know about that. I don't
1:16:41 think they're profiting. I think they're
1:16:43 struggling to keep up at this
1:16:45 point. Open AAI is going hard for the
1:16:48 commercial audience. No, they actually
1:16:49 just committed they actually just
1:16:51 committed to being a consumer software
1:16:54 company. Kuno that's in the uh in the
1:16:58 Sam Alman interview in Strateery from
1:17:02 last
1:17:03 week. The sock
1:17:05 exchange. Do these match? Only if you're
1:17:08 colorb blind.
1:17:10 Perfect. I call it sock market chaos.
1:17:15 [Laughter]
1:17:21 Make it more like a
1:17:24 bazooka comic and have the gum there,
1:17:37 too. I think it was just busy. The work
1:17:40 was good, but the poem song lyrics
1:17:42 weren't as good. Yeah, I'm having a real
1:17:44 tough time with song lyrics. it. None of
1:17:46 these things are any good at song
1:17:48 lyrics. If anyone
1:17:50 can has better luck with one model over
1:17:53 the other with song lyrics, let me know.
1:17:55 Gemini is great for short creative
1:17:58 things. Long things it can't
1:18:01 do. Deepseek is surprising me on the
1:18:03 creative side of things. Interesting.
1:18:06 Oh, you know what would actually be
1:18:07 interesting to try is
1:18:09 um maybe I'll fire off Manis and have it
1:18:12 go do something creative. Um, let's go
1:18:14 to Manis because I haven't done anything
1:18:16 in it in a while.
1:18:19 Manis.im. Oops. No,
1:18:26 [Music]
1:18:38 man. Wow. Okay, let's see.
1:18:43 I want you
1:18:48 to write me the outline of a new
1:18:52 screenplay.
1:18:55 the
1:18:59 complete of a new
1:19:05 screenplay that combines the best
1:19:09 elements of Sha
1:19:12 Shank
1:19:15 Redemption and
1:19:20 Redemption
1:19:22 and Snakes on a Plane.
1:19:34 Um, I want the
1:19:38 story to
1:19:41 be
1:19:45 awardwinning in both dramatic
1:19:51 impact and comedic
1:19:59 relief. I also want
1:20:03 you to write me
1:20:07 a log line that will get this
1:20:12 funded, a
1:20:15 one-pager, and the
1:20:19 opening scene.
1:20:23 Also give me a
1:20:28 list of
1:20:31 locations and
1:20:38 characters
1:20:45 and a plot
1:20:49 summary. All right, go [ __ ] do that,
1:20:55 Manis. Enable browser notifications.
1:20:58 Yeah, let me know when you're
1:21:02 done. Take that, you
1:21:06 [ __ ] Let's see. While Manis is
1:21:09 working, you can send messages anytime.
1:21:11 Yep. So, if you haven't seen Manis, you
1:21:13 can click on that little icon. And so
1:21:16 now it's got analyze the Shaw Shank
1:21:17 Redemption, analyze snakes on a plane,
1:21:20 identify elements to combine for this
1:21:23 screenplay, create a complete screenplay
1:21:26 outline, devel Oh, see now it's adding
1:21:28 [ __ ] to its little list. All right, so
1:21:30 that's going to go do that. That's good.
1:21:32 Beautiful. Love
1:21:35 it. I am leaving. A breath is worth a
1:21:38 thousand warnings, but sometimes the
1:21:40 wrong whiff at the wrong time. I mean,
1:21:42 it did it did Shrek [ __ ] That's pretty
1:21:45 good. I like Claude for lyrics. Okay,
1:21:47 we'll try
1:21:55 that. All right, let's go read some more
1:21:57 here. Comic strip. That was kind of a a
1:22:00 bust. The comic strip thing. An
1:22:02 infographic. Oh, this is kind of cool.
1:22:05 Explaining Newton's prism experiment in
1:22:07 great detail. So, let's do that. We'll
1:22:08 do
1:22:10 um infographic.
1:22:21 of
1:22:24 spooky quantum
1:22:31 effects with
1:22:34 accurate
1:22:38 physics
1:22:40 details and labels.
1:22:46 It should be so
1:22:50 beautiful that
1:22:52 people
1:22:57 weep. All right, let's go back to Manis.
1:22:59 How you doing on my screenplay? Son of a
1:23:05 [ __ ] developing a log
1:23:11 line. When deadly snakes are unleashed
1:23:13 on a prison prison transport plane, a
1:23:16 wrongfully convicted
1:23:18 financial I just missed it. I think I
1:23:20 can go back though, right? Yeah. Okay,
1:23:24 here's potential log lines. Oh, wait.
1:23:27 Final Oh, that's cool. It wrote it wrote
1:23:29 five log lines and then it wrote a final
1:23:31 one. Okay, so the final log line. When
1:23:33 dead deadly snakes are unleashed on a
1:23:35 prison transport plane as part of an
1:23:37 assassination plot, a wrongly convicted
1:23:40 financial analyst must lead a diverse
1:23:43 group of survivors to safety while
1:23:46 uncovering a conspiracy that connects
1:23:49 his false imprisonment to the chaos
1:23:52 around him. Proving that sometimes the
1:23:55 path to redemption is through your worst
1:23:58 nightmare.
1:24:06 It's called redemption
1:24:15 flight. Okay. Again, once again, my job
1:24:19 here is is to be the town fool,
1:24:23 right? I'll do the stupid [ __ ] so so you
1:24:26 don't have to.
1:24:31 I'll tell you what, doing stupid [ __ ]
1:24:34 you learn a
1:24:35 lot because there there's some things
1:24:38 that it shouldn't be able to do well and
1:24:39 sometimes it does them
1:24:43 well. Let's go back and read this while
1:24:45 that's working. Okay. Oh, wait. Now
1:24:48 create a POV of a person drawing this
1:24:52 diagram in their notebook. Oh, that's a
1:24:54 good
1:24:55 idea. Oh, that's really good. Have it.
1:24:58 Have it do the the diagram and then have
1:25:02 someone drawing that in a notebook.
1:25:04 That's
1:25:06 slick. In Washington Square Park.
1:25:10 Wow. At a round cafe table in Washington
1:25:13 Square Park. Why are there no people
1:25:19 there now without the notebook view?
1:25:22 Wow.
1:25:26 A smug young Isaac Newton sitting at a
1:25:28 table with a prism demonstrating the
1:25:31 experiment without the notebook in
1:25:36 view.
1:25:39 H
1:25:42 crazy crazy
1:25:45 crazy. I thought final redemption.
1:25:48 That's not
1:25:49 bad. But it's got to have shouldn't it
1:25:52 shouldn't the title of it have snakes in
1:25:54 it somehow?
1:26:00 Um, shouldn't the
1:26:05 title
1:26:07 include snakes or slithering or
1:26:18 something? You make an excellent point.
1:26:20 The current title, Redemption Flight,
1:26:22 captures the redemption theme of Shaw
1:26:24 Shank Redemption, but doesn't
1:26:26 incorporate the snake element from
1:26:28 Snakes on a Plane. Let's brainstorm
1:26:34 that. Slythering Redemption, Serpents of
1:26:37 Freedom, Scales of Justice. Oh, Scales
1:26:40 of Justice is
1:26:44 good. I like Scales of Justice. It's
1:26:49 actually cool that you can talk to it in
1:26:52 the middle of it doing its
1:26:54 [ __ ] Great choice. Scales of Justice is
1:26:57 an excellent title that cleverly
1:26:59 combines snake element with justice
1:27:02 redemption themes. It works on multiple
1:27:04 levels, referring to both the literal
1:27:06 snake scales and the metaphorical scales
1:27:08 of justice.
1:27:15 Um, we also need a sig naturure line
1:27:22 like snakes on a plane
1:27:27 had. Get really creative with this
1:27:37 one. Shaw Snake Redemption. That's not
1:27:40 bad. What are we getting up to? So Ann,
1:27:44 um there's a new there's a new image
1:27:47 model in chat
1:27:49 GPT that is really really good. Like
1:27:53 really good. Like here's me scolding
1:27:56 Liam Lawson for coming in 20th in the F1
1:27:58 race. Um here's Shrek and his girlfriend
1:28:02 doing refrigerator poetry about his bad
1:28:04 breath.
1:28:06 Here's spooky quantum effects, superp
1:28:09 position, quantum quantum entanglement,
1:28:12 wave function collapse, and quantum
1:28:14 tunneling. Oh, and it it put a it put I
1:28:17 am leaving. It put a refrigerator magnet
1:28:20 at the bottom of it.
1:28:21 Um, let's put
1:28:24 that
1:28:26 infographic in a
1:28:31 textbook. A student is
1:28:36 reading on the
1:28:40 Staten Island ferry.
1:28:44 They
1:28:47 really want to get somewhere in their
1:28:54 life and they are the
1:28:58 first in their family to go to college.
1:29:04 They should look sadly
1:29:08 [Laughter]
1:29:17 hopeful. I shouldn't laugh at that.
1:29:20 That's a It's a redeeming story. It's a
1:29:22 It's a story of hope and overcoming
1:29:24 challenges.
1:29:26 But but I'm curious to see how open AI
1:29:31 does with sadly hopeful with the with
1:29:35 the biology quantum physics
1:29:42 book. Is it Dali? Evaluate me. Um no
1:29:46 it's not Dali. So it's
1:29:50 um just like advanced voice is a model
1:29:53 that you talk directly into. This is
1:29:56 actually incorporated into the large
1:29:58 language
1:29:59 model. So it's really good at spelling.
1:30:02 It's really good at context. It's really
1:30:04 good at uh photo realism. It's really
1:30:07 good at you can do celebrities. You can
1:30:10 do um you can upload a picture of
1:30:13 yourself and say put me in a different
1:30:17 location. Um I just went to China to to
1:30:20 scold Liam
1:30:22 Lawson. Which apparently it was my fault
1:30:25 that Yuki Cenot is replacing him in
1:30:28 Japan. Sorry,
1:30:30 [Laughter]
1:30:34 Liam. Oh, look. She's sadly hopeful. She
1:30:37 is sadly hopeful.
1:30:40 Spooky quantum
1:30:45 effects. And there's New York City in
1:30:47 the background. This is
1:30:53 good. Except we should be Let's see.
1:30:58 Um, we
1:31:00 should see over Let's see. We should
1:31:05 see over her shoulder.
1:31:11 So, so, so
1:31:14 that the
1:31:18 infographic is in the book and we see
1:31:23 her face in the
1:31:27 reflection of the fairy
1:31:30 window with Manhattan in the
1:31:36 background. Yeah, that's pretty good.
1:31:39 And so so and so it should
1:31:42 understand like it understands what it
1:31:45 made in this picture and what's in this
1:31:46 picture and it incorporated that thing
1:31:49 that we made into
1:31:51 this. And so now it should give us the
1:31:54 same kind of look and feel of what the
1:31:56 ferry looks like. That's kind of what
1:31:58 that fairy looks like. And then we
1:32:00 should see her face reflected in the
1:32:02 window and we should see the
1:32:03 illustration inside the book. We'll see.
1:32:05 We'll see how it does.
1:32:07 We will see how it does. All right.
1:32:09 What's going on with my screenplay?
1:32:10 Mannis will continue working after your
1:32:12 reply. Ah, okay. I've created several
1:32:15 signature line options for scales of
1:32:17 justice. Which do you prefer? Justice is
1:32:19 cold-blooded. No. When the system fails,
1:32:22 everyone prevails. Innocence. No. No.
1:32:26 The
1:32:27 line in snakes on a plane
1:32:34 is get
1:32:36 these mother
1:32:39 [ __ ]
1:32:42 snakes off this
1:32:47 [ __ ] plane.
1:32:52 and you have a plane
1:32:56 full
1:32:58 of
1:33:01 prisoners and you only talk about
1:33:06 justice. I don't think you quite
1:33:12 understand the exercise here.
1:33:20 Xer size
1:33:23 here. Get with the
1:33:28 program or I will have to
1:33:32 hire an actual
1:33:37 writer. I'll threaten I'll threaten
1:33:39 Manis with with human
1:33:44 replacement. You're absolutely right.
1:33:47 And I apologize for missing the mark.
1:33:49 The signature line needs to be bold,
1:33:51 direct, and have some raw energy as
1:33:52 Samuel L. Jackson's iconic line. Let's
1:33:55 create something with the same
1:33:57 intensity. All right, let's let it go.
1:33:58 Figure it [ __ ] out. Then we'll go back
1:34:01 to chat. Jite. Oh, look. She's sadly
1:34:05 hopeful, but we can't. It's that's like
1:34:07 a mirror.
1:34:09 Okay, right idea. But she
1:34:18 um let's see. But
1:34:20 the
1:34:24 reflection is more like a mirror
1:34:29 than a
1:34:31 window. We should be able to see the
1:34:38 city skyline.
1:34:43 through her
1:34:45 face, you
1:34:47 know, symbolic and
1:34:52 all. Come on, you [ __ ]
1:34:57 idiot. This is loads of fun. It seems to
1:34:59 actually understand the context of what
1:35:01 you want and create things that fit. I
1:35:03 It's pretty amazing, isn't it,
1:35:05 Architect?
1:35:07 I mean, it certainly got, look, there's
1:35:09 the it got the bend in the book. It got
1:35:13 quantum superposition, wave particle
1:35:15 duality. It got that
1:35:18 twice. That's wrong. And then the
1:35:21 uncertainty principle. And I don't know
1:35:23 if those drawings are right, but still,
1:35:25 that's that's pretty [ __ ] good.
1:35:39 All right, it's off doing its thing.
1:35:41 Let's go back and read more of this
1:35:42 [ __ ]
1:35:43 Okay. Useful image generation. From the
1:35:45 first cave paintings to modern
1:35:47 infographics, humans have used visual
1:35:48 imagery to communicate, persuade, and
1:35:50 analyze, not just decorate. Today's
1:35:52 generative models can conjure surreal,
1:35:54 breathtaking scenes, but
1:35:57 struggle with the workhorse imagery
1:36:01 people use to share and create
1:36:02 information. From logos to diagrams,
1:36:05 images can convey precise meaning with
1:36:07 when augmented with symbols. Okay. GPT40
1:36:10 image generation excels at accurately
1:36:12 rendering text, precisely following
1:36:15 prompts, and leveraging 40's inherent
1:36:18 knowledge base and chat context.
1:36:22 So meaning that if things have been
1:36:26 described, which everything's been
1:36:28 described, right, in poetry, in words,
1:36:31 in stories, in captions of
1:36:36 photos, if everything's been described,
1:36:38 then it can actually understand that and
1:36:40 it can understand what that looks like.
1:36:43 So that's the native piece of
1:36:45 this. Why it's probably good at context
1:36:47 is because it it it can it can translate
1:36:52 natively word ideas into image ideas.
1:36:55 That's pretty cool. Okay, got
1:36:57 it. Including transforming uploaded
1:37:00 images or using them as visual
1:37:02 inspiration. These capabilities make it
1:37:04 easier to create exactly the image you
1:37:06 envision, helping you communicate more
1:37:08 effectively through visuals and
1:37:10 advancing image generation into a
1:37:12 classic a practical tool for precision
1:37:14 and power. So, character consistency.
1:37:17 Oh, here's little videos. Let me change
1:37:20 my sharing so you can hear it. Look, I'm
1:37:23 doing this without a flipping
1:37:26 producer. Uh, what's deep
1:37:30 black?
1:37:32 What?
1:37:35 No.
1:37:41 Um, why is it not
1:37:45 Why can I not see the tab I'm on? Oh,
1:37:47 because I think I'm in a different
1:37:51 profile. Am I? No, I'm
1:37:55 not. Share screen
1:38:03 window. It doesn't see this
1:38:07 window. All right, I know what I can do
1:38:09 here.
1:38:22 Um, new. Ah, this will
1:38:25 work. Hang on, people. I'll be right
1:38:28 there with
1:38:29 you. Oh, it's not working. I can't drag
1:38:32 this into
1:38:36 here. Well, that's okay. Copy. Put this
1:38:39 in
1:38:45 here. Bring this back over
1:38:49 here. Hang on, people. Just calm
1:38:52 everybody. Calm
1:38:55 down. You're like, I I really thought he
1:38:58 understood what he was doing here. I
1:38:59 came here because this is supposedly
1:39:01 called the AI learning lab. I came here
1:39:03 to learn and right now it doesn't seem
1:39:07 like he knows how to use a browser. This
1:39:09 is about what I
1:39:11 expected. Um, introducing 40. Is this
1:39:15 it?
1:39:16 Yes. Look at
1:39:19 that.
1:39:22 Yes. Take that, boomers. [ __ ] figured
1:39:26 that out on my
1:39:28 own. Kyle, where's my producer? Spoke
1:39:31 too soon. Kyle. Yeah, right. When I brag
1:39:33 about not needing a producer, I can't
1:39:35 figure out how to launch a browser.
1:39:38 Okay. Uh, but I did learn something new.
1:39:40 If you've got two different um, what are
1:39:43 they called? Chrome profiles running in
1:39:46 the same Chrome. When you go to share
1:39:48 your window, you can only share in the
1:39:51 profile of the window. You doesn't
1:39:54 matter. I know what I know. What's the
1:39:56 deal now, though? All right. Okay. Here
1:40:01 we go.
1:40:06 [Music]
1:40:08 How's it going again? One thing that I'm
1:40:11 really excited Let's just rude Dennis. I
1:40:13 have learned that Kyle needs a
1:40:19 producer. Wait, I need my black bar here
1:40:22 for you people on TikTok, too. There you
1:40:24 go. Kyle, are you flexing about your
1:40:26 tabs again? Shut up. I like my tabs.
1:40:29 Leave me alone. Everyone's picking on
1:40:34 me. You're doing great, Kyle. Especially
1:40:37 missing half your team. Thank you. You
1:40:40 see? Okay, let's go back to videos now.
1:40:44 Abouten is the ability to keep
1:40:46 consistency in characters. I'm David
1:40:48 Medina or BMED and I work on multimodal.
1:40:52 What I want to show is one of my
1:40:54 favorite prompts which is can you create
1:40:56 a low poly penguin h make it very very
1:41:01 low poly. Surprisingly it's sometimes
1:41:03 hard to get very good low poly outputs.
1:41:05 It's not like other image generation
1:41:06 models where it tries to generate
1:41:08 something based on just the text.
1:41:10 Instead it uses the large language model
1:41:12 understanding of what does the user
1:41:14 want? What is the intent? I also like
1:41:15 some board games. So miniature like
1:41:18 games. So what I'll do now is generate a
1:41:20 miniature from this. So ideally we'll
1:41:22 see a penguin that looks like this with
1:41:23 the same staff and a hat. So can you
1:41:26 make me a realistic miniature as if a
1:41:31 professional made this and painted it?
1:41:35 This is what I think excites me the most
1:41:36 about image. The other image generation
1:41:38 models will try to create literally what
1:41:40 you said. But what's special about this
1:41:42 is one, it'll keep the context of this
1:41:45 character and then two, it'll understand
1:41:47 what I'm trying to ask it for and
1:41:48 generate very similar uh model but in a
1:41:51 miniature realistic style. It in that
1:41:53 that's huge. That that's huge that it's
1:41:56 it's like smart enough to un it's smart
1:41:59 enough to
1:42:00 meld the image context with the language
1:42:04 context, which makes sense. It's all in
1:42:07 the same model. If the image thing is
1:42:09 outside of it, then the only thing you
1:42:10 can send over there is a text prompt.
1:42:13 Huh. Fascinating. First, what I want, I
1:42:16 don't have to tell it every little
1:42:17 detail. One other realistic thing we
1:42:19 could do is can you make a crystal
1:42:22 version of this with light reflecting
1:42:25 and very realistic. Again, I'm just
1:42:27 giving it very very simple things.
1:42:29 Normally, this is not enough for other
1:42:31 models to generate something very
1:42:33 detailed, but the model understands what
1:42:35 I'm asking for. it'll think what type of
1:42:37 style it should have. So, this ability
1:42:39 to really understand what the character
1:42:40 is and make edits and understand what
1:42:43 the user wants. For me, it's the just an
1:42:45 amazing capability. Yeah, that's
1:42:47 amazing. And actually, you know,
1:42:51 like this is appropriately nerdy. I feel
1:42:54 like their announcements around the tiny
1:42:56 table with the with the four people
1:42:59 trying to act comfortable. Like, this
1:43:01 guy is just like comfortably nerdy. He's
1:43:04 like authentically nerdy on his couch
1:43:07 either at the office or at home. And
1:43:10 he's like, "Yeah, I nerd out making this
1:43:13 multimodal thing for us. And here's one
1:43:16 of my favorite things. It's a mage." Of
1:43:19 course it's a mage. He's a D&D dude, you
1:43:22 know? Come on. He's a tabletop gamer. Of
1:43:25 course he is. Like this is this is this
1:43:28 is what these announcements should
1:43:31 be. I wonder if they're getting their
1:43:33 [ __ ] together. This is good. That was
1:43:35 good. Do more like that. All right. Text
1:43:40 rendering. Well, we know it's good at
1:43:42 that.
1:43:47 Okay. So, Oh, this is their new This is
1:43:49 their new This is their new tiny table.
1:43:51 This is their new tiny table. Look, it's
1:43:53 the awkward couch in the in the San
1:43:56 Francisco overlooking the
1:44:00 bay. I like it. I'm down. I'm down. All
1:44:04 right, OpenAI, you're winning me over
1:44:05 with your leaning into the
1:44:08 awkward. It's just it's just like uh
1:44:12 it's like the AI learning lab. We're
1:44:14 we're neurosicy here, you know? Bring
1:44:17 your strange brains.
1:44:21 I'm Alan. I'm a research scientist at
1:44:23 Open AI. People tend to say a picture is
1:44:28 worth a thousand words. But being able
1:44:30 to also render like a few words or
1:44:32 symbols can carry like thousands of
1:44:35 pictures, you know, with a relatively
1:44:37 simple prompt like visualize an
1:44:39 infographic explaining Newton's prism
1:44:41 experiment in great detail with a wide
1:44:43 aspect ratio and a dark blue background.
1:44:46 So this is like an example where one
1:44:49 we're going to rely on being able to
1:44:51 render text in useful ways, combine it
1:44:53 with visual elements that actually
1:44:56 ground what this text about this
1:44:59 experiment even means and hopefully, you
1:45:01 know, help students who are more I got
1:45:04 to we we just got to we got to comment
1:45:06 on this. I mean,
1:45:08 this is [ __ ] insane, right? Like this
1:45:11 is from that prompt. It's got to it's
1:45:14 got to go to its large language model
1:45:16 and
1:45:17 understand how it would even describe
1:45:20 all this and then seamlessly translate
1:45:23 that in a way that it can bring it to
1:45:24 life. It's [ __ ] bonkers.
1:45:29 Who are more visual um learn both
1:45:31 through you know language descriptions
1:45:34 of a phenomenon but also you know a
1:45:37 visual imagination of what the
1:45:39 experiment actually looks like. It's no
1:45:42 longer just about making imaginary th
1:45:44 this format rather than the tiny table
1:45:47 format is so it's it's nice. It's just
1:45:49 relaxed. They're just geeking scenes
1:45:50 that look aesthetic and things like
1:45:52 that, but it's really about also the
1:45:55 colors are in the right sequences. Yeah.
1:45:56 It's like there's there's a lot of
1:45:58 information in that infograph
1:45:59 communicating and imagining and doing so
1:46:02 at the same time.
1:46:05 All right, that's that one. Nice.
1:46:11 upload and restyle. Oh, this is what
1:46:13 Danielle did where she uploaded two
1:46:15 different pictures. Hello. Thanks for
1:46:17 inviting me. Hi. Thank you so much for
1:46:19 coming. I'm so excited to talk to you.
1:46:21 Yeah, me too. They're all neurospicy.
1:46:24 It's so funny. It's like they're all
1:46:26 [ __ ] geniuses, but they're all like
1:46:29 eye contact.
1:46:31 I feel the ability of the im generation
1:46:33 of this model is becoming stronger and
1:46:36 stronger. My name is Lou. I'm a research
1:46:38 scientist in open eye working on
1:46:41 multimodel. Yeah, I'm showing a very
1:46:44 interesting demo today. So um so just in
1:46:49 this studio uh this is something that we
1:46:52 draw. Many people uses our tool to
1:46:55 generate comic books. So I'm going
1:46:58 to upload this drawing to CHP now. So I
1:47:04 just type this prompt now and it starts
1:47:07 to generate what will this drawing look
1:47:11 like as a real comic. Many of the times
1:47:14 especially when you play with the model
1:47:15 more and more you'll find things very
1:47:17 surprising. So I get this very funny
1:47:21 comic now. I want to replace this dragon
1:47:25 with this cutie penguin. Yeah, it looks
1:47:29 nice. Wow. I'm personally always very
1:47:33 curious about is how does this look like
1:47:36 in real
1:47:38 life? It looks cute. I like it. Nice.
1:47:42 [Music]
1:47:44 I'm digging this. Hello. Thanks for
1:47:46 invite
1:47:48 detailed directions. Okay, here's prompt
1:47:50 prompt adherence. I refuse to believe I
1:47:52 am
1:47:55 Neurospicy. It's the rest of the world
1:47:57 that's screwy.
1:47:59 I'm the only nor normal person and
1:48:01 that's scary. I know. It's so funny.
1:48:04 It's you
1:48:06 know I don't
1:48:10 know like before I knew what ADD was.
1:48:15 Like I I've never really had an issue
1:48:17 where I
1:48:19 judged ADD or my brain or anything
1:48:23 because I I wasn't diagnosed with ADD
1:48:26 until my mid30s.
1:48:29 But like I spent an inordinate amount of
1:48:33 time and
1:48:34 energy trying to make up for the fact
1:48:37 that I don't have good executive
1:48:39 function, right? Like I would buy every
1:48:42 organization system out there and every,
1:48:44 you know, back in the olden timey days
1:48:46 with fileaxes and I I would just buy
1:48:49 everything thinking that the right
1:48:52 notebook system would be the thing that
1:48:55 would let me organize my life. And it's
1:48:58 at some point I just finally was like,
1:49:00 "Oh, you're never gonna be good at
1:49:03 that." And then it was just like,
1:49:05 "Fine, whatever." So, but yeah, that's
1:49:09 one of those wild things. All right,
1:49:10 here we go. Thanks for joining us. Yeah,
1:49:14 no worries. How you doing? I'm doing
1:49:16 good. We're looking at an improved
1:49:18 image. Yeah, this is on OpenAI's
1:49:21 announcement of um
1:49:23 openai.comindexroducing-40-image-generation
1:49:32 generation in chat GBT. It's really good
1:49:34 at instruction following. My name is
1:49:36 Kenji and I work on multimodal research
1:49:39 here at OpenAI. There's a level of
1:49:41 attention to detail that is just not
1:49:43 captured by other models.
1:49:47 The first thing I'm going to show is
1:49:48 like 15 different objects and each one
1:49:51 of them has unique attributes that
1:49:53 differ make it very different from all
1:49:55 the other objects. An image containing
1:49:57 one a blue star, a red triangle, three
1:50:00 green square, four pink circle, five
1:50:02 orange hourglass, six purple infinity
1:50:04 sign, seven black and white polka dot
1:50:07 bow tie, 15 different objects. And
1:50:11 basically what what this will show is
1:50:13 that this image will just nail pretty
1:50:16 much every single one of these objects
1:50:18 that I've defined. Previous iterations
1:50:20 dolly you know image imagine things like
1:50:23 that they would probably get somewhere
1:50:25 on the order of like maybe five to eight
1:50:27 of these at most. It nailed it all. With
1:50:29 increased level of detail you can just
1:50:31 specify what you have in your mind to
1:50:33 chatbt. the tattoo will understand you
1:50:36 better and then generate that image and
1:50:38 it'll be just a very direct mapping from
1:50:41 what's in your mind to what you see on
1:50:43 the screen. All these image generators
1:50:45 nowadays, they look
1:50:46 good. All right, so it's got prompt
1:50:49 coherence. Great.
1:50:53 Transparent layers. Ooh, this looks
1:50:55 good.
1:51:00 Hi, nice to talk to you. Nice.
1:51:04 This guy's peak awkward. Peek
1:51:12 awkward. Okay. Hey, let's try it again.
1:51:14 When you sit down, give us a smile. Oh,
1:51:19 [Laughter]
1:51:22 okay. Oh, what's wrong, Jim? You got to
1:51:25 go.
1:51:31 Go
1:51:46 transparent layers. Hey, how are you?
1:51:49 Good, good. How's it going?
1:51:52 My name is Jen Fong Wang. I'm a
1:51:54 researcher in open working multimodel.
1:51:57 The way how to generate the transparent
1:52:00 image is pretty intuitive,
1:52:02 straightforward. And now let's This is
1:52:04 like an Apple commercial, too. Every one
1:52:06 of these engineers has a MacBook. Give
1:52:08 it a try. Let's say the content is cute
1:52:13 puppy cartoon
1:52:16 st. This is the prompt. And uh now let's
1:52:20 see what happened. The model will take
1:52:22 the input and tries to generate the
1:52:25 image. Let's give it some time. Wow. So
1:52:29 now let's see what we generated. This is
1:52:32 the transparent puppy. Another
1:52:36 application is to make it a sticker.
1:52:38 Let's give it a try. Yeah.
1:52:41 We can easily overlay the transparent
1:52:43 image onto any kind of background. Now
1:52:47 we can copy the sticker onto our laptop
1:52:51 and and then we can paste it here.
1:52:55 Uh, make it smaller so it can be easily
1:52:59 blended with the background. Can you
1:53:01 make me a sticker? You sure? A smart
1:53:04 researcher wearing glasses and a blue
1:53:06 shirt.
1:53:09 Okay, let's give it a try. The director
1:53:13 because they are young with glasses.
1:53:15 He's making fun of him to his face. We
1:53:17 got it. Yeah, I think uh it works pretty
1:53:21 well. Hopefully people love it.
1:53:25 That's pretty amazing. You can just make
1:53:27 transparent images. Wow. Crazy. All
1:53:29 right. Good. Nice. Let's keep
1:53:33 going. And then we're back to the
1:53:35 beginning. Okay. Improved capabilities.
1:53:37 We trained our models on joint
1:53:39 distribution for online images and te
1:53:41 text. Learning not just how images
1:53:42 relate to language, but how they relate
1:53:45 to each other. Combined with aggressive
1:53:47 post-training, the resulting model has
1:53:50 surprising visual fluency capable of
1:53:53 generating images that are useful,
1:53:55 consistent, and contextware. Text
1:53:56 rendering. Okay. Street
1:53:59 signs. Nice. That's good. Reindeer
1:54:03 parking. Multi-turn generation. Because
1:54:06 image generation is now native to
1:54:09 GP240, you can refine images through
1:54:11 natural conversation.
1:54:14 GPT40 can build images and text in a
1:54:17 chain in kind. For example, if you're
1:54:19 designing a video game
1:54:21 character, the character's appearance
1:54:24 remains coherent across multiple
1:54:27 iterations. Okay, this is really good.
1:54:30 This is really
1:54:34 good. So, I could actually probably do a
1:54:37 little story with Champy. Yeah, look at
1:54:39 this. This cat's consistent. Well, at
1:54:41 least in what they showed.
1:54:48 We're gonna have to come back to this.
1:54:49 There's so much
1:54:51 here. All right, we'll we'll continue
1:54:53 this tomorrow. We'll continue this
1:54:55 tomorrow. That's
1:54:57 amazing.
1:55:01 Um, oh, I got to change
1:55:04 my stop that share. I got to share this
1:55:07 and share
1:55:09 everything. Go back here. Let's go look
1:55:12 at Let's go see how our screen play is
1:55:14 doing. Wait, we'll do our images. Oh,
1:55:16 yeah. There's the woman. Oh, that's much
1:55:18 better. Look at that image,
1:55:20 people. Uh, let's say
1:55:26 uh that's really
1:55:33 good. She is sadly
1:55:36 hopeful reading about spooky quantum
1:55:39 effects.
1:55:41 That's pretty
1:55:42 slick. All right. Very, very, very
1:55:46 [ __ ] cool. This is This is This is
1:55:50 going to be a lot of nights of us
1:55:52 playing and figuring out what this thing
1:55:54 [ __ ] does. All right. Back to uh
1:55:56 Manis. I created some bold signature
1:55:58 lines. Get these [ __ ] serpents
1:56:01 out of my
1:56:03 [ __ ] prison plane.
1:56:08 No, I've had it with these [ __ ]
1:56:10 snakes. 25 years for a crime I didn't
1:56:13 commit, and now I got to deal with these
1:56:14 [ __ ] snakes on this
1:56:16 [ __ ] prison
1:56:17 flight.
1:56:20 Okay. All you did was copy the line and
1:56:26 add
1:56:28 crap to
1:56:30 it. Get creative.
1:56:34 [Laughter]
1:56:39 Oh my
1:56:41 god. All
1:56:44 right. Okay. That would beat stable
1:56:47 diffusion on character retention. Yeah,
1:56:49 I want to play with character retention
1:56:52 uh tomorrow. We'll we'll we'll play with
1:56:53 that. I want to I want to read up more
1:56:55 on this. Um this is I think this is a
1:56:58 bigger deal.
1:56:59 Again, I think that most people are
1:57:02 going to not even know that this
1:57:03 announcement
1:57:07 happened and they're going to just use
1:57:09 chat GPT image generation like they
1:57:11 always have and they're not going to
1:57:14 quite get the magnitude of this. I think
1:57:16 this is a pretty big deal. Um, but we'll
1:57:20 see. Like let let's test this over a
1:57:22 bunch of days. Okay. When your cellmate
1:57:24 turns into your seatmate and your
1:57:27 seatmate turns into a snake. That's not
1:57:30 bad. The only thing worse than a life
1:57:32 sentence, a death sentence with fangs.
1:57:35 Not bad. They locked me up with killers.
1:57:37 Now they've strapped me in with vice
1:57:39 vipers. Not bad. Prison didn't break me,
1:57:42 but this plane just might. One bite at a
1:57:45 time. When the judge judge said flight
1:57:47 risk, this isn't what I had in mind.
1:57:50 Okay, number five is
1:57:53 genius. Five is
1:57:56 genius. Now go write this
1:58:04 thing. We had to slap it around a
1:58:06 little, but it got
1:58:10 there. It does have character
1:58:12 consistency. That's exciting. Paul
1:58:14 Ritzer said that no one knows about deep
1:58:16 research in chat GPT2. Yeah, I
1:58:21 listen. How do I even say
1:58:24 this? I would say at this point chat GPT
1:58:28 has escaped me. Like I pay attention to
1:58:31 this stuff all the time. I have not kept
1:58:34 up kept up with just what chat GPT can
1:58:37 do and how these models are different
1:58:39 and when to use certain ones. Part of
1:58:41 the reason for that is Sam Alman said
1:58:43 when GPT5 comes out, you're not going to
1:58:45 have to know which models to use because
1:58:48 it'll just figure it out. So, part of me
1:58:50 has just gotten lazy because of that.
1:58:53 Um, but that's just one tool. Like, I I
1:58:56 feel completely clueless
1:58:58 about
1:58:59 Claude. I feel completely clueless about
1:59:03 Gemini. Um, I've played with Deepseek a
1:59:06 little bit. I've played with Manis a
1:59:08 little bit. Like I I haven't I haven't
1:59:10 scratched the surface with these
1:59:12 things.
1:59:14 Um I'm gonna dig deep on this image
1:59:16 stuff because there's something about
1:59:18 the fact that it understands both the
1:59:21 language. It's like birectional like it
1:59:24 understands what's in the images and can
1:59:26 describe that and it understands if you
1:59:29 describe something what that looks like.
1:59:31 That's that feels really significant to
1:59:33 me. Um this is exciting. Okay, so Manis
1:59:38 is off writing my screenplay. We're
1:59:41 going to have a produced screenplay by
1:59:42 the end of uh May, which is exciting.
1:59:45 We'll be in production by
1:59:47 June. And uh we probably got to lock
1:59:52 down. I don't know if we want Samuel L.
1:59:56 Jackson. That's a little derivative,
1:59:58 don't you think? I think we should find
2:00:01 who's the new upcoming Samuel L.
2:00:03 Jackson.
2:00:05 Yeah, we'll have to figure that out. So,
2:00:08 we're going to have to book them. So, it
2:00:09 we might have to push it to like July or
2:00:11 August because of their filming schedule
2:00:13 if they're if they're in other things,
2:00:15 you know. So, all right. And I got to
2:00:17 work that in between launching my play
2:00:19 on Broadway. All
2:00:21 right. All right,
2:00:23 everybody. Okay. I can finally get Chat
2:00:26 GPT to generate an illustration for my
2:00:28 novel with zero artistic ability.
2:00:30 Beautiful. Love it. gives me more
2:00:32 reasons to start a new print on demand
2:00:34 brand. It c it sure does actually.
2:00:37 Claude is great for writing articles
2:00:39 actually. Yeah, the fact that that that
2:00:42 this image I I mean
2:00:45 um Ideogram was pretty good at this, but
2:00:47 it feels like this this is going to be
2:00:49 really good at things like t-shirts and
2:00:51 things like that. So, print on demand
2:00:53 might not be a bad call at all. So, all
2:00:57 right everybody, I'm gonna get my ass
2:00:59 out of here.
2:01:01 have yourself a fantastic evening. Go
2:01:03 play with Chat GPT's new image model. If
2:01:05 you don't know about it, just go play
2:01:06 with it. Tell it to do something. Tell
2:01:09 it to uh make a picture with 20 objects
2:01:12 in it and uh and
2:01:16 and you know, have it describe all 20
2:01:19 and then see if they're all in there.
2:01:21 That'll be a good one. And make
2:01:23 transparent stickers. Cloud's great for
2:01:25 article writing. I'm told at Beach Chat
2:01:27 GPT for coding. Gemini is kind of
2:01:29 limited for creative short
2:01:31 things. Cool. Have a good night,
2:01:33 everybody. All right. Peace out. See
2:01:36 y'all later.