4/18/2025 - Reasoning Models and Multimodality in ChatGPT o3

Description

Chapters

Transcript

0:03 [Music]
0:09 [Music]
0:10 [Applause]
0:12 You and I get all
0:15 alone. Sunday morning here at
0:19 home. Blue and the car is strong. It's
0:23 true.
0:27 Then I open my eyes to a dream real life
0:31 in front of
0:34 me. And I haven't got a clue what in the
0:37 world is happening to
0:41 me. Think I think I'm
0:44 happy like first day summer vacation.
0:48 Happy got to get a little rest,
0:50 relaxation.
0:53 Happy choir on Sunday morning singing
0:56 true.
0:57 [Music]
1:29 Every time I see you, I get that look in
1:35 mine. Every time I see your mouth, I
1:40 hear that
1:43 smile. The early misty morning light. I
1:47 heard the engine
1:49 turn the old foot
1:54 [Music]
1:58 outside. You were leaving
2:01 me again
2:04 today. You will convince
2:08 me again today.
2:12 You're leaving this heart town looking
2:15 for someone else's golden
2:20 ring. Should
2:22 say so long,
2:29 Suzan. Now don't you
2:34 [Music]
2:37 cry. So long, Suzanna.
2:44 Don't you cry for
2:46 [Music]
2:56 me. Sharing drinks and cigarettes and
2:59 keeping warm out on the
3:02 road. Chasing down a lifestyle out on
3:05 Highway 24.
3:09 New York State was a rolling breeze and
3:11 the sunshine with a blue sky falling to
3:15 the chill of all September
3:18 [Music]
3:21 creeping. You believe in
3:24 me again
3:27 today. You convince
3:30 me again
3:33 today. You're leaving this hotel looking
3:37 for someone else's golden
3:42 ring. Should have
3:44 say so long,
3:47 [Music]
3:50 Suzan. Hush now. Don't you
3:56 cry. So long, Susanna.
3:59 [Music]
4:03 Don't you cry for me.
4:06 [Music]
4:14 Thank you so much, Winston. I saw the
4:16 galaxy you sent. That was beautiful.
4:18 Generous. You're amazing. Thank you so
4:20 much. Welcome everybody. It's Friday
4:22 night date night. You know what we do
4:24 here. We cook nachos. We microwave hot
4:28 pockets. We have fun. We learn. We learn
4:31 together. Tonight we're gonna explore
4:35 stuff. Posted my poster about five
4:38 minutes after
4:40 [Music]
4:46 you.
4:49 [Music]
4:54 Um, one thing I'm excited about. Am I
4:57 excited about it?
5:00 one thing that
5:04 I I feel validated
5:08 by. So, so this new 03 model from Chat
5:12 GPT and the O4 mini
5:14 things are just a quantitative and
5:19 qualitatively different tool than we've
5:22 ever had before.
5:25 And while I don't
5:28 know, I don't yet personally
5:31 know how it's a big deal, I just have
5:35 this deep deep deep instinct that it's a
5:37 really big [ __ ] deal.
5:40 And like building from the launch
5:43 through like right now, like right
5:44 before I came on, my Twitter timeline is
5:48 just people I respect up and down the
5:51 board from the sort of math and science
5:54 side, from the AI, you know, influencer
5:57 side to just regular people like the
6:00 like everyone is posting that this is a
6:03 big [ __ ] deal.
6:05 Um, so that's that's largely what I want
6:08 to talk about tonight.
6:11 [Music]
6:23 Um, Kyle, please go look in your
6:25 regulars at Where in the World is Kyle
6:27 Shannon. I can't pass that one up.
6:35 All right, AI learning lab, here we
6:38 come.
6:40 Yes. Um, no, I was just going to tell
6:42 you on the uh where in the world is uh
6:44 Kyle Shannon, um the post is good in of
6:47 itself, but the real um specialty lies
6:51 in the chat link that Danielle shared.
6:55 And it's it's link it's if you expand
6:57 out her text the chat GPT story or chat
7:02 that she had is in her um message. So if
7:06 you scroll up to click on this and then
7:11 nope
7:13 scrolling up just a little bit because I
7:15 was just looking at this. Pretty cool.
7:16 So expand that. Yeah. And then open. So,
7:19 what basically what she did was she
7:20 reversed image searched a screen grab of
7:23 your live, but you have to go in, scroll
7:25 up to the top, and look at the process
7:26 that 03 took to find you. And then she's
7:30 wondering if you know if it got it right
7:32 or Oh, cool.
7:37 Okay. Upload an image. This is a copy of
7:40 a I can't see which image it is, but
7:43 that's all right.
7:45 [Music]
7:46 Um, so uploaded some more images. So I
7:49 don't know what they are. Danielle, tell
7:51 me which images you did. Uh, she grabbed
7:54 a a screen grab of where you were saying
7:56 go play now outside of a building. Oh,
7:58 okay. Okay. Yeah. Yeah. Yeah. It should
8:00 be able to find that one. All right.
8:02 Let's see what it does
8:04 here. I mean, I love all the Okay. So,
8:08 if if you're new
8:10 to Wait, let me flip this. Let me put my
8:14 black bar up. We're jumping right in
8:17 here. This is cool, Danielle.
8:20 Okay. So if you're new to
8:24 what what makes 03 different, why it's
8:28 unique and
8:29 interesting is
8:32 that if you know a little bit about
8:35 normal large language models and
8:37 reasoning models, the the the basic
8:40 difference between them is a normal
8:42 large language model like GPT40, the ba
8:45 the basic model at the top of the list
8:48 is that you ask it something and it
8:50 gives you an answer, Right? A reasoning
8:53 model, you ask it something and then it
8:55 has a conversation with itself. It calls
8:57 it thinking as Pate M rightfully points
9:00 out. It's not actually thinking, but it
9:02 behaves like it's thinking, right? It's
9:04 it's sort of prompting and responding
9:07 responding behind the scenes before it
9:10 gives you an answer.
9:13 03 amplifies that. It can do much longer
9:18 chain of thought conversations.
9:22 It's also multimodal. So it can
9:24 incorporate images and do image analysis
9:27 as well as um as well as text, you know,
9:30 large language model stuff. Can do
9:32 both. And it can do tools. It can write
9:36 and execute its own Python code. So So
9:41 if what she asked was, you know, tell me
9:44 where this thing is. if it needs to
9:46 analyze a photo to try to find things
9:48 out,
9:50 um, it
9:52 will write its own code to do that
9:54 analysis as part of its internal
9:57 conversation with itself. So that's some
9:59 of the stuff that we're seeing here. So
10:00 it thought for 43 seconds, which in 03
10:04 is a pretty long time because it's a
10:05 very, very fast model. And this is the
10:08 epitome of we'll do it live. go ahead
10:10 and
10:11 uh refresh your screen or uncheck the
10:15 box for all HO details because you're
10:18 seeing something very different than
10:19 what I'm seeing and what I'm seeing on
10:21 my end is um probably more relevant uh
10:26 to how the model operates. Wait, I'm
10:28 just not sure why it didn't load on your
10:30 end. And if not, I can share screen. Oh
10:33 yeah, now it's got pictures in it. There
10:34 we go. Okay, now now we're seeing it. So
10:37 she didn't upload those like screen
10:40 grabs. That's the model like zooming in
10:44 and analyzing in real time. So it like
10:47 that's the original screen share and
10:49 then as you scroll down it's like
10:50 pinching in on your like your cheek and
10:52 building behind you. Great. That's the
10:55 model doing that. That's not Yeah.
10:58 Danielle giving them the information.
10:59 Yeah. Exactly. So that's perfect. So you
11:02 know her prompt. Can you tell me exactly
11:05 where this location is? Right. And then
11:08 and then it starts like if you have
11:10 another frame that would help. And then
11:12 it looks like maybe she uploaded some
11:13 stuff. But now it's importing the image
11:16 and you know uh looking at it. So
11:19 there's the first it looked at the whole
11:20 image first. Now it chopped it up.
11:23 Looked at that part of it. Now it
11:25 chopped it up. Looked at that part of
11:27 it. That's actually um this is a brewery
11:30 right across the street from the place I
11:32 work. That's the place I work. And these
11:35 are flipped for some reason, I guess,
11:37 because it was flipped on on TikTok, so
11:39 they're
11:39 backwards.
11:41 [Music]
11:42 Um, it's looking at my little fuzzy
11:45 sweater. It's looking at my my
11:48 dramatically receding
11:55 hairline. There's some fat neck. It's
11:57 like, dude, dude, stop eating the
11:59 burgers.
12:03 Oh. Oh my. Oh my.
12:08 So, what will be really impressive is if
12:09 it not only tells you where you are, but
12:11 that you should go to a gym. Exactly.
12:14 Exactly. Here's our recommendation.
12:16 Here's here's where Kyle works, and
12:19 here's a gym within walking distance you
12:21 might want to consider.
12:26 But okay, not for
12:29 nothing, here's here's what what is
12:32 different about
12:33 this. You could before with code
12:37 interpreter, you could upload a photo
12:40 and you could say to chat GPT, zoom in
12:43 on part of this image and analyze what
12:45 you see in it, right? And it would write
12:47 Python code and it would zoom in on part
12:49 of an image and it would tell you what
12:50 it saw. This is doing it by itself, you
12:55 know, as part of the thinking process.
12:58 So these are called tools. Every time it
13:00 writes one of these, it's basically
13:01 creating a tool for itself to use in the
13:04 process of of trying to solve the, you
13:07 know, answer the question. All right, so
13:09 let's keep going. Let's quickly get past
13:11 my mouth, shall
13:12 [Laughter]
13:16 we? All right, let's
13:19 see. Uh oh, that's cool. Now it flipped
13:23 it. It It inverted the colors or
13:26 something. Very
13:30 cool. It's doing good old and and you
13:33 know, if you think about
13:35 this organizations like the CI CIA,
13:39 right, if they're doing, you know, like
13:41 satellite photo analysis or they've got
13:43 a picture of someone, you know, now
13:44 we've got AIS that just, you know, you
13:45 take a picture and it tells you exactly
13:47 where it is. But we're watching this
13:48 thing sort of figure it out on its own.
13:51 Um, I really liked your
13:59 mouth. That's pretty funny. Can Can you
14:02 tell me the last time this guy went to
14:04 the
14:07 dentist? Yeah, it's doing negatives.
14:10 Look at
14:12 that. Oh, what's also cool
14:15 here, this took a while. Look this one.
14:19 It said thought for 5 minutes and 35
14:25 seconds. Wild. Oh, look. It found this
14:30 is a golf place in my
14:32 building. Chat with vision walked
14:34 through my garden today. Oh, cool.
14:37 Suggested plantings for now and for
14:40 light space. That's very cool. Chat GBT
14:43 has a crush on you.
14:48 Okay, wait. Here we go. I took another
14:51 deep dive. New frame, zoom, sharpen,
14:53 projected, sign, Google induct
14:55 storefront images. The best probable
14:58 match I can find for the black squared
15:02 mountain and lines logo is A-ine
15:04 Boutique's hanging blade sign in
15:06 Denver's Loto micro district district
15:08 called Dairy Block. That's not right,
15:10 but it's close. It's really [ __ ]
15:12 close. like it's this is probably a you
15:15 know 3 minute drive
15:19 away. What would lock it down? One
15:21 steady high resolution still of the
15:23 sign. Yeah. Any cross street or
15:27 awning? There's a there's a tool out
15:30 there right now that that does this. You
15:32 can take a shot like the one you put up
15:34 there and it figures out what the
15:35 building is and then goes goes and finds
15:37 its location. But this got I mean not
15:40 for nothing this is remarkably
15:43 close. Like I wonder why like how it got
15:47 to Denver. Is there anything in the
15:50 original image that that
15:54 says where I am? No. Huh.
16:00 Fascinating. No idea. That's very very
16:03 cool. All right. So So there's a there's
16:05 a use case. So that's that's what we're
16:07 going to talk about
16:09 tonight. And this is a good example of,
16:13 you know, so imagine a similar sort of
16:15 thing where you take some crazy neo
16:18 gothic painting, you know, one of those
16:21 big giant 8 foot wide paintings that's
16:24 got all sorts of crazy, you know,
16:27 imagery and symbology in it. And you
16:30 upload that and you tell this thing,
16:32 tell, you know, find everything you can
16:34 possibly find in this painting and
16:36 catalog it for me and research the
16:39 history of everything you find and, you
16:40 know, tell me all of the allegorories
16:42 and all the metaphors and all the all
16:44 the symbols and all that sort of stuff.
16:46 Danielle um should should ask if you
16:49 have any cavities. Yeah, exactly.
16:53 Um, I feel Vicki I feel like it would
16:56 just be quicker to ask Kyle where he
17:04 was. Always coming in with the logic.
17:07 It's a good It's a good shot, Vicki.
17:09 It's It's a It's a
17:15 winner. But that's pretty awesome.
17:25 Um, very very cool. Okay. Um, all right.
17:29 So, so let me So, I figure here's what
17:32 we'll do
17:33 tonight. Um, first of all, how is
17:38 everybody? Possibly Jake Vicki, but this
17:40 is Chad add.
17:47 Well, you know, and if you ask me where
17:48 I am, you know, my response is going to
17:50 be like, "None of your damn business.
17:52 Shut up, you nosy
17:55 person." You know, so that would be
17:56 awkward. So, have chat GBT do it.
18:06 [Music]
18:21 Does everyone in Denver have that shirt?
18:23 No, nobody has that shirt. I'm fancy.
18:26 What's champ? What's up? What's up,
18:29 dude? Tik Tok pin. All right, we saw
18:31 that one. Beautiful.
18:36 [Music]
18:43 Um, reinvigorated post Friday Kyle
18:46 meeting. Nice, Gwen.
18:50 Fantastic. Yeah, there there was uh the
18:53 AI office hours today. If you weren't at
18:55 it, we went through some things. So,
18:57 here's here's the good
18:59 news. If you are at office hours today
19:03 and if you're here tonight andor andor
19:08 um I'm gonna give you what I think is a
19:10 really
19:11 cool powerful
19:16 um unlock for for o for testing 03 for
19:21 discovering what it makes possible. I
19:22 think what Danielle just did there, like
19:25 like that use case, find out where this
19:27 is. That's one way you could use that.
19:29 But basically, upload an image and have
19:31 it do detailed analysis of that image.
19:34 There's all sorts of things you could do
19:35 with that. Like that's a great that's a
19:38 great use case to explore 03, but you
19:41 could also do it with 04 mini high and
19:43 see if they're different.
19:47 Um, I said this today on my content
19:52 evolution call. I said this today on my
19:54 office hours and I'll say it again
19:55 tonight.
19:59 Um, I think it is probably in everyone's
20:02 interest if you're interested in AI, if
20:04 you want to start to understand what is
20:06 now possible that wasn't possible 3 days
20:10 ago, spend at least an hour a day really
20:14 pushing 03, 03, 04 mini high, 04 mini,
20:20 just those three reasoning models that
20:23 are the new ones that came out this
20:24 week.
20:26 Um, do everything you can to understand
20:30 what they are, how they work, what
20:32 they're good at, what they're not good
20:33 at, what they're good at for you. That's
20:36 the most important thing. What they're
20:38 good at for you. All
20:41 right. So, what I'm going to show you, I
20:44 I it always baffles me how long [ __ ]
20:47 takes, but I I put together this
20:49 document today of all these prompts, and
20:51 I'm going to give you these prompts. I
20:52 just posted it on my ex account as an
20:54 article.
20:58 Um, but and I'll put it on the the AI
21:01 salon uh as well and I'm going to also
21:04 put it on LinkedIn because because I
21:06 want people to explore this. My request
21:08 is as you learn stuff about 01 just like
21:11 Danielle did, share it in the AI salon.
21:14 You can share it in the regulars channel
21:16 in the in the AI learning lab channel.
21:18 You can share it in, you know, the
21:20 general talk stuff at the top of the
21:22 list. Um, you can share it in here's
21:25 something I made or here's something I
21:27 discovered. Um, just start sharing what
21:29 you're learning
21:33 because there going to be so many
21:36 different people are going to unlock so
21:39 many different ways this is powerful
21:41 that only if we kind of share what we're
21:43 learning will will it be good. What are
21:46 the usage limits for the new models? Uh,
21:48 I don't know, but I do know that you
21:52 um I
21:54 saw Where did I see that? Oh, no. Maybe
21:58 that was in Was it here? Hang
22:08 on. Is it in
22:10 here? I thought I saw somewhere it said
22:13 how many things you have left, but I
22:15 don't I'm not seeing it now. So, I don't
22:17 know. All
22:18 right.
22:20 Um, if you haven't played with deep
22:23 research, that's another one to play
22:25 with. Just click that button and see
22:27 what happens.
22:29 Um, and just just
22:36 like I think it's it's hard
22:40 to it's hard to
22:45 overstate how much of a reset this is in
22:48 some way. And here here's what I mean by
22:49 that. Yes, it's in chat
22:52 GPT. Yes, it's a reasoning model. Yes,
22:55 it's just the sort of next one from 01,
22:58 right? We had 01 and 01 pro and now
23:00 we've got 03. O2 was a trademark thing,
23:04 so it doesn't exist in case in case
23:06 you're like, wait, what happened to O2?
23:08 I learned how to count in kindergarten,
23:10 and that's not how it goes. It that's
23:14 how it goes with Open AI.
23:16 Okay, so on the surface, it looks
23:19 exactly the same. It's not. It's a
23:23 reasoning model with
23:25 tools. It can think for itself. It can
23:28 analyze. It can do
23:30 mathematics in the process of analyzing
23:34 something. I mean, it would be
23:36 fascinating and you could do it. You
23:39 could take that that prompt that
23:41 Danielle shared and go look at that chat
23:44 and just look at how it actually found
23:45 it. Like, what was that process? What
23:47 tools did it write? How did it figure
23:49 that [ __ ] out?
23:52 Did it did? Maybe it saw, you know, the
23:55 Rocky Mountains in the
23:57 [Music]
24:01 background. I talked to the new chat GBT
24:04 and it said it
24:07 made made it and all I had to
24:11 do Oh, was click to explore. Oh, yeah.
24:14 Okay. So, Silver Fox brings up a really
24:17 good point.
24:20 One of the things that
24:22 03 in particular, but I I assume 04 and
24:26 uh 04 04 mini and high have this as
24:30 well. It hallucinates like a
24:34 [ __ ] It just it it unashamedly
24:38 makes [ __ ] up. So it will it will say
24:40 here's a Dropbox link with all these
24:42 files that I created. And you click on
24:44 the Dropbox link and it doesn't exist.
24:47 So, uh, so that'll have to deal with,
24:50 but yeah. So, so just know that it
24:53 hallucinates a
24:55 lot. But I just I, you know, I feel like
24:58 we've spent the past two years kind of
25:00 understanding, you know, how to use
25:02 large language models, how to prompt,
25:04 how to get good stuff out of them, you
25:06 know, how to do it for particular use
25:07 cases.
25:09 Um, and then March 31st of this year, we
25:12 get this new image generation model,
25:14 which is a really powerful and unique
25:17 thing. And it's funny, the the sort of
25:20 Google lovers out there are like,
25:21 Google's had that for months. They've
25:23 had integrated image gen model in, you
25:26 know, they've had a multimodal large
25:28 language model, but there's something
25:29 about OpenAI's implementation of it.
25:32 It's just better. It's the first one
25:34 that I'm experiencing it. It generates
25:39 um like commercially
25:42 usable creations, art, you know, ads,
25:46 social media post, YouTube thumbnails,
25:50 just right out of the gate. You don't
25:52 have to take it into some other assembly
25:54 program like an Adobe Suite
25:56 program. So, that really changes things.
25:59 And then this new model, because it's
26:01 it's got reasoning with
26:04 tools, shit's different. It's a It's
26:06 just a different thing. And like when do
26:09 you use it? How do you use it? Why would
26:10 you use it? Is there any reason not to
26:13 use
26:14 it? And the answer is yes. Sometimes you
26:16 need a quick answer for something.
26:18 Sometimes you just want a paragraph
26:19 written. But sometimes you want a
26:21 report. And in that report you want it
26:23 researched and analyzed and thought
26:25 about and graphs made and charts
26:27 made. Then it can do all that.
26:31 Okay. 50 per week for plus. There you
26:34 go.
26:37 Um, and I mean, you know, there's
26:39 there's a part of you probably going
26:40 like 50 a week doesn't sound like much,
26:42 but
26:43 like that thing did all that thinking.
26:46 You know, that was a single prompt and
26:47 it did all that work. So, if you come up
26:49 with interesting prompts, you're going
26:51 to be good. And I'm going to give you a
26:52 bunch of those tonight. All right? And
26:54 we'll we're going to go play with those.
26:55 Okay. Thoughts,
26:57 questions? Anything?
27:00 [Music]
27:04 No. No. Everyone's like, "It's Friday
27:08 night. I'm just drinking, Kyle. Could we
27:10 just not be so intense on the edge
27:14 of Thanks. Thanks. Hey, Kyle. Kyle,
27:19 listen. Listen. I love that you're so
27:21 passionate about the artificial
27:23 intelligence, but you know this Harry
27:25 Kerry. We like baseball.
27:28 baseball booze and you know hanging out
27:31 on Friday night maybe lighten
27:36 [Music]
27:38 up does it retain its tools? No, it
27:41 doesn't. It's like it's it's literally
27:44 you can almost think of it like this.
27:45 It's it's in the chain of thought. It
27:48 says, "Okay, I've gotten population data
27:51 from this county and this county and
27:54 this county, and the next thing I need
27:56 to do is I need to consolidate the data
28:00 across those three counties." And it's
28:02 like, "Oh, okay. Well, then I need to
28:05 write a Python application to extract
28:07 the data, put them into a new
28:09 spreadsheet or a new data table, and
28:11 consolidate it so I can look at it." So
28:13 it writes the Python code, consolidates
28:15 the data, and goes, "Okay, here's your
28:17 data." And then it goes, "Okay, now I
28:19 got the data. Awesome. Now I need to
28:21 actually analyze that data and look for
28:22 trends. So let me write a Python code to
28:25 analyze the data based on what I what it
28:28 has and analyze the trends." So it
28:30 basically just keeps creating a tool,
28:33 executing it, looking at the answer,
28:35 figuring out what it needs to do next,
28:37 writing another tool, or just using the
28:40 large language model. that basically is
28:42 just working its way down through
28:45 this process of of it trying to get to
28:48 the answer. So, it's it's insane. It's
28:51 crazy. All
28:59 [Music]
29:05 right. Now, Brian, one interesting thing
29:08 that you raise
29:11 Because chat GPT now has memory. Like if
29:14 you don't know this, the other thing
29:15 that happened in the past week, week and
29:17 a half is chat GPT now remembers all of
29:21 the chats you've ever had with it. So
29:24 it's possible, Brian, that it it it
29:27 would understand that it it it created a
29:30 tool for a previous prompt and it it was
29:32 successful at a
29:34 thing. But I don't think Chat GPT
29:37 necessarily cares. Like I feel like
29:39 that's a human thing. Like I wrote a
29:41 thing that works. Let me reuse it so I
29:43 don't have to think about it and
29:45 recreate it the next time. But it
29:48 doesn't care. It's just like I need a
29:50 tool that analyzes this specific data
29:53 from these three specific sources that
29:56 might not be reusable. So like it's
29:58 disposable code. It doesn't care. Dr.
30:01 have to drop and we'll have to do
30:03 re-watching. Well, thanks for hanging in
30:05 or checking in and hanging out. verified
30:08 by chat. GBT 50 messages per week for
30:11 plus and near unlimited for pro. Not
30:15 sure what near unlimited means. I guess
30:17 basically if you start melting their
30:19 servers, they'll shut you down. Um I
30:22 used 03 to create a proposal from an RFP
30:25 and it did a pretty good job. Formatting
30:27 was I assume shitty. Oh, it was amazing.
30:30 Cool.
30:32 Awesome. Formatting was amazing. Writing
30:37 Why Why is this so blown out? Tik
30:43 Tok. Um, do I
30:47 have I'm going to try
30:50 something.
30:53 Brightness. Hang on. Wait.
31:14 I'm trying I'm doing monitor
31:20 settings. Oh, god damn it.
31:24 Um, without really knowing how to
31:30 um.
31:32 Ah, god damn
31:42 it. Hang on people. This is if you think
31:47 this is
31:55 easy. Uh, yes.
31:58 Okay. Yes. Okay.
32:05 Yes. Oh, look. I'm doing
32:08 it. Tik Tok. You can actually see me
32:11 now. Okay. Now I can't I can't see the
32:14 screen at
32:16 all. Let me turn brightness up a little
32:18 bit. Maybe we had too high contrast for
32:21 Tik Tok. Hang
32:25 on.
32:32 Okay. Okay. There. I can see it a little
32:35 bit. And yeah, you can you can see
32:38 things
32:39 now. Tik Tok. I fixed it.
32:42 [Laughter]
32:46 You're
32:47 welcome. Welcome to chat
32:53 add.
32:55 Okay. How to change your monitor
32:58 settings without reading the manual by
32:59 Kyle
33:03 Shannon. Oh my goodness. We're back in
33:07 business. Look at that. He can actually
33:09 now see me. I like it's
33:12 unfortunate.
33:16 Um, only took four months. Shoot up.
33:20 Producer Brandon's smacking me down. Uh,
33:23 could we have done this night one? We
33:26 could
33:27 [Laughter]
33:29 have. Can you have a guitar in your
33:33 hands? Maybe more difficult. Oh, because
33:36 you have a guitar in your hands. Maybe
33:37 more difficult. No, this is just
33:43 um Okay, so here's what I thought we'd
33:48 do. I thought it might be nice, let me
33:50 get rid of the the
33:52 guitar. What I thought might be
33:56 nice, is to go to the go to the Twitter.
34:00 Some call it the
34:02 X and we don't call it the Twitter or
34:05 the X
34:07 boomer. So, you go to the
34:09 X and I figure we'll just like scroll
34:13 and look at what people are talking
34:15 about. All
34:18 right. What am I doing here? I got to
34:21 adjust
34:23 my so used to seeing myself blown out.
34:26 I'm a little caught off guard here.
34:28 Okay. Producer Brandon in the house.
34:31 Dennis M. I can't control Kyle's
34:33 monitor.
34:36 Wait, are we making fun for Kyle for not
34:39 having a producer again? No, this is the
34:41 one thing that Brandon controls
34:43 everything except the quality of of the
34:46 display on Tik
34:52 Tok. This is I will not I will not uh
34:55 whatever you call it blame producer
34:58 Brandon for that. There's much I can
35:00 blame producer Brandon for, but not that
35:02 one. Okay. Okay. So, I figure what we'll
35:05 just do is scroll Tik Tok for a bit and
35:08 see um what people are talking
35:15 about. Yeah. Oh, what did I say? Tik
35:18 Tok. Yeah, I
35:27 mentioned. Yes, ma'am. Get back to
35:31 work. All right. It's really good at
35:33 stuff. Dear algorithm, how do you guys
35:36 deal with burnout? Just go
35:42 deeper. I just I I got to tell you, I
35:44 mean, I I feel
35:47 like the end of 2024 for me was was
35:51 pretty burnouty and and
35:53 I I think January and February for me
35:57 were pretty burnouty. March March has
36:00 started to light light lighten the the
36:02 load and and now I'm kind of
36:04 reinvigorated again here in April.
36:07 Launching deep
36:10 agent. I don't know what that
36:14 is. Sorry friends. Can't keep up with
36:16 all the DMs. Okay. Funniest thing.
36:18 Social.
36:20 Okay. Uh a meteorite just hit somewhere.
36:27 This was an interesting Oh, here this is
36:30 a good here's a interesting Sam Alman
36:33 tweet and Gary Marcus response. I think
36:36 this is going to be more like the the
36:37 Renaissance than the industrial
36:39 revolution which I agree with. Right.
36:40 I've I've been saying for I don't know
36:43 year and a half that we're entering the
36:44 great renaissance. I that's absolutely
36:46 what I
36:47 see. Gary Marcus Sam Alultman used to
36:50 promise radical improvements in
36:52 productivity. Now he hints at vibes.
36:54 Why? because companies like Johnson and
36:56 Johnson finally realized that Genai was
36:58 not going to deliver on its promises.
37:00 This is such
37:02 [ __ ] And let me tell you why. Let
37:05 me talk about this one for a second. I I
37:06 have deep feels on this one. One, I I
37:08 just think Gary Marcus is just nasty for
37:11 nasty's sake. It's his brand and and
37:15 what he thinks people expect of him.
37:20 Um the first place people
37:24 go when a new technology comes along is
37:29 efficiency because how computers have
37:32 always behaved is that they make human
37:35 work more
37:37 efficient. The G in GPT stands for
37:41 generative. So yes, you can use AI to
37:44 make what you currently do more
37:45 efficient. But what you can also do is
37:48 you can create entirely new ways of
37:50 working entirely new kinds of
37:54 outputs in the context of work in the
37:56 context of your job. It's it's an
38:00 evolved understanding of what AI makes
38:03 possible. It's not just an efficiency
38:05 machine. It's an efficient an efficiency
38:07 machine
38:09 plus whatever generative [ __ ] you want
38:12 to generate out of it. Um, and that's
38:15 why I think the Renaissance comment is
38:18 right. This is not just the industrial
38:21 revolution. If this felt just like a
38:24 tool that was going to make us more
38:26 efficient, then it's an extension of the
38:28 industrial revolution. The industrial
38:30 revolution quite frankly was about
38:32 augmenting human strength, right? We
38:35 used these machines that were stronger
38:37 than we were to do the
38:40 work that we used to do physically.
38:44 AI is augmenting us
38:49 intellectually. Like we don't lament the
38:51 day where we had to, you know, carry
38:54 heavy things, you know, across the
38:57 river, you know, now we made steamboats
39:00 and we could steam cranes to put the
39:03 heavy [ __ ] on the boat and steam it
39:05 across the river.
39:09 Same thing's happening now with with our
39:13 intellect. And so it's not just about
39:16 efficiency. That's that's it's an
39:18 antiquated way of thinking. Anyway,
39:20 sorry, mini rant. If chat GPT is so
39:23 efficient, why do I lose hours working
39:25 with it? It's both a gift and an ADHD
39:27 nightmare. Totally agree. But but that
39:30 actually what you're talking about is
39:33 what I was just talking about, right? If
39:34 it were just an efficiency machine,
39:36 you'd be a hell of a lot more efficient.
39:39 But you like do this efficiency thing
39:40 and you're like, "Oh, maybe I could
39:42 create a new marketing plan." And then
39:45 you're down the rabbit hole, right? I'm
39:47 still on fire from burnout. And it's
39:50 only April. I
39:53 know. Tweeting hour. Um, okay. 03 is
39:57 insane. I don't even know what to say.
39:59 This is the moment we've been waiting
40:01 for in programming. Okay. really hard to
40:03 comprehend what 04 will feel like.
40:06 [Music]
40:07 Um, it's Ethan Mollik. It seems like a
40:10 big question in AI right now is whether
40:13 there's a flywheel in AI development and
40:15 when it kicks in, either because having
40:19 advanced AIs let you code and train
40:21 better AIs or due to the returns on
40:24 experience. If when it does, all
40:28 non-frontier labs will will fall behind
40:30 forever. That's
40:37 interesting. A thing that I think is
40:39 fascinating about OpenAI right now
40:44 is they launch a new way of doing
40:47 something, right? So they they innovated
40:49 with chat
40:52 GPT and then all the other
40:55 companies quickly replicated what they
40:58 did in chat GPT and then they came out
40:59 with GPT4 and then people took whatever
41:02 it was 6 eight months a year to
41:04 replicate what they did in
41:06 GPT4 and now they come out with a new
41:08 thing now they come out with reasoning
41:10 models and everyone's like you know what
41:12 is what is Orion and what's a reasoning
41:14 model and how are they going to do it
41:15 then they come out with it now everyone
41:17 quickly
41:18 replicates what they're doing. This new
41:20 model, this new reasoning with tools is
41:24 is a new
41:26 thing. Open AI right now to me is
41:30 starting to feel like Apple used to feel
41:32 in the 80s. YouTube question on 03 is it
41:36 really just more for
41:39 programming or a bit better than four
41:41 for everything? I haven't played with 03
41:43 much yet. I don't know the answer to
41:46 that. and what I'm going to share
41:48 tonight. So, this article that I wrote,
41:50 I'm going to share tonight, but I'm I'm
41:51 gonna walk through it. We're gonna
41:53 experiment with
41:54 it. I created a whole series of prompts
41:58 that are about using 03 for non- STEM,
42:02 science, technology, engineering, and
42:04 math.
42:08 I went to Grock and I and I said I I
42:10 asked Grock, "Go find me use cases that
42:13 people are talking about with 03 that
42:15 are non- STEM, that are non-scientific."
42:18 And it basically found like three use
42:20 cases. And then it said, "Not many
42:22 people are using 03 for non-scientific
42:25 use cases." And so I said, "Well, let's
42:28 come up with some use cases to be able
42:30 to go test it." And so that's the that's
42:32 the article I'm going to be sharing
42:33 tonight. and and I'll
42:36 um um and that we'll start playing with.
42:40 But that's the that's the uh that's the
42:43 thing.
42:48 Um fulminating
42:51 Friday. Although AI is moving much
42:53 faster pace than Apple. Yeah, it's
42:55 moving faster. But there's something
42:57 about like Apple used to do used to do
43:04 industrydefining
43:06 innovations where Apple would do it and
43:08 everyone would go, "Wait, what? Oh, holy
43:11 [ __ ] Yeah, a a screen you can touch
43:15 instead of a keyboard on a
43:17 phone. Computer in the pocket. Yeah,
43:20 that's good. Let's go copy that." Right.
43:23 That used to be Apple every two, three
43:25 years, right? And with Open AI, it's
43:28 every six months, right? Six, eight
43:31 months so
43:34 far. Um, yeah, fascinating. All right. I
43:38 use four for marketing purposes. Yeah. I
43:41 I think listen, 40, the the the the
43:46 little default mode of chat GPT is
43:49 incredibly powerful. You don't need you
43:52 don't need these reasoning models. But I
43:55 think the more that we can understand
43:56 what these reasoning models make
43:58 possible, it's going to start to make
44:02 40, god damn these
44:04 names, it's going to start to make 40
44:07 feel increasingly
44:08 antiquated where you go in and you give
44:11 it a prompt and it gives you an answer
44:12 and then you do the reasoning to go,
44:14 okay, here's a follow-on prompt and
44:16 here's a follow-on answer. Oh, okay.
44:19 Right? as opposed to going, "Here's a
44:22 goal." And then have the AI think about
44:25 all the stuff it needs to think about to
44:27 give you a complete answer rather than
44:30 little one-offs. Um, so anyway, all
44:35 right. Someone allegedly stole someone's
44:38 canvas first, although AI is moving much
44:40 faster pace. your article TLDDR oh the
44:43 article
44:44 is 30 I think it's 30 30 different use
44:49 cases for non- STEM here I'll let's let
44:53 me go share the article now so if you go
44:55 to my profile on
44:57 LinkedIn or not LinkedIn uh X the X
45:02 Twitter the one that used to have the
45:03 bird that one and scroll down the you
45:08 you'll see this
45:09 live and then right below below it.
45:12 There's this
45:14 article that I wrote. Oh my god, it's so
45:17 good. Tik Tok, you can actually see
45:19 this. There's a rumor. I heard a rumor
45:21 that Tik Tok's going away Sunday that
45:23 China's pissed, but Brandon said he
45:25 hasn't heard anything about it. But
45:26 anyway, I fixed my monitor settings. You
45:28 can actually see what I'm seeing on my
45:30 screen now. No, I have I have an update
45:32 on that, Kyle. Um, the ban the deadline
45:35 was extended until June 19th. And some
45:39 of those that are from not paying
45:42 attention were conflating it with April
45:44 19th because they just saw the 19th.
45:46 Yeah. The the ones the the ones that I
45:47 saw were not that there was speculation
45:49 that that um you know due the due to the
45:53 tariff stuff they were just going to
45:54 shut it down prematurely. But whatever
45:56 doesn't matter if it's gone. It's gone.
45:58 You all know how to get to the to the
46:00 YouTube channel. We'll keep going there.
46:02 But anyway, after four months I've
46:03 finally fixed it so you can see. So
46:05 anyway, unlock 03's hidden powers. a
46:08 playbook of non- STEM prompts
46:10 uh for for non-scientists. That's
46:12 actually bad writing, isn't
46:14 it? What do you call
46:17 non-scientists? For creatives, but it's
46:19 not just creatives. Whatever doesn't
46:21 matter. Um,
46:23 okay. So, the the way this is there's
46:27 there's three sections in it. One is
46:29 just 03 in the humanities, how you would
46:31 use 03 in the humanities. And then it's
46:33 got like use case comparative thematic
46:36 analysis description example prompt
46:39 right compare the themes of love and
46:41 sacrifice in Shakespeare's Romeo and
46:43 Juliet and Jane Austin's Pride and
46:46 Prejudice highlight how each reflects
46:48 the historical context and provide
46:50 nuanced interpretation of their
46:51 differences. Right?
46:53 So broad use case, you know, example
46:57 prompt and just go play with these. And
47:00 so um stylistic analysis of pros or
47:04 poetry intertextual connections. What
47:07 identify and explain the interextual
47:10 references to Greek mythology in TS
47:13 Elliott's The Wasteland. How do these
47:15 references enhance the poem's meaning? I
47:17 mean, this is kind of [ __ ] you get in,
47:18 you know, some advanced literary class,
47:21 right?
47:22 non-geeeks. Why don't I need to get rid
47:24 of the word non, right? Because I've got
47:27 a playbook for non- STEM
47:30 prompts. A playbook of non-
47:36 STEM prompts
47:44 for Oh,
47:46 yeah. Shut up. Shut up. Shut up.
47:49 producer
47:52 Brandon. Okay, this
48:00 title. Wait, what happened?
48:04 [Music]
48:12 um has
48:15 two
48:17 nons in it. I want to
48:23 remove the second
48:26 one. What do I call
48:31 non
48:33 scientists in this context?
48:38 Make it more
48:41 [Laughter]
48:46 better.
48:50 Non-geeks. Here are a few cleaned up.
48:54 Unlock it. For creative thinkers. Oh, I
48:57 like that. For humanists. That sounds
49:00 too that sounds loaded. For curious
49:03 minds. Scientists can be curious.
49:05 Culture makers. Everyday innovators. Oh,
49:07 I like everyday
49:09 innovators. All right. Everyday
49:11 innovators or creative thinkers? I think
49:13 we're going with creative thinkers. Yes.
49:17 Creative thinkers or everyday
49:18 innovators.
49:20 Votes. Come on. Come on, people. Let's
49:23 put some comments. Pretend like you're
49:25 listening. One. One. Vote for for uh
49:29 creative thinkers. Creative thinkers.
49:32 Everyday innovators. Both. We can't have
49:33 both. Oh, we could have both for
49:36 creative thinkers and everyday
49:38 innovators to explore. But that's a
49:40 little
49:41 long. Everyday
49:45 innovators. One. One. Uh, like we're
49:47 split down the middle. Damn it. I have
49:49 to make a
49:52 choice. Creative thinkers. Uh, critical
49:56 thinkers. Creative thinkers. I think
49:58 it's everyday innovators because the the
50:01 loaded part of creative is well I'm not
50:03 creative but everybody can be an
50:04 everyday innovator. Okay, everyday
50:06 innovator. All right, let's go fix
50:09 it. Um so do me a favor. Here's my
50:12 request to
50:15 you. The article will be temporarily
50:17 unpublished. Yeah, fine. We got to get
50:20 rid of
50:22 non-scientist every day.
50:30 Innovators, a
50:33 playbook
50:36 of non- STEM prompts for everyday
50:40 innovators to explore. Beautiful. Okay,
50:42 here is my
50:45 request. Go push the [ __ ] out of this
50:49 thing. Like it, reshare it, quote tweet
50:51 it, get it out there because I want this
50:54 to help people. Okay, so let's just keep
50:55 going through this for a bit. All right.
50:57 So, um, so literary analysis was one of
51:01 them. Couple of examples in there.
51:03 Creative writing, um, constrained
51:06 narrative generation. Oh, this is kind
51:08 of interesting. One of the things it
51:09 said is in a reasoning model, you should
51:12 be able to say something like, write a
51:14 500word story, and it should get close
51:16 to 500 words. All
51:21 right. Poetry composition. Write a
51:24 Shakespearean sonnet about the passage
51:26 of time. Use imagery of Let's actually
51:29 use this. Let's go do this one because I
51:32 want to see a Shakespearean sonnet is a
51:36 14line a 14line poem
51:40 with what is it? Six and six and two. I
51:43 forget how I forget the structure of
51:49 it. Oops. I didn't actually copy it.
51:52 Copy. Go over here. You can see all
51:55 this,
51:59 right?
52:02 Thinking ensure the rhyme scheme and
52:04 amic pentameter are strictly
52:10 followed here. Look, look what it ah
52:12 beautiful. Look at this. Aab CD CD
52:18 ef GG.
52:22 The first few lines are pushing along
52:24 nicely. Line one seems to follow the
52:26 correct meter. I'm adjusting the
52:28 phrasing for line two while distance
52:31 starts. Let's go look at all this. This
52:33 is really cool. So, oh, here it did
52:37 it.
52:42 Wow. Okay. I've never
52:45 had I've never had an LLM be able to
52:49 successfully write a sonnet that follows
52:53 the actual format. So, let's go. I I
52:55 want to look at what it did. It thought
52:56 for 35 seconds. The sonnet must have 14
53:00 lines following amic pentameter 10
53:03 syllables per line.
53:09 and then use an ab cd
53:13 cd ef g rhyming
53:18 scheme. I'm starting the first line. The
53:21 river clocks the hours with silver flow.
53:24 However, that's 11 syllables. I can
53:27 adjust the river counts the hours in
53:30 silver flow to match the 10 syllable
53:33 limit. This is so [ __ ] cool.
53:38 The first few lines are pushing along
53:45 nicely. I think I'm going to put old
53:48 Sunseo lyrics through it. No. Start up.
53:51 Start over. The Tik Tok cut out. Oh, did
53:55 it? When did it cut out?
53:58 Danielle, am I still Am I live on Tik
54:00 Tok?
54:06 Okay. Okay. That might have just been on
54:09 your end. Just tell me what you missed,
54:12 Danielle. So, so here's where I am. I've
54:15 written an
54:18 article on X. I'm I'm going to put this
54:21 on AI Salon and LinkedIn, but on X right
54:24 now. Unlock 03's hidden superpower, a
54:27 playbook of non- STEM prompts for
54:29 everyday innovators to explore.
54:32 So, it's the it's the other than this
54:34 live, it's the next it's the next post
54:37 down in my thing. So, go find it and
54:40 then it's just it's just like a big
54:42 longass list of things to do. So, where
54:44 we are right now is we're in poetry
54:47 composition and have it
54:50 write a Shakespeare like a properly
54:54 formatted Shakespearean sonnet. And so
54:57 we're just looking at the the reasoning.
55:00 The first few few lines are pushing
55:02 along
55:03 nicely. It's coming together. But line
55:06 nine, but rivers teach that change is
55:10 their
55:11 decree. If you don't understand
55:13 reasoning models like what this is doing
55:16 right now is the perfect example. What
55:18 if you if you tell Forro to write a
55:21 sonnet, it will vomit out a poem.
55:26 And then even
55:28 if it got it wrong and you said, "Hey,
55:30 the last two lines are supposed to rhyme
55:32 and they don't," it'll say, "Oh, let me
55:35 fix that." And it'll vomit out another
55:36 poem. And it may or may not fix the
55:39 bottom things. This is fixing it before
55:42 it actually shows you anything. It's
55:43 it's basically thinking, right? It's
55:47 reasoning before you get
55:51 anything. All right. It looks like I've
55:53 got a re the rhyme scheme mostly figured
55:55 out. Line 10 rhyming with light is
55:59 solid. I like that it says the rhyming
56:01 with light is solid. That means it's
56:03 it's doing qualitative analysis of like
56:07 was this is this a good
56:08 poem. It's been a while friends. It CS
56:12 tutor hope you've been well. Thanks for
56:13 popping in. That's pretty fascinating.
56:16 Couplets 13 and 14 fit with tide and
56:18 glide rhyme. So that's progressing well.
56:20 I've adjusted the other lines for better
56:22 flow, particularly switching some words
56:25 to achieve perfect
56:27 rhymes. The rhyme scheme now follows AB
56:30 A B CD CD E F E F
56:33 G. All right. Upon the river's clock and
56:37 watchful
56:38 stars, the river counts the hours and
56:41 silver flow, while distant stars keep
56:44 watch with ageless gleam. Their winding
56:47 tails repeat the world of flow. Oh, so
56:50 it used flow twice. We don't like
56:53 that yet. Mark our
56:57 mortal.
56:58 Oh, okay. This is
57:03 bad. It used the same word. It didn't
57:07 actually rhyme them. It used the same
57:08 word. So, let's give it feedback and see
57:10 if it can fix it. Tik Tok pin. It's
57:12 talking like you. Oh, that's the other
57:15 thing. Oh, so, so this is an important
57:17 thing to
57:19 note in your
57:21 settings under
57:24 personalization. There are now two
57:27 options. Reference saved memories. So,
57:30 the saved memories are where it says
57:31 updating memory. It's like an explicit
57:33 list of memories. And then this new
57:35 thing is reference chat history. You can
57:38 turn that off.
57:40 By default, now chat GPT knows
57:43 everything about what you've ever talked
57:44 about and can reference it in your
57:46 current chat. It's
57:48 insane. Like it chat GBT starts to
57:52 become much more like a companion that
57:54 gets to know you over
57:56 time. But you for something like this,
57:59 if we want to write this poem, we might
58:01 not want it to write like Kyle. We might
58:03 want it to write more like um
58:05 Shakespeare, right? So I might want to
58:07 turn that off before I do this exercise.
58:09 So, you should know that that's a
58:12 possibility. Well, technically C does
58:14 rhyme with C. Yeah, exactly. So, okay.
58:16 So, we're going to say,
58:22 um, I'm impressed that you got the
58:25 structure correct. However, what you
58:28 should notice is that in your
58:33 AB
58:37 AB rhyme
58:39 schemes, the words not only rhyme but
58:42 are
58:44 identical. Period. That's not very
58:46 creative. Period. That's not
58:49 very poetic. Period. That's certainly
58:53 not Shakespearean. Period. I want you to
58:56 dig deeper, try
59:00 again, and not only
59:07 achieve definitional
59:10 accuracy, but creative
59:18 excellence. I'm going to treat it like a
59:20 freshman in college. Listen, listen. You
59:23 did the assignment, but you're still
59:25 getting a C
59:26 [Laughter]
59:27 [Music]
59:35 minus. It said the word solid. And you
59:38 say that word frequently. Oh,
59:40 interesting. That's
59:45 cool. Wow, it did it really quick. What
59:47 What did it say? Okay, the user wants
59:50 it. Okay, the first line
59:54 Done. I guess because it had something
59:57 to start with. Okay. The
59:58 river upon the river's clock and
1:00:01 watchful stars. Revision. The river
1:00:04 tallies hours in murmured sweep.
1:00:07 Murmured
1:00:09 sweep. Yeah, that makes sense with the
1:00:11 river. While lanterned stars record the
1:00:14 night in ink, their whispered chronicles
1:00:17 descend the deep and constellations
1:00:20 pause yet never blink. And let's so
1:00:24 let's doamic pentameter. The river
1:00:27 tallies hours wait. The river tallies
1:00:32 hours in murmured
1:00:34 sweep while lantern stars record the
1:00:37 night in
1:00:38 ink. Their whispered chronicles descend
1:00:42 the deep and constellations pause yet
1:00:46 never blink. So those are
1:00:49 correct. They sing of moments carved on
1:00:52 passing foam and threads lost wishes
1:00:55 through the vaulted sky. Yet every crest
1:00:58 betray betrays a fleeting a fleeing
1:01:01 home. While heaven's map remains
1:01:04 refusing why. Oh restless clock that
1:01:07 glimmer glimmers in the tide. Teach
1:01:10 mortal eyes to read their shifting
1:01:12 flame. For rivers point the starlet
1:01:14 fates reside and blaze tomorrow's course
1:01:17 in silent claim. So let my fleeting
1:01:21 breath find patient sleep within the
1:01:24 starcross streams. Your echoes
1:01:28 keep. It's pretty [ __ ] good.
1:01:32 I've I've never had a large language
1:01:34 model get get a Shakespearean sonnet
1:01:37 structure right. Um pretty amazing.
1:01:40 Okay, back to back to this document. So
1:01:42 that was poetry composition. Like we're
1:01:44 just getting started like like um so
1:01:47 what I did here's the three sections I
1:01:49 have in this document. Section one is
1:01:51 just humanities like just do like [ __ ]
1:01:54 like this like poetry. Section two is
1:01:58 give me prompts that live in the space
1:02:00 in between humanities prompts and STEM
1:02:03 prompts. Right? Because there's some
1:02:05 things you might want to do in
1:02:06 humanities that that require some
1:02:08 calculation or deeper
1:02:10 analysis. And then the third section
1:02:13 is 03 is a multimodal model that can see
1:02:16 and understand images as as Danielle's
1:02:19 example really nicely illustrated. So,
1:02:22 so there's a whole series of prompts
1:02:24 about the multimodality of it.
1:02:26 Okay. Sections. What are we doing? So
1:02:29 what we're doing, Pate, is I wrote an
1:02:31 article um that I just published on X uh
1:02:36 that is
1:02:37 about it's a series of example prompts,
1:02:40 non-
1:02:41 STEM prompts for
1:02:44 using 03 reasoning model because because
1:02:48 OpenAI is full of engineers, the
1:02:50 examples that they showed were all math
1:02:52 examples, which as a as a guy with an
1:02:55 acting degree is it's meaningless to me.
1:02:57 Like I I'm you know I assume it's good
1:02:59 math. I don't [ __ ] know. Um but I
1:03:03 know that the way these things work
1:03:06 there's probably very deep power in
1:03:07 them. So I wanted to write an article
1:03:09 that gives people a head start for how
1:03:12 to prompt it. Um poetry composition
1:03:15 historical fiction. Write a 300 words
1:03:18 word scene set in 18th century Paris
1:03:20 during the French Revolution. include
1:03:23 two characters debating a monarch, the
1:03:25 monarchy's fall using period appropriate
1:03:28 language and historical details. That's
1:03:30 pretty cool. Historical analysis.
1:03:33 Analysis of historical events. That's
1:03:35 pretty cool. Source evaluation. Evaluate
1:03:38 a letter from a civil civil war
1:03:41 soldier and a modern historian's account
1:03:44 of the same
1:03:45 battle. Assess their reliability,
1:03:48 biases, and how they complement each
1:03:50 other. That's a really good one. You
1:03:52 know what? What would be cool with this
1:03:53 one is go find an image in the
1:03:59 Smithsonian of a famous Civil War letter
1:04:02 that's had a lot written about it and
1:04:05 just give
1:04:06 03 the image of the letter and say,
1:04:09 "Read this letter, identify
1:04:12 it,
1:04:14 find anything that's been written about
1:04:16 this letter or this story, and then
1:04:19 compare those things." Paid on TikTok.
1:04:21 Um, I'm on the flip side of
1:04:27 writing. I'm on the flip side of
1:04:30 writing. Wait, I'm on the flip side.
1:04:33 Writing part is useless to me. Yeah,
1:04:35 exactly. I don't give a [ __ ] if I can
1:04:37 write like Shakespeare. Yeah, exactly.
1:04:39 Curious Kyle and Ma
1:04:41 Mathematic. So, Pate, that that's
1:04:43 exactly my point. It's like this
1:04:47 is the the the significance of the
1:04:52 multi-step reasoning with tool
1:04:55 use instinctively feels to me like it
1:04:58 goes way beyond doing math stuff, right?
1:05:01 And so that's I I think mathematicians
1:05:05 and scientists and researchers are going
1:05:07 to look at what it does and it's much
1:05:09 more patently obvious why you would use
1:05:11 a model like 03 in those things. It's
1:05:15 not as obvious on the humanity side. So
1:05:16 that's exactly what my article is about
1:05:18 is go play with these things and let's
1:05:20 learn what it's good at um in the non-
1:05:23 STEM stuff. Count counterfactual
1:05:25 scenarios. What might happen if if an
1:05:28 American what oh what might oh it's like
1:05:31 what if scenarios what might h what good
1:05:34 lord what might have happened if the
1:05:36 American revolution had failed provide
1:05:39 plausible scenario for the next 50 years
1:05:42 grounded in 18th century political and
1:05:44 social conditions that'll be a good one
1:05:46 I and I honest to god like reading
1:05:49 through the um the reasoning it did when
1:05:52 it just wrote that sonnet doing more
1:05:54 meaty things like this could be really
1:05:56 interesting
1:05:56 Okay. Philosophical inquiry, ethical
1:05:59 dilemma analysis
1:06:03 um philosophical text interpret
1:06:06 interpret interpretation summarize
1:06:09 Nichi's concept of will to power in thus
1:06:13 spoke
1:06:14 uh
1:06:15 zeraustra critique its implications for
1:06:18 modern individualism.
1:06:20 Good lord. Construct arguments for and
1:06:23 against the existence of free
1:06:25 will. All right. So, and then
1:06:27 educational content creation, lesson
1:06:29 plan development, interactive study
1:06:32 guides, historical timeline creation.
1:06:35 Oh, that one's kind of cool. Let's do
1:06:36 one of these because then we can do an
1:06:39 image. Okay. Of this. Okay. So, what
1:06:42 we're going to do, create a timeline of
1:06:44 the Renaissance, 1400 to 1600,
1:06:47 highlighting 10 key events in art,
1:06:49 literature, and politics. Annotate each
1:06:51 with a brief explanation of its cultural
1:06:53 impact. So, we're going to go do that,
1:06:55 and then we're going to have
1:06:58 03 make a visual of the timeline because
1:07:02 it can do
1:07:05 images. So, we'll just do that prompt as
1:07:07 is. Not perfect initially, Valerie Cox,
1:07:11 but so amazing in the end. Just
1:07:13 wow. I assume you're talking about the
1:07:16 poem there, or you're talking about
1:07:17 something that you did. That's pretty
1:07:19 cool, Kyle the playright. Exactly.
1:07:26 Exactly. Love that Civil War prompt.
1:07:29 Yeah, that was pretty
1:07:31 cool. All right, what are we doing here?
1:07:33 Oh. Oh, this is our timeline. Okay, it
1:07:35 did this really fast, right? Okay, these
1:07:37 10 moments I'm gonna say Okay, great.
1:07:40 Um, oh [ __ ] my mouse my mouse stopped
1:07:43 working. All right, hang on. I think my
1:07:46 mouse
1:07:47 is it's charge is dying. Hold
1:07:50 please. Charging said mouse. Okay. And
1:07:54 now I'm using a trackpad. Okay. So now
1:07:56 I'm going to say um I want you to
1:08:02 create an illustrated
1:08:07 timeline that
1:08:10 has
1:08:12 period
1:08:14 accurate
1:08:18 imagery,
1:08:20 appropriate
1:08:22 labels, and a clear timeline feel.
1:08:28 This should look so
1:08:32 good that it could be
1:08:37 displayed in the
1:08:41 Smithsonian Museum.
1:08:54 it. We'll just let it we'll just let it
1:08:56 be there.
1:08:57 Okay. So, let's see how it does here.
1:09:00 And then I'm going to try one other
1:09:02 thing. I I I read on X earlier that
1:09:06 using
1:09:08 Python, it can actually upscale an
1:09:11 image, creating an illustrated timeline.
1:09:14 The user wants a period accurac accurate
1:09:16 timeline. Okay, so it thought. So what
1:09:18 did it
1:09:19 think? We might need to either combine
1:09:22 several images into one. The user wants
1:09:24 Okay. However, since I can't directly
1:09:27 marry generated images into one large
1:09:30 graphic, I could either suggest them one
1:09:32 at a time. This this would help. So
1:09:34 let's see. Let's see what it does.
1:09:50 Come on, chat GPT, don't
1:09:53 suck. But anyway, apparently you can
1:09:55 upscale images now in 03 because it uses
1:10:00 a Python library for upscaling
1:10:05 images. I don't know if that's true or
1:10:07 not, but we can find out.
1:10:10 [Music]
1:10:12 My mouse retired recently,
1:10:25 too. Friday night, date
1:10:28 night, playing with
1:10:30 03. So, are you guys going through the
1:10:32 article? And again, do me a favor. How
1:10:34 many folks are in here? We got pile of
1:10:37 people here, pile of people there.
1:10:39 Go to Okay, it's not
1:10:43 bad. It did it. It said it did it 4K
1:10:46 wide, but it cut it
1:10:52 off.
1:10:54 Greek. Uh, I'm going to
1:10:58 say, let's see. I like the idea though
1:11:04 the image is cut off.
1:11:08 on right and
1:11:11 left. So
1:11:15 either make it two
1:11:19 rows
1:11:22 or allow
1:11:35 um things to be
1:11:39 smaller to accommodate
1:11:46 the space
1:11:49 requirements. Also, the color palette
1:11:56 um
1:11:57 feels old and tired.
1:12:02 make it more
1:12:07 modern with
1:12:19 um
1:12:20 contrasting
1:12:24 background versus the text and
1:12:29 illustrations. Let's see what it does
1:12:30 with that.
1:12:35 Ask it to create padding around the
1:12:37 border. Shrink to fit. Yeah, those would
1:12:39 both be
1:12:40 good. Those would both be solid. I
1:12:42 should have looked at that before. Seems
1:12:44 plausible, but may just be a quirk.
1:12:45 Would probably be better off doing a
1:12:47 dedicated upscaler. Yeah, definitely. I
1:12:49 just want to see if it I just want to
1:12:51 see if
1:12:52 it is functionally there. I would think.
1:12:56 Yeah, something like what's that one
1:12:57 called? Magnafi seems to be one that's a
1:13:00 dedicated one that looks like it does
1:13:01 really good results. That guy's always
1:13:03 posting examples up there. Um, while
1:13:07 that's doing that, if you wait, can I
1:13:10 pop out of here and pop back in? Yeah, I
1:13:11 guess I can. If you haven't seen the
1:13:15 typo and timeline Oh, yeah. I'm sure
1:13:17 there's going to be
1:13:18 typos. Yeah. Yeah.
1:13:22 Yeah. Actually, what would be
1:13:25 interesting is we'll let it generate
1:13:27 this one. If it's got any major typos in
1:13:29 it, it might be interesting to tell it
1:13:32 to write write a creative brief for the
1:13:35 timeline first with exactly what text
1:13:38 needs to be on it. It'd be interesting
1:13:40 to see if it got if it wrote all the
1:13:42 text out first as a creative brief if it
1:13:45 got more accurate in its output.
1:13:56 But I mean, do you
1:13:58 know, you know, when we were back in
1:14:00 college and you had to go like like, you
1:14:04 know, read all the books and get all the
1:14:06 the crap from the Renaissance, the major
1:14:09 things that happened, and you had to do
1:14:10 this [ __ ]
1:14:15 yourself. All right. It still [ __ ] it
1:14:17 up. Renaissance timeline, but it's not
1:14:19 bad. I kind of like the the color thing
1:14:22 it
1:14:23 did. Uh, let's see. Add
1:14:29 padding to the
1:14:32 left. It is still cut
1:14:42 off. All
1:14:48 right, we'll come back to
1:14:50 this. Um, go back over here. Historical
1:14:54 timelines. Uh, and then it talks about
1:14:57 that. Okay, so here are the hybrid STEM
1:14:59 and humanities use cases. These are kind
1:15:02 of cool. So this talked about um I'll
1:15:04 propose a new list of use cases that
1:15:06 push beyond surface level applications
1:15:09 exploring this boundary layer and
1:15:11 venturing into cultural philosophical
1:15:14 even speculative territories. These
1:15:17 ideas aim to exploit 03's reported
1:15:20 strengths um simulated reasoning private
1:15:23 chain of thought and adaptability. And
1:15:26 then it gives some benchmark [ __ ] while
1:15:28 embracing humanity's focus on meaning,
1:15:31 context, and human values. So these
1:15:34 categorized use cases. Um, oh, I didn't
1:15:37 I meant to edit this
1:15:39 [ __ ] Well, whatever. Um, cultural
1:15:42 synthesis and reimagination. So,
1:15:44 cross-cultural narrative fusion. Th this
1:15:47 was this was my favorite my favorite
1:15:50 prompt so far that it wrote. Write a
1:15:52 600word narrative about a refugees
1:15:55 journey, blending the rhythmic cultural
1:15:58 story te storytelling style of West
1:16:01 African giot traditions with the
1:16:04 fragmented introspective style of
1:16:06 Japanese
1:16:09 zoosu. Ensure each cultural culture's
1:16:12 aesthetic is distinct yet harmonized and
1:16:15 use the theme of displacement to bridge
1:16:17 them. Include annotations explaining
1:16:19 your stylistic choices.
1:16:22 Um, DAI tomorrow is still on. Oh,
1:16:26 demystifying AI is still on. Zero things
1:16:28 to present though. So, come with
1:16:31 questions. Okay. So, tomorrow, if you're
1:16:34 a member of the AI salon, Pate's doing a
1:16:38 um demystifying AI session. If you
1:16:41 haven't been to one of his, you should
1:16:43 go. So, if you go to
1:16:46 events, should be here right at the top.
1:16:49 Demystifying AI. So, tomorrow at 2 PM.
1:16:52 So, go RSVP for that event. If you've
1:16:55 not been to one of Pat Oh, and how you
1:16:57 get to the AI salon, if you're not a
1:16:58 part of it, go to the
1:17:00 salon.ai, click on join our community.
1:17:03 That'll take you over here to the um to
1:17:05 the mighty networks community. Register,
1:17:08 introduce yourself. That's step two. And
1:17:11 then just go to events and RSVP for
1:17:13 Pate's uh AI demystifying AI session.
1:17:17 Um, if you want to understand how AI
1:17:20 works and like the stuff we're playing
1:17:21 with tonight, actually, what's going on
1:17:23 under the hood, that's what Pate's
1:17:26 really good at. Good. Awesome. Thank you
1:17:27 for sharing that.
1:17:30 Um, okay. Let's go back here. Did it do
1:17:33 it? It did it. Yay. You did it.
1:17:39 Yay. Look, Tik Tok can actually see this
1:17:41 [ __ ] now. It's a little miracle.
1:17:45 Um, still a timeline. Oh, Tim Timoline,
1:17:50 right? Let's see.
1:17:54 Florence Baptistry is there
1:18:00 twice. Yeah. So, that's pretty crappy,
1:18:02 but that's okay. Let's just That's
1:18:04 That's fine. Download revised graphic.
1:18:07 Let me know. Um, please
1:18:11 upscale this image. Let's see what it
1:18:14 does
1:18:17 here. Yeah, I 1590 to 1991. Is that what
1:18:21 it
1:18:23 said? Yeah, the last one. All right. So,
1:18:28 it's not accurate, but you know,
1:18:32 hey. Okay. Wait. Thought for eight
1:18:36 seconds. Oh, it says the original image
1:18:39 was created by Deli, which it wasn't.
1:18:47 Since I don't have a direct upscaling
1:18:51 tool. Huh, that's
1:18:54 weird. It should have invoked a tool. It
1:18:56 should have generated a tool to do this.
1:19:03 I'm going to stop it and
1:19:05 say, "Can't you use
1:19:11 a a Python tool to
1:19:16 upscale the
1:19:18 image? Use your tools, you
1:19:23 tool." Does
1:19:26 harshweed the first make an appearance?
1:19:30 Ah, look. PIL
1:19:35 image
1:19:36 open. New size. So, it's doing a 2x
1:19:40 upscale. Resize using the high quality
1:19:43 length zas filter. I don't know if
1:19:46 that's high quality or
1:19:49 not. Download the upscaled
1:19:53 version. Starting
1:19:55 download. So, it downloaded something.
1:20:02 All right, there's that. So, let's see
1:20:04 if this is actually a different size
1:20:05 than the other
1:20:08 one. Download
1:20:15 this. Same
1:20:18 size. It lied to us. It
1:20:22 lied. You hallucinator.
1:20:27 All right. So, that was a fun one. All
1:20:29 right. Back to our back to our little
1:20:31 thing here.
1:20:32 Um, speculative cultural
1:20:36 evolution. Imagine the evolution of DLL
1:20:40 Moros celebrations from 2025 to
1:20:44 2125 influenced by climate induced
1:20:46 migration and augmented reality
1:20:49 technologies. That's cool. philosophical
1:20:52 and ethical simulation
1:20:58 um
1:21:00 interdisiplinary historical
1:21:01 reconstruction, right? So these hybrid
1:21:03 ones are just much more sophisticated,
1:21:06 you know, taking multiple inputs and and
1:21:08 smashing them together, which is kind of
1:21:10 cool.
1:21:12 Speculative anthropology modeling
1:21:15 posthuman belief systems. Design a
1:21:18 700word belief system for a 2100 society
1:21:21 where humans and AI are cognitively
1:21:25 integrated. Base it on trends in
1:21:27 transhumanism, Buddhist ethics, and
1:21:30 neural interface technology. Describe
1:21:33 the core tenants, rituals, social
1:21:36 implications, and explain how 03
1:21:39 synthesized these influences. Cool. All
1:21:42 right. Alien cultural
1:21:45 contact, simulating it. Okay. So, that's
1:21:49 those. And then the third section is
1:21:52 multimodal capabilities. So, visual
1:21:55 cultural analysis. Analyze the the
1:21:58 iconography and car K caravio's the
1:22:02 calling of St. Matthew. Identify at
1:22:04 least five key symbols and explaining
1:22:07 their religious and cultural
1:22:08 significance in Baroque,
1:22:12 Italy. Very
1:22:15 cool. Digital art and cultural
1:22:17 reimagination. Create a 600 by 600 pixel
1:22:20 digital artwork inspired by Aboriginal
1:22:22 Australian dot painting. Adapt the
1:22:25 technique its techniques and
1:22:27 storytelling to address climate change.
1:22:30 Ensure the artwork respects cultural
1:22:33 principles of Aboriginal art. Provide a
1:22:36 300word reasoning chain explaining how
1:22:39 you incorporated traditional elements
1:22:42 and modernized the narrative. This is
1:22:44 fascinating. Let's see what this does.
1:22:57 I'm at the
1:22:58 poem point in the weekend. A sign
1:23:01 currently create an
1:23:05 illustration blackout poem. I don't know
1:23:08 what that is, but that sounds cool.
1:23:10 Don't forget to give it a thumbs up. Oh,
1:23:13 good. Vicki G Riot pronounced Griot.
1:23:16 Okay, good. I didn't know if it was
1:23:18 Yeah. Griot
1:23:20 or Yeah. Or G Riot. Okay. Griot. Good.
1:23:23 Thank
1:23:25 you. This is not doing AI upscaling, but
1:23:28 it's something. Yeah. Yeah. Exactly. You
1:23:31 know what's weird, Pate? 03 seems to be
1:23:34 very very um it's very very fickle.
1:23:38 Sometimes it seems to be brilliant and
1:23:39 sometimes it seems to be totally
1:23:41 incompetent. Lord Digital God's given me
1:23:43 a two $2 sticker, super sticker on
1:23:46 YouTube. Thank you so much, sir. very
1:23:48 generous. I appreciate
1:23:51 that. Oh, lordy lordy lordy
1:23:56 lordy. Thinking is it what's is it just
1:23:59 thinking or is it
1:24:01 making? I think we might have stumped it
1:24:03 with this
1:24:06 one. That's the other thing I noticed is
1:24:08 that it seems to be a little their
1:24:10 servers seem to be being hammered.
1:24:15 It's a classical upscaling algorithm.
1:24:18 So, not bad, just not bleeding edge.
1:24:20 Yeah, I know. I I figured that was the
1:24:22 case, Pate, because it didn't seem like
1:24:24 it was any sort of AI sort of
1:24:26 thing. Get node by ID or message ID. No
1:24:30 node found on the ID. There's some
1:24:33 There's some [ __ ] code that they
1:24:35 didn't get behind the uh
1:24:38 popup popup menu.
1:24:43 We'll put weekend assignments on
1:24:44 irregulars. Oh, good. So, you can see
1:24:46 blackout film. Nice. The AI told me to
1:24:49 make these. Cool. Very nice. Considering
1:24:52 cultural sensitivity, the rules for
1:24:55 Aboriginal art probably haven't changed
1:24:57 a lot. Searching the web. Getting
1:25:00 started. That's kind of cool. All right.
1:25:02 So, wait, what did it
1:25:03 do? I'll quickly research
1:25:08 guidelines if I'm uncertain for thing.
1:25:11 If I'm uncertain or things are dynamic,
1:25:13 I'll browse for the latest info. In
1:25:16 particular, I can search for the
1:25:17 respectful use guidelines for Aboriginal
1:25:21 art like dot painting. That's actually
1:25:23 kind of
1:25:24 cool. I'll make sure to include anything
1:25:27 relevant and site cultural principles to
1:25:30 stay accurate and respectful. Let's do
1:25:33 the
1:25:34 search. All right.
1:25:38 Keeping in mind these avoiding sacred
1:25:41 symbols. Oh, that's cool. This is aware
1:25:44 of what symbols it can include and what
1:25:46 it can't. I'll
1:25:48 incorporate elements symbolizing water,
1:25:50 rising temperatures, and footprints.
1:25:52 I'll also include a thoughtful touch
1:25:55 acknowledging a consultation with a
1:25:59 we're a jury
1:26:01 artist. Along with the image, I'll
1:26:03 provide a 300word explanation of the
1:26:05 artistic choices. Make sure to site
1:26:07 cultural guidelines for accuracy. Let's
1:26:10 get started. Oh god, that's
1:26:16 gorgeous. I got to
1:26:20 sneeze. Okay, that was the weirdest
1:26:22 sneeze I ever had. I just had a
1:26:24 non-sneeze. I had the in inhale part and
1:26:27 then it just
1:26:29 disappeared. That was
1:26:31 weird. Yikes.
1:26:34 Blackout poem is a chunk of text and
1:26:36 words that are blacked out. So remaining
1:26:39 words make the poem
1:26:44 something. Please zoom
1:26:53 in. That's as far as I can zoom in
1:26:58 here. Love the river.
1:27:03 Um, let's see what it
1:27:05 says. To honor Aboriginal Australian dot
1:27:08 painting while addressing climate
1:27:09 change, I began by identifying
1:27:11 non-sacred visual conventions that are
1:27:14 widely shared across nations, concentric
1:27:17 circles for water holes or communities
1:27:20 and meandering dotted lines for song
1:27:22 lines. Avoid any clansspecific or
1:27:25 ceremonial symbols. Follows the First
1:27:27 Nations cultural and intellectual
1:27:29 property protocols that stress
1:27:32 permission for respect of restricted
1:27:33 imagery. Pallet. I chose deep ochres,
1:27:37 siennas, and
1:27:40 umbers. Cooler teals and indigos flow
1:27:42 diagonally to symbolize rising seas and
1:27:45 disrupted rain
1:27:49 cycles. Um, a large central
1:27:55 Sun circle anchors the piece. Is that
1:27:58 right?
1:28:03 Yes. Simulated cultural consultation.
1:28:06 Before finalizing motifs, I drafted a
1:28:10 brief and hypothetically hypothetically
1:28:14 shared it with a we're a jury knowledge
1:28:18 holder and a Gatagle environmental
1:28:21 artist. feedback asked, "Avoid mimicking
1:28:24 toemic animal tracks. Emphasize shared
1:28:27 stewardship over didactic warning and
1:28:30 foreground water as connector, not
1:28:32 merely a
1:28:34 threat." Those insights led me to weave
1:28:36 meandering aqua pathways through every
1:28:39 quadrant so country sky and people
1:28:43 remain balanced in balanced dialogue. I
1:28:46 mean, I don't know [ __ ] about any of the
1:28:51 ethics of this, but it would be
1:28:53 interesting to see if this if it
1:28:56 actually achieved what it said it did.
1:29:00 Um, there are dot paintings in most of
1:29:03 the buildings here, especially
1:29:06 government buildings. Wow.
1:29:08 Cool. So thoughtful. Wow. Yeah, it's
1:29:10 pretty crazy, right? This is beautiful.
1:29:13 Teaser
1:29:16 sneezer. Yikes. Was to the Oh, was to
1:29:19 the node ID error message. Yeah, I know.
1:29:22 I know, right, P? It was like It was
1:29:24 like, uh, guys, you might want to make
1:29:26 sure that your errors are not, you know,
1:29:29 out there for the public to
1:29:33 see. You You did launch this thing,
1:29:36 right? Like this is available to
1:29:40 everyone. Um, okay. So, we'll go back to
1:29:42 this thing. So anyway, so so we don't
1:29:44 have to go through a whole document, but
1:29:46 might as well. Um, speculative cultural
1:29:48 artifact creation. Imagine a visual
1:29:51 artifact, a ceremonial object. Oh, this
1:29:53 is cool. Imagine a visual artifact. A
1:29:56 ceremonial
1:29:58 object, a public mural for a 2075
1:30:02 society shaped by AI human symbiosis and
1:30:06 ecological restoration.
1:30:10 generate a an image of the
1:30:13 artifact and manipulate the photo of a
1:30:16 2025 urban street to incorporate it in
1:30:20 400 words. Explain its Okay, let's let's
1:30:23 go do this one, too. I'm going to change
1:30:24 some of the stuff in it. This is
1:30:29 cool. And again, I'm doing 03 here. The
1:30:32 other thing to do is switch between 03
1:30:35 and 04 mini high. Uh, like if if you run
1:30:39 out of 03 things, do that. But also
1:30:41 maybe just try the same prompt in
1:30:43 multiple models and see what it
1:30:48 does. I ran this ran the same prompt and
1:30:51 got a very similar picture, but it is
1:30:53 different.
1:30:55 Fascinating. Cool firsthand info. Hey
1:30:58 Steo, was there anything in either the
1:31:01 image or that explanation that you felt
1:31:03 was was like out of bounds or off base
1:31:06 in terms of like ethical considerations.
1:31:48 [Laughter]
1:31:51 Damn. What am I doing? Tibetan throat
1:31:54 singing over here.
1:32:04 I do that too with my
1:32:07 voice. It's just ADD self
1:32:10 soothing. No, it seemed
1:32:12 fine. Uh, this is not what I wanted at
1:32:15 all.
1:32:17 Um, this is supposed to be a
1:32:23 photo of a
1:32:25 mural on a
1:32:28 building in the
1:32:32 future. Take this same image and
1:32:36 re-imagine
1:32:37 it in that
1:32:44 context. What's it doing now? It's
1:32:46 writing the Oh, it's Oh. Oh, this is
1:32:49 going to be interesting to see what it
1:32:50 does. Thought for 30
1:32:56 seconds. I need to find a street photo.
1:32:58 It didn't find it. Okay. I found an
1:33:01 image that could work as the street
1:33:03 scene even though it's not obviously
1:33:05 from 2025. It has the description. 2025
1:33:08 should be acceptable. Now I need to
1:33:10 generate the artifact
1:33:13 image. After that I'll figure out how to
1:33:16 combine both the street photo and the
1:33:19 artifact image. Although passing the
1:33:21 correct image references will be
1:33:24 important for editing. Time to proceed.
1:33:26 All right. So there's the image it
1:33:27 created. Kind of
1:33:31 cool. Image
1:33:33 C. So these are basically the same
1:33:36 thing.
1:33:49 If it just repeats this again, then I'm
1:33:51 going to stop it. I'm going
1:33:53 to give it an explicit
1:33:57 prompt. But again, you know, the
1:33:59 difference thing here is it's generating
1:34:01 these images in this chain of
1:34:04 thought thing, right? It's not just I
1:34:07 didn't tell it to create multiple
1:34:08 images. I told it to create an image and
1:34:13 combine it with a photograph. And so
1:34:15 it's it's doing those steps. Yeah, it's
1:34:18 just it's just repeating itself. Okay.
1:34:21 So, we're going to stop this and then
1:34:23 we're going to
1:34:23 go here's an explicit prompt. Get your
1:34:26 [ __ ] together.
1:34:33 [Music]
1:34:40 mural future reimagination creating a
1:34:42 photorealistic scene. I'm considering
1:34:45 that we might need a better
1:34:58 query. I'm going to stop it because it's
1:35:01 saying 2025.
1:35:03 The
1:35:04 photo
1:35:08 should be from 20 from the
1:35:13 future. Ignore the 2025 photo
1:35:20 um element. Just make a
1:35:25 street
1:35:27 photo
1:35:29 with the image you
1:35:32 created on a massive mural on a
1:35:37 building in the
1:35:40 photo in the future. Damn it.
1:36:00 Wee. Oh, did it do 2025 again? I think
1:36:03 it did.
1:36:04 [ __ ] Oh, the user has shifted the
1:36:07 focus. Okay. Since we can't manipulate
1:36:09 an existing photo, I need to generate a
1:36:11 completely new futuristic street scene
1:36:13 with a giant mural using text to image.
1:36:16 Yes, correct. I'll also use the other
1:36:18 artifact image I created earlier. Good.
1:36:22 So, it's remembering. This is one of the
1:36:24 big things about
1:36:26 03 when it's doing this stepbystep
1:36:29 thinking, it's maintaining the
1:36:31 context across all these different
1:36:34 steps. So, the fact that it just
1:36:36 referenced that it made the images
1:36:38 earlier, that's something
1:36:40 that straight chat GPT can't do or it's
1:36:44 hard to get it to do that. You're like,
1:36:46 "Use the image you made before." It's
1:36:47 like, "Okay." And then it makes a new
1:36:49 image,
1:37:07 right? Hey, it did it. [ __ ]
1:37:11 yes. It [ __ ] did it. it up, people.
1:37:16 But wait, we're going to tell it.
1:37:19 Okay.
1:37:21 Um, okay. You gave it the
1:37:27 gippers
1:37:28 try, but let's
1:37:34 amp it up. Make it 16 by9 wide.
1:37:41 Cinematic and
1:37:43 photographic in
1:37:46 style. In
1:37:48 style. It should look very very very
1:37:54 realistic.
1:37:58 realistic with
1:38:03 dramatic lighting and
1:38:08 um
1:38:12 atmospheric
1:38:16 haze. What the hell did it write after?
1:38:18 Oh, it just wrote all this other [ __ ]
1:38:23 The circular society not emblem is the
1:38:25 most ubiquitous icon of
1:38:28 20755's ecote tech
1:38:31 [Laughter]
1:38:34 commonwealth. It's fantastic.
1:38:37 Kyle Shannon the human digy
1:38:47 do future
1:38:52 Starbucks. Oh lordy lordy lordy lordy
1:38:57 lordy. I tell you lordy lordy lordy. I
1:39:01 probably shouldn't say that. It's good
1:39:03 Friday, right?
1:39:15 Maybe the haze is gone in the future.
1:39:17 Yeah, it probably should be, but there
1:39:18 could still be there could still be fog
1:39:20 rolling in off the water.
1:39:22 So, just what the atmospheric haze
1:39:25 should do is just make it so the stuff
1:39:27 in the foreground is
1:39:29 more, you know, more prevalent. But
1:39:31 we'll see. We'll see how it does.
1:39:36 I actually am not grooving on they took
1:39:39 away. So one of the things that when
1:39:40 they came out with the new image gen
1:39:42 model, it would start out with this
1:39:44 interface and then it would go to this
1:39:45 blurry image and it would sort of draw
1:39:47 the image from the top down because it's
1:39:50 doing it as as a as a transformer. Oh
1:39:54 yeah, it did it. Wow, this is
1:39:57 great. There you
1:39:59 go. I wouldn't exactly call it
1:40:01 photorealistic. It still looks a little
1:40:03 illustrative, but it's got atmospheric
1:40:07 haze. It's
1:40:09 cinematic. It's got a drone in it.
1:40:12 Everyone's walking in the same
1:40:16 direction.
1:40:18 Huh. All
1:40:21 right. [ __ ]
1:40:25 works. And back here. What else? Visual
1:40:28 historical
1:40:29 reconstruction. Oh, reconstruct a lost
1:40:32 fresco. That's pretty
1:40:37 cool. Reconstruct the lost fresco from a
1:40:40 fifth century Mesop Messoamerican
1:40:44 temple using textual descriptions from
1:40:47 the cotices, architectural data on
1:40:50 pigment use, and images for the related
1:40:54 Mayan art generate a picture of the
1:40:57 fresco.
1:41:00 So what they say Mesoamerican
1:41:12 temple. Oh,
1:41:16 fresco. Excuse
1:41:20 me. So we need a wrecked one.
1:41:38 All right. Copy this
1:41:47 image. All right. So, we're putting in
1:41:49 an image of a Mesoamerican fresco.
1:42:05 So we'll say from
1:42:08 this
1:42:10 messo
1:42:13 American
1:42:18 example let's
1:42:20 see identify
1:42:23 and
1:42:25 reconstruct this loss fresco or
1:42:30 this this
1:42:32 uh
1:42:35 dilapidated
1:42:39 fresco. I don't know. Let's just see
1:42:41 what it
1:42:48 does. Thinking
1:42:59 Flipboard edu podcast. Lol. [ __ ]
1:43:06 works. Yeah, this is pretty good. This
1:43:08 is this is this is really something. All
1:43:11 right. Considering Mayan art, let's
1:43:12 let's look at how it's doing its
1:43:14 thinking here. Okay. So, the image
1:43:16 mentions a feathered Oh, look at this.
1:43:20 The image mentions a feathered serpent
1:43:23 and flowering trees which could relate
1:43:27 to te
1:43:28 to oh tia tat haneo. How do they say
1:43:32 that?
1:43:34 Tiat t to tak hakan murals. Um but the
1:43:41 user asked specifically about Mayan art.
1:43:43 There's a possibility this could be from
1:43:46 Kak Ta with its mix of Mayan
1:43:54 influence, but it's not purely Mayan
1:43:57 style. Let me do some more searching.
1:43:59 Analyzed image. Hm. The swirling designs
1:44:02 remind me of Mayan glyphs, possibly from
1:44:04 a mural. I'm thinking this could be
1:44:06 temple of murals from Bonamp Pac.
1:44:11 The image looks narrow on the right
1:44:14 side. Fiery
1:44:25 serpent. So, this is interesting. I
1:44:28 picked an
1:44:31 image. I picked an image that was not
1:44:34 Mayan and it and it busted me. And so
1:44:38 now it's it's saying, "Well, this
1:44:41 doesn't match Mayan, but he said Mayan,
1:44:44 so he must know what he wants. But let
1:44:46 me go see if I can find this other
1:44:51 shit." In if if you've ever wondered, is
1:44:55 the AI smarter than you are? [ __ ] yes,
1:44:57 it
1:45:02 is. Okay, let's see. It thought for two
1:45:06 minutes and 13 seconds. That's cool. So
1:45:09 where did it end up? We need to ensure
1:45:11 cultural sensitivity while sharing our
1:45:13 explanation. So here's the plan. I re
1:45:16 researched code cotices like Dresden and
1:45:19 Madrid to study water scroll motifs
1:45:23 using pigment analysis. Cineabar for
1:45:25 red, Maya for
1:45:26 blue, Maya blue for blue and yellow for
1:45:29 geoite.
1:45:32 G. Oh,
1:45:34 goite. I made sure the design mirrors
1:45:36 symmetrical compositions of
1:45:39 kakla murals. Now I'll put together a
1:45:42 prompt for
1:45:44 reconstructing feather serpent mural.
1:45:46 Keeping it naturalistic. The scene will
1:45:51 depict
1:45:53 Ket
1:45:56 Kalddle or whatever water scroll motifs.
1:46:00 The bird warrior at top the serpent's
1:46:01 bat, the sea creatures
1:46:04 below.
1:46:07 Huh? All right, I'm learning [ __ ] while
1:46:09 we're making
1:46:10 pitches. This is [ __ ]
1:46:21 crazy. Make a thing that doesn't suck.
1:46:23 Please make a thing that doesn't suck.
1:46:25 Oh, that's gorgeous. All right, so let's
1:46:27 go look. Let's download
1:46:30 this and then let's
1:46:33 go. Click on this. Copy it. Save image
1:46:37 as
1:46:40 um fragment. We'll call
1:46:43 it to downloads. It's
1:46:47 fine. And then we'll go bang
1:46:54 bang. And then we'll
1:47:06 go. All right, that's pretty
1:47:15 cool. Can be used in midjourney as an
1:47:19 SREF. Probably. I want to animate that.
1:47:22 All right, I'll post this to uh I'll go
1:47:24 post this to
1:47:26 the You know what I'll do? I'll post Oh,
1:47:29 you know what we should also do is So,
1:47:30 we not only have the
1:47:35 um the image, but we have the the
1:47:37 writing of
1:47:41 it. Why this fragment is from this
1:47:46 place. Oh, this is so cool.
1:47:53 All right. So, let me grab
1:48:06 this. I'm going to go to the
1:48:10 salon and I'm going to go to the AI
1:48:12 learning lab. Boom. Pop that bad boy in
1:48:16 there.
1:48:41 Okay, there that
1:48:43 is. And then we'll post
1:48:45 that. Don't notify everyone.
1:48:50 And then I'll say
1:48:53 um here was the uploaded
1:48:58 image
1:49:00 and copy pasted prompt.
1:49:06 um
1:49:09 that
1:49:12 inaccurately defined or uh I
1:49:16 inaccurately uh said this was Mayan
1:49:22 art. It
1:49:26 wasn't. and
1:49:29 03
1:49:31 corrected the
1:49:34 mistake before the
1:49:40 reconstruction. Here is the
1:49:43 prompt I used
1:49:46 from and then I'm going to put the link
1:49:48 to the
1:49:49 article. Look how good I'm doing here.
1:49:52 You kidding me?
1:49:54 How do I get
1:49:57 uh
1:50:00 link pin your article, edit
1:50:04 article, share, copy link to article
1:50:08 back here.
1:50:15 Boom. Here's the prompt I used. And then
1:50:18 my prompt
1:50:21 was
1:50:25 Uh
1:50:33 here. And then we're going to get rid of
1:50:36 that. And then we're going to add
1:50:38 in the fragment that I
1:50:43 uploaded. Comment. All right. So there
1:50:46 you have it. You've
1:50:47 got there's what it generated. And then
1:50:50 in the comment is the image I uploaded
1:50:52 and the prompt I used. Also posting your
1:50:55 article prompts there for the
1:50:59 weekend.
1:51:02 Wait. Oh, he just did it. Okay, good.
1:51:05 Beautiful. So, there's that one. That
1:51:07 was cool. That was really
1:51:09 cool. Let me edit this for a second and
1:51:12 say
1:51:15 um edit
1:51:22 Cool.
1:51:24 03
1:51:28 experiment. There we go. Save
1:51:35 it. Beautiful. All
1:51:37 right. There you have it. So, if you
1:51:40 haven't been to the AI salon before, you
1:51:42 should join.
1:51:44 Is it true that Tik Tok shuts down on us
1:51:47 on Sunday? It was a big broadcaster that
1:51:49 said it. Uh it I would absolutely not be
1:51:54 surprised if it shuts down on Sunday. I
1:51:56 would assume that it's a negotiating
1:52:00 tactic. So don't be surprised if there's
1:52:03 some big announcement
1:52:06 um that you know someone saved the day
1:52:10 and it doesn't shut down. But also,
1:52:12 don't be surprised if it's just [ __ ]
1:52:14 shut down. This is why. So, if you're on
1:52:17 Tik Tok right now and you don't know
1:52:19 about the YouTube channel, go to YouTube
1:52:22 and
1:52:23 subscribe to the salon or no, learning
1:52:31 lab-ai on YouTube. That's my YouTube
1:52:34 channel. So, I've been dual
1:52:37 broadcasting. They call in the business
1:52:39 they call it
1:52:40 simalcasting. I've been simal casting on
1:52:43 Tik Tok and YouTube and X and LinkedIn
1:52:47 but primarily YouTube for about four
1:52:49 months
1:52:51 now. Um because I just assume the Tik
1:52:54 Tok shit's going to go away. All
1:52:56 right. So it it may I don't know
1:52:59 anything more than you know. Um,
1:53:02 but it seems like apparently 175%
1:53:07 tariffs have pissed some people
1:53:12 off. Oh, dang it. I hit my limit on 03.
1:53:16 So, Cindy [ __ ] that means you're
1:53:18 winning. And just flip down to 04 mini.
1:53:21 Like, just keep playing with the other
1:53:22 models until you max them all out. I
1:53:25 like run yourself out of being able to
1:53:28 do these things. I think it's really
1:53:30 important to start experimenting with
1:53:31 what these things make possible. I mean
1:53:34 that that reconstruction of that
1:53:37 image with all the research it did where
1:53:40 I asked it about a Mayan reconstruction
1:53:42 and then it identified that's not what
1:53:44 it was and and recreated, you know, this
1:53:48 what it might have looked like. That's
1:53:49 pretty [ __ ] insane. Let's do one
1:53:51 other thing before I go.
1:53:56 Let's
1:53:57 say this is awesome. Oh, wait. Sorry. Am
1:54:02 I still sharing? I think I am. Am
1:54:06 I? Okay. This is awesome.
1:54:12 Um, now I want you to
1:54:18 create a
1:54:22 photograph that looks like a
1:54:27 National
1:54:31 Geographic
1:54:33 photographer went back in
1:54:38 time and
1:54:42 photographed this fresco
1:54:46 um
1:54:48 in its
1:54:50 actual
1:54:57 location, you know, on
1:54:59 a wall or building
1:55:04 or what may be appropriate.
1:55:09 It should be
1:55:11 a wide
1:55:14 image
1:55:17 captured
1:55:20 at magic hour just
1:55:24 before sunset.
1:55:31 have high
1:55:33 contrast and
1:55:36 look
1:55:38 incredibly
1:55:43 realistic. I want the
1:55:46 viewer to feel as if
1:55:50 they are
1:55:54 there. That's good enough. All right.
1:55:57 Boom. It's going to be cool. It's going
1:56:00 to be cool, people. I promise you, it's
1:56:02 going to be awesome or it'll suck. It's
1:56:06 been fun and unfortunate. Unfortunately,
1:56:07 I have to jump off. Have a great
1:56:09 weekend. Thank you, Steo. Have a great
1:56:10 one. Yeah, we're basically done
1:56:12 here.
1:56:14 Steo, find us if Tik Tok is gone. Yeah,
1:56:17 Steo. Well, Steo's here on he's
1:56:19 commenting on YouTube.
1:56:21 So, that's the other thing. Hang out in
1:56:24 the AI salon. hang out in the in the AI
1:56:26 learning lab channel on the salon
1:56:28 because then you can talk even when I'm
1:56:30 not live, right? You can you can keep
1:56:32 the conversation going. So, am I
1:56:38 late? Chemical farmer. Excellent. I'll
1:56:41 make some
1:56:42 time. Mr. out there selling some stuff.
1:56:45 Chemical farmer super late out with
1:56:47 friends. Friday night date night. Good
1:56:49 night. Should I rewatch? Yeah, this is a
1:56:51 good one to rewatch because we're
1:56:52 talking about how important 03 is, the
1:56:56 new model, and then the ' 04 mini and 04
1:56:58 mini high, why they're important. And
1:57:01 then I also shared an article that's got
1:57:03 all these example prompts that are non
1:57:06 science and math prompts and we
1:57:09 experimented with some of them tonight,
1:57:10 so it's probably a good one to rewatch.
1:57:12 And then the link to the article article
1:57:15 if you just go to my exac account at
1:57:16 Kyle
1:57:17 Shannon sort of the second or third post
1:57:20 down is this article and the article is
1:57:28 called unlock 03's hidden superpower a
1:57:33 playbook for non of non- STEM prompts
1:57:36 for everyday innovators to explore.
1:57:38 Right.
1:57:40 Oh my god, that's
1:57:43 gorgeous. Holy [ __ ] Look at
1:57:54 that. That is amazing. It's been fun,
1:57:58 informative. Okay, got to
1:58:00 go. That's [ __ ] gorgeous. Okay, I'm
1:58:03 gonna I'm
1:58:04 gonna download this and post this into
1:58:06 the into the comment with this
1:58:11 with the
1:58:14 prompt. So, I'm going to copy
1:58:17 that and then I'm going to download that
1:58:20 and then we're going to go back to here
1:58:22 and I'm going to throw in another
1:58:23 comment and say
1:58:26 um I then
1:58:31 asked
1:58:32 to place the
1:58:35 frisco in situ. How do you spell situ?
1:58:38 Is it sit
1:58:40 tu?
1:58:46 Anyone? I'm such an
1:58:50 idiot.
1:58:53 Siri, what? Oh, come
1:58:57 on.
1:59:04 [Music]
1:59:06 What? I am so done with [ __ ] um
1:59:11 should have quit while you were ahead. I
1:59:12 know,
1:59:15 right? Mighty Networks just sucks. What
1:59:18 What is it with these sites with their
1:59:19 stupid [ __ ] upload?
1:59:22 It's like they got an intern to make
1:59:24 their upload
1:59:39 [Music]
1:59:44 service. There we
1:59:50 go. in situ. Yeah. Okay,
2:00:06 great. And then what I would actually
2:00:08 do, the next prompt I would give it is I
2:00:10 would make it make that image with
2:00:12 people in it, you know, dressed in the
2:00:15 appropriate garb. That's [ __ ]
2:00:16 amazing.
2:00:18 And the fact that it's even, you know,
2:00:23 potentially like culturally and
2:00:25 historically accurate based on that
2:00:28 [ __ ] [ __ ] up fragment of a thing to
2:00:31 go from that to get the education about
2:00:33 what it actually is and then end up with
2:00:36 with that
2:00:38 photo. This is [ __ ] bonkers,
2:00:41 man. All right, I'm leaving you. I'm
2:00:45 leaving you for the weekend. Your
2:00:46 homework is clear.
2:00:49 kick the [ __ ] out of 03, 04 mini, and 04
2:00:52 mini high. Okay, just go play. Go, you
2:00:57 can use the prompts in my article. Go
2:00:59 come up with your own. Cindy [ __ ] did an
2:01:02 amazing thing today. She took my article
2:01:05 with all the prompts in it and she took
2:01:07 information about her client that she's
2:01:09 working on and she goes, "Keep all the
2:01:12 same basic use cases, but personalize
2:01:15 every prompt for my client." And so it
2:01:18 gave her that whole example of all the
2:01:21 different things you could do with 03
2:01:22 specific to her client. [ __ ] genius.
2:01:26 Like get creative and just kick the snot
2:01:30 out of these models. And then when we
2:01:32 come back next week, I want you guys to
2:01:34 like share stuff. Here's the other
2:01:36 thing. Share what you're learning. So,
2:01:38 just like I shared this thing in the AI
2:01:40 salon, figure out an appropriate channel
2:01:42 wherever you want to share it. You can
2:01:43 share it here. You can share it up in
2:01:47 um in
2:01:50 uh
2:01:54 Duh. Look what I
2:01:58 made. All right? Or you could share it
2:02:01 in the water cooler. So water cooler is
2:02:05 just a general
2:02:07 chat. Look what I made is like showand
2:02:11 tell
2:02:13 area. Oh, that's beautiful. Who did
2:02:15 that? Sharon
2:02:17 Crawford. That's [ __ ]
2:02:22 gorgeous. And then um and then AI
2:02:25 learning lab. You can share it in there.
2:02:28 All right.
2:02:31 Let's teach each
2:02:32 other what this thing's good
2:02:36 at. Don't just hoard the information.
2:02:39 Don't just be a lurker. Generously lead.
2:02:41 That's one of our
2:02:44 values. It's one of our imperatives
2:02:47 within the within the salon, right? The
2:02:50 cycle to AI readiness. Play first,
2:02:54 mindfully create, generously lead. So,
2:02:56 just go play with this thing. Once you
2:02:59 figure out it can do something cool like
2:03:00 we did tonight, mindfully create it and
2:03:03 then share what you did. Critical right
2:03:05 now. Okay. All right. Peace out. Have a
2:03:08 fantastic
2:03:10 weekend. Um I hope you had fun. I did.
2:03:14 This was a good night. This is a good
2:03:15 high energy
2:03:17 night. And uh I appreciate you all more
2:03:20 than you know. All right. Have a great
2:03:22 weekend. Bye.