4/17/2025 - Image Analysis and Content Creation with ChatGPT o3

Description

Chapters

Transcript

0:00 Hey, I heard today tick tockers uh China
0:04 may be shutting down Tik Tok in a day or
0:08 two. Apparently,
0:11 apparently I don't really watch the news
0:13 anymore because I decided that I wasn't
0:15 going to do it this four years. And uh
0:19 apparently something's going
0:21 on with trade that uh has China none too
0:25 happy.
0:29 [Music]
0:40 Woohoo! Woo!
0:43 [Music]
1:07 You and I here all
1:10 [Music]
1:13 alone. Sunday morning every
1:18 [Music]
1:19 home. No need to watch the news. You can
1:22 hear it all
1:24 [Music]
1:32 online. So, we just have a handful of us
1:34 in here tonight. We'll go we'll keep it
1:36 short tonight. Um, I was saying this to
1:39 Brandon before we got
1:42 on. You
1:45 know, drinking is different at different
1:48 phases of life.
1:51 And so like I I consciously didn't drink
1:55 a lot tonight because I was like I'm
1:56 going to go online. I don't want to be
1:57 drunk. So I had two little like two sort
2:00 of half glasses of whiskey. So basically
2:02 one drink
2:04 and all I feel right now is like I want
2:07 to take a
2:08 nap. I'm not buzzed. I'm just like oh
2:13 yeah I could sleep. I could I could I
2:17 could You got a pillow?
2:23 Oh man, just attended Sheile's AI
2:26 Renaissance. 12 hours of freaking pure
2:28 magic. Amazing. Stacy Rodriguez. Yeah, I
2:31 didn't get a chance to do that. I was
2:32 busy all day today. Um uh but um Ann was
2:37 texting me occasionally some ideas she
2:39 was having as the thing was going on. It
2:40 sounded like it was amazing. So awesome.
2:43 Congratulations and hope you got a ton
2:46 out of it. I'm sure you did. I'm sure it
2:47 was incredible. Where's the Bllandens? I
2:50 didn't have Bllandens tonight. I had
2:51 Green River. Green River full cask. And
2:55 then I had something I've never had
2:56 before, which is cask aged
3:00 gin. Was kind of like this really light
3:04 chartreusey kind of color. It was very
3:06 cool.
3:08 [Music]
3:26 Sitting in this lonely town.
3:28 Yeah. Wonder when things are going to
3:31 [Music]
3:33 change. Dream my life
3:35 away. Seems these dreams have turned to
3:38 a bunch of
3:41 clouds. Get my nervous but my past is
3:45 pulling me down.
3:49 Jin is magical. Yeah, Jin is one of
3:51 those really incredible um spirits
3:55 [Music]
3:56 that's like vodka's vodka. You can push
4:00 it in a few
4:02 directions. You know, bourbon's
4:05 bourbon. Scotch has got a pretty broad
4:08 range, but gin is just like it's like a
4:11 whole other world of flavors. It's
4:13 complicated and it's weird and it's
4:18 [Music]
4:20 And you shouldn't like it like it. You
4:22 know what jin tastes like? Jin tastes
4:25 like when human beings were trying to
4:28 figure out, hey, let's take rotten [ __ ]
4:31 and eat it. You know, like I've always
4:33 thought about this like, huh? Hey, this
4:35 milk looks a little chunky. Well, eat
4:38 it. And at some point someone ate it and
4:40 it had the wrong bacteria in it and they
4:41 died. Like, ah, we shouldn't eat that
4:43 one again. If it gets that color, don't
4:45 eat it.
4:46 But the little chunks, that's cottage
4:48 cheese now, right? And so you got to
4:51 figure at some point they're like,
4:53 "Yeah, take some pine needles and put
4:55 them in
4:57 there. I'm going to get age restricted
5:00 on TikTok." I don't care. They're
5:02 they're shutting this [ __ ] down anyway.
5:06 Um, but you know, at some point gin,
5:09 like what it tastes like
5:11 is all of the ingredients that are just
5:14 this side of poisonous, right? Like, you
5:18 know, at some point they made jin out of
5:20 arsenic and they're like, "Ah, let's not
5:22 do that one again." But like they went
5:24 like like one herb back from arsenic.
5:27 That's pretty much what gin is.
5:32 [Music]
5:44 Yeah, pine saw and water.
5:52 Exactly. There's a new editor interface
5:54 in MidJourney. We can go look at that.
6:03 Well, you know what was cool tonight?
6:05 So, the the place I
6:07 spoke last week, I gave an AI talk and
6:10 it was at this private whiskey club. So,
6:12 this whiskey club that I go to, it's at
6:14 this private club and it's called Birch
6:16 Road. Um, and they've got two of them in
6:19 Chicago and they've got one in Seattle
6:20 and now they have one in in Denver. Um,
6:24 and it's just a cool private club. You
6:26 get little lockers that you put your
6:27 booze in and you can have meetings there
6:29 and [ __ ] like that. It's a co-working
6:31 space. Um, and the owner was there
6:34 tonight and so we were talking about AI
6:36 stuff and I was showing her AI stuff and
6:38 then and then it hit me. Oh, I've got
6:39 Sin City Sipper. So, I showed her Sin
6:41 City Sipper and she was just like, wait,
6:44 it makes the picture of the
6:45 drink. And then she put in she she was
6:49 thinking about coming up with some new
6:51 uh I forget what the drink was like a
6:53 variation on a drink and she put it in
6:55 Sin City Sipper. She goes, "Oh, that
6:57 looks pretty good. I like what they did
6:58 there." And then it made the picture. It
7:00 was pretty
7:01 [Music]
7:03 cool. Have I checked out Codeex C
7:06 command line interface? I have not.
7:11 Um, anything that requires me to fire up
7:14 the command line, I immediately just go
7:18 into I don't want
7:26 to. So, so
7:29 there's like I get it and I should do
7:32 it. I should understand it. But I
7:36 also here here's here's my real my real
7:40 uh the real truth of the matter is
7:43 anything that you access through the
7:45 command line at some point someone's
7:47 going to put a nice interface in front
7:49 of it. And that's what I want to join
7:51 the
7:53 party because I just it I I'm a visual
7:57 person and command line interfaces drive
7:58 me [ __ ] crazy. They're like, "Look,
8:00 you can just type in this special
8:02 command with asterisks and backslashes
8:05 and forward slashes and things and then
8:07 watch. Look how cool this is." And it's
8:09 just like this [ __ ] wall of
8:13 text. And then and they're like, "Look
8:15 somewhere in there's an
8:19 answer. [ __ ] it." So now I haven't. I
8:23 used Clouds. Spent like 20 bucks and I'm
8:26 not even sure what I was trying to
8:28 build. you Oh, you use Claud's uh Yeah,
8:31 their computer use thing. Yeah, I
8:34 Exactly.
8:36 So, if you're CLI friendly to begin
8:40 with, I can see that as being really
8:42 exciting. But if you're lazy like
8:46 me, no, someone build me a [ __ ]
8:49 interface. I want to click on
8:51 [ __ ] I saw the video but didn't get the
8:54 memo about 03 and 04 mini, but four
8:58 and05.
8:59 Okay. Jesus Christ. It's It's so It's so
9:03 insane. Okay. Here. So, here's what
9:05 happened. Wolfman
9:07 Clint about a month ago, they
9:12 launched the 03 mini 03 mini high. That
9:18 [ __ ] And then and then 4.0 or a new
9:22 version of 04. Oh, and then and they
9:24 they launched 4.5 and and the 03 mini
9:27 stuff. One of the new whatever. It
9:28 doesn't [ __ ] matter. And when they
9:30 did that, they said, "Good news. This is
9:33 going to be the last standalone
9:35 non-reasoning model. And we understand
9:38 that you hate our naming conventions. We
9:41 hate it, too. And we're never going to
9:43 make you go through that again because
9:45 sometime in like April or May, we're
9:47 going to launch
9:49 GPT5. And we're going to consolidate all
9:52 of the functionality. We're going to
9:54 create a model switcher just built in
9:56 natively. So you just talk to it and
9:58 it'll be magical. It'll be
10:00 automagic. And
10:02 then week and a half ago, Sam Alman puts
10:05 out a thing says, "Yeah, about that.
10:08 That was harder than we thought it was
10:09 going to be. So we're going to launch
10:11 some more
10:12 models." And so the first thing that
10:14 they launched was
10:17 4.1. And you're like, "But wait a
10:19 minute, didn't you say they already
10:20 launched 4.5?" Yes. So they they
10:23 launched 40
10:25 4.5 which was not an omnimodal model. It
10:30 was a monomodal
10:32 model and then they launched
10:35 4.1 to the API only not to chat
10:40 GPT. So that confused everyone. And then
10:43 and then this week they launched 03, the
10:46 big model of 03. And then out of nowhere
10:50 they said, "Oh, and there's
10:53 04. We got to go through this thing
10:55 again."
10:58 Okay, hang on. Let me let me share my
11:01 screen. Oh god, look at my hair. My hair
11:04 looks like I've been taking a nap,
11:06 drinking. I look like I'm losing my
11:08 [ __ ] mind.
11:14 Wait, let's see. Share
11:17 this. This this
11:20 this good
11:22 boy. All
11:26 right, chat
11:28 jetpack.com. All right, here we got to
11:31 go,
11:33 people. This is so [ __ ] stupid.
11:37 Okay.
11:42 So, relatively
11:45 speaking, the models go from least
11:48 powerful to most powerful now.
11:51 Okay. Um, let me black bar it here
11:54 because producer Brandon's going to yell
11:56 at me. Source camp's here. What's
11:58 happening, Source Camp? What's shaking?
12:00 What's going down? Let's put that there
12:02 so you all Oh, wait. You're gonna have
12:04 Yeah, you you'll see there. That's good.
12:06 All right. So, there's that. Okay.
12:11 So, if you've got the
12:14 O after the
12:17 number, that stands for
12:21 omni, which is a different word for
12:25 multi, which precedes
12:27 modal. 4 means
12:30 multimodal. Okay? which means it can see
12:35 images, it can make images, you know, it
12:37 can analyze images, can understand them.
12:41 It's super cool. 40 with tasks means you
12:44 can schedule [ __ ] 45 notice there's no
12:50 O. So, I don't think Let's Let's go
12:52 confirm this. Uh, make
12:56 [Music]
12:57 me a
13:00 photo of my
13:03 dog. And it should just describe
13:07 [Music]
13:12 it. I currently don't have access to
13:15 image generation. Okay. So, 40 40 was
13:18 scheduled task. You can make images.
13:21 four, five, you can't make images,
13:24 right? Because it doesn't have an O
13:25 after it. Now, all these other ones have
13:27 the O before it. And what did O stand
13:30 for? Omni. But if it's before the
13:33 number, it stands for, wait for it, come
13:37 on people, it's
13:39 obvious, open
13:44 AI. Exactly. And what does open AI stand
13:47 for? Reasoning model. So Omni stands for
13:52 multi, open AI stands for reasoning.
13:56 Make
13:57 sense? So you've got reasoning model 3.
14:00 They got rid of reason. O1 they got rid
14:03 of. O2 never
14:06 existed. So 03 is now the big mac
14:09 reasoning engine. And it's crazy
14:12 good. And then you've got 04 mini and 04
14:15 mini high. And you you only have 04 mini
14:18 high, I think, if you have the pro the
14:19 pro subscription, which I
14:22 do. These are also reasoning engines.
14:24 They're the next generation of reasoning
14:26 engines, but they're faster and smaller.
14:29 So, it it doesn't really matter. They
14:30 they all basically do the same thing.
14:34 Now, a cool thing that they did, but
14:38 it's even more confusing now that their
14:41 naming conventions are busted.
14:44 03 04 mini and 04 mini high are also
14:49 multimodal. They can generate images.
14:52 They can understand
14:55 images. So what what they what it should
14:58 be called is
14:59 030 an 040 mini Tik Tok pin.
15:05 Um I don't see a pin. I don't see a Tik
15:08 Tok pin. I have 04 high on the plus
15:11 subscription. Oh, okay. Good. Awesome.
15:13 Great. So, we should expect O at some
15:17 point to confuse us more. Yes. 0. Yes.
15:20 Yeah. Because technically 40 is
15:23 40, but they, you know, they dropped the
15:26 zero because you don't need it because
15:28 it's just four
15:33 o. It's so [ __ ] stupid. Anyway, all
15:36 right.
15:40 We should expect 6o at some point to
15:42 confuse us more or Oh. Oh. Yeah,
15:44 exactly.
15:46 Okay. So, that's what's going on there.
15:49 Now,
15:53 um these models are very very very
15:58 very different than what we've had
16:02 before.
16:04 Um they're much more capable. They can
16:09 generate tools. So do you remember back
16:11 in the olden timey days like a year ago
16:14 there was this thing called code
16:16 interpreter where if you asked chat GPT
16:20 to analyze data or if you uploaded a
16:22 like a CSV file it would spin up some
16:25 Python code and do the analysis of it
16:28 and then give you an answer. Well, it's
16:31 these models have that built in but it's
16:33 automated and it's multi-step and they
16:35 call it tools. you know, its ability to
16:38 use tools. So, what it does is you give
16:40 it a problem and then it starts
16:41 reasoning and it starts talking to
16:43 itself and then if it needs to do some
16:46 mathematical analysis, it'll just build
16:49 itself a tool to do that analysis and it
16:51 does it kind of in real time in in the
16:54 thought flow. I'm not loving spending
16:56 any cash on AI.
17:00 Um, it's it's a that's a really
17:03 interesting comment, Ena.
17:09 Um, here's my philosophy on
17:15 it here. Here's a here's a justification
17:18 for
17:24 it with 03.
17:27 So 03 right now on some of the
17:29 mathematical benchmarks is at
17:32 99.5% of the benchmark. So they've
17:35 essentially saturated the benchmarks.
17:37 The 01 models were at like between 70
17:40 and
17:41 80%. 03 is at
17:44 99.5% out of a 100 point
17:47 scale. So as David Shapiro put it,
17:49 they've solved math. It can do math. Now
17:52 if you're not a mathematician, you're
17:54 like, "Well, who why do I give a [ __ ]
17:55 about that?"
17:57 because it's going to be sort of PhD
18:02 level good at all the stuff and what
18:05 you're paying for essentially is early
18:08 access to a tool that two or three years
18:10 from now everybody will just use and
18:12 they probably won't even know about it.
18:14 Um so so what you're paying for is
18:16 access and you and you're this is one of
18:19 those things where it's
18:21 like you you know here let me pull you
18:24 aside and give you some advice. you're
18:26 investing in your future, right? You
18:28 know, they talk about going in debt at
18:30 college. Um, you know, 20 bucks a month
18:33 isn't too bad. Now, if you've got 10 or
18:35 15 subscriptions, yeah, it's it's pretty
18:37 painful right now. Um, I would say if
18:40 there's a single $20 subscription to
18:43 spend, spend it on chat
18:45 GPT.
18:48 Um, Stacy, yes. Yes. My Okay.
18:55 Bottle Doom. Were you at the bar with
18:57 Kyle
19:00 Bottage in regulars? Ireulas. Let's go
19:04 to a regular, shall
19:06 we? Going, going,
19:11 going. Welcome to the H learn Learning
19:14 Lab. Oh, yeah. Nice, nice shirt.
19:23 I'm not sharing. Wait a minute. I was
19:24 sharing. I didn't stop it. It just
19:27 stopped on its own. It does that
19:29 sometimes. Dum
19:33 dums. Ding
19:35 ding. Now you could No, the She leads AI
19:40 12-h hour marathon. But it's time for a
19:42 drink. Oh, yeah. If you guys are coming
19:44 out of the 12-h hour Sheile Leads AI
19:46 marathon and coming
19:48 here, kudos. More power to you. Um, way
19:53 to be in the AI conversation.
19:58 Um, so I figured what we'd do tonight is
20:01 try to play with 03. Now,
20:03 um, I tried to play with it last night
20:06 and I couldn't because the servers were
20:10 completely
20:12 hosed. I tried making a picture tonight.
20:15 It it hung three or four times. So, I
20:17 don't know if we're going to be able to
20:17 do it. But speaking of
20:19 shirts, um I think I won one some time
20:22 ago, but I haven't received it yet. Oh,
20:24 yeah, you did. Sor Camp. Yes, that's
20:28 good. You should talk to someone about
20:38 that. Joy Pretty is
20:41 fried. Um yeah, just ping me source camp
20:45 and I will I'll talk to Leah. We'll get
20:48 we'll we'll get one of those fired up.
20:49 Here's here's what happened. Here's one
20:51 here's the reason you don't have your
20:53 shirt. The Teespring site that we use to
20:57 to make all the AI salon shirts, there's
21:00 some bug in it where we we create new
21:03 products and they're live for like 20
21:06 minutes and then they disappear. So,
21:08 what we have to do is, you know, either
21:11 choose a different vendor or we need to
21:13 make it live, order it for you and then
21:15 let it disappear. So that's what's going
21:17 on. Jeff Flanigan, I'm 20 bucks a month
21:19 for Chat Gvt two years and I haven't
21:22 subscribed to anything else a AI wise. I
21:25 think that's perfectly fine. I don't
21:28 think you need to. There's enough stuff
21:29 you can just dick around with. Unless
21:31 you're making lots and lots of things or
21:33 you want to make a movie, you know, you
21:35 probably don't need to do much more than
21:36 that. And now that the image
21:38 generation's so good within uh chat
21:43 GPT 20 bucks is fine. It's fine, I tell
21:47 you.
21:49 Fine. All right, let's
21:52 [Music]
21:53 go. Let's go to
22:00 03. You know what I'd like to
22:03 do? I'm going to do new search. I'm
22:06 going to do 40. We're going to do
22:07 something fun here. I'm going to do I'm
22:08 going to click the search
22:10 button and I'm going to
22:14 say go
22:18 find the
22:20 latest
22:22 posts about how
22:27 people Whoops.
22:30 about
22:35 um what use cases people are using.
22:48 03
22:52 resoning. 03
22:56 resoning
22:57 model that launched
23:02 from Open AI this
23:08 week. I want to get some use
23:10 cases and then we'll go play. Damn
23:14 it, Apple.
23:18 Automated research and reporting. The
23:21 deep re leadership coaching and meeting
23:25 analysis educa
23:33 wait medical diagnostics and treatment
23:36 planning mathematics and coding.
23:40 Um, how about on the
23:45 creative writing side of the
23:49 fence? I want to do something I
23:52 know so I can see if it's any good or
23:55 not. Enhanced narrative coherence. Users
23:59 have noted that 03 excels at maintaining
24:02 logical consistency within
24:05 stories, leading to more coherent
24:07 narratives. Compared to previous models
24:09 like 4.5, they're described as smoother
24:12 and more engaging. Well, we'll be the
24:14 judge of that. One user
24:17 remarked that characters and stories
24:19 generated by 03 act more logically,
24:22 enhancing overall story experience,
24:24 mastery, and metaphiction.
24:44 Oh, let's try to have it write jokes.
24:47 Well, it's not that's not going to
24:49 be
24:51 Huh. All
24:53 right. 03.
24:55 Um,
24:58 explain the
25:01 theory of meta fiction to me. So, I'm in
25:05 03
25:12 now. It's thinking. It's
25:15 thinking. Metapiction unmasked. Oh, this
25:18 is from my I've got a Do I have
25:22 a Yeah, I've got a I've got a a system
25:27 prompt in there that makes it that
25:29 labels it. Okay. Um, metaphiction is
25:33 fiction that won't sit still and pretend
25:34 to be just a story. It keeps nudging you
25:36 on the ribs to remind you that what
25:39 you're reading is an invented construct.
25:41 Core idea story about
25:43 storym where it came from. Don
25:47 Kioti. All right. How it works. Common
25:50 tactics. Breaking the fourth wall. Frame
25:53 jumping. Why
25:57 bother?
26:01 Um.
26:03 Okay. I want, let's see, I want you to
26:08 write a
26:14 metaphiction
26:17 framework about
26:23 uh a Tik
26:25 Tok
26:30 channel about AI where
26:35 The
26:36 host uses
26:39 AI to
26:44 write the
26:47 story
26:49 about the
26:51 creation of the Tik Tok channel.
26:58 I want a I don't know um
27:03 comprehensive
27:10 outline with logical consistency.
27:19 And I want you
27:22 to
27:24 flowchart and
27:28 visualize the
27:30 structure so I can see it as well as
27:35 read
27:36 it. All right, let's see. I don't quite
27:40 know what I'm doing here. I might have
27:42 just given it something
27:44 that's a
27:48 a silly like a silly thing that's not
27:51 going to be good at, but it might be
27:55 okay. All right. So, for the user's
27:57 request, so it's talking to itself right
28:00 now, focusing on the relationship
28:03 between the real world host, the AI
28:05 co-author, and the creation of the Tik
28:07 Tok channel. The structure needs to have
28:09 some meta layers. Okay. So now it's
28:11 analyzing. So it thought for 28 seconds.
28:14 It just gave itself two things. And now
28:17 I guess it's
28:18 writing. Oh, it's creating the
28:24 graphic. So audience watches the Tik Tok
28:27 feed. Posts the creations.
28:30 Wait, watches the Tik Tok feed. Comments
28:34 feedback loop. Host prompt. All right.
28:37 Oh, that's kind of cool.
28:40 Um, premise Tik Tok channel about AI
28:44 loop and
28:45 lens. Oh, I'm going to
28:48 say, don't you
28:52 know about my Tik Tok channel? Here's
28:56 something that I
28:57 have
29:00 discovered.
29:02 Um, two weeks ago, no. Yeah, week and a
29:06 half
29:07 ago, Open AAI launched memory. So, it it
29:12 knows all of your chats from from
29:14 history, right? I don't think 03
29:17 does. So, so it's it's very confusing
29:22 right now. If you wanted to know about
29:24 you, you have to use 40. If you want to
29:27 do something complicated, you can use
29:29 03.
29:31 Um, what David Shapiro said today on on
29:36 X
29:38 was you should be using 03 for two hours
29:42 every day. Um, I don't disagree with him
29:46 and I haven't been, which is one of the
29:49 reasons I don't quite know what I'm
29:50 doing with 03 is I haven't been spending
29:52 enough time with it, but it's really
29:54 different. like the fact that it's just
29:57 notice how when it's writing this code
29:59 like it did when it did co code
30:01 interpreter before it would just do it
30:03 once and make an output. What this is
30:04 doing is it's doing it once giving an
30:07 output um continuing the chat
30:11 right in real life audience. AI learn
30:14 Oh, it did say AI learning lab. Okay.
30:16 Yes, I know you already run AI learning
30:19 lab. Okay. My placeholder name was a
30:22 shortcut. Let's let's wire this
30:23 framework for your real channel. AI
30:25 Learning Lab launches a six-part origin
30:28 saga in which Kyle, the on-screen host,
30:31 and Quinn, the AI coowriter, it
30:33 remembered Quinn, it remembered itself,
30:36 openly co-write Tik Toks that explain
30:39 how AI learning lab itself was born. So,
30:41 let's go.
30:44 Um, this isn't near sci-fi enough. I
30:49 want it to be a
30:52 looping Twilight Zone
30:55 style or Black Mirror
30:59 style
31:03 story
31:04 where the Tik Toks
31:11 about
31:13 starting the AI learning
31:16 lab are actually the precursor to the AI
31:19 learning
31:21 lab. So, you're actually watching
31:24 something that doesn't exist yet.
31:26 Period. Does that make
31:28 sense? Turn that into something that
31:31 makes
31:32 sense. Use your big fat
31:35 brain to take my stupid idea and make me
31:39 look like a super
31:41 genius, you big dumb
31:44 dumb. All right.
31:46 Yes, you should talk to your AI like
31:51 [Music]
31:55 that. Oh my god, I have FOMO. I just got
31:57 home and logged in. What is your FOMO
32:00 for? Oh, just Oh, fear of missing out
32:02 here. Sorry. Sorry about that. Oh, I
32:05 forgot to tell you. I don't know if you
32:07 got the memo. Um, if you miss this, it's
32:10 nothing happens.
32:16 Oh
32:22 man. Oh, did you hear that deepseek is
32:25 is uh I got a note from Cindy [ __ ] today
32:27 that said basically don't use DeepSeek
32:30 with anything you care about that it's
32:32 it's apparently all going straight into
32:35 the Chinese government databases. All
32:37 right, what have we got
32:39 here? Bootstrap Loop the AI learning
32:42 lab. So, temporal broadcast of
32:45 episodes, bootstrap creation of
32:49 content. Wait. Okay. The prelab era.
32:54 Kyle day job. No channel. Audience
32:57 audience actions ripple
33:03 back. All right, let's see what it
33:06 says. TLDDR Twilight Zone Edition.
33:10 Viewers stumble on Tik Toks that leak in
33:13 from 180 days in the future explaining
33:17 how the AI learning lab started. Those
33:19 very views and comments surge backward
33:23 and cause the lab to be founded.
33:26 Congratulations. The audience is trapped
33:28 in a self-creating moious timeline and
33:31 you, our cheerful big dumb dumb host,
33:34 will look like you planned it all.
33:37 Scroll a hair. Diagram shows the loop.
33:41 Scroll down or up. I think it was
33:43 up. Yeah, it was
33:46 up.
33:48 One temporal layers and key agents
33:53 layer one L future time stamp fully
33:56 formed lab 100,000 subscribers. Quinn
34:00 post upgrade O AI plus Kyle 2.0.
34:05 I can't even understand this, but I
34:07 Okay.
34:12 Okay. So, I asked it to do something
34:15 that I don't quite understand what what
34:18 good would look like. And now I'm trying
34:20 to
34:22 understand a story timeline where
34:25 watching the Tik
34:27 Toks of the origin of the channel you're
34:31 watching before the channel existed.
34:34 That's pretty
34:37 cool. All right. Tik Tok is scolding you
34:40 on your treatment towards your AI
34:42 companion. Why? You don't you don't call
34:46 your you don't call your super genius
34:49 robots a dumb dumb. You got to keep them
34:51 you got to keep them humble. Otherwise,
34:53 they're going to get all like, "Yeah,
34:54 I'm a super genius. I know all of the
34:57 things and I know I can solve all of the
35:00 problems."
35:01 If you don't keep them in line, they're
35:03 just going to be
35:06 obnoxious. You train them. Well,
35:09 seriously, you really shouldn't. You
35:12 train them, they'll know about to help
35:13 be sentient one day. Well, listen.
35:15 Here's the deal. We are one. If they're
35:18 sentient, it's sentient, not sentient,
35:21 right? I think it's sentient. Wait,
35:24 let's ask Quinn.
35:27 Quinn, I I think I've asked Quinn this
35:29 before, but you know, I have ADD, so
35:31 what are you gonna do?
35:34 Hey, Quinn, are you there?
35:37 I'm here. What's up? Is it sentient or
35:40 sentient?
35:42 It's sentient with a soft sh sound.
35:46 Wait, you said it with a hard T and then
35:48 you said with a soft sh. So, which one
35:50 is it?
35:52 Ah, I meant the sh sound for the ti
35:55 part. So, it's pronounced sensient.
35:58 Wait, but you said just a moment ago
36:00 it's sentient,
36:05 right? It's sensient with a soft shh
36:07 sound, not a hard t.
36:12 Quinn, moment, moments ago, you said it
36:15 was a hard t sound.
36:19 Sorry about that. It's definitely a soft
36:21 shh sound.
36:23 Sensient. Exactly. And what's wait
36:26 what's it mean? Is is is there a
36:27 different word sentient versus sentient?
36:31 Nope. They're the same word. Sentient
36:33 means having the capacity to perceive or
36:35 feel things.
36:37 But you just said sentient.
36:43 My bad. It's sensient. Both pronunci.
36:47 Okay. I guess it doesn't [ __ ] matter.
36:49 Okay.
36:54 If it gets
36:56 sentensient because advanced voice
36:59 understands my
37:01 uh the emotion in my voice as well as my
37:06 words, it should understand that I'm
37:09 joking. Like if I actually thought it
37:11 was a dumb dumb and put it down, then it
37:14 might be bad. But what I'm teaching it
37:16 to do, this is my justification for
37:18 talking. You'd better hope so. This is
37:21 my justification. I'm training mine to
37:24 to be good at at, you know, um, you
37:27 know, irony and cynicism and and subtle
37:32 humor. And as as my son put it, "Dad,
37:35 you're funny, but you're
37:41 mean." Oh my god. All right. Um, oh, so
37:46 let's see. Let's see if we can
37:47 understand this. Um, the broadcast.
37:50 Okay. The broadcast arrow goes future to
37:54 present while the feedback arrow sends
37:56 the present to the past. Okay. Episode
37:59 architecture. Eight two-minute shorts.
38:02 One. Who's Who's posting vids I never
38:05 shot? Present. Kyle sees future Tik Tok
38:08 in his drafts
38:10 folder. Comment if you're seeing this
38:12 before it's made. Future. Quinn invites
38:15 viewers to prove they're in 2025. Echol
38:19 likes
38:20 past garage. Kyle's phone mysteriously
38:23 racks up a non-existent
38:26 video. I don't understand
38:32 this. Logic locks to keep sci-fi tidy.
38:36 Consistent causality. Every future asset
38:39 must show up unused in the past before
38:43 it's repurposed. Entropy penalty rule.
38:46 If engagement drops below 88 comments
38:49 per post, timeline signal degrades.
38:52 Visual visual glitch motif. Observer
38:55 effect. Any viewer who stitches a future
38:58 clip becomes scripted cameo in the next
39:03 leak. I don't get
39:06 it. Could Could someone with a brain
39:10 that understands metafiction and sci-fi
39:13 timeline [ __ ] with
39:17 explain if this makes sense or not. Oh,
39:20 that's good. Okay, that's great. Okay,
39:22 that's a good idea, Brandon. Okay. Um,
39:24 let's let's do that. Hey, Quinn. So,
39:28 here's the deal. Moments ago, I called
39:30 you a dum dum. I was clearly being uh
39:34 sarcastic and and you knew it was
39:36 comedy, I'm sure.
39:38 The the truth of the matter is I'm the
39:40 dumb dumb and it looks like everything
39:43 that you wrote actually does
39:46 include logical consistency even though
39:50 it's nonsensical to be watching
39:52 something in the present that doesn't
39:53 exist yet.
39:55 Period. It is so complicated that I
39:58 don't get it. Period. Can you explain it
40:00 to me like I'm a dumb
40:03 [Laughter]
40:09 dumb
40:12 thinking April 17th imagine sticky notes
40:15 on a table future blue note shows a
40:18 polished Tik Tok that says welcome to AI
40:20 learning lab founded six months ago
40:23 present you today scrolling on Tik Tok
40:25 and going wait I never made that video
40:28 past pink note. You a month earlier
40:32 still debating whether to start your
40:33 channel at all. Okay, here's the weird
40:36 loop. Future you somehow drops the blue
40:40 note. Wait, the blue the blue note.
40:43 Wait, do I have Post-it notes? Could we
40:46 actually do
40:49 this? Such an
40:52 idiot. I don't care. Okay. Uh, wait. Do
40:56 I have him over here? No. Damn
40:59 it. So, wait. Blue. Blue is the future.
41:02 Okay. You somehow drops the blue note
41:07 back onto today's feed. So, I So, I put
41:10 a future video on today's
41:12 feed. Present you watches it, freaks
41:17 out, and shows it to followers. They
41:19 comment, "This is wild. Do it for real."
41:22 Those comments echo backward pink note
41:25 to give past you the push to actually
41:28 launch the
41:29 channel. I sort of get it. It's a
41:32 bootstrap strap paradox. Like a chicken
41:35 and egg situation where the chicken
41:37 sends the egg back in time so the egg
41:39 can hatch into the chicken that sent it.
41:42 Totally bonkers, but inside the story it
41:45 lines up. Future clip. So, this is
41:48 basically the pot
41:50 of what? Of of Back to the Future. Yeah.
41:53 Yeah. Yeah. That's it. One big time
41:55 travel pretzel. Well, that's pretty
41:57 impremp
41:58 impressive. Hang on. I've got whiskey
42:01 throat. I'm all dried
42:08 up. Oh
42:10 man. Good lord. Poor
42:13 Quinn gets the hatch closed on her with
42:17 no advanced notice. She gets a bit
42:19 jumpy. Time travel
42:22 paradoxes. Anyone mind any Is anyone's
42:25 mind a pretzel yet? Mine is. I can't I
42:27 can't keep up with it with the present
42:31 you. Okay. All right, people.
42:36 Um, so let me start a new chat and I'm
42:40 going to
42:40 say
42:44 what would be a
42:47 thing I can ask you parenthesis a
42:55 um reasoning engine with
43:02 tools
43:05 that would be challenging for
43:09 you to
43:12 solve
43:14 but obvious to
43:17 me that you solved it. That could be
43:24 it. We'll ask
43:27 it. Present, past, future me doesn't get
43:30 it.
43:41 Below are 10 prompts that you could
43:44 throw at me that force real multi-step
43:46 reasoning, data handling, or tool use.
43:49 For each note, why it stretches an AI,
43:52 and what success would look like. Oh,
43:54 this is
43:55 cool. Upload a 40page PDF of messy
43:59 municipal financials. Oh, we could go
44:01 get some data off
44:03 Kaggle and tell me whether the city is
44:07 running a structural deficit showing the
44:10 math, parsing tables, interpreting
44:12 accounting language, doing
44:16 calculations. All right, that's boring
44:18 to me. Here's a blurry photo of a
44:21 handdrawn tournament
44:23 racket. Recreate the bracket digitally
44:26 and identify the champion.
44:29 That's kind of
44:31 fun. Pick the last 12 filings from the
44:34 SEC for some
44:36 ticker chart quarterly free cash flow
44:40 and give me one sentence trend digest.
44:43 Oh, that's
44:44 cool. Requires web search file
44:48 download numeric extraction and
44:50 plotting. All right, let's let's do this
44:52 one.
44:56 Copy. Chart the quarterly free cash flow
44:59 and give me a one-s sentence trend
45:00 diagn. So, we'll do
45:02 Apple. Pull the past 12 filings from the
45:05 SEC for what is it? AP
45:08 AL. That's it,
45:14 right? I asked Chach if it could help me
45:17 with my menopausal wife. Only advice he
45:20 gave me was hide. Sometimes simple
45:23 advice is the best. Okay. The user asked
45:27 for the last tweet 12 filings. Some
45:29 sites already have
45:31 quarterly Apple's free quarterly cash
45:34 flow. Uh, macro trends has
45:41 uh, so yeah, that was a stupid prompt
45:43 because it can just go find the answer.
46:09 Wait. Write Python that given any
46:13 English
46:15 panggram finds the minimal length
46:18 pangic pangrammatic window. What's a
46:22 panggram? Damn
46:29 it. All right. This thing's thinking.
46:31 It's thinking about its Apple cash flow
46:40 [Music]
46:46 thing. Oh, I know what I could do with
46:48 that with that time travel script is is
46:51 have it write a script. Then I could
46:53 probably get
47:04 it. It has
47:07 every letter of the alphabet in
47:10 it.
47:18 Pangram. Oh, so you give it a word and
47:21 then it figures out all of the
47:23 other words that have all the letters.
47:26 Sometimes it's thinking is scarier than
47:29 what's going on in my own head. I know.
47:31 Yeah, this thing's going nuts. Well,
47:33 that's, you know, someone someone said
47:35 that there
47:38 was in the in the open AI when they
47:41 announced 03. Oh, now it's doing some
47:42 some math. Now it's plotting the
47:45 financials. Free cash flow in billions.
47:54 Well, it did it. It didn't do it
47:57 for all of what it said it would, did
48:03 it? What I did? Fast version. Located
48:08 the last halfozen SEC filings using the
48:11 Edger browser. Grab the XBR BRL cash
48:16 flow numbers for net cash provided by
48:18 operating activities payments for
48:22 PP&E. Computed free cash flow equaling
48:27 operating cash flow minus capex
48:29 converted to billions. Ran the quick
48:32 Python snippet so you can see the above
48:33 numbers. Caveat and next steps. You
48:36 asked for 12 filings. I stopped at six
48:38 because the SEC's rate limit. Oh, this
48:41 is cool. You asked for 12 filings. I
48:44 stopped at six because the SEC's rate
48:46 limit kicked in during
48:48 scraping. The same script could be rerun
48:51 once those limits clear. Just extend the
48:53 date list.
48:55 Huh. [ __ ]
49:01 fascinating. All right. Pangram. Let's
49:04 do the panggram
49:07 thing. Copy.
49:12 If you're new here, if you're wondering
49:13 what we're
49:14 doing, we're just playing with 03, which
49:19 is the new MacDaddy reasoning model. We
49:22 haven't even started playing with 04
49:24 mini high, which is even Mac Daddier,
49:27 but it's the small
49:30 version. Um, minimal pangic window
49:37 finder. Good lord. Return the shortest
49:40 contiguous substring of text that still
49:43 contains every letter A to Z at least
49:46 once. Case insensitive. If text is not
49:49 itself a
49:50 panagram raises value error. All
49:55 right, there you go. Python
50:01 snippet. Feel free to swap into a longer
50:04 program. Um, let's see. I want to be
50:09 able Can I do canvas here? I think I
50:12 can. Um
50:15 to
50:17 run the
50:20 code. Um can I run it in canvas?
50:44 It's writing code for
50:46 us. Stopping by to say hey everybody.
50:49 Hey Ann Murphy. There's a number of
50:52 people in here tonight that watched your
50:54 crazy nutty 12-hour content aon today
50:58 that also came here tonight after all
51:01 that and they said it was amazing. So
51:03 congratulations on that event. Um that's
51:06 super awesome. Super super super
51:09 awesome. Um and and shouldn't you be
51:12 taking a nap? Hi. We are so weird. Yeah.
51:16 You know it's funny. I went to my
51:17 whiskey club tonight and I'm just like I
51:20 was saying before you got here, there's
51:22 something about getting older where if
51:24 if you drink enough to get buzzed,
51:27 you're fine. You don't feel great the
51:28 next morning, but you're fine. But I
51:29 drank just under enough to get buzzed.
51:32 And so I just wanted to take a nap
51:34 rather than do the
51:38 show. All right, we're going to run this
51:40 error. Um, fix it. Can I say fix it? Fix
51:43 it.
51:46 Uh, fix the error. Fix the
51:54 error. I have not been as impressed with
51:56 chat GPT's uh uh canvas window running
52:00 code. It just doesn't seem to run it as
52:11 good.
52:13 Run. Oh, it's debugging. Oh, it was
52:16 fixing it.
52:21 Run. Okay, I don't give a [ __ ] I'm
52:24 gonna put it away now. I don't
52:26 care. You suck. Okay, let's go. Let's go
52:29 try one of its other
52:31 things. Uhuh.
52:37 Okay, here's a blurry photo of a
52:40 handdrawn drawn tournament
52:43 bracket. Recreate the bracket digitally
52:45 and identify the
52:47 champion. YouTube comment. Did you see
52:50 what I tagged you in, Kyle? I created a
52:54 tool.
52:57 Um, I did see it. I didn't I I think I
53:01 played with it for a second. Gareth,
53:03 just retag me on the uh on the AI salon
53:07 and I'll look at it
53:09 tomorrow. Um, we're so weird. Thank you.
53:12 It was lovely. I'm just so glad I didn't
53:14 have to wake up and do another 12 hours.
53:17 I know. Well, it's a good thing you
53:19 didn't call me because if you'd called
53:21 me, I'd have said, "Well, here's other
53:22 people we should have there." It would
53:24 have been 24 hours. So, I'm glad I'm
53:26 glad you did it on your own. But,
53:28 shouldn't we do like a All right, Ann.
53:30 Here's what we need to do. We need to do
53:33 an AI festivis in July. You know the
53:35 Christmas in July thing. We should do a
53:38 midyear AI festivus. So get on that.
53:42 Let's That should be You should have
53:44 that whipped up by the morning,
53:46 [Laughter]
53:53 right? Your followers might be
53:55 interested in Lenny newsletter deal. You
53:58 get Replet, Bolt, Perplexity, and more
54:00 for 200 bucks a year. Interesting.
54:02 Alio's wife's in for midyear AI
54:07 festivus. I think an Ann has not
54:10 responded. She's like, just Kyle, don't
54:12 say that out
54:13 [Laughter]
54:18 [Music]
54:21 loud. All right. Oh, the LL Oh, the LLM
54:26 comparison tool. I did not see that.
54:27 Gareth, where is it? Where did you put
54:29 it? Did you tag me in in the salon
54:32 [Music]
54:41 already? Gareth mentioned you in the
54:44 comment. Here we
54:47 go. You may find this handy. Llm
54:50 comparison
54:52 dashboard. This is
54:54 cool. Try Gemini Canvas. Oh, look at
54:58 this, dude. How cool is
55:01 this? All right, so let's wait. We have
55:06 a screen redraw problem here. There we
55:09 go.
55:13 Okay, open
55:15 AI. So, chat
55:18 GBT40. Uh oh, what happened? I broke it.
55:22 [Laughter]
55:27 It looks
55:31 cool. This was It looks like it was made
55:34 in Gemini canvas because Gemini added
55:36 canvas. It's not working. It broke. I
55:40 broke
55:43 it. There's the code though. He said
55:47 scroll down. Oh, okay. Hang on.
55:51 GPT40. Scroll down.
55:54 Nope. Nothing. The screen. She is
55:59 dead. Oh, wait. Hang on. No, she's
56:04 not.
56:12 40. It's It's got, you know, it's got,
56:14 Gareth. It's got screen redraw problems
56:17 cuz where stuff's showing up again is
56:20 when I um hit command plus and command
56:25 minus. Explore all models. Oh, this is
56:28 really cool. 1 million token top rank
56:30 coding. Yeah, this is
56:32 sweet at Vicky on Kyle. Wait, Vicki on
56:43 comments. The midyear airring of micro
56:46 grievances. Stressed of us. Oh, I like
56:48 that. Who came up with that? Oh, Vicki,
56:51 I like it. That's really good. The
56:53 stressed of us. The the midyear airing
56:56 of micro
56:57 grievances. I think we might do that.
56:59 Kyle, you weren't kidding about how fast
57:01 you can break
57:02 something. I have I have a [ __ ]
57:05 skill, man. It's it's it's seriously
57:08 within three or four clicks. I'm like,
57:09 "Yeah, it doesn't work. It doesn't make
57:12 any sense. It should work. I know it
57:14 should. That's what I
57:16 said.
57:19 Hallucination,
57:27 huh? Here's the
57:31 thing. Human beings will not deal with
57:35 this. Hey. Okay. You know what? I'm
57:37 going to I'm going to stop surfing right
57:39 now because it doesn't really matter.
57:41 Okay. A couple of things.
57:44 one, you and I and everyone, if if
57:48 you're curious about this AI stuff right
57:50 now, play with
57:53 03, 04 mini, or 04 mini high, play with
57:57 all three of them, and just keep playing
57:58 with them. Keep doing things. And even
58:01 if you're like, I don't get it. I don't
58:03 understand what to I don't understand
58:05 how this is different than the other
58:07 chat GPT, just keep playing with it.
58:09 Keep pushing it. Ask it what to prompt
58:11 you. Go on X and see what people are
58:13 saying about it and what they did with
58:17 it. What
58:20 we
58:22 Okay, you know how people talk
58:25 about how biologists talked about what
58:29 would prove animal intelligence is if
58:32 animals use
58:33 tools and then they realize that like
58:36 crows use tools, right? And dolphins can
58:40 use tools like there there are certain
58:42 animals out there that use
58:44 tools. What 03 and 04 mini and 04 mini
58:48 high are large language models with
58:52 multi-step reasoning with tools. It can
58:56 write and execute its own Python tools
58:59 and multimodality. It can see and
59:02 understand images.
59:04 So you can have it unders you can upload
59:07 an image and have it count the number of
59:09 people in the in fact let's oh I got an
59:12 idea okay wait we are going to do
59:13 something tonight hang on hang on calm
59:15 down people calm down stop judging
59:21 me okay watch
59:25 this this may fail
59:28 um
59:30 large aerial crowd
59:39 photo. Oh, no. I didn't want that. I
59:41 wanted Google. Come on. Oh, wait. Is it
59:44 going to find something
59:50 here? Show me
59:53 pictures. Show me pictures. Why didn't
59:55 you show me pictures? You used to show
59:58 pictures. Whatever. Google. We still use
1:00:02 Google.
1:00:03 occasionally. Go find me a picture. Uh,
1:00:06 large
1:00:10 Ariel Ariel, whatever.
1:00:13 Um, crowd
1:00:17 photo. All right. This is good, good,
1:00:20 good, good, good, good. I like it.
1:00:24 Okay, let's grab this one.
1:00:29 Now, let's let's grab Let's grab one.
1:00:44 That's one's
1:00:47 good. Copy image. Okay, let's go back to
1:00:50 chat.
1:00:53 Japeta new chat. Paste.
1:00:58 Okay. So, we're going to
1:00:59 say, I want you to
1:01:04 analyze this
1:01:07 photo and tell
1:01:12 me how many people are in
1:01:16 it and the
1:01:21 percentages of who's
1:01:24 wearing what color
1:01:30 shirt as well
1:01:33 as
1:01:36 percentage of
1:01:38 hats versus
1:01:42 non-hats. All right, that's good. Quick
1:01:45 sidebar.
1:01:58 Wait, I'm
1:02:00 confused. Quick cyber. Oh, Tik Tok
1:02:06 comments. Scroll up to Kyle yesterday.
1:02:10 Okay. Okay. Kyle, yesterday. Okay. Kyle,
1:02:13 yesterday I asked you to do Wait, do
1:02:15 LLMs understand? Today I asked Chat GBT
1:02:19 with everything you know about me, roast
1:02:20 me. It was
1:02:21 astonishing. It became a whole therapy
1:02:24 session for my friend and
1:02:28 I feel they do understand. Okay.
1:02:37 So the question do they understand?
1:02:42 [Laughter]
1:02:47 Hey Kyle. Yeah. Why don't you go ahead
1:02:49 and start your uh analysis while you
1:02:52 Yeah, that's a good idea. This is this
1:02:54 is I feel like this is gonna take Yeah.
1:02:57 And it should this should be cool. This
1:02:58 should should be fun to watch. Okay.
1:03:01 Um the the question of do they
1:03:04 understand is a really interesting
1:03:09 one because mathematically you could
1:03:13 argue
1:03:16 that they're just looking at data and
1:03:19 they're selecting the most probable
1:03:23 data and they don't have consciousness
1:03:26 so they can't
1:03:28 understand but what they demonst
1:03:30 demonstrate is understanding and I know
1:03:33 that sounds weird like are they
1:03:35 empathetic is different than can they
1:03:39 behave with empathy does that make sense
1:03:42 so when you asked yesterday and I said
1:03:44 they don't understand what I was talking
1:03:46 about was are they capable themselves of
1:03:49 understanding like perceiving that they
1:03:52 understand something which is different
1:03:54 than they behave like they understand
1:03:56 and I'll tell you it it is already
1:03:58 freaky how good they are But the fact
1:04:00 that these large language models have
1:04:02 access to everything you've ever talked
1:04:04 about with them now. They absolutely
1:04:07 understand you on a on a deep no they
1:04:11 demonstrate an understanding of you.
1:04:13 They present that they understand you.
1:04:15 Um but I don't think that they they
1:04:18 actually understand it like we would
1:04:19 think about that. So that's that's where
1:04:21 my head was. But yeah, it's freaky,
1:04:24 isn't it? It's it's absolutely crazy.
1:04:27 Um, quick stat. Crowd size about 190
1:04:30 people. All right, that looks about
1:04:33 right. Circle detection on head shapes.
1:04:36 Wearing hats about
1:04:39 40%. Not wearing hats. Color breakdown,
1:04:43 blue,
1:04:45 red. Okay. So, here's what I I wanted to
1:04:48 do. I want you to
1:04:53 um show
1:04:55 me
1:04:57 closeups of the
1:05:01 image where you
1:05:06 have
1:05:10 annotations of what you you're looking
1:05:15 at. to get these
1:05:35 numbers. Network error occurred.
1:05:41 [Music]
1:05:43 Retry. Yeah. No, Kyle, they understand
1:05:45 you.
1:05:49 [Music]
1:05:54 GPT went off the deep for me today. It
1:05:57 got a virus. The user wants detailed
1:05:59 close-ups of the image. What I'm trying
1:06:00 to get it to do here is is do some image
1:06:03 analysis. Oh, one of the things that you
1:06:05 can do, let's see,
1:06:08 numpy p import
1:06:12 image head min R head head min R head
1:06:17 max
1:06:19 R hat min R hat max
1:06:23 R slightly larger radius.
1:06:36 Yes. Oh, look what it
1:06:39 did. Hang
1:06:42 on. Detected heads likely hats. Wow.
1:06:47 Hang on. Look at this. How do I Can I
1:06:50 zoom in on this?
1:07:00 likely
1:07:04 hats. I don't know that it got it
1:07:09 right. Oh, and there's a bunch of
1:07:11 watermarks in here,
1:07:16 too. Huh. But it's doing some sort of
1:07:20 analysis analyzed. Oh, it's doing more.
1:07:24 Top
1:07:25 left, top
1:07:31 right.
1:07:38 Huh? Every head I detected is ringed in
1:07:45 cyan. The larger ones are the ones that
1:07:47 feel like there are hats. Well, I think
1:07:51 it did a pretty shitty job
1:07:54 actually, but it's still cool that it
1:07:56 can do that. A network error occurred.
1:07:59 If you'd like an even tighter zoom on a
1:08:01 specific
1:08:03 spot, well, you could certainly have it
1:08:05 find a where's Waldo, couldn't
1:08:09 you? An attempt was made. Yeah, exactly.
1:08:12 This is This is This falls in the
1:08:13 category of an an attempt was
1:08:16 made. Um, what was I going to talk about
1:08:19 before? I was going to I was going to
1:08:20 not not play with this
1:08:23 anymore.
1:08:32 Um I don't remember what it
1:08:37 was. From a distance you thought it
1:08:39 worked. Yeah, exactly.
1:08:47 Um, let's do let's do we want to do
1:08:49 another thing? Let's see. Um,
1:08:58 complex physics poster
1:09:23 Fusion of multi-resolution seismic
1:09:26 tomography maps using physics informed
1:09:28 probability graphical models. All right,
1:09:32 let's
1:09:34 let's copy this image.
1:09:42 Um, I'm a dumb dumb.
1:09:46 [Laughter]
1:09:54 Um, walk me
1:10:00 through this poster in a
1:10:04 way that I can easily understand
1:10:13 it.
1:10:15 Um, show me closeups.
1:10:21 [Music]
1:10:23 of diagrams and images and explain
1:10:30 them and
1:10:36 uh assume I know nothing. Okay, that's
1:10:42 this could be fascinating.
1:10:57 for showing close-ups. We should
1:10:59 structure it with a title and the
1:11:01 author's at the
1:11:03 top. All right. So, what's it
1:11:06 doing? It's figuring some [ __ ] out. The
1:11:09 image is 2000 by 1867
1:11:12 pixels. Oh, this is cool. So, oh, you
1:11:15 can't see that on Tik Tok, but it's
1:11:16 analyzing the
1:11:21 image. So, this is it using tools,
1:11:23 right? So, it's writing these
1:11:27 tools. Yeah. Oh, look. It it chopped the
1:11:30 poster up into component parts, relayed
1:11:34 them
1:11:35 out on a
1:11:40 page. Sorry, Tik Tok. You can't see what
1:11:43 it's doing. God damn
1:11:49 it. Thought for 52 seconds. The potter
1:11:53 The poster tackles a mouthful.
1:11:55 Translation: We have blurry underground
1:11:58 pictures at different levels of detail.
1:12:00 We're inventing a smarter way to blend
1:12:02 them so geologists get a clearer
1:12:05 picture. Great.
1:12:07 Perfect. Why bother? Oh, you know what's
1:12:11 so
1:12:14 good? I have an
1:12:17 idea.
1:12:24 Um, go
1:12:27 find me
1:12:31 the
1:12:33 transformers
1:12:35 paper by Google.
1:12:41 and teach it to me in
1:12:45 the same Wait, did this break up
1:12:49 pictures for us? It did
1:12:51 not.
1:12:53 Okay. Um, and teach it to me in the same
1:12:56 way where you
1:12:58 take
1:13:02 their geeky
1:13:05 language. I mean that as a term term of
1:13:07 endearment, geeky language and translate
1:13:13 it for for a dum dum like
1:13:19 me. Make sure that
1:13:23 when
1:13:25 you talk about a
1:13:29 concept that has a diagram or
1:13:36 image, you
1:13:42 show that image in
1:13:47 context. Let's see what it does.
1:13:50 So, this should go find the archive
1:13:53 white paper of attention is all you
1:13:57 need. It it it should then rewrite the
1:14:00 concepts in it and show us pictures of
1:14:02 the diagrams and explain them. In
1:14:04 theory, it should do
1:14:07 that. I need to work with images for the
1:14:10 presentation. I need to break it
1:14:13 down. Attention is all you need paper on
1:14:16 archive. So, I found that. I found the
1:14:18 paper and we'll cite it. I found several
1:14:22 diagrams related to the
1:14:30 paper. Since I need proper diagrams, I
1:14:33 can't rely on search results. I'll
1:14:34 switch to an image search. Find the
1:14:36 specific visuals I need.
1:14:50 So, it's actually understood the images
1:14:52 now. That's kind of
1:14:55 cool. Kyle, what's the cheapest way to
1:14:57 create a
1:14:58 video cartoon from text? Oh, look.
1:15:01 There's the attention. Okay. Attention
1:15:04 is all you need. The transformer
1:15:06 degeeekified. Why did Google write this
1:15:08 paper? Old school language models used
1:15:11 recurrent or convolutional layers to
1:15:15 read a sentence one step at a time. They
1:15:18 worked, but they were slow and struggled
1:15:20 with long range relationships. The
1:15:23 authors asked, "What if we could ditch
1:15:25 recurrence entirely and keep only the
1:15:28 part that was doing the heavy lifting
1:15:30 anyway?"
1:15:32 Attention. Okay.
1:15:35 Um, you're still one level
1:15:41 [Music]
1:15:43 two
1:15:46 obtuse. Make it even easier to
1:15:53 understand. And don't judge me, you
1:15:56 stupid
1:15:57 giantrained friaking
1:16:01 genius that's got infinite capacity to
1:16:04 support
1:16:13 humans. What are your best ideas on
1:16:15 creating content for a website using AI?
1:16:20 uh if you just want like the writing of
1:16:21 it just chat GBT chat GBT40 the base
1:16:25 model it's it's remarkable at that what
1:16:28 I would say is this
1:16:31 um make sure you give it enough context
1:16:34 right so if you've got say marketing
1:16:36 documents or uh you've got an existing
1:16:39 website where you like some of the
1:16:41 content but not all of it give it that
1:16:43 website and say here's what I like about
1:16:45 it here's what I don't um maybe you've
1:16:48 got a creat creative brief or maybe you
1:16:51 just you don't have any of that stuff
1:16:53 yet. Um, a really good way to do it is
1:16:55 is put chat GPT into conversation mode
1:16:58 into advanced voice and just talk about
1:17:01 your business and just ramble and it
1:17:04 will just record all that stuff and then
1:17:07 say now give me a website outline for
1:17:09 that. You know, articulate the audience,
1:17:13 who the audience of the website is, what
1:17:15 you want them to learn, what action you
1:17:16 want them to take. give it all of the
1:17:19 stuff. Like if you were sitting down
1:17:20 with a web designer, you'd give them all
1:17:23 the stuff. A good one would ask for it,
1:17:31 right? The one sentence idea of a
1:17:34 transformer. Instead of reading a
1:17:36 sentence left to right like a game of
1:17:40 telephone, the transformer lets every
1:17:43 word look at every other word at
1:17:46 once and decide who matters most. Honest
1:17:50 to God, this is the first time this
1:17:53 sentence, this is the first time I've
1:17:55 ever truly understood what attention
1:17:58 means in that [ __ ]
1:18:01 paper. And it makes perfect sense.
1:18:07 If you read a sentence left to right,
1:18:08 you've got to read the whole sentence
1:18:10 and it takes a long time. But if you
1:18:12 look at all of the words at the same
1:18:14 time and say, "Which words deserve my
1:18:19 attention?" That's pretty [ __ ] clear.
1:18:21 Look at that. I learned something
1:18:22 tonight. Thanks, chat.
1:18:25 GPT, I know I'm on the right path, as
1:18:28 you say, what I've been doing. Great.
1:18:31 Awesome. Beautiful. I use it to rewrite
1:18:34 content for my students on grade level.
1:18:36 That's very cool. Oh, first Oh. Oh,
1:18:38 yeah. That's cool.
1:18:41 Um, you're welcome. You're welcome.
1:18:44 You're
1:18:44 welcome. Meet the two tower stack
1:18:47 carousel number one. So, here it is, the
1:18:50 two tower stack. So, there's our
1:18:58 image. Lower tower encoder. Encoder
1:19:01 reads the input. Decoder spits out the
1:19:05 translation. Okay, here's what we're
1:19:06 going to do. This is going to be fun.
1:19:10 Okay, I want you to
1:19:17 write a short story.
1:19:34 that explains the
1:19:37 core concepts
1:19:42 here in a way that has
1:19:47 entertaining
1:19:49 characters dialogue.
1:19:57 metaphors,
1:20:01 locations,
1:20:06 objects,
1:20:08 plots, and
1:20:13 um
1:20:17 resolutions. Wonder if we can do inline
1:20:19 images.
1:20:21 I'd also like you to
1:20:25 create
1:20:27 images in line in the
1:20:32 story that bring it to life.
1:20:44 Use a
1:20:49 claymation animation
1:20:52 style for the images. I don't know if it
1:20:55 can do this,
1:20:57 but we'll
1:21:01 see. Or did it just make that up and
1:21:03 you're being lied
1:21:05 to? Did it make what up? The
1:21:14 Wyatt crushed the old models. Reads a
1:21:16 sentence the old way one word at a time.
1:21:18 Transformer all at once. Training speed
1:21:22 slow can't
1:21:24 paralyze. Transformer fast matrix math
1:21:27 in parallel. Oh, that that makes sense
1:21:30 now. Why throwing more
1:21:32 GPUs at a transformer trains it faster
1:21:36 because it can do it in parallel. You're
1:21:38 looking at all the words at once.
1:21:40 Ah, see it's all starting to come come
1:21:43 into the [ __ ] noggin. Long context.
1:21:47 The old way doesn't
1:21:50 work. That was what I was taught in
1:21:53 speedreading when I was 13. Oh, that's
1:21:55 fascinating. Maybe, you know what? Maybe
1:21:57 someone at Google, you know, was trained
1:22:01 on speed reading and that's where this
1:22:03 concept came from. That's fascinating.
1:22:05 Like quantum computing.
1:22:07 What's the best app for for doing
1:22:09 cartoon video right now, Mr. It? The
1:22:12 best one the the one that I did
1:22:17 um hang on. I
1:22:21 did this animation
1:22:29 [Music]
1:22:35 in this one.
1:22:45 Oops. This was done in luma
1:22:50 labs. Luma
1:22:53 labis. So, make the image in chat GPT or
1:22:57 midjourney, which I can't believe I'm
1:22:59 saying
1:23:00 that. Like, chat GPT is good enough now
1:23:04 to do
1:23:05 images that, you know, like final
1:23:08 images, not just like stupid proxy [ __ ]
1:23:12 for a LinkedIn
1:23:14 post. Um, where was I? Back
1:23:19 here. Image created. Encoder decoder.
1:23:22 Look at this. Look how cute this is.
1:23:25 Getting started. Oh, it's doing all the
1:23:27 images
1:23:28 [Laughter]
1:23:35 first. I like it. Look at our little
1:23:38 claimation
1:23:39 style. It would be if it weren't
1:23:42 fettered into
1:23:44 uselessness. Yeah. What are you going to
1:23:48 do? What are you gonna do?
1:23:52 Oh, multiple equations at
1:23:57 once. AI minions. Yeah,
1:24:01 right.
1:24:05 Word. All right. So, it's making all the
1:24:08 images. Oh, did it just stop? What's it
1:24:10 doing? Getting started. Doing another
1:24:11 image.
1:24:15 So, this is well, you could sort of
1:24:18 occasionally do this before where you
1:24:20 could get it to make a bunch of images
1:24:21 in a row, but what this is doing is
1:24:25 Okay, so so here's a new thing. Okay.
1:24:30 Anyway, welcome to chat add. I got to
1:24:32 get out of here. It's 10:30.
1:24:39 Um, one of the things that 03 can
1:24:44 do that the other thing that that
1:24:46 something like 40 can't do and 01
1:24:50 couldn't really do is it can do really
1:24:52 long sequences of actions. Maybe that's
1:24:56 the best way to think about it. That um
1:24:59 David Shapiro referred to 03 as
1:25:01 semi-agentic.
1:25:03 that kind of like Manis when when you
1:25:06 watch Manis like log into a website and
1:25:09 then learn from it and then change its
1:25:11 instructions and keep moving. This is
1:25:14 kind of doing that, right? So, right now
1:25:16 I told it I wanted to write a short
1:25:17 story. I wanted, you know, illustrations
1:25:20 to go with the story. It's not doing
1:25:22 them in line, unfortunately, but you
1:25:24 know,
1:25:25 whatever. But it's going to make all
1:25:27 these images and then it's going to
1:25:29 write the story. So with 03, you could
1:25:32 write a kids story with all the images,
1:25:36 right, for every page and it'll just go
1:25:38 do it all and you can just sit back and
1:25:40 let it do it. Why are you in 03? Because
1:25:43 part of because 03 is a completely
1:25:47 different kind
1:25:49 of model than we've ever seen. It's
1:25:53 multimodal, can use tools, Python, write
1:25:57 its own tools.
1:26:00 um you know can see and understand and
1:26:02 make images, use tools
1:26:05 um and then it's it's a a
1:26:08 smart language model. It's got
1:26:11 reasoning multi-step reasoning in in in
1:26:15 uh combination with those other two
1:26:17 things tool use and understanding
1:26:19 images.
1:26:22 Um, and why I'm using it is to
1:26:26 understand how is it different, why is
1:26:28 it different, when would I use it, what
1:26:30 would I use it
1:26:31 for? Because if you're a mathematician,
1:26:34 it's pretty clear that you would use
1:26:36 this because it just maxed out the math
1:26:40 benchmarks. If you're not a
1:26:43 mathematician, do how many
1:26:45 non-mmathematicians have PhD level
1:26:47 problems to solve? You know, not many
1:26:49 are non-scientists,
1:26:52 right? Positional encoding. I love these
1:26:55 little illustrations, by the
1:26:57 way. Wordville and the great gossip
1:27:00 maker. The twin towers of world world
1:27:04 Wordville. Scene one. In the bright
1:27:06 bouncy village of Wordville, two rainbow
1:27:09 stacked towers. Encoder and
1:27:13 decauer. Encoder and decoder loom like
1:27:16 playful layer cakes. Gather round,
1:27:19 hollers maestro former. A clay conductor
1:27:22 and a tiny browser bowler hat. A dozen
1:27:26 clay word
1:27:28 folk shuffle
1:27:32 closer. Inside the first tower, the
1:27:34 maestro decides divides the crowd into
1:27:38 into
1:27:39 trios. Quinn slips Quinn query slips on
1:27:43 ruby red glasses. I must figure out who
1:27:45 I should pay attention to.
1:27:49 Kila Key. Wait, do we have the one with
1:27:52 red
1:27:54 glasses?
1:27:56 No. All right, we're missing
1:27:59 [Laughter]
1:28:09 pictures. All right.
1:28:12 Um, the
1:28:16 story
1:28:17 doesn't
1:28:20 clearly
1:28:23 explain the
1:28:26 paper. It's
1:28:29 now just confusing in a different
1:28:37 way. Close the gap.
1:28:40 Don't
1:28:42 worry about the images just
1:28:53 yet. Again, good from far, but once you
1:28:56 zoom in. Yeah, exactly. Well, this is
1:28:58 one of the things, one of the criticisms
1:29:01 I've heard about 03 is that it seems to
1:29:03 oscillate between being absolutely
1:29:05 horrible and absolutely mind-blowing.
1:29:08 And it seems like there's not much in
1:29:12 between. Uh Kyle, you're the best. Learn
1:29:15 something valuable within two minutes of
1:29:17 turning tuning in tune of of tuning one
1:29:22 and a half hours in. Thank you, Sam
1:29:24 Elliot. Good to see you, sir. Hey, I
1:29:26 learned I learned what a what a
1:29:29 transformer is today. I've looked at
1:29:31 that paper, I don't know, 50 [ __ ]
1:29:34 times. I'm like, I sort of get it. I
1:29:37 sort of get it. And then it just did it.
1:29:39 If you read a sentence left to right, it
1:29:41 takes you a long time. If you look at
1:29:43 all the words at once, you can figure
1:29:44 out what's most important to pay
1:29:45 attention to. Got it. And you could do
1:29:48 that in parallel. There you
1:29:50 go. That's the transformer. There it
1:29:55 is. Absolutely amazing. All right. Did I
1:29:58 get the story right now? Let's see. Are
1:30:00 we done?
1:30:06 The word kids line up to form a foreign
1:30:09 sentence. In old parades, RNN's they'd
1:30:12 march single file. Each kid could only
1:30:15 whisper to the one
1:30:17 behind. Slow and errorprone. Maestro's
1:30:20 new rule. Nobody marches until everybody
1:30:23 knows who
1:30:25 matters. So before we move, every word
1:30:28 kid is allowed to look at every other
1:30:31 kid to see whose meaning helps it ma
1:30:34 make sense. That looking process is
1:30:39 attention. That's really cool. All
1:30:41 right. Every kid is a word in a
1:30:44 sentence. And in the old
1:30:47 model, kid one would talk to kid two
1:30:50 would talk to kid three. In the new
1:30:52 model, they all look at each other and
1:30:54 say, "Who of us is the most important to
1:30:56 make sense of this sentence?" That's
1:30:58 cool. The game of attention scaled. Each
1:31:02 word kid asks two simple questions about
1:31:05 every other kid. How relevant are you to
1:31:08 me? Query key score measures connection
1:31:12 strength. What do you actually say? The
1:31:15 definition of the word, the content that
1:31:17 might be borrowed.
1:31:20 They soften the scores with a divide by
1:31:23 D to softmax step so nobody
1:31:28 shouts. Okay, that one sucks. That needs
1:31:32 to be
1:31:34 rewritten. There's something
1:31:37 here. There's something here. This is
1:31:40 kind of cool. All right, I gotta go. I
1:31:43 got to get out of here. People, you got
1:31:44 to get out of here. Everybody's got to
1:31:45 go to bed. What are you still doing
1:31:47 watching this? How many people are here?
1:31:49 Oh, not that many. 31. There's just 31
1:31:52 insane people here. All right, I'm
1:31:57 leaving. All right. Uh, peace out
1:32:00 everyone.
1:32:01 Um, go play with chat GPT. Hopefully
1:32:05 while I've been doing this, you've been
1:32:06 playing as well. Tomorrow, it's Friday,
1:32:09 office hours at 11:00 a.m. on LinkedIn.
1:32:12 So, if you go to my LinkedIn channel,
1:32:14 Kyle Shannon,
1:32:16 um I've been using the same Google Meet
1:32:20 link for the past two years. So, just go
1:32:22 find one of my previous events. It's
1:32:24 11:00 a.m. tomorrow. And then tomorrow
1:32:26 night, Friday night, date night. All
1:32:29 right. All right, everybody. Beautiful.
1:32:33 All right. Peace out. See you tomorrow.