4/16/2025 -ChatGPT's New Reasoning Models o3 and o4-mini-high: A Deep Dive

Description

Chapters

Transcript

0:02 You
0:06 ready? Are you ready? Yes, I'm ready.
0:10 Are you ready? Yes, I'm ready. I'm
0:15 falling in love.
0:19 [Music]
0:30 [Applause]
0:32 [Music]
0:34 There's been something baby I've been
0:36 trying to
0:41 say. I don't know
0:46 how past and a future now surrounding
0:50 [Music]
0:52 me. surrender to an earth through can be
0:56 [Music]
0:58 found. There's been a little there's
1:00 been a little
1:03 trouble since you came to my
1:06 [Music]
1:10 rescue. And if you like all of the rest,
1:14 I would have quit you long ago, but I
1:17 couldn't do that.
1:20 [Music]
1:22 Oh, tell me now. Women and wine never
1:26 went
1:29 to make a man and make him go as
1:36 hell. I know one night you wish me
1:40 well. But in spite of your
1:43 trying, still going to have to find my
1:47 way through.
1:55 morning. My dream
1:59 remains hanging in the burnt fields of
2:02 my
2:03 [Music]
2:04 memory. Time is a chorus of your
2:08 [Music]
2:12 disdain. Left behind me.
2:16 I'm searching solutions on the double.
2:22 on his two ton
2:28 telephone. My friend says, "You ain't
2:30 got no real
2:32 troubles. There's too many choices that
2:35 you left
2:36 [Music]
2:40 undone." My friend says, "Put an action
2:42 to those words." Put an action to those
2:45 words. They're
2:51 all tell me now
2:55 women never went too well. Champy sounds
2:59 like he's been watching the news. You
3:01 know what? I have not been watching the
3:03 news. So, I am in ignorant bliss. I I
3:06 just assume it's all burning
3:10 down. There's lightning on fire. So,
3:13 here we go.
3:16 Right about the time they light it on
3:18 fire, AGI shows up. Oh man. All right. I
3:24 made a website using Claude and Chat GPT
3:26 that users can
3:28 learn
3:30 about Hang on. Christian music
3:34 composers. Very cool. That's very
3:39 cool. You know, it's it's really funny.
3:42 We're gonna
3:45 [Music]
3:50 There's going to be like there's just
3:51 going to
3:52 be some amazing things are going to
3:55 happen
3:57 and there's going to be a lot of
3:59 [ __ ] a lot of crap out there
4:02 um and a lot of
4:03 noise. But what's going to happen is
4:06 people who are very passionate about a
4:08 very specific
4:10 thing because they're going to be able
4:12 to wield such powerful tools are going
4:15 to be able to make really incredible
4:17 resources, applications, games, music,
4:20 all that sort of stuff for their little
4:23 specific niche. And I I think just like
4:28 I've talked for two years now that I you
4:30 know we're about to enter the great
4:32 renaissance.
4:34 um kind of Renaissance
4:36 2.0. Um and
4:39 uh you know that that's a really good
4:42 example of it, I
4:45 think. You know, something you're
4:47 passionate about, you put it out there,
4:48 bang, done. Love it. It's very cool.
4:52 It's very cool, people. I like it. I
4:55 like it a
4:59 lot. at 88
5:02 day.
5:06 Um, who's that? Silverf Fox sharing the
5:09 live with her friends. Thank you so
5:10 much, Silver
5:12 Fox.
5:14 Um, tonight we're going to be digging
5:17 into chat jpai.
5:20 You can make money with JP. You can
5:22 [ __ ] solve math with chat GPT.
5:26 Um, we we're at a we're at
5:29 a there's some crazy [ __ ] You know
5:31 what's
5:33 funny? I put out um I put out a video on
5:37 Tik Tok I put out a couple of videos on
5:39 Tik Tok today. I was a little productive
5:40 beaver.
5:44 Um but the last one I did at the end of
5:46 the day was um you know get the toe to a
5:49 nunnery. Get your ass to chat GPT and
5:53 start playing. Start playing with these
5:54 new tools. do it today. It's no longer
5:56 optional. I kind of [ __ ] up that part
5:58 when I was getting in the car. I didn't
6:00 quite remember what I was
6:02 saying, but um some of the comments are,
6:08 "Wait, so it's 03 and
6:11 04, but that's better than
6:15 4.5?" It's like, oh [ __ ] Yeah, cuz the
6:18 naming conventions are so [ __ ] up that
6:21 people are completely confused, right?
6:24 They launched 4 40 4 then then 03 then
6:28 4.5 then 4.1 to the API but they still
6:33 announced 4.1 after 4.5 and then today
6:38 03 and 04. So if you're not paying
6:40 attention if you didn't watch my video
6:42 explaining the naming conventions and
6:45 memorize what I told
6:47 you then you may be stuck in this
6:49 situation where you're like well I
6:50 thought 4.5 was the best. And if you
6:52 were off playing with
6:55 4.5 thinking it was the best and not
6:58 really understanding what a reasoning
6:59 model
7:01 was, these new reasoning models that
7:03 came out today are going
7:06 to shock
7:08 you. They're going to shock you at what
7:11 they can do.
7:13 [Music]
7:23 Hey Pate, I'm curious. I just saw that
7:25 you came in. I'm curious if you played
7:27 with 03. I think I understand, but no.
7:31 I'm curious if you played with 03 or 04
7:33 mini or mini
7:36 high because you know the maths. I don't
7:39 know the maths, but David Shapiro seems
7:42 to think they have solved
7:43 math, which is I that feels hyperbolic
7:47 to me,
7:48 but 99.5 on one of the math benchmarks
7:52 is pretty pretty [ __ ] pretty [ __ ]
7:54 close to 100.
7:57 [Music]
8:36 Oh, you don't really do anything with
8:37 chat GPT. Yeah. Oh, and then and then uh
8:42 uh Google today said they're going to
8:44 they've got a new something coming out
8:46 that's going to So, so what happened
8:48 today? So, Gemini 2.5 Pro was at the top
8:52 of the the comparison leaderboard,
8:54 whatever that leaderboard thing's
8:55 called. Um and then 03 and 04 Mini just
8:59 topped that leaderboard. And then and
9:02 then someone at Google said, "Yeah,
9:04 we're coming out with a new one next
9:05 week. It's going to be better. So, the
9:08 the com the the wars are hit heating up
9:12 people. Math benchmarks are tricky
9:14 because you it's usually in the training
9:16 data for a while. Yeah. Yeah, I
9:19 know. But still going from like 70s, you
9:24 know, high7s to 99.5 on that one
9:27 [Music]
9:28 benchmark. But I understand your point.
9:31 the the thing the thing that that's
9:33 really different. I don't know if you
9:34 saw the announcement, you know, the
9:35 stream today, Pate, but um probably not.
9:40 Um they were talking about one of the
9:44 one of the things they had it solve it
9:47 did 600 tool calls. So in the new model
9:52 it it will spin up um little Python uh
9:56 windows and write code, execute it and
9:59 then use that in part of the reasoning.
10:01 And they said it for one one
10:03 particularly hairy problem. It it you
10:06 know it invoked a tool more than 600
10:09 times which is pretty
10:11 staggering and then it got it right.
10:15 So So we're going to go play with that
10:18 today. And it's also it's they're also
10:20 multimodal. So they do the image
10:22 generation stuff. I was playing around
10:24 with some of that. They understand
10:27 images. They will do things like
10:31 um split images up to analyze parts of
10:35 them.
10:38 Um it's pretty crazy. So we we're just
10:41 going to go play. I I haven't played
10:43 enough to know what's really in there.
10:45 So we're going to go do that.
10:47 There's probably other things that we
10:49 should be looking at, but this this
10:51 feels like a pretty
10:52 significant situation. Situation
10:57 [Music]
11:02 situation. I know what I'm going to do.
11:04 Hang on a sec.
11:07 [Music]
11:10 And my the the other kind of indicator
11:12 that this is a big deal is
11:16 my my Twitter timeline like every single
11:21 post of anyone that I you know know or
11:25 respect with you know points of view on
11:28 AI like every single one is saying this
11:31 is a big
11:32 deal. It's it's like without exception.
11:35 Normally, normally it's like, you know,
11:37 I think I saw one that was
11:39 like04 mini sucks because it's not good
11:42 at creative writing. So, that very well
11:45 may be the case. But,
11:49 um, what am I looking for?
11:57 [Music]
12:07 [Music]
12:12 Come
12:13 on. Must be
12:24 [Music]
12:26 a
12:28 media.
12:29 Here's
12:32 the
12:35 Okay, this is where we're going to go. I
12:38 figure we'll start out with
12:40 this. So, as you know, I'm a fan of this
12:43 guy named David Shapiro. I know I know
12:46 Pate is not a super fan of his,
12:49 but the the only two people that I know
12:53 well enough to like respect their
12:56 opinions about math are David Shapiro
12:58 and
13:00 Payton. I don't I don't run in math
13:03 circles. Let's just say let's just put
13:06 it this way. at Penn State, the theater
13:09 building must have been, you know,
13:12 geographically opposite of wherever the
13:14 math building was because I don't think
13:17 I ever met a mathematician at Penn
13:21 State. Anytime I would say it would not
13:24 be hyperbolic to say I know they're
13:25 about to say something hyperbolic. It's
13:27 the one lesson I've learned in life is
13:29 is when someone tells you what they're
13:32 not, it is
13:35 absolute fact that that's who they are.
13:38 Listen, I'm not going to be an [ __ ]
13:40 but followed by
13:44 [ __ ] Did Theater X on the X account
13:47 engage with the dog Kyle? Nope. I got no
13:50 love from the uh Penn State theater
13:52 social media people because they're los
13:55 theater people. What do you expect them
13:57 to be good at social
13:59 media? Um all right. What am I doing?
14:02 Oh, I'm ch I'm sharing. Calm down. I
14:05 know. Listen, I know that I am a little
14:09 slower than
14:10 most. I understand
14:12 that it just resized my window. Why does
14:16 it do that?
14:18 Good
14:20 lord. Okay, black
14:23 bar. We got our me. I got
14:26 it. Producer Brandon is particularly
14:29 snappy tonight. I think he had a little
14:31 extra cup of coffee in between. So, this
14:34 is this is this is producer Brandon's
14:37 third live stream with me today. Twice
14:40 as participants or as as audience
14:42 members and tonight as the producer. So,
14:44 I think he's a little sick of me. He's
14:46 jumping in there. All
14:50 right, let's watch Yan
14:56 [Music]
15:02 video. Yeah, those tend to perform way
15:05 worse on unseen math
15:07 problems. But let's see. 01. So, it's
15:11 the it's the AIM 2025 and the AIM
15:15 2024. And the the model that did really
15:17 good was 04 mini with tools. And what
15:21 was fascinating today is they talked
15:26 about they talked about these these
15:29 models having access to
15:31 tools. And I I guess how
15:36 they're how they're defining tools in
15:39 this context is multiple Python
15:41 libraries. So it's basically chat GBT
15:45 reasoning engine with Python and it's
15:49 multimodal. And it was interesting Greg
15:52 Brockman even said in this conversation
15:55 he said
15:56 um he goes you know what's really
15:58 remarkable? He goes, "This is still just
16:02 um next token prediction, but what
16:05 they've done is they've put in this
16:06 multi-step reasoning and they've given
16:08 it access to to write Python and then
16:11 all of it is just next token prediction,
16:14 but it's doing it in this in this, you
16:16 know, thoughtful sequential way where
16:19 it's improving on itself within a within
16:22 its thinking, I guess, or within its
16:24 answering process."
16:27 Um, and they've gotten it sort sort of
16:30 fast, efficient, and it's it's pretty
16:33 cool to watch, too.
16:36 Um, all right. So, let's let's get
16:40 going. Yeah, that's my guess based on
16:41 how you described it. Python calls or
16:44 subprocess calls. Yeah, it's a lot of
16:46 Python calls. And what's cool, P, I I'll
16:48 show it here in a second.
16:50 Um, you know how it used to be, um, if
16:54 you if you had on a code interpreter, it
16:57 would start answering and then it would
16:59 say whatever process analyzing and then
17:01 you could open the window and you'd
17:02 watch the the Python write itself. Well,
17:05 it's now showing all that work and and
17:08 it's showing if it's doing image
17:10 analysis, it's showing that it's doing
17:11 image analysis. Um, and so you're seeing
17:14 it sort of think, process, get the
17:17 answer, think some more, process some
17:19 more, think some more. It's pretty cool.
17:21 So anyway, so let's let's see what Dave
17:23 has to say here. That OpenAI just made
17:26 history. Now, to state that in as plain
17:30 and clear and simple terms as possible,
17:33 I want to call your attention to just
17:35 two numbers on their 03 and 04
17:38 performance, specifically 04 mini with
17:41 Python. Now this is 04 mini not 04 full.
17:46 So on the AIM or AIM 2024 competition uh
17:51 math benchmark they got
17:54 98.7% with 04 mini and Python tool use
17:58 only. I want to just I just want to
18:00 respond to Pate. Pate says sounds cool.
18:02 I think we're getting to the point where
18:03 it's all just preferences at this point.
18:06 I agree. I this is something that's been
18:09 kind of flying around my head for the
18:11 past 3 weeks or
18:12 so
18:14 that in in essentially every instance of
18:19 output whether it's video, music, large
18:22 language model, coding, math,
18:26 um any tool that does any of the things
18:29 that we're trying to get it to do at
18:31 this point are so good that you can
18:34 effectively pick any one and then I
18:37 think I think to your point the ones
18:39 we're going to choose are the ones
18:41 that's like yeah that feels like me
18:43 right if if you sort of imprinted on
18:45 Gemini you're going to do Gemini if you
18:47 imprinted on Chad GPT or you like
18:49 Claude's personality I I think that's
18:52 that's ultimately what the tools we're
18:54 going to use and it and you know it's
18:56 funny even though there's all of this
18:59 stuff happening and there's 12 different
19:01 video tools and
19:02 then this is going to be just like any
19:05 other technical ical um birthing, right,
19:10 where you're going to have dozens and
19:12 dozens and dozens of major players at
19:14 the phase we're in right now. And so
19:16 keeping up with it is impossible, right,
19:18 that we're
19:20 experiencing. And
19:22 then people will start to imprint as as
19:24 the tools start to stabilize and and
19:27 start to, you know, get good enough to
19:29 use for everything we're going to do.
19:31 Like the fact that Open AAI was the
19:35 first one to give access to all of your
19:37 previous chats, that's a reason to use
19:41 it. Even if you might like Claude's
19:43 personality a little better, but you've
19:45 got all this history in chat GBT, you're
19:47 like, well, I'll just do this one in
19:48 chat GPT. And at some point, what'll
19:50 happen is it'll just consolidate down.
19:52 All of the all of the sub players will
19:55 either get acquired or they'll just
19:56 they'll just peter out because their
19:59 usage will drop off. Um, now how long is
20:02 that going to take? I don't know. That
20:03 normally
20:04 takes 10 years. Seven to 10 years. With
20:08 AI moving as fast as it is, that might
20:11 hap happen in, you know, three to five.
20:14 But but we'll see. Fascinating.
20:17 Um, I felt like when I made that
20:21 website, any model can do HTML coding.
20:23 Yeah, exactly. Yep. Exactly. Okay, let's
20:26 watch Dave late on AIM 2025. They got
20:30 99.5%. Now again this is with 04 mini
20:33 with Python only. Now you might say okay
20:35 you know Sam Alman said we're going to
20:37 saturate the benchmarks and you know
20:38 they delivered on
20:40 that. Now a benchmark is one thing but
20:43 what is a benchmark a proxy for? In this
20:45 case it's a proxy for mathematical
20:48 performance. Now I as a human being will
20:51 never even get to 01 level mathematics.
20:54 Uh so there is that. But what I want to
20:56 do is I want to show you the reactions
20:58 of some other people and then I'll kind
21:00 of unpack my full reaction as to why
21:02 these numbers are just astounding and
21:05 incredible. So the first reaction I want
21:07 to show you is Aiden. So Aiden is an
21:10 OpenAI employee. He's a really chill
21:12 guy. Um he and I talk uh not all the
21:14 time, but like we exchange messages and
21:16 like he's just a super chill dude. But
21:18 anyways, one of the things that he says
21:20 is ignore literally all the benchmarks.
21:22 The biggest 03 feature is tool use. Of
21:24 course, it's smart, but it does deep
21:26 research quality in 30 seconds, debugs
21:28 by googling docs and checking Stack
21:30 Overflow, writes whole Python scripts in
21:32 its chain of thought for firmy
21:33 estimates. It does a lot. So, one of the
21:35 things that they mentioned in the live
21:37 stream announcement of 03 and 04 is that
21:40 in some of the tests, it did 600 tool
21:42 use calls in a row. Now you might say,
21:45 well a human could probably do it in
21:47 fewer, but the fact that it was able to
21:49 keep track of a task that complex means
21:52 that it is more and more agentic. So
21:54 that's reaction number one. And that's
21:57 important, right? What he's saying. So
22:00 the the the reasoning models, you give
22:03 them your question and they talk amongst
22:06 themselves and you know do prompt prompt
22:08 prompt prompt prompt internally and then
22:10 answer. This is now doing prompt tool
22:14 use, prompt tool use, web search, tool
22:16 use, prompt, right? So, it's using
22:19 multiple things and it's keeping track
22:21 of all that. So, what that starts to
22:23 look like is you give it an increasingly
22:26 complex problem and then you basically
22:28 just sit back or just go to another tool
22:30 and do something else. Go make some
22:32 images in midjourney while it slogs
22:34 through all the stuff. Uh, reaction
22:36 number two comes from Daria. So, Daria
22:38 is big on Twitter. Um he's a researcher
22:41 of chronic fatigue and and immune
22:43 systems. He's actually helped me with
22:45 burnout. So Daria is a really good guy.
22:47 Um he also often gets early access to
22:49 open AI stuff. Um Sarah, which which app
22:54 chat GPT or PO? Not which one to not
22:56 sure which one to use or spend money on.
23:02 Okay. The advantage of PO is that you
23:05 have access to lots and lots and lots
23:07 and lots of models.
23:12 Um, the disadvantage to Po is that,
23:20 um, I think they may have their own
23:23 model, but I'm not sure. I think they
23:25 do. Um, they're not as well funded.
23:28 They're they're they're
23:30 they're more like a convenience app than
23:33 than a core frontier lab. Um, if you
23:37 want to be able to play with, you know,
23:40 all of the latest models from whether
23:43 it's Gemini models, llama models, open
23:46 AI models, anthropic models, if you want
23:48 to use them all in the same interface,
23:51 um, PO is an okay way to go. I
23:54 personally would go with chat GPT
23:57 because what they what they launched
23:59 today is staggeringly powerful and what
24:03 they launched on March 31st Image Gen is
24:07 staggeringly powerful. So if you were
24:10 going to spend 20 bucks a month for me
24:12 I'd go with OpenAI. Um so it looks like
24:15 he had this this post on deck. Um but
24:18 anyways he's he is a power user of these
24:20 things. He he actively uses these tools
24:23 to help him with his research and I can
24:26 verify that by using these reasoning
24:29 tools uh 03 mini 01 pro and those sorts
24:31 of things I have gotten I'm not going to
24:33 say that I'm above his level but I can
24:35 talk to him at near his level on things
24:38 like burnout chronic fatigue and immune
24:41 related issue. I was on a uh I was in a
24:44 Twitter spaces last week or over the
24:46 weekend um and there was a a Yale um
24:51 neurosurgeon who was in there and he was
24:52 talking about 01 pro, right? So that's
24:57 not as good as 03 and 04 mini um and he
25:00 was saying for him it you know it was
25:04 absolutely above his capability. He was
25:07 he said he was with 01 Pro he was
25:11 submitting a patent application a month.
25:15 Um and he said he you know he said we're
25:17 we're we are in a you know we've passed
25:20 the tipping point according to him. He
25:22 you know he he claimed it was AGI but it
25:25 depends on
25:26 the on which definition you go for. Sam
25:30 Open AI is no longer compute constrained
25:32 after Microsoft lost its exclusive cloud
25:35 provider status. Oh, that's fascinating.
25:37 Yeah, fascinating
25:39 issues, at least the ones that I've that
25:41 I've uh studied. And what I the that's
25:45 not that's not really a flex. What I'm
25:46 saying is these models are smart and
25:48 they can teach you very quickly and
25:50 they're operating at the level where
25:52 people like Daria, who is a worldleading
25:54 expert in this kind of stuff, is also
25:56 impressed. And by the way, he was
25:58 impressed with 01 and 03 Mini. Now
26:00 they've taken it to the next level.
26:02 Yeah. So, those are two other reactions
26:04 that I wanted to share before sharing
26:06 mine. And my reaction is here. So, I'll
26:09 just read this to you real quick because
26:12 it'll come out better. So, yeah, if you
26:14 go to his ch channel, he's got this this
26:16 post is pinned on his channel. It's Dave
26:22 Shapi. Dave Shappy. So, what I said is
26:25 AI has solved math. Open AAI did it with
26:27 04. not as close to solving math, not as
26:31 competitive at math solved. This is far
26:34 bigger than anyone realizes. Let me
26:35 explain why. First, you need to
26:38 understand some historical context.
26:40 Typically, with a IML, you know that
26:42 you're getting close to fully
26:43 generalizing a problem space when you
26:44 get to the 70% and 80% solution range,
26:47 which real quick, that's where we were
26:49 with the previous generation. in eight
26:52 months. I want to I want to emphasize
26:54 that these models came out in September
26:57 2024. In eight months, they closed the
27:00 gap. So, just keep that in mind. So, uh
27:03 however, as we often see, reality is in
27:05 the edge cases, meaning that the last
27:07 mile jump from 80% to 99% is often far
27:10 harder. It usually takes much much
27:12 longer to get from like from 0% to 70%
27:17 when you're starting with a new AI
27:18 technology than it takes to get from 70%
27:21 to 99%. They got to
27:23 99.5% in 8 months. But OpenAI did that
27:27 not in years but in months. Remember 01
27:29 and 03 were announced in September of
27:31 last year. It's been just over eight
27:32 calendar months and they closed the gap.
27:35 Just from a research and development
27:36 perspective, this is a remarkable
27:38 velocity. But I'm not talking about
27:39 benchmarks. I'm talking about real world
27:41 implications. This puts a world-class
27:43 mathematician in every pocket on every
27:45 team. Do you know what math underpinn?
27:48 This puts a worldclass mathematician in
27:51 every pocket on every
27:53 team. It's like, wow.
27:56 Including your little single startup,
27:59 including, you know, the 10erson company
28:02 you're a part of, right? Pretty much
28:04 everything. The first order consequences
28:07 of a semi-agentic AI system that has
28:09 conquered math is pretty obvious.
28:11 Anything that requires math, it can
28:12 likely solve on its own or with very
28:14 little d redirection. For for example, a
28:17 good friend of mine is in CFD,
28:18 computational fluid dynamics, which is
28:20 used for uh oceanography and meteorology
28:23 extensively. He's been using these
28:25 reasoning models since they came out uh
28:27 and they were helpful to him, but he's
28:29 they still needed an expert guiding
28:31 hand. These might these new models might
28:33 nuke that. Second order consequences
28:36 otherwise known as downstream impacts
28:38 from this are difficult to predict but
28:40 not difficult to
28:42 overstate. Let's put a second order
28:44 consequence let's
28:46 put let's put second order consequences
28:48 in practical terms. This will accelerate
28:51 AI research itself. AI research is among
28:54 other things math. It's also code. Guess
28:56 what? These models crushed math and
28:58 code. Beyond that they are
29:00 semi-aututonomous i.e partially agentic.
29:03 They require less human direction,
29:05 correction, and oversight. In practical
29:07 terms, it means they can use more tools
29:09 without help, work on larger, longer
29:11 problems without supervision, and are
29:13 less likely to make mistakes around user
29:15 intent. Guess what else is math
29:16 intensive? Biochemistry, robotics,
29:19 spaceflight, cryptography, nuclear
29:21 physics,
29:22 blockchain. So, this is right. This now
29:27 starts to be the thing where
29:30 advancements in those um in those
29:33 verticals start to accelerate
29:35 blockchain. All of this has just been
29:37 solved. Now take this uh now to make
29:39 this even more impressive these models
29:41 did this with one tool Python. Not
29:44 series of tools not mat lab not
29:46 supercomputers Python. Now let me
29:49 underscore what this means in the long
29:50 run. Your smartphone will be if you
29:51 don't know what Python is it's a it's a
29:53 programming language. And so when these
29:55 things are reasoning, if they need to do
29:57 math, they will spin up what's called a
30:00 container
30:02 um which is like a little, if you've
30:03 ever heard of Docker, it's think of it
30:05 like a little computer instance running
30:07 in a window inside the chat. It'll write
30:10 Python code, execute it, get the answer
30:12 from that, and then continue on with its
30:14 reasoning. Be a math genius before too
30:16 long and a coding genius and a lingu and
30:18 a linguistics genius. and and and third,
30:22 fourth, and fifth order consequences of
30:23 this technology alone are impossible to
30:25 overstate. And this technology will only
30:28 get better. You know that scene where
30:29 Tony Stark is figuring out time travel
30:31 in his kitchen? Yeah, that's the level
30:33 of AI math we're talking about in just
30:34 one or two more generations. If warp
30:36 drive is possible, these machines will
30:38 help us figure it out. So, I hope that
30:41 this gives you some context as to what
30:43 this means. Thanks for watching. All
30:46 right, enough Dave Shappy. All right, I
30:50 gotta change my tabs or you can't see
30:52 [ __ ] So, what we'll do here is I'm just
30:56 going to start playing. Now, what you
30:58 need to know in case it's not patently
31:01 obvious, I am not a
31:07 mathematician. Sir, may I inquire? What
31:10 were your qualification? What pretail
31:12 are your qualifications? Sir, talking
31:15 about mathematics and
31:17 superhuman approximation, I wonder what
31:21 what
31:22 what your qualifications make such a
31:25 statement,
31:29 sir. Uh yeah, I got none.
31:33 But we can do cool [ __ ] with pictures
31:36 and storytelling. I I did have a fun
31:38 idea. Oh, actually, you know what? Let's
31:40 just do it. We'll start with
31:41 it. Um
31:44 Okay. So, so if you're new here, I'm at
31:47 if you're Sarah, if you're still here,
31:48 I'm at chat
31:50 GPT. So, one of the things you get when
31:52 you get chat when you have chat
31:54 GPT is you get all of the models. Now,
32:00 people are making fun of chat GPT as
32:02 they should because the naming
32:03 conventions here are [ __ ] insane.
32:07 But here's all you really need to know
32:09 know
32:11 um f for
32:14 99% of what most people are working on
32:18 the first option is fine. It's fine.
32:21 It's fine. It's good. It's really like
32:23 it's really
32:24 good. Um archetypal architect dropped
32:28 his 50th book
32:30 tonight. That must have hurt. Did it hit
32:34 your foot?
32:37 Oh. Oh, you mean you you published your
32:39 50th book. That was it was it was
32:42 publishing
32:43 comedy. Get it? Because you dropped
32:47 it. Like a book like if it was a heavy
32:49 book and it fell on your toe. That So
32:52 that's where I This is also kind of a
32:55 comedy channel, like an entertainment.
32:56 So, I like to entertain while I'm I
32:59 well, I can't really teach because I
33:00 don't have qualifications, but and I'm
33:02 not really funny,
33:04 but anyway, I just because that was the
33:07 funny part, the the joke part of the
33:10 dropping was the gravity. Well, if it
33:13 hit your
33:16 [Music]
33:24 tongue, the comedy gets so much better
33:26 when explained.
33:28 Exactly. Exactly. Those comedians that
33:32 think that jokes are okay to just stand
33:34 on their own, I've never understood
33:36 them.
33:38 Anyway, all right. Fantastic. So in this
33:42 list the first three models are not
33:45 reasoning models. So what they are is
33:48 you give them a prompt they give you an
33:51 answer. Reasoning models you give them a
33:54 prompt and they go huh the user asked me
33:57 for this. Let me think about that. I
33:59 wonder if I should go search the
34:01 internet. Let me go search the
34:03 internet. It gets back answers. It's
34:06 like, well, now based on what I've
34:08 learned, I had made this assumption, but
34:10 now I do that assumption. So, they're
34:11 they're literally kind of conversing
34:14 with themselves the way you talk in your
34:16 crazy straw of a brain about stuff and
34:19 work through problems. That's kind of
34:21 what they're doing.
34:23 Um, so 03 is the is the
34:29 um the the uh the evolution or the the
34:33 next big model. Uh after 01, you're like
34:39 what happened to O2? It was a trademark
34:40 issue. They just went from 01 to 03.
34:43 Then 04 mini and 04 mini high are the
34:47 are the Okay, how do we do this?
34:52 Okay, if this is 03, a box this size is
35:00 03 like in terms of like compute
35:04 required and size to keep it in memory
35:06 and size to keep it on hard drives and
35:09 amount of compute required to answer one
35:11 of your inane questions about I don't
35:14 know fileo dough.
35:18 That's that's the that's the the base
35:20 model, the big model. And then they make
35:22 these mini models which have less
35:24 parameters. This one's probably got a
35:26 trillion parameters or 100red trillion,
35:28 doesn't [ __ ] matter. And this one's
35:30 got some less billion, right? And so
35:33 this is the mini model. So what we've
35:34 got is we've got 03 MacDaddy and then we
35:38 have 04 mini and then 04 mini high just
35:40 basically takes more time. It's this
35:42 model just thinking longer. Okay, so
35:45 that's how they work.
35:47 Um what's different about these models
35:51 than previous reasoning models? Previous
35:53 reasoning models were not
35:56 multimodal and I don't I don't think
35:59 they had access to to the Python the the
36:02 the tool use. Um so they were just
36:05 thinking along. So, so what the what
36:08 these are is very very um optimized,
36:13 fast, cheap, incredibly capable models
36:17 that can use tools and are multimodal.
36:20 Multimodal just means it it's been
36:23 trained on images as well as words and
36:26 then audio. So, it can do voice. So,
36:28 it's it's a it's a very sophisticated
36:30 model. Okay. And so what we're going to
36:32 be playing with tonight is 03 and 04
36:35 mini
36:37 high. All right. So let's
36:40 see. Let's go with
36:44 03. And then I think what I'm going to
36:46 do Well, let's see. No, I'll do I'll do
36:49 um a thing that I did before which was
36:51 kind of fun. Hang on. Is this
36:55 it? No.
37:05 [Music]
37:11 which yeah, this is it. Okay. So, I'm
37:14 going to copy this
37:16 image. So, let me show you this
37:20 image. See that
37:22 image? I went to
37:25 Google and I said
37:29 um I searched for complicated image and
37:33 that came up. It's a
37:34 doodle, right? Like some dude or dudet
37:38 got stoned and you know this was the
37:42 [Laughter]
37:45 weekend. All right. So then I said, what
37:49 did I do it in? I think I did it
37:51 in in
37:55 uh Oh, I did it in 03. Okay. So, so
37:59 let's do a new We're going to do new
38:01 chat. We're going to do 03. And then I'm
38:05 going to say I want
38:08 you
38:11 to
38:14 catalog
38:17 any item you can
38:22 identify in this drawing.
38:28 And I want you to
38:32 create
38:35 a
38:37 hex color palette.
38:48 um that
38:51 features the
38:54 10
38:57 most
38:59 prevalent colors in the
39:03 drawing. Okay.
39:08 So, so in order for it to analyze the
39:11 drawing, it's going to have to zoom into
39:13 certain pieces of it and look at it and
39:15 analyze it. It can understand drawings
39:17 because it's natively it's multimodal.
39:20 It can understand images. And then I
39:22 spelled pallet wrong. Is that two L's in
39:27 pallet? I guess so. No. How do you spell
39:30 pallet? Oh, one one L1 T palate. Fine.
39:37 Um,
39:41 prevalent. Jesus.
39:45 Okay. Um, and then identifying
39:49 like it's got to do an analysis of the
39:52 whole
39:53 image, figure out the percentage of all
39:56 the colors in all the pixels and sort of
40:00 condense them down into the top 10. And
40:02 that's going to take some Python. So, we
40:04 should get some some good action here
40:06 where we can watch it do it. Bring out
40:08 the V vin diagram. Yeah, exactly. Okay,
40:12 here we go. Let's see what this thing
40:14 does. And I'll expand the thinking when
40:17 it starts doing
40:20 it.
40:25 [Music]
40:30 Thinking GBT, what's the matter?
40:33 Everybody using
40:35 you? The user provided an image. For the
40:38 image
40:40 analysis, I'll use Python. Look. So
40:42 there it is. So So in line of it doing
40:45 the conversation, it's now spun
40:47 up Python. So it's imported some
40:51 libraries. Um it's here's the image
40:54 path. It's analyzing it. It's doing some
40:56 arrays, some image ni shape something or
41:01 other shape of pixels. It's using a k
41:04 means of 10 clusters. So I guess k means
41:07 is how they're going to do the analysis
41:08 of the colors.
41:10 It seems like there might be an issue
41:11 with the ninit auto setting in an older
41:14 vision version of
41:19 scikit-learn. Um, let's see. Let's
41:21 proceed. Wait, this should work smoothly
41:23 without causing any errors. It's I guess
41:26 it's still thinking, but I don't know.
41:28 Maybe it might have got stuck. So, we'll
41:33 see. It's creating its own app to get
41:35 the answer. Yeah, it's it's creating
41:37 multiple apps. So it's actually what
41:39 it's actually doing is it's using Python
41:42 to write a tool. So it's basically
41:44 saying, okay, the user wanted some
41:46 colors. I know from because I'm smart
41:48 that we could use this K whatever it
41:50 called it, the K K means clustering. It
41:54 knew it could use K means clustering.
41:56 And so what it does is says I've got to
41:58 write a K means clustering tool. And so
42:00 it'll write that in
42:02 Python. Now it does look like it's
42:04 stuck. Let me stop it here and let me
42:06 just say um try
42:09 again. We'll see see what it does. It it
42:12 that that might have been a premature
42:14 abort abort but you know
42:17 whatever. Yeah, it's creating its own
42:20 app to get the answer. Yes,
42:21 apps skarn cluster import k means import
42:26 pandas as pd.
42:30 So, so it's doing some
42:41 stuff. Just
42:43 nuts. It's thinking. It's doing whatever
42:45 it's doing.
42:57 [Music]
42:58 And if you wanted to learn code, you go
43:00 figure out all the things it's doing in
43:02 its little
43:04 tools, which you know when you watch it
43:07 when you watch it write code, you're
43:08 like, I don't want to learn that [ __ ]
43:10 Um, does anyone know why AI movie why AI
43:14 is moving so extremely fast?
43:22 Um part of it part of it is because
43:27 um okay so the so a couple of
43:34 things almost all of the
43:37 AI the the consumer AI large language
43:40 models and the tools that we're using
43:42 today are leveraging the same core
43:46 technology. It's a thing called the
43:47 transformer that Google invented in
43:51 2017. One of the attributes
43:56 of how the transformer works is that the
43:59 more data you throw at it and the more
44:02 compute you throw at it, the more
44:04 thinking time you get give it, the
44:07 better it gets. Sam Alman was the one
44:13 that sort of got this idea first and
44:17 biggest and said, "Hey, if I can, you
44:20 know, raise enough money, we can throw
44:24 we can go get all the data and and throw
44:26 all the compute at it, right?" And so he
44:29 cut a deal with Microsoft
44:31 um to be able to use their big data
44:33 centers and you know whatever the
44:36 initial investment was turned into a
44:39 billion dollar investment and then that
44:40 turned into a $10 billion investment.
44:43 They recently just closed a $40 billion
44:46 investment on a on a valuation of I
44:49 don't know $320 billion. So these
44:53 relatively
44:55 um low employee count companies are
44:58 raising billions of dollars. They're
45:00 spending most of that money on GPUs on
45:05 the things that are that are making
45:07 these
45:08 models. Couple other things that that
45:11 are that are why there's so much money
45:13 here. So why the acceleration is
45:15 happening is because there's so much
45:16 money. Why is there so much money here?
45:22 AI's been around for ages, right? AI's
45:25 been around for 30, 40, 50, 70 years,
45:28 you know, depending on on your
45:29 definition of it, neural networks and
45:31 things like that, for a long, long time.
45:34 Um, AI historically has been very
45:38 specific to a specific kind of data and
45:40 a specific kind of use case, right? So,
45:42 if you want to do map analysis, you
45:45 train it on maps with, you know,
45:47 different kinds of AI. The transformer
45:49 kind of unlocks this
45:51 generalpurpose capability if you use it
45:54 right. Sam Alman figures this out and in
45:58 2020 comes out with GPT3. So they did
46:01 GPT1, GPT2. When GPT3 came out, this was
46:05 the first thing that started to act like
46:08 what we're used to seeing in chat bots
46:10 today.
46:12 And then November 30th, 2022, they take
46:15 that basic core technology GPT3. They've
46:18 got the next version which is
46:20 GPT3.5. And what they did is rather than
46:23 having an engineering interface which is
46:25 designed for engineers to use the tool,
46:29 they launched chat
46:31 GPT which was the first consumer
46:35 interface for AI and and machine
46:37 learning.
46:39 And there's there's if you don't know
46:41 this, the history of this is absolutely
46:43 [ __ ]
46:44 fascinating.
46:46 The
46:49 instructions for how to create chat GPT
46:53 were in the documentation of
46:56 GPT3. So in the
46:59 documentation, anyone could have built
47:03 using
47:04 GPT3, anyone could have built chat GPT.
47:07 So, OpenAI cuts this deal with Microsoft
47:10 in in October of
47:14 2022. Sam says, "Hey, we're going to
47:16 announce a billion dollar investment
47:18 from Microsoft at the end of October or
47:21 at the end of
47:22 November. Tech team, can you guys whip
47:25 up that chat GPT thing that's in our
47:27 documentation? It would be cool to have
47:29 something to
47:32 show." And so they did. So like in a
47:35 month they whipped up the chat GPT
47:39 interface because the instructions were
47:41 in the in the documentation. They just
47:43 had to execute
47:45 it and then they launch it and within
47:48 five days they had a million users. They
47:50 went out for drinks the night before.
47:51 They they they said they would have been
47:52 happy with like you know couple of
47:55 thousand people using it and maybe 5,000
47:57 comments on whatever on X or whatever.
48:01 A million people used it within five
48:04 days. A hundred million people used it
48:07 within six
48:09 weeks. Six weeks from zero to 100
48:12 million users. Let me put that in
48:15 perspective. The early days of the
48:17 worldwide web, I started one of the
48:18 early uh digital agencies in 1994,
48:21 company called
48:22 agency.com. When we started agency.com,
48:25 like barely anyone was using worldwide
48:27 web. Like not not everyone had internet
48:29 all the time. You still had to your
48:31 computer had to dial up the internet. It
48:33 was it was just jankier than crap. The
48:36 worldwide web to get from when it
48:38 started to 100 million users was six
48:42 years. Chat GBT was six weeks. It was
48:46 the fastest adoption of technology in
48:49 history. And it could code and it could
48:51 write and it could do all this stuff.
48:53 And so what happened was because of the
48:55 runaway success of that, everyone turned
48:59 from anything else they were investing
49:01 in and they just came came to AI. So So
49:05 basically starting in 2023, all of the
49:08 money, all of the engineering talent,
49:11 all of the research labs, all of the
49:13 schools just all went the same
49:15 direction. We got to do transformers.
49:18 We've got to create frontier labs. Meta
49:20 said, "We've got to get in on this game.
49:22 we're going to open source ours.
49:23 Anthropic spun off. So, so you've just
49:26 got everyone running in the exact same
49:28 direction because of the runaway success
49:30 of chat GPT. So, that's the that's the
49:32 seinal piece of of technology that
49:36 kicked everything off. And quite
49:37 frankly, it was just a simple interface
49:40 like interface
49:42 matters. You could do the exact same
49:44 thing you could do with chat GPT on the
49:47 developers
49:48 playground. But if you were on the
49:50 developers playground, you had to choose
49:52 a model. There were all these different
49:53 models and you had to choose temperature
49:55 and you had to understand what a system
49:57 prompt was versus a chat prompt and like
50:01 it was just like it was confusing. You
50:04 had to be an engineer to get it. And so
50:06 chat GPT just made it super simple. Tik
50:08 Tok question. Major ethical cons
50:11 concerns concerning AI can't tell you
50:15 what time it is.
50:18 Who's going to say it's broken? I don't
50:21 know what that means.
50:23 Labor. AI can't tell you what time it
50:28 is. I just made a post which I regret
50:31 not using AI for. That's a good
50:33 [Laughter]
50:41 one. Um, if you can't trust it with the
50:44 basic question, how do you know? I
50:47 labor. I don't know. Wait, it means
50:48 whether you work on a project for 5 days
50:52 or 5 years. It's not
50:54 tracking your project. I still don't
50:57 understand what the question is. Yeah,
50:59 it like I it sounds like I do. Are you
51:02 using um AI
51:05 labor
51:07 because what you're saying was true in
51:12 2022. Like chat GPT is connected to the
51:15 internet now. So it knows what time it
51:17 is. It can write its own Python
51:21 code.
51:24 So oh the clocks is are you talking
51:27 about the clock thing? Well there
51:28 there's a clock thing which we should
51:30 talk about. Network error retry. Um here
51:34 let let me let me do a new let me let me
51:36 show you something. Um chat
51:39 GPT. So, so you are on to a point that
51:42 it is maddening, which is this. Make me
51:45 a clock, a picture of a cuckoo clock.
51:50 We'll do cuckoo clock. Cuckoo clock. Um
51:54 at uh
51:58 with with the
52:02 hands reading uh
52:07 7:30 7:35. No, because that they would
52:11 be overlapping.
52:13 7:25. So they'll be on the 7:25. They'd
52:16 be on the bottom of the clock like this.
52:18 Right.
52:20 Okay.
52:22 Reading. Let's make this image.
52:37 talking about how it doesn't know when
52:39 you worked on something this week or a
52:41 year
52:42 ago. Um, well that I mean that could
52:46 just be a software issue. So chat GPT
52:49 did just get access to all of your chat
52:52 history. I don't know like it knows when
52:55 it had a chat with you. I don't know if
52:58 it's got access to that
53:00 metadata but like that feature is a week
53:04 old. So, so part of what your part of
53:08 what your question is is, if I can
53:12 paraphrase your question, basically what
53:14 you're saying is, if this [ __ ] [ __ ]
53:16 ain't perfect, why would I ever use it?
53:19 Um, because it's one of the most
53:21 profoundly powerful technologies in the
53:24 history of humanity. And the
53:27 opportunity whenever there's a new
53:29 technology, the opportunity is when you
53:31 get in early. If you want to wait for it
53:34 to get perfect, you can, but but then
53:37 everyone will be using it and the
53:39 opportunity will it'll be much harder to
53:41 capitalize on the opportunity.
53:45 Um, so so be very very careful not to
53:50 use the fact that it's not perfect as an
53:52 excuse to sit on the
53:54 sidelines like understand.
53:58 Yes, Brandon.
54:00 Kyle. Um, yeah, I'm trying I've been
54:03 having some back and forth with Labor in
54:05 the Tik Tok comments. Um, so we we still
54:07 haven't gotten it. Um, I'm taking a
54:09 different approach. I think what, um,
54:13 Labor is saying, Labor, please correct
54:14 us if we're wrong still. I think what
54:16 we're saying here, um, is that the
54:20 overall operating system isn't stable.
54:23 You know, when you're working in coding
54:25 and logic, you expect very consistent
54:27 output. And Chad GBT will occasionally
54:31 say things like, "I can't make
54:33 pictures." And you're like, "No, you can
54:35 just try again." Oh, yeah. I completely
54:37 forgot to make pictures. There's a lot
54:39 of inconsistency. Like sometimes it just
54:43 You've never talked about Blotato on
54:46 this program. Oh, wait. We have That's
54:48 right. I forgot you sponsored it last
54:50 week. Yeah. Yeah. Yeah. Okay. All right.
54:52 So, really good question then if if
54:55 that's the if that's the if that's the
54:57 question. Okay. So, here's what you're
55:00 coming up against, Labor, which is um
55:03 something we're we're all going to come
55:05 up against. Oh, this says image created.
55:07 Oh, they're they're um they're try
55:10 again. Um their their servers are hosed.
55:14 Every everybody's using it. Their
55:16 servers are hosed. Okay.
55:20 Um Okay.
55:22 For all of my lifetime, for all of your
55:24 lifetime, for all of all of our
55:27 lifetimes, computers behaved a very
55:30 specific
55:31 way. You would write logical
55:34 code that had logic, which is why they
55:37 called it logical code. Well, they just
55:39 call it code, but it's got logic in it,
55:41 so it's
55:42 logical. And you would put in rules,
55:44 right? If this, then that loop here,
55:47 right? you you would explicitly code in
55:51 logic so that when a user used it, if
55:54 the user typed in Bob, it would go into
55:57 its code and say if user types in Bob,
56:00 then you know, make up a last name or
56:03 whatever, whatever it was, whatever was
56:04 programmed here, the output was
56:08 consistent and predictable based on the
56:10 logic, right?
56:13 Okay. Generative AI changes all of that.
56:19 For the first time in our lifetimes,
56:21 computers do not behave the way they've
56:24 always
56:26 behaved because of the G in GPT stands
56:30 for
56:31 generative. It's generative. It's not
56:34 just saying, I know Python code. I'm
56:37 going to use Python code like I'm going
56:39 to copy a block of code. You ask for an
56:41 Asteroids game. I'm going to go look for
56:43 Asteroids code. I'm going to copy the
56:45 code. and I'm going to come over here
56:46 and paste it. No, that's how that's how
56:49 it would have worked in the olden timey
56:51 days 3 years ago. Now, in the in the new
56:55 days, it's actually generating it's
56:58 actually writing the Python
57:01 code. So, you ask for whatever, it
57:04 writes the code one token at a time or
57:08 maybe blocks of tokens, but whatever.
57:11 But it's writing it. It's generating it
57:12 each time. And if you ask for it a
57:14 second time, it may write it different.
57:17 If if you've ever used any of these vibe
57:19 coding
57:20 tools, it makes you an app and then
57:22 you're like, "Oh, could you change the
57:24 background color from blue to white?"
57:26 And then it completely breaks the app
57:28 because it went in and just wrote the
57:30 code different. Now, part of that's just
57:32 bad software design. That's something
57:34 they could fix. Uh, new Tik Tok pin.
57:36 Yes. Also, it discredits someone's life
57:40 work if they're spending years on a
57:42 project. I I don't I don't know that it
57:44 discredits it. Labor. I think
57:53 that well, it doesn't discredit it. The
57:58 person that's been working on a project
58:00 for a
58:01 lifetime may feel discredited. They may
58:04 feel like, "Wait a minute. What the [ __ ]
58:06 have have I been doing for 30 years? If
58:09 if I take pride in my coding and I take
58:12 pride in making sure it's clean and
58:13 clear and all that sort of stuff and
58:15 then some [ __ ] punk like the dude at
58:18 the AI learning lab can come along with
58:19 a theater degree and write code and
58:22 claim that he's a coder now. Well,
58:24 that's an affront to my value as a
58:26 human, right?
58:28 Every single every single knowledge
58:32 worker on the planet is about to face
58:36 the crisis that you're
58:40 describing. And here's the only thing I
58:42 know like this. Here's my philosophy on
58:44 this. There's you can absolutely judge
58:47 it, but the reason this channel exists
58:50 is here's what I know. I know that AI
58:54 isn't going away.
58:57 And I think one of the things that we're
58:59 talking about here tonight with 03 and
59:01 the fact that it just made this sort of
59:03 quantum leap in quality,
59:06 um, it's getting better really
59:09 fast. And so then you only have two
59:13 choices. It's actually quite binary,
59:15 which it sounds like you like code, you
59:18 like predictability. Here's the binary
59:20 choice. If AI is not going away and it's
59:23 getting exponentially better, you have
59:25 two
59:25 choices. You can embrace it or you can
59:28 ignore
59:30 it. And if you ignore it, you're going
59:33 to be the victim of
59:35 it. It ju you just
59:38 are. And if you embrace
59:41 it, you may not like that how you write
59:46 code is changing. You may choose to
59:48 write some of your code the way you've
59:50 always written it and then use AI for
59:52 other things to make other efficiencies
59:54 in your work. But what you also might
59:57 discover is that maybe you got really
1:00:00 good at one particular programming
1:00:01 language and you always you wanted to
1:00:03 move to some more modern platform but
1:00:05 because of your career you you were kind
1:00:07 of locked into this into this platform.
1:00:10 Now you could be masterful at 10
1:00:12 different programming languages and you
1:00:14 could take your expertise as a
1:00:16 developer, as an engineer and apply that
1:00:19 to multiple
1:00:22 different programming languages. Or
1:00:25 maybe you always wanted to write a
1:00:28 novel, but you were always coding and
1:00:30 but now you can use AI to write or make
1:00:33 images. One of the one of the guys in uh
1:00:36 in the AI salon, a guy named Roger
1:00:38 Scott, he was a software engineer for 30
1:00:40 years, 35 years. Like real keragin, like
1:00:44 he's just like he he's just very
1:00:46 matterof fact.
1:00:49 Um and he discovered Midjourney. And you
1:00:52 know how he discovered
1:00:54 Midjourney? We were in the Discord early
1:00:56 on in the salon. We were on Discord
1:01:01 and all these artsy fartsy types were in
1:01:04 there making midjourney images and Roger
1:01:07 kept asking questions like, "Well, what
1:01:09 specific prompt did you give it to get
1:01:12 that to to get that look and feel?"
1:01:14 Because he would take the prompt that
1:01:16 that that person did and he would get
1:01:18 something different and he was really
1:01:20 frustrated by that, right? He's like,
1:01:22 "Why is it different? Why is it
1:01:23 different? Like, why is it doing that?"
1:01:25 He was trying to figure it out.
1:01:27 And so people within the salon said,
1:01:29 "Well, just go play. Just go over to
1:01:31 MidJourney and start playing." And you
1:01:33 know what Roger discovered? He [ __ ]
1:01:35 loved making art. And because he had
1:01:38 programming skills, because he had
1:01:40 logic, he approached midjourney
1:01:43 prompting like coding and he got really
1:01:46 [ __ ] good at it. And he ended up with
1:01:48 some of his artwork in an art show.
1:01:51 So on the other side of the existential
1:01:54 angst that I hear in your
1:01:57 question is there's the possibility that
1:02:01 it can free you to do things you you
1:02:03 know in a much more amplified way than
1:02:05 you could have historically but yeah it
1:02:09 people who've been doing something for
1:02:11 30 years they are absolutely going to be
1:02:13 confronted by this. So, um, yeah, and
1:02:16 and just I was following along in some
1:02:18 of the back and forth comments and yeah,
1:02:20 you know, I think there's definitely
1:02:21 something to be said for the the value
1:02:23 of the human in the loop. I mean, this
1:02:25 stuff was trained on all of humanity and
1:02:28 you're you're really working with
1:02:29 humanity on this, but to that point,
1:02:32 you're working with it on a platform
1:02:35 that two and a half years ago was a
1:02:38 research institution that fumbled their
1:02:40 way into being this juggernaut of a
1:02:43 company. And so when you know they are
1:02:46 definitely playing catch-up. I always
1:02:48 try to give them a little bit of grace
1:02:49 in the fact that they don't have all of
1:02:52 their ducks in a row in terms of having
1:02:54 uptime reports and you know fail safes
1:02:57 for when their systems go down because
1:02:59 the research company you know cosplaying
1:03:02 as a uh actual company. Um before you
1:03:06 hit enter on the this um I was a couple
1:03:10 of folks in the comments have noted that
1:03:12 maybe 03 doesn't have image capability.
1:03:16 It does. It does well or or right now
1:03:19 because it keeps but if you change the
1:03:21 model you might have better uh luck with
1:03:24 generating images at least for the
1:03:26 purposes of this demo. Yeah. Yeah. Yeah.
1:03:29 Um try again.
1:03:33 Yeah. I I think 03 looks like it's
1:03:36 buried. Try again.
1:03:40 Yeah. And of course, we we encourage
1:03:42 labor and everybody to join us at the
1:03:44 salon.ai. Yeah. Pop the pop the URL up
1:03:47 there if you haven't. So So if if you're
1:03:50 not a part of the the salon, go to the
1:03:52 salon.ai. Um click on the button that
1:03:55 says join our community and that'll pop
1:03:58 you over to a Mighty Network site where
1:04:00 the where the community is. And there's
1:04:02 like a really simple sevenstage um uh
1:04:05 whatever you call it uh welcome you know
1:04:08 welcome area and and you know step two
1:04:12 in that process is introduce yourself.
1:04:14 So please please introduce yourself uh
1:04:17 and just tell us where you are and and
1:04:18 like you know if you're frustrated with
1:04:20 this like I totally get the frustration.
1:04:23 Um it it looks like the image generation
1:04:25 here is not is not working out. I might
1:04:27 have um an image here of a clock. Let me
1:04:30 see.
1:04:31 Clock
1:04:39 image. C
1:04:43 clock. Cuckoo. Clock is generating
1:04:46 artwork. Wait 9 minutes. Clock
1:04:49 gears. Steampunky
1:04:53 readiness. Beach
1:04:57 ball. No, I don't have one here. Yeah,
1:05:00 it looks like their servers are
1:05:03 hosed. Let me let me start a new chat
1:05:06 because maybe I just ended up in some
1:05:07 sort of weirdass loop. Uh, make a
1:05:11 picture of a
1:05:13 clock with
1:05:17 hands
1:05:25 indicating 7:30.
1:05:29 Dennis, I see no issue with using AI as
1:05:33 we all build on top of past research. It
1:05:36 just gives us knowledge faster and wider
1:05:38 access to it. Yep, I would agree with
1:05:40 that. I was going to say we have the the
1:05:43 fire we have the firewolf. Why the hell
1:05:46 not? Oh, the dire do. That's funny.
1:05:49 YouTube spotlight.
1:05:51 Um, yeah. I I'll tell you one of my
1:05:57 um like you know I talked about my
1:05:59 philosophy on this site is AI is not
1:06:01 going away and so you you know then you
1:06:04 have two choices right deal with it or
1:06:06 not and
1:06:07 listen I have if you choose not
1:06:13 to at least try
1:06:15 AI don't think there's a good excuse for
1:06:18 that if you choose to try AI and then
1:06:21 after really giving Even at a good
1:06:22 gipper go, you decide, nope, I hate it.
1:06:24 I never want to use it. Um, I have
1:06:27 respect for
1:06:28 that.
1:06:31 Um, I lost my train of thought. I don't
1:06:34 know where I was
1:06:35 going. I was thinking about my clock
1:06:38 clock
1:06:40 hands. Oh, whatever. Oh. Oh, wait. It
1:06:42 was Oh. Oh. Oh, yeah.
1:06:45 So when I first started learning about
1:06:47 AI, like I started, one of the reasons I
1:06:49 started this channel from a selfish
1:06:51 standpoint is I wanted to learn
1:06:53 everything I could about AI. And so I
1:06:56 thought if I if I learn out loud, if I
1:06:58 learn stuff and then as I'm learning it,
1:07:00 I share it, I'll learn it better.
1:07:03 And for about the first year when I
1:07:05 really got my head around how the
1:07:07 transformer model worked and and how
1:07:09 embeddings worked and how the latent
1:07:11 space worked and how token prediction
1:07:13 worked, it's absolutely incredible. Like
1:07:16 mathematically it is stunningly elegant
1:07:20 and it's actually weirdly simple, right?
1:07:25 You you
1:07:26 you figure out the semantic groupings of
1:07:30 tokens and you put them in thousand
1:07:32 dimensional mathematical space and then
1:07:35 when someone writes a prompt you use the
1:07:37 same basic basic math to identify a
1:07:41 probability region in thousand
1:07:43 dimensional mathematical space. It's
1:07:45 it's beautiful. It's elegant but when
1:07:47 you understand it it's a calculator.
1:07:53 Transformers are a
1:07:55 calculator and yet your experience of it
1:07:58 is this weirdly human poetry writing
1:08:03 programming mathy image generating
1:08:06 thing and I couldn't reconcile those
1:08:10 two. I couldn't figure out what why is
1:08:12 it that this thing is got this humanity
1:08:14 to
1:08:16 it. And someone on Twitter they they
1:08:20 they wrote this very simple tweet. I'll
1:08:22 never forget it because it was it was
1:08:23 this revvelatory moment for me. They
1:08:25 said artificial
1:08:27 intelligence is the collective
1:08:29 intelligence of
1:08:31 humanity or AI is the collective
1:08:33 intelligence of humanity was the
1:08:36 tweet.
1:08:38 And what what finally dawned on me which
1:08:41 is patently obvious like like all good
1:08:43 revelations they're obvious once you
1:08:45 have them. The reason it feels so human
1:08:48 is what it was trained on. The P in GPT
1:08:52 stands for
1:08:53 pre-trained. So what what OpenAI did
1:08:57 first well and then what every
1:08:59 subsequent frontier model has done since
1:09:01 is they've gone out and they've gotten
1:09:03 as much data as they they could and yes
1:09:06 you can absolutely argue that that's
1:09:09 probably unethical and possibly illegal.
1:09:11 Yes, all of that. What I can tell you is
1:09:13 if they had asked for permission, we
1:09:15 wouldn't have these tools today. So they
1:09:18 went out and they did the San Francisco
1:09:19 thing and they said, "Fuck it. We're
1:09:20 just going to go get get all the data.
1:09:22 We're going to throw it at all these
1:09:25 GPUs, but what they actually did was
1:09:27 they took the collective output of
1:09:29 humanity and they compressed it into
1:09:32 this beautifully elegant math ball that
1:09:36 we all get to tap into.
1:09:40 And what it is in a very literal sense
1:09:43 is is you put in a prompt and it's like
1:09:45 you're collaborating with everyone
1:09:48 that's come before
1:09:51 us and it's it's quite elegant. It's
1:09:55 quite
1:09:56 amazing.
1:09:58 Um we as humans still have agency. We
1:10:02 can choose when to use these tools. We
1:10:04 can choose what to use them for. A lot
1:10:07 of creative people that I know, they've
1:10:09 got hermetically sealed parts of their
1:10:12 work that they're like, "I'm never going
1:10:14 to use AI for
1:10:16 this." Awesome.
1:10:19 Perfect. A good buddy of mine, a
1:10:21 creative director, will never use
1:10:25 another artist's work in his prompt
1:10:28 because that's an ethical line that he's
1:10:30 choosing to draw. He's like, "I want to
1:10:33 take advantage of these remarkable
1:10:34 tools, but I won't invoke an artist's
1:10:37 name. I'll figure out how to describe
1:10:39 their work, but I won't invoke their
1:10:41 name because I don't want to do
1:10:43 that." Like, I have total respect for
1:10:46 that. Um, but that's what the that's
1:10:50 what this is. This
1:10:52 is, you know, kind of imagine it like
1:10:54 this. Imagine
1:10:56 taking all of the artifacts and all of
1:10:59 the knowledge of every museum, every
1:11:03 research library, every university,
1:11:05 every university
1:11:08 library and giving everyone who uses
1:11:12 this tool instant access to all of that
1:11:14 knowledge. That's what AI is.
1:11:18 So, so you know when I hear individuals
1:11:22 that are alive in in this time window
1:11:24 going, "Oh, they stole my work."
1:11:26 Well, they trained it on sort of
1:11:29 everything and your work likely got
1:11:31 swept up in that. Yes.
1:11:34 Um, but to identify when when this
1:11:37 outputs something, to identify which
1:11:40 piece of which artist's work is in
1:11:43 there, the the the the closest
1:11:45 equivalent I can come is if you're
1:11:47 making a cake and you take one piece of
1:11:50 one grain of
1:11:51 sugar and you throw it in the cake
1:11:54 batter and you stir it up and you put it
1:11:57 in the oven and you bake a cake and then
1:11:59 someone says, "Well, I want my grain of
1:12:01 sugar
1:12:03 back. It is not in that form
1:12:06 anymore, right? It's it's it's
1:12:09 molecularly completely different, which
1:12:12 when you embed data into a large
1:12:14 language model, the original works cease
1:12:16 to exist. They get turned into these
1:12:18 little fragments like little grains of
1:12:20 sugar and blasted into
1:12:22 thousand-dimensional mathematical space.
1:12:24 The original works cease to exist. The
1:12:26 only thing that remains is the semantic
1:12:29 relationship between the
1:12:31 tokens. Anyway, um Becky Rue has a good
1:12:34 prompt for Grock. Create the six for
1:12:37 Grock. Oh, the six flag the six flags
1:12:41 guy AI image who's really Jeff Bezos.
1:12:45 Oh, with a new ride called Blue Origin.
1:12:47 Oh, okay. Yeah, because Okay, because
1:12:49 Rock will do things. Yeah. Uh it looks
1:12:51 like we it looks like uh yield uh chat
1:12:56 GPT is kind of useless tonight. So I
1:12:58 guess we'll go do other things. Um let's
1:13:00 go over to we'll go over to
1:13:02 Twitter. So anyway, does that answer
1:13:04 everyone's question, Kyle? Irregulars
1:13:06 ASAP. Okay,
1:13:09 good. Oh, what's this? That looks cool.
1:13:12 Am I in a
1:13:14 regulars? I posted this to another
1:13:17 channel, but it should be in here for AI
1:13:18 brand search. What is
1:13:24 this? Oh, that's
1:13:26 cool. Huh. All right.
1:13:31 [Music]
1:13:32 Um, nice. They say duct tape has limited
1:13:36 usage. Wait, I don't need a plan. I need
1:13:39 caffeine. Caffeine duct tape and a mild
1:13:42 disregard for the consequences. That's
1:13:44 really good. Nice one, Dennis.
1:13:49 Too old for Tik Tok trends, but too
1:13:50 young for bingo night. Too crumpy for
1:13:52 both. That's me. Pretty much me. I I
1:13:56 relate to that one story behind the
1:13:59 picture. Amid the beautiful chaos.
1:14:01 Beautiful. Love
1:14:02 it. My dog is a human.
1:14:09 Nice. Oh man, that was using chat GPT
1:14:14 this morning. That's cool. Let's see.
1:14:15 Valerie, I believe our way of life will
1:14:18 exponentially change. The hope is for
1:14:20 the better. Here here's my here here's
1:14:24 my opium for the for the radical change
1:14:26 that's coming. Um, and I I do think it's
1:14:29 going to be quite radical. Um, I don't
1:14:31 think it's going to happen as fast as
1:14:32 the the tech bros think it is because
1:14:36 just humans humans are humans take a
1:14:39 while and they resist change. Um, but I
1:14:42 think it's going to be profound and and
1:14:44 and profound in ways we can't
1:14:47 anticipate. So, if you think about when
1:14:49 the steam engine came out, we we were
1:14:51 largely an agrarian society. Everyone
1:14:55 worked on a farm and everyone worked
1:14:57 with horses and everyone worked with
1:14:59 their hands and everyone worked with
1:15:00 their
1:15:02 muscles. And the steam engine came along
1:15:06 and it
1:15:08 essentially
1:15:10 replaced human
1:15:13 strength. Now imagine being a strapping
1:15:17 back then you you probably wouldn't be
1:15:19 strapping until you're 50, right? You
1:15:20 probably only lived till you're
1:15:22 45, but you're a strapping 35y old man,
1:15:26 which is like old in those days, but
1:15:28 you're still still good for your age,
1:15:29 right? You're a strapping 35y old man.
1:15:31 Along comes this thing called the the
1:15:34 steam
1:15:35 engine. And we don't need you anymore.
1:15:37 We don't need the We only need two
1:15:39 strapping men and we've got 10. So eight
1:15:42 strapping men and their
1:15:46 families are displaced,
1:15:50 right? And if you sort of take in the
1:15:52 very real fear of today, AI is going to
1:15:56 take your job and it's going to replace
1:15:58 a lot of people. We only need two. those
1:16:00 other eight, we just don't need them
1:16:02 anymore. Yes, that's going to happen.
1:16:05 What happened in steam engine days?
1:16:07 Those people that lived on farms moved
1:16:10 to
1:16:11 cities and the cities got bigger and as
1:16:14 the cities got bigger an entirely new
1:16:17 kind of dynamic was created and service
1:16:20 economies were created and manufacturing
1:16:23 and you know all sorts of things
1:16:25 happened as a result of that shift that
1:16:27 you you just literally couldn't even
1:16:29 imagine well what would I do without a
1:16:32 horse I don't know live in a place where
1:16:35 all the stores are within walking
1:16:37 distance and you don't need one. Oh,
1:16:40 wait. Is that
1:16:41 possible? That's that's the kind of
1:16:44 shift we're at the beginning of. So, the
1:16:48 only thing it feels like right now is
1:16:50 [ __ ] existential
1:16:53 fear, right?
1:16:58 But what
1:17:02 if for the first
1:17:05 time since computers were
1:17:09 invented, they did
1:17:13 more than make our work more efficient.
1:17:16 Right? If you look at the history of
1:17:18 computing, it is largely an extension of
1:17:21 the industrial age where machines
1:17:24 replaced our muscles and made us more
1:17:26 efficient at doing physical
1:17:29 stuff and then computers came along and
1:17:32 made us more efficient at doing, you
1:17:34 know, knowledge work and and you know,
1:17:36 information processing stuff. But it was
1:17:39 still us doing the work and the computer
1:17:41 was there making it more efficient.
1:17:44 Which is why when people first start
1:17:46 using AI, one of the first things that
1:17:48 they do is they're like, "Oh, I could
1:17:50 automate all the [ __ ] I do and make it
1:17:53 more efficient." That's what computers
1:17:56 have always
1:17:57 done. But they're missing the the
1:18:01 relevance of the G in
1:18:04 GPT
1:18:07 generative. Yes, you can make it to make
1:18:10 your work more efficient. You could also
1:18:13 use it to completely reimagine how that
1:18:15 work gets done that we've never had
1:18:18 before. So
1:18:20 imagine, and I feel decently confident
1:18:24 this is where we're headed, imagine that
1:18:26 AI systems are going to be able to do
1:18:29 the bulk of the work, the bulk of the
1:18:34 work that crushes people's souls.
1:18:40 filling out forms, sitting on hold for 4
1:18:42 hours, trying to get your insurance paid
1:18:44 for, um, doing data entry because your
1:18:49 company still uses fax systems and
1:18:51 someone has to type that [ __ ]
1:18:53 in. Imagine if all of the
1:18:57 drudgery was done by the machines and
1:19:00 our only
1:19:02 job was to ask the question, how do I
1:19:06 want to impact the world? what kind of
1:19:08 difference do I want to
1:19:11 make? And then because of AI, you have
1:19:13 the resources to be able to execute
1:19:18 that. That starts to look like a pretty
1:19:21 [ __ ]
1:19:25 enlightened
1:19:27 future where where anyone who wants to
1:19:30 do something for the world, for their
1:19:32 neighbors, for their family can do it at
1:19:35 any level.
1:19:37 I've got chronically ill kids with Lyme
1:19:41 disease. No one's putting a dime into
1:19:43 Lyme
1:19:44 research because the insurance companies
1:19:47 won't pay for
1:19:48 it.
1:19:50 Well, we're probably within five years
1:19:52 of someone going, "Fuck it. I'll cure it
1:19:56 myself."
1:19:59 Right? So, all
1:20:03 right. People worried about jobs at the
1:20:06 bank with ATMs in the 80s. Yeah. It went
1:20:09 on to create thousands of new jobs.
1:20:11 Every single time. Every [ __ ] single
1:20:14 time. Every time there's a new
1:20:18 technology, people resist it. Every
1:20:21 time. Every time. Every single time. And
1:20:24 the more profound the technology, the
1:20:26 stronger the resistance.
1:20:29 Danielle Mike
1:20:33 Myers. Oh, wait. You can't see that.
1:20:36 Hang on. Can you pop the screen up
1:20:38 there, Brandon? That's pretty
1:20:44 funny. Yeah,
1:20:46 [Laughter]
1:20:51 baby. That's
1:20:53 [Laughter]
1:20:55 awesome. Oh.
1:21:00 [Music]
1:21:04 Gorgeous. Gorgeous. Gorgeous. What's
1:21:06 that? I've been vibe
1:21:08 coding mural application for muralist
1:21:11 for the metaquest here. Oh, that's cool.
1:21:13 Made some progress. That's
1:21:17 neat. I love making these crowd
1:21:19 renaissance images. I made this with
1:21:21 Google's whisk. That's really
1:21:24 cool. I took a photo. I've got a photo
1:21:26 that I took in the shuttle train at at
1:21:30 Denver
1:21:31 airport where it's a photograph of
1:21:34 people in the other car, but the glass
1:21:36 reflection shows the people in my car
1:21:39 and everyone's just looking in a
1:21:41 different direction and it very much
1:21:42 looks like a Renaissance kind of
1:21:44 composition. It's really
1:21:49 cool. Oh, Steo did the Austin Powers
1:21:51 thing. That's good.
1:21:54 That's really
1:21:56 good. Oh man. Not in your children's
1:22:00 grandchildren's
1:22:03 lifetime. I don't know what that's in
1:22:05 reference
1:22:07 to. Alex, I'm a truck driver. I would
1:22:10 love for the industry to go completely
1:22:12 driverless. It'll never
1:22:15 happen. Um, well, okay. Let me give you
1:22:20 another. I don't know if this one's
1:22:29 um this is this is me trying
1:22:33 to put a put a silver lining on on some
1:22:37 of the some of the changes that are
1:22:40 happening right now in our government.
1:22:47 Um, in order for there to be radical
1:22:50 change, there needs to be radical
1:22:52 change. So, I would I would agree with
1:22:54 you that if everything stays the way it
1:22:57 is right
1:22:59 now, unions won't allow, you know,
1:23:03 driverless trucking, right? Um, we saw
1:23:07 the we saw the the dock workers do do
1:23:10 their strike. Um, if what's happening
1:23:14 right now in government ends
1:23:16 up
1:23:19 obliterating the the infrastructure of
1:23:22 how businesses run. Um there there is a
1:23:26 very real likelihood it will be
1:23:29 recombined over the next I don't know 10
1:23:32 or 20 years um in radical new ways and
1:23:36 and because this technology is
1:23:38 accelerating and and because you know
1:23:41 capitalism is capitalism um people are
1:23:44 going to look for ways to do things more
1:23:46 more and more efficiently. Um and and if
1:23:49 if everything changes, right, if if just
1:23:52 everything breaks apart, um then then it
1:23:55 can be much more radically reassembled,
1:23:57 right? Um if that doesn't happen, then I
1:24:01 agree it's going to take some amount of
1:24:02 time for that change to happen. But but
1:24:05 but you know, saying that there will
1:24:08 never be driverless trucks is kind of
1:24:10 like saying, you know, steamowered
1:24:13 tractors will will never be able to farm
1:24:16 farms as good as horses. Well, for a
1:24:19 generation maybe, right? And then at
1:24:22 some point it's like, well, yeah, the
1:24:24 tractors are actually
1:24:27 better at the farming part or at the
1:24:31 large. You going to Tik Tok going to
1:24:34 throttle you for opium? Yeah,
1:24:36 [Laughter]
1:24:41 exactly. Change is inevitable, but
1:24:44 change will also be resisted.
1:24:54 What do you mean except from a vending
1:24:58 machine? Oh, change is
1:25:01 inevitable. Yeah, you'll never get your
1:25:03 change back from a vending machine. Musk
1:25:06 says, "Nobody will have work in 10 years
1:25:09 and 90% of cars will be autonomous in 5
1:25:12 to 10."
1:25:13 Um the the like something I do take
1:25:18 umbrage with is Musk didn't
1:25:22 say 90% of people won't have
1:25:26 work. For most people work will be
1:25:29 optional. Those are two very different
1:25:32 things. Human beings want to connect and
1:25:36 want to contribute. Like we want
1:25:39 to have have purpose. Some people don't
1:25:43 have purpose and when you lose your
1:25:44 purpose, you tend to die, right? You
1:25:46 tend to just fade
1:25:48 away. There's a there's a very like like
1:25:52 I find this incredibly
1:25:54 strange that one of the fears of AI is
1:25:57 that people will lose their purpose.
1:25:59 Now, they may if they're if they remain
1:26:01 completely attached to how we used to do
1:26:03 it, they will absolutely lose their
1:26:06 purpose.
1:26:08 But what? So, so if you've got a 30-year
1:26:10 career as a coder or as an illustrator
1:26:12 or as a as a art
1:26:17 director and then AI comes along and it
1:26:19 changes how you do that. It doesn't
1:26:22 change what you do, but it changes how
1:26:24 you do
1:26:25 it. Again, you have two choices. You can
1:26:28 hang on to the way you've done it
1:26:30 before. You can
1:26:31 quit and do
1:26:34 nothing or you can find a new way to
1:26:36 express yourself. And what you have the
1:26:38 opportunity to do is say, well, what was
1:26:41 the reason that I got into art in the
1:26:43 first place? Do I still want to do that
1:26:46 here? Maybe I got into art because mommy
1:26:50 was an artist and she told me to be an
1:26:52 artist and I went into it, but I always
1:26:53 resented it and I always wanted to be a
1:26:55 lawyer.
1:26:57 Well, now go be a lawyer, right? Now go
1:27:00 be a mathematician. Yeah, but I can't go
1:27:02 to Stanford. And it doesn't
1:27:04 m The new chat GPT got 99.5 on a
1:27:10 kick-ass math benchmark. It's probably
1:27:13 going to be better at math than you
1:27:15 would ever be if you went and did four
1:27:17 years at a
1:27:19 college. So, everyone's going to have
1:27:21 that opportunity to do that. the the
1:27:23 sort of trope that if people get
1:27:26 displaced then they're purposeless and
1:27:28 they're jobless. I just don't buy
1:27:33 it. I mean the
1:27:36 tools the tools are going to go from
1:27:39 being able to do [ __ ] for us to being
1:27:42 able to do to to kind of spin up virtual
1:27:45 employees that work on our behalf 247.
1:27:52 So, a way you could think about this
1:27:57 is if with your little laptop sitting on
1:28:01 the
1:28:02 beach, you could have a company of a
1:28:05 hundred employees that were all PhD
1:28:08 level
1:28:10 quality just working for
1:28:12 you. What would you
1:28:16 do? Like that's the question we get to
1:28:19 ask. And that's like why I've got such
1:28:23 optimism for this
1:28:25 [ __ ] is because what a [ __ ]
1:28:28 remarkable time to be alive in
1:28:31 history. Are you [ __ ] kidding me?
1:28:35 Really? We get to do that?
1:28:39 Yes. But if you just sit on the
1:28:42 sidelines with your arms crossed like
1:28:43 harump. Well, that's not how I used to
1:28:46 code or that's not how I used to art
1:28:49 direct. The [ __ ] the jet is taken off
1:28:53 the
1:28:57 carrier. You're going to be on a big old
1:28:59 rusty boat in the middle of the [ __ ]
1:29:00 ocean that runs out of gas at some
1:29:03 point. Do your
1:29:05 kids Wait, do your kids the future
1:29:08 believe like you do? Um, well, my kids
1:29:11 are chronically
1:29:13 ill, so their relationship with the
1:29:16 future is very different than mine. I
1:29:18 like I have from a very young age,
1:29:21 there's something about the way my brain
1:29:23 is wired, you know, maybe it's the ADD.
1:29:27 I don't know what it is. There's
1:29:28 something about the way that my brain is
1:29:29 wired that I don't need much information
1:29:32 in to be able to to pattern match and
1:29:34 and see see what's coming.
1:29:38 So, I've always just had a really good
1:29:39 instinct about what's coming. And so,
1:29:42 for
1:29:43 me, one of the one of the things that
1:29:46 gives me the the most joy in the world
1:29:49 is like imagining what it's going to be
1:29:51 like 5 years down the road and
1:29:54 experiencing that in the
1:29:56 present. And so, that allows me to go,
1:29:58 oo, if that's going to be the future,
1:30:00 here's what I need to do right now to
1:30:01 get
1:30:03 there. Um, they don't have that. They're
1:30:06 like, "Am I going to be in pain
1:30:08 tomorrow if I go out with my friends
1:30:11 tonight?" Like, their their their
1:30:13 worldview is very, very
1:30:15 limited. Um, and also they've got a dad
1:30:18 who's like, "AI is the best. You guys
1:30:21 should love AI." Well, what do you do as
1:30:23 a kid? [ __ ] you,
1:30:25 dad. Whatever dad wants, I want the
1:30:28 opposite of that.
1:30:34 Every new software I've helped it
1:30:35 implement has come with negativity.
1:30:38 Yep. I used AI to learn. Wait, what was
1:30:43 that? Someone just had a post up there.
1:30:46 What was it? I used I used AI to learn
1:30:50 digital drawing and I draw by hand. It's
1:30:53 the most amazing tool for learning.
1:30:55 Yeah. Like there's a there's a great
1:30:57 example, right?
1:31:02 If you're an artist and you look at AI
1:31:04 and you see someone go, "Give me a
1:31:07 picture of whatever and it spits out a
1:31:09 picture that's kind of like the picture
1:31:10 you
1:31:11 draw." Your immediate instinct is that's
1:31:14 evil. [ __ ] that. It's taking my job.
1:31:18 Another way to look at it is what if you
1:31:20 used it as an ideation tool?
1:31:23 What if you could spit out lots and lots
1:31:26 of images to inspire
1:31:29 you and then you draw them because you
1:31:32 enjoy drawing? Like like we as human
1:31:34 beings, we get to choose where we put
1:31:37 our time. But just because a tool can do
1:31:40 something as good or better than you
1:31:41 can, it doesn't mean it doesn't
1:31:43 invalidate that you enjoy drawing. What
1:31:47 it might do is
1:31:49 devalue the thing you used to be able to
1:31:51 make money on.
1:31:54 And that's what's going to [ __ ] a lot of
1:31:56 people up because then you've got to
1:31:58 force, right? You're being forced out of
1:32:01 the thing you've you love to make a
1:32:03 living by. And that we're all going to
1:32:05 have to confront. Tik Tok question.
1:32:08 Kyle, could you explain what Musk means
1:32:10 when he says people will have the option
1:32:12 to work? Yeah.
1:32:16 Um so one of the things Sam Alman
1:32:21 um Sam Alman has talked about so so
1:32:24 there's this concept of universal basic
1:32:26 income the ba the basic theory is this
1:32:29 the AI is going to get so good and so
1:32:33 autonomous that assuming it doesn't kill
1:32:35 us all which you know that's still
1:32:38 possible but it seems you know
1:32:40 increasingly not likely but whatever we
1:32:43 may get there. But assuming it doesn't
1:32:44 kill us all, in in which case then we
1:32:46 don't have anything to worry about if it
1:32:50 does.
1:32:53 Um the computers are going to do the
1:32:56 bulk of the work. And you're like, well,
1:32:58 not drive trucks or not uh dig ditches
1:33:01 or not. Well, yeah, with
1:33:03 robots, right? If Musk has his way,
1:33:06 there will be at least a robot for every
1:33:08 human on Earth. And he said probably
1:33:11 two. So physical work and knowledge work
1:33:14 will just be
1:33:19 done and it will create so much
1:33:21 prosperity and this is a thing I have a
1:33:24 tough time with. It will create so much
1:33:26 prosperity that the capitalists can't
1:33:28 keep it all that they will actually have
1:33:30 to give some back right because
1:33:32 otherwise people would revolt and you
1:33:34 know eat their babies. I probably can't
1:33:37 say that on Tik Tok either. So if if I
1:33:40 disappear from Tik Tok, you'll know why.
1:33:44 Um so if the computers are going to do
1:33:46 all the work, one of the concepts Sam
1:33:48 Alman put out there was
1:33:52 um the idea of universal basic compute.
1:33:55 There's a concept of universal basic
1:33:57 income, which is if you have enough
1:33:58 prosperity, you take some of that
1:34:00 prosperity, you give it back to the
1:34:01 people as a salary and you get paid to
1:34:05 do nothing. Right?
1:34:08 Sam Alman says, "What we should do is
1:34:10 rather than give people money, we should
1:34:12 give them compute cycles. We we should
1:34:14 give them compute time, which on the
1:34:17 surface sounds just [ __ ] stupid and
1:34:19 weird, right? But let's
1:34:22 say that plays out and every person on a
1:34:26 monthly basis is given enough universal
1:34:29 basic compute that they could run an AI
1:34:33 that's the equivalent of a 100
1:34:35 employees.
1:34:37 So that means that every person could
1:34:40 create a company with up to a hundred
1:34:43 employees running that company or
1:34:45 executing the vision of that
1:34:48 company. Well, some people might not
1:34:50 want to be an entrepreneur and they
1:34:51 might want to might not want to do that.
1:34:54 Okay, so maybe they sell all of their
1:34:57 universal compute to someone who wants a
1:34:59 thousand
1:35:01 employees, right? And so then they get
1:35:03 money and they can do nothing. So, the
1:35:05 optional part is you could choose to
1:35:08 just do
1:35:10 nothing. I think that's appealing for
1:35:12 people that are overburdened with work
1:35:14 and overwhelmed with work. Oh, I'd love
1:35:16 to just do [ __ ] nothing until you're
1:35:19 doing [ __ ] nothing for a month or two
1:35:20 and then you're like, "Oh, this sucks."
1:35:23 And so the work is optional part is then
1:35:26 then we get to say so if we all have
1:35:29 enough resources to say I I can choose
1:35:32 to do
1:35:33 nothing at the point at which that
1:35:35 becomes boring for you what do you
1:35:38 choose to
1:35:39 do knowing that you've got this machine
1:35:42 that can execute anything you put your
1:35:44 mind
1:35:45 to so that that's where he said that's
1:35:48 that's what he's talking about when he
1:35:49 says work is
1:35:51 optional that there will be you know so
1:35:54 much prosperity that there'll be more
1:35:56 than enough for everyone. It's it
1:35:58 effectively all goods and services
1:36:01 become infinitely
1:36:04 inexpensive. So everyone can have
1:36:06 everything they want. And then you start
1:36:08 to ask the question, well what has
1:36:10 meaning? Well, that's our job as humans.
1:36:14 Go make
1:36:17 meaning. Tik Tok question. Why are they
1:36:22 dictating the way the world works? Who
1:36:25 are they? I don't know what that
1:36:28 means. Why are they Sam
1:36:38 Alman? They're not dictating the way the
1:36:40 world
1:36:42 works. Well, Musk is.
1:36:48 He's doing that for power, right? You
1:36:50 know, why do why do rich people do what
1:36:53 they
1:36:54 do? Why do rich people try to gain and
1:36:58 expand their power? I don't know.
1:37:00 Because that's that's what rich people
1:37:03 do.
1:37:07 Um, I would argue that that Sam Alman is
1:37:10 not dictating the way the world works. I
1:37:12 think what Sam Alman is doing
1:37:16 is has has recognized that that this is
1:37:19 a profoundly
1:37:21 important that AGI artificial general
1:37:24 intelligence which people are arguing
1:37:27 that the model that they launched today
1:37:29 is really the first model that that is
1:37:33 that where where it's as good or better
1:37:36 than most humans at most jobs. I would
1:37:38 probably argue that it's not that yet,
1:37:40 but we're pretty [ __ ]
1:37:42 close.
1:37:46 Um, I think what these companies are
1:37:49 trying to do
1:37:51 is compete to be the two or three
1:37:55 companies that survive the shakeout
1:37:57 that's coming in the next five years,
1:37:59 which every time there's a new
1:38:00 technology, you start out with 20
1:38:02 companies and it whittleles down to
1:38:03 three
1:38:04 logos. There's always three. Like when
1:38:07 it was browsers, it was Netscape and AOL
1:38:10 and something else, right? It was always
1:38:12 three brow. It's always three logos and
1:38:15 it always distills down to three logos.
1:38:17 So So they're all fighting to be one of
1:38:19 those three logos.
1:38:24 um what what Open AI and quite frankly
1:38:28 what Google's doing right now, what Meta
1:38:31 is doing right now, what Anthropic's
1:38:32 doing right now is they're trying to get
1:38:36 this technology to the point that it's
1:38:38 good enough that it can sort of live
1:38:41 into this vision that I just described.
1:38:43 It's not good enough right now. It's all
1:38:44 jankier than [ __ ] So, technology will
1:38:48 always evolve. It always
1:38:50 has.
1:38:53 Um, if they succeed in making the
1:38:56 technology that
1:38:58 good, the existence of the technology is
1:39:01 what's going to change the world. It's
1:39:02 not them saying you have to use AI.
1:39:06 It's them saying, "I want to create a a
1:39:09 a piece of technology that is so
1:39:11 compelling and
1:39:13 powerful that people change their
1:39:16 behavior to do things this new
1:39:21 way." Like, could we choose to not use
1:39:24 AI and just keep [ __ ] sitting on hold
1:39:28 for four or five hours waiting for a
1:39:30 customer service representative to not
1:39:32 be able to help us and then hang up on
1:39:34 us? We
1:39:36 could. Do you want
1:39:38 to? I would like to have that not be in
1:39:41 my life
1:39:44 anymore. Wouldn't it be nice
1:39:47 that the the IRS creates these tax codes
1:39:50 that are so complicated they're designed
1:39:52 for only the rich people to be able to
1:39:54 leverage the loopholes because they have
1:39:57 the best accountants in the world. Well,
1:39:58 what if we all had the best accountants
1:40:00 in the world? [ __ ] them. And now taxes
1:40:04 become infinitely
1:40:06 simple, then they'll change how taxes
1:40:08 work, right? So, so we could hang on to
1:40:11 the way we do things now. I just I like
1:40:14 I, you know, I I don't see this as as
1:40:18 maniacal as as a lot of people think it
1:40:21 is or or perceive it is. I think this is
1:40:26 people who truly believe, you know,
1:40:30 listen, I was there in the early days of
1:40:32 the worldwide web. And what I can tell
1:40:34 you is if you were there in '94, '95, 96
1:40:37 in New York City, in San Francisco, in
1:40:40 in sort of the major hubs where the web
1:40:43 developed, there was only a handful of
1:40:45 people that were doing anything in that.
1:40:47 And like what we knew was that this
1:40:50 technology was going to change the
1:40:51 world.
1:40:53 And if you look at some of the 60
1:40:55 Minutes um episodes from the era
1:41:00 around 98 99 2000 right before the
1:41:05 bubble burst when it was at its maximum
1:41:07 hype and everyone was throwing all this
1:41:10 money at the worldwide web and
1:41:11 businesses were getting funded that had
1:41:13 no right being funded. Everyone was
1:41:16 like, "Oh, this is just a bunch of
1:41:18 [ __ ] suckass punks trying to change
1:41:20 the way tra change change trying to they
1:41:22 believe they can change the world and
1:41:24 they're just a bunch of punks trying to
1:41:25 [ __ ] up the way we do
1:41:28 things.
1:41:31 Um, it got very perverted, like the
1:41:33 perception of it got very perverted once
1:41:36 all the money came in." And so what
1:41:38 we've got with AI is we've got a
1:41:40 profoundly powerful tool that like the
1:41:44 amount of money that's come
1:41:47 into AI
1:41:49 development makes like it's almost like
1:41:52 the web wasn't even funded. Like as
1:41:54 hyped up as as as the the.com bubble
1:41:57 was, that is [ __ ] nothing compared to
1:42:00 what's happening in AI right now.
1:42:03 So when you take that much money and
1:42:05 then you put it into a tool that the
1:42:07 more money you throw at it, the better
1:42:09 it gets, you're going to get this crazy
1:42:11 ass acceleration. And if it gets as good
1:42:13 as it looks like it's getting, it's
1:42:15 going to change everything. So I I don't
1:42:17 know. I mean, I I like I I just I don't
1:42:21 buy the, you know, there's there's two
1:42:23 or three people that can just change
1:42:25 everything. Um I
1:42:28 think current administration is trying
1:42:30 to do that.
1:42:32 Um, it'll normalize at some
1:42:34 point. I think we grew up with sci-fi
1:42:37 movies that portrayed cool technology
1:42:38 that seemed unattainable. We're now
1:42:40 beginning to see come true. I
1:42:43 know I talk about we're living in a
1:42:46 postci
1:42:47 world. Like, we grow up with movies that
1:42:50 that a lot of the stuff we're playing
1:42:52 with today. That was my very first my
1:42:55 very first reaction when I used chat
1:42:58 GPT was
1:43:04 I'm in a sci-fi
1:43:06 movie. And the second thing that popped
1:43:08 into my head was the Turing test just
1:43:12 just got
1:43:13 passed, which it probably hadn't, but at
1:43:16 least in my head, like I'm interacting
1:43:18 with this thing that's just [ __ ]
1:43:19 talking to me. It's not programmed with
1:43:22 logic.
1:43:23 Like I I don't know if you paid
1:43:26 attention to this. People have been
1:43:27 talking about AI chat bots for about 20
1:43:31 years cuz AI's been
1:43:33 around and chat systems have been around
1:43:37 and chat systems that have AI underneath
1:43:39 them have been
1:43:42 around. But
1:43:44 the the way they worked was a logic tree
1:43:47 that human beings went in and said, "If
1:43:50 the user says this, then answer this."
1:43:54 And then at some point it might use AI
1:43:56 to go figure out, you know, based on
1:43:58 this user, which movie should I
1:44:01 recommend to them, right? Like AI was
1:44:04 used as as this sort of enabling
1:44:05 technology underneath shitty logic
1:44:09 trees. So for 20 years, chat bots and AI
1:44:14 chat bots were these [ __ ]
1:44:16 horrifically disappointing things.
1:44:20 And then Chat GPT comes out and all of a
1:44:22 sudden it's like holy
1:44:25 [ __ ] This is This is like the
1:44:30 movies. They They did it. Like that that
1:44:33 was like my immediate response was they
1:44:36 did
1:44:36 it. This is that this is that thing.
1:44:40 This is what they've been promising.
1:44:44 You know, the week after Chat GPD
1:44:46 started, I started the AI salon and I
1:44:48 started this Tik Tok channel. I think I
1:44:50 started them, you know, like the day or
1:44:52 two after Chat GPT launched cuz cuz I've
1:44:56 been waiting for this [ __ ] for my whole
1:44:59 life and like it's
1:45:03 here and we've got a lot of people like
1:45:05 sitting on the sidelines like I don't
1:45:07 like it. It makes it makes
1:45:10 typos. Okay. Do do you know it can write
1:45:14 code for
1:45:16 you? Do you know there's a thing called
1:45:18 vibe coding now where you can just sort
1:45:20 of sit there like Rick Rubin with your
1:45:22 big [ __ ] beard and your big long hair
1:45:24 just going I think I'll make an app that
1:45:27 you know talks about the growth of
1:45:29 fingernails today. Okay, it'll make that
1:45:31 for
1:45:32 you. That's a reality
1:45:36 today. People are bitching about it.
1:45:38 Well, you know, I I don't know. I don't
1:45:41 get it. But I do. Change is scary. I got
1:45:45 my bot on Discord. Then I was able to
1:45:47 give it conversational voice with 11
1:45:49 Labs. Oh, that's super cool. And this is
1:45:51 just the beginning. Once you add
1:45:52 advanced robotics and quantum Yeah.
1:45:54 quantum computing, it gets insane. It
1:45:56 gets insane.
1:46:00 We we as a
1:46:03 species have
1:46:05 never ever
1:46:11 ever lived in a
1:46:13 time where change is
1:46:17 exponential. So, so you know how that
1:46:20 works. Do you do you know how that
1:46:22 works? So, linear change is what we've
1:46:25 always experienced.
1:46:28 You have a thing here, you improve it
1:46:33 and it's, you know, it's this much
1:46:36 better and then the next improvement is
1:46:38 this much better and it's linear and
1:46:40 it's
1:46:42 predictable. Exponential change starts
1:46:44 out really flat. It seems like nothing's
1:46:46 happening at all. David Shapiro talks
1:46:48 about this. He goes, it's nothing,
1:46:50 nothing, nothing, and then all at
1:46:52 once. We're in this part of this curve,
1:46:56 the curve that goes
1:47:00 Right. 50 years of neural networks and
1:47:03 somewhere in here was big blue beat Gary
1:47:06 Casper off and then we did the, you
1:47:08 know, there, you know, there there were
1:47:10 sort of specific AI milestones along the
1:47:13 way, but it's basically flat. It has
1:47:15 nothing to do with
1:47:17 society. And then GPT3 comes out and
1:47:20 it's kind of like it's pretty good. And
1:47:22 then chat GPT comes out and that's the
1:47:25 first thing where we've got a
1:47:26 generalurpose tool. This is really
1:47:28 important. General purpose
1:47:30 tool that can do everything from recipes
1:47:33 to kids books to physics to mathematics
1:47:36 to
1:47:37 programming to business plans to
1:47:39 marketing
1:47:40 plans to analysis, data
1:47:45 analysis. And and so
1:47:49 November 30th,
1:47:52 2022, there's a little blip in the
1:47:55 flatness, right?
1:47:57 And then we're just slowly doing this.
1:48:00 And what happens is you don't just go
1:48:02 from like one to two, two to three,
1:48:05 three to four. You go from like 1 to two
1:48:09 to 4 to 8 to 16 to 60 or to 32 to 64,
1:48:15 right?
1:48:18 The changes are
1:48:21 accelerating and they're going to
1:48:23 accelerate so fast that it'll be
1:48:26 impossible to even comprehend them. And
1:48:28 I think that's starting like that's why
1:48:33 this feels so [ __ ]
1:48:35 weird is we've never experienced this
1:48:38 before. If you like checked in on chat
1:48:40 GPT like in 2023 and you're like, "Oh
1:48:43 yeah, I'll do chat GPT." It's whatever.
1:48:44 It's like Google. Yeah, I get it. And
1:48:47 then you went away for two years and now
1:48:49 you're coming back because maybe you saw
1:48:50 that image gen thing. People are making
1:48:52 those studio Giblly pictures and they're
1:48:54 making the the
1:48:56 the blister packed dolls of themselves.
1:49:00 Oh, how are they doing that? Let me go
1:49:01 check that out. And then you go into
1:49:03 chat GPT and you're like, holy [ __ ]
1:49:05 what is all this
1:49:09 stuff? And then if you look at what they
1:49:11 launched today, even though it doesn't
1:49:13 work now because everyone's overusing
1:49:15 it. Um um where it's it starts thinking
1:49:20 for
1:49:21 itself. How do you process that? This is
1:49:24 one of this is one of the things that
1:49:25 that I just it's why I started this
1:49:27 channel. I want as many people as
1:49:28 possible to try this [ __ ] so that when
1:49:31 the the exponential starts to take off,
1:49:34 we're at least
1:49:36 aware that it's taking off. Most people
1:49:39 don't most people haven't even heard of
1:49:41 chat GBT. They don't know what it does.
1:49:44 They're like, "Hey, it's that AI stuff
1:49:45 that's that's for the geeks." They don't
1:49:47 get that it's not for the geeks. It's
1:49:49 for all of us. You don't have to be a
1:49:52 mathematician to use chat GPT. You
1:49:55 can How
1:49:57 long do you think they will have to
1:50:00 throttle this back on us? When When will
1:50:02 they turn it loose?
1:50:05 Um, listen, I think
1:50:10 that I think the frontier models are are
1:50:13 in a kind
1:50:15 of they're in a bit of a bind, right? If
1:50:19 they just turned it loose in all of its
1:50:22 power and it really could sort of make
1:50:26 itself autonomous and cut us out of the
1:50:29 loop and shut down our power grid
1:50:31 because it doesn't like the way we do
1:50:33 stuff and it makes its own power off
1:50:35 somewhere else. Like that could be bad,
1:50:38 right? So I think they've got to put
1:50:40 some guard rails in there. But then we
1:50:42 as humans were like we don't want any
1:50:44 guard rails. We're Americans. We want it
1:50:46 to do
1:50:47 everything. So, I don't I I I I think
1:50:52 because these tools are new and because
1:50:55 computers are acting in ways they've
1:50:58 never acted before, I think having some
1:51:01 controls right now is
1:51:02 fine.
1:51:04 Um, is it really frustrating if you want
1:51:07 to make like a horror film and it's like
1:51:11 you can't use the word bloody, you have
1:51:13 to use a guy covered in red
1:51:15 paint. Yeah, that's a bit of a pain in
1:51:18 the ass, but at the same time, like, you
1:51:21 know, you got to balance the safety
1:51:24 while we're figuring out what these
1:51:25 tools even are. Start comments and
1:51:27 StreamYard rapid fire. Okay, cool. And
1:51:29 then time check. Yeah. Okay. So, let's
1:51:30 do
1:51:31 some Okay. And then, holy
1:51:36 [ __ ] I don't know what that was in
1:51:38 reference to, but I like it. Um, okay.
1:51:42 Valerie, we are at the point in the very
1:51:44 near future where if it's possible, it
1:51:46 will be reality. You know, I was really
1:51:49 struck. Brandon the other night talked
1:51:51 about his four-year-old child um who
1:51:54 made a song inso
1:51:57 um and now says, "Daddy, I want to hear
1:52:00 the song I
1:52:01 made."
1:52:03 Um that kid is growing up in a world
1:52:08 where you can ask for what you want and
1:52:11 it will it will be manifested, right?
1:52:16 You can you can have an idea in your
1:52:17 head and bang it's in the
1:52:22 world. We're get you know what's you
1:52:25 know what also is really remarkable
1:52:26 right now we are
1:52:30 in we are the last
1:52:35 generation that will know what it's
1:52:38 like to not have AI.
1:52:43 Brandon's four-year-old kid won't know
1:52:46 that
1:52:49 world. I mean, think about it. When
1:52:52 Brandon's fouryear-old
1:52:54 is
1:52:56 12, eight years from
1:52:59 now, eight years from now, like we've
1:53:01 been we've been doing this, you know, if
1:53:03 you count chat GPT is the start. Two and
1:53:06 a half years. two and a half years we've
1:53:08 gone from chat GPT to all this crazy
1:53:12 image gen stuff, video stuff, 3D world
1:53:15 building stuff, math, programming,
1:53:18 talking to dolphins in in in [ __ ]
1:53:22 three
1:53:24 years. Imagine eight years from
1:53:27 now what his 12-year-old boy is going to
1:53:29 be able to ask for and
1:53:33 get. Oh, you know what would be cool?
1:53:36 Hey, Dad. You know what would be cool?
1:53:38 What if we started up like a micro
1:53:40 micromanufacturing plant where um we
1:53:43 took we took garbage from from the store
1:53:47 and we converted that into plastic and
1:53:49 then we turned that plastic into these
1:53:51 cool sneakers that I have and then we
1:53:53 sell them to NBA
1:53:56 players. And then Brandon's like, "Yeah,
1:53:59 cool. Yeah, that sounds good. All right,
1:54:02 cool. I'm going to go start that." heads
1:54:04 off to his room, doesn't type. Typing is
1:54:07 for old people, just talks. Hey, uh, I
1:54:10 want to start a new business today.
1:54:11 Okay, what's it called? I don't know.
1:54:13 Give me 30
1:54:15 names. And then a month later, he's
1:54:18 selling sneakers to the
1:54:23 NBA. I I like I don't
1:54:26 know. Kid's going to own his own robot
1:54:28 for his 16th birthday. You damn right he
1:54:30 will. All right, Brian.
1:54:33 I've been waiting for AI since Eliza in
1:54:35 the 80s. I know. Honest to God, like
1:54:37 every time Google talked about
1:54:39 TensorFlow or the the AI thing. It was
1:54:42 talked about on our podcast today about
1:54:44 2018, the the Google whatever event
1:54:47 where Sundar Pchai showed the thing
1:54:50 where the the AI called the restaurant
1:54:52 and we're like, we want that and like we
1:54:54 but we don't have it. It's been decades
1:54:57 where we've been promised AI chat chat
1:54:59 bots and cool AI [ __ ] and AI is going to
1:55:02 change the world. Well, now we're
1:55:07 here. Now we're here and people are
1:55:10 bitching about it. Yeah, but it doesn't
1:55:11 spell right all the time.
1:55:18 Well, still waiting for the GPT store. I
1:55:21 think we'll be waiting that for a while.
1:55:22 True, Tom. It's been said that even if
1:55:24 they changed nothing with AI going
1:55:26 forward, it would take us decades to
1:55:28 figure out the things it could
1:55:29 do.
1:55:31 I just with with GPT40 with the base
1:55:35 model, the non-reasoning
1:55:37 model, if they stopped it today, they
1:55:40 said, "Okay, no more advancements in AI.
1:55:43 Everyone's just going to use chat GBT40.
1:55:47 We're we're going to we're g we're going
1:55:48 to liquidate and and erase all of the
1:55:51 code for all the image generation tools.
1:55:54 Midjourney's gone. Song things are g
1:55:56 everyone everything's gone. Just chat
1:55:58 GPT 40 basic chat model. Yeah, I think
1:56:02 we could explore it for 10 years and not
1:56:04 hit the bottom of it.
1:56:07 Absolutely. Oh man, I typed Eliza into
1:56:10 my Commodore 64. Nice. Okay.
1:56:14 My first imp impression when I use it
1:56:16 was, "Oh my gosh,
1:56:19 uh, I'm ahead of my colleagues." Yep.
1:56:22 Yeah. Yeah.
1:56:25 Exactly. Exactly.
1:56:30 it. I I find it
1:56:33 really
1:56:36 depressing that so many people are not
1:56:40 only like if people are ignorant of AI,
1:56:44 I don't hold that against
1:56:46 them. It's the people that are
1:56:52 willingly, it's beyond staying ignorant.
1:56:55 They're
1:56:57 willingly resisting it. It just makes it
1:57:00 It's really depressing to
1:57:03 me because I kind of feel like if they
1:57:06 took that passion and put it into
1:57:09 exploring AI and what it makes possible,
1:57:12 imagine the [ __ ] difference they
1:57:14 could make in the
1:57:16 world. And what's going to happen
1:57:19 is the AI is going to accelerate away
1:57:22 from all of us. And those people are
1:57:24 going to be like, "Oh, I get it now.
1:57:26 Wait, can
1:57:29 I how do
1:57:31 I Now, it's also possible that the that
1:57:34 acceleration is it gets so easy to use
1:57:37 AI that everyone just uses it. Anyone
1:57:39 who resisted is like, oh, [ __ ] it.
1:57:40 That's definitely a better way to do the
1:57:43 [Laughter]
1:57:45 world." Oh, man. Okay. All right,
1:57:49 people. Some people need to take more
1:57:51 time to process. Absolutely.
1:57:54 Everyone's a prompt genius all of a
1:57:56 sudden. Exactly. Uh transformational
1:57:59 information available through AI.
1:58:02 Willing resistance is just human nature.
1:58:04 I know everybody resisted. I know every
1:58:07 single technology has been resisted. It
1:58:09 just makes me sad because this one this
1:58:12 is like nothing I've ever experienced in
1:58:13 my life. And like I want everyone to
1:58:15 experience it. I on this channel the the
1:58:18 folks that are here all the time know
1:58:19 this. Like my hope is that every person
1:58:22 I touch experiences their first Kevin
1:58:25 Mallister moment with AI which is that
1:58:28 was from Home
1:58:29 Alone. That
1:58:34 moment you can't unhave that moment with
1:58:38 AI. And you also can't explain
1:58:41 it. You can tell someone, "Oh, it's so
1:58:43 amazing." But you just sound like a
1:58:45 [ __ ] cult member.
1:58:50 Please don't touch the
1:58:58 viewers. All right, I'm going to get out
1:59:00 of here. So, what's today? Today's
1:59:02 Wednesday. Um, I've got a whiskey night
1:59:06 tomorrow night. Um, a whiskey club, so
1:59:11 I'll probably be late tomorrow and I
1:59:13 might be a little tipsy.
1:59:16 Isn't he drunk now? He seemed a little
1:59:19 drunk. What? What did he say? What was
1:59:22 he talking
1:59:24 about? I need a
1:59:26 [Laughter]
1:59:31 drink. So, it'll be a little late
1:59:33 tomorrow. And then we got Friday night
1:59:35 date night. Uh, I'm going to try
1:59:36 tomorrow. I think I've got some cycles,
1:59:38 but I got some other [ __ ] to do. But I'm
1:59:40 going to try to
1:59:45 uh at least get a couple of other use
1:59:47 cases going for 01 and 03. And hopefully
1:59:50 they'll have their their servers
1:59:52 reconfigured or rebalanced so so that
1:59:54 the [ __ ] actually runs.
1:59:57 Um anyway, that was kind of fun tonight.
2:00:00 I mean, I like going on rants. That was
2:00:02 a good rant. Good ranty rant. Hope you
2:00:04 had fun. Hope you learned something. Um
2:00:07 join the AI salon. Go to the salon.ai.
2:00:09 AI and click on join our community.
2:00:13 That's the one right there. Go to that
2:00:16 URL. Click on join our community.
2:00:22 Um yeah, that's all I got right now.
2:00:24 Just keep coming back to this. Keep
2:00:26 coming back. Here's the whole point of
2:00:28 this channel, by the way. If you're new
2:00:29 here and and you're just like, is it
2:00:31 always like that? What is this like? Is
2:00:34 it Does he ever It's the learning lab.
2:00:36 Does he ever do the teaching? Does he
2:00:38 have a a chalkboard? What? How does the
2:00:40 Is it a
2:00:41 whiteboard? It's just chaos. Every night
2:00:44 it's chat
2:00:46 add. There's no plan. There's just show
2:00:50 up and then what we
2:00:53 do is just talk about AI. Why do that?
2:00:58 Because be in the conversation. be in
2:01:01 the conversation with other curious
2:01:03 people with people asking really good
2:01:05 like the questions tonight were really
2:01:07 good. They were really
2:01:10 good. Especially the hard ones like wait
2:01:12 a minute isn't this [ __ ] up people's
2:01:14 careers?
2:01:17 Yes. Isn't that bad?
2:01:25 Yes. Is it also a massive opportunity
2:01:29 also? Yes.
2:01:33 Right. And this stuff is changing so
2:01:35 fast that if you're not in the
2:01:37 conversation, it it will accelerate away
2:01:41 from
2:01:41 you. And that's not good. So, just keep
2:01:46 coming back. Okay? So, do me a favor. If
2:01:49 you're on YouTube, please like, you
2:01:51 know, do the do the like thing over
2:01:53 there and subscribe to that channel.
2:01:54 We're trying to grow a subscriber base
2:01:56 there. Okay. Um, you can subscribe
2:01:59 there. There are two paid subscriptions.
2:02:01 If you want to support the channel, you
2:02:03 can do that on YouTube. You can also do
2:02:04 that on Tik Tok. Don't know how long Tik
2:02:06 Tok's going to be around. So, you know,
2:02:08 do it there. And if it goes away, then
2:02:10 that comes off your monthly bill. I'm
2:02:13 new. I've selected 04 mini. Can you give
2:02:15 me a fun prompt to try, please? Uh, love
2:02:18 being in the conversation. Thank you,
2:02:19 Jeff. Um, a fun prompt to try for 01
2:02:22 mini is,
2:02:27 um, let's
2:02:29 see, find me the latest data
2:02:35 on population growth
2:02:38 globally and write a report with pretty
2:02:42 graphs in
2:02:44 it
2:02:45 that talks about the top 10 trends.
2:02:50 and the the things behind
2:02:53 them. Something like that. Just do
2:02:56 something or or take something that you
2:02:58 know and make a ridiculously complicated
2:03:02 or ridiculously ambitious request of it.
2:03:06 Like just give it something that in your
2:03:09 mind you know this thing shouldn't be
2:03:11 able to do and then just see if it can
2:03:13 do it. If nothing else, it'll be
2:03:15 fascinating to watch how it how it tries
2:03:18 to get there.
2:03:20 Um,
2:03:22 yeah, the one that I wanted to show you
2:03:25 tonight where it analyzed the image and
2:03:27 pulled all the the did a catalog of the
2:03:29 [ __ ] it found in it, I did it this
2:03:31 afternoon. It was really cool. So, we'll
2:03:33 do that tomorrow night. All right. I'm
2:03:35 going to watch Celebrity Jeopardy. Good
2:03:36 night. Good night, Harshwede. Perfect.
2:03:39 And that's the perfect the perfect
2:03:41 followup to an AI learning lab.
2:03:43 Celebrity Jeopardy.
2:03:48 Awesome. The cuckoo clock was a great
2:03:50 example. Well, it was, except it didn't
2:03:55 work. Son of a [ __ ] Did it Did I ever
2:03:58 get it? No. Let me see. Try We'll try it
2:04:00 one more time. Try again and then we'll
2:04:02 get out of
2:04:05 here. Oh
2:04:08 man, that's the example.
2:04:12 Yeah, that's the example. Come on, man.
2:04:15 Oh, by the way, labor, if since you're
2:04:17 still here, the reason it can't tell
2:04:19 time, it looks like this is going to
2:04:21 work now. The reason it can't tell
2:04:24 time
2:04:27 is, you may or may not know this, but
2:04:31 every clock ad, every watch ad in the
2:04:34 history of watch ads, the hands are
2:04:36 always at 10 and two.
2:04:39 They're always at 10 and two. Why? I
2:04:42 don't know. It's probably some
2:04:43 psychological geometric whatever that
2:04:47 it's pleasing to humans. So, the hands
2:04:48 are always at 10 and two. And so, you'll
2:04:51 notice as it's generating this clock,
2:04:52 the hands are at 10 and two. You cannot
2:04:56 you cannot get it to not do
2:04:59 this. Now, what it did do here, which is
2:05:02 interesting, is it put the second hand I
2:05:05 told it to do 725. So, it put the second
2:05:08 hand at 25 25 seconds, but it didn't get
2:05:11 the seven. That's because in the
2:05:14 training data, there's no clocks with
2:05:16 hands in different
2:05:17 positions. It's just one of those
2:05:20 things. Now, could someone train a
2:05:23 model? Could they explicitly train a
2:05:25 model with a bunch of different pictures
2:05:27 of clocks and watches and things with
2:05:28 hands at different positions? Yep, they
2:05:30 could. And that would fix it. And so,
2:05:33 there's all sorts of things like that.
2:05:35 Wine glasses are another one. You can't
2:05:37 you can't get it to produce an image of
2:05:39 a wine glass that's overflowing because
2:05:41 pictures of wine glasses, they're always
2:05:42 two/ird
2:05:45 full. But then some things it can do
2:05:48 brilliantly, right? Like like some of
2:05:51 the psychology stuff it does because
2:05:53 like all of the writings of all of
2:05:55 psychology are have just been out on the
2:05:58 internet for decades and so it knows all
2:06:02 of that stuff and it's brilliant at
2:06:04 that.
2:06:05 So that's what it is. All right. I've
2:06:08 got a Becky Rue image and a regulars.
2:06:10 One final image. We'll go one final tour
2:06:13 through a
2:06:15 regulars. Becky Rue.
2:06:20 Nice. Is that Jeff Bezos with my
2:06:23 glasses? That's pretty funny. Nice dodo
2:06:25 bird. Solid. Direwolves. Oh yeah. They
2:06:29 they just un unextincted the direwolf,
2:06:32 so might as well bring back the dodo
2:06:33 bird.
2:06:36 All right, everyone. I'm gonna get out
2:06:37 of here. Have a good night. Enjoy
2:06:38 Celebrity Jeopardy. Thanks everybody.
2:06:40 Thanks for the really good questions,
2:06:41 too.