aiLL 7/6/23 - The Generative AI Revolution is Here! ChatGPT, Stable Diffusion, and More!

Description

Chapters

Transcript

0:01 foreign
0:23 welcome to the AI learning lab I'm just
0:27 going to
0:29 get a news story
0:48 exciting news coming there's big things
0:52 happening
0:54 welcome everybody we'll let people get
0:57 in here it's the AI learning lab my name
0:59 is Kyle Shannon
1:01 let me pop up my little
1:04 I generally start out with this looks
1:06 like Ron's here's that you were on maybe
1:08 I think so
1:12 um
1:14 what was I gonna say oh the website
1:18 everyone get in here
1:22 share it far and wide people share far
1:25 and wide we just have a handful of
1:26 people in here now
1:28 we gotta we gotta get this in here
1:30 hey yellow what's happening
1:33 Victor Troy good evening Mr Decker
1:36 was shaking
1:39 oh Silver Fox shared the live if you
1:41 haven't shared the live you're you're sl
1:43 you're losing ground it's very sad if
1:45 you haven't shared it yet
1:49 many people are sharing this live
1:51 [Laughter]
1:57 welcome welcome welcome it is what day
2:00 is today Thursday night it's a hot
2:02 Thursday night it's actually kind of
2:05 cool here in Denver
2:06 there's a chance of tennis ball sized
2:09 hail or nothing it's like 30 percent
2:12 30 percent that it's going to be
2:14 catastrophic and then you know 70 that
2:18 it's gonna be just fine everything's
2:20 gonna be just fine
2:21 so I don't know what was it last time
2:24 two nights ago I was on live I guess it
2:26 was July 4th and uh yeah the garbage
2:28 cans washed into the intersection it's
2:30 pretty pretty nasty uh welcome everybody
2:34 Silhouettes hope you're having a great
2:35 day too appreciate it uh welcome my name
2:38 is Kyle Shannon this is the AI learning
2:39 lab let me explain what all these URLs
2:41 are on screen for
2:43 I'm not associated with any of them
2:46 um they're really just there more as a
2:47 public service announcement
2:49 um
2:50 the channel is about talking about AI
2:52 sort of going on this Grand adventure
2:55 together of whatever the [ __ ] this
2:57 generative AI thing is going to be I'm
2:59 an entrepreneur of many years I've
3:00 started 12 or 15 companies in my day
3:02 some of them got really big some of them
3:04 were just abject failures that shouldn't
3:06 have been started in the first place my
3:07 current company Story vines 11 years old
3:09 and
3:10 profitable and we do a lot of good work
3:13 in the healthcare and Pharma space so
3:14 that's kind of fun
3:16 and uh the past year I have been
3:19 obsessed with this AI stuff and when
3:21 chat GPT came out I had an epiphany
3:24 about it
3:25 that was not dissimilar to an epiphany I
3:28 had about the World Wide Web back in
3:29 1994. where I thought holy [ __ ] the
3:33 world just changed and no one knows it
3:34 yet
3:35 and that's that's a feeling about this
3:37 stuff so what
3:39 what's here if you haven't played with
3:41 chat GPT
3:43 um what I will do most of the night is
3:44 just tell you to go do that so check
3:47 that openai.com that's the official chat
3:49 GPT website if you haven't played with
3:50 it go there bing.com is also chat EBT if
3:54 you click on their chat button that's
3:55 Microsoft Microsoft's version of it and
3:59 it's gpt4 and it's free
4:02 um and you can generate images there and
4:05 you can it's connected to the internet
4:07 so it's it's fairly sophisticated not as
4:11 good in my opinion as the gpt4 up there
4:13 but whatever it's free and then
4:14 poe.com's got a whole bunch of different
4:16 things you can play with so uh go play
4:19 with those prompts.chat is a a little
4:22 document that'll explain to you what
4:25 prompting is and how to do it how to
4:26 talk to these things and then it'll give
4:29 you a giant long list of personas where
4:31 you can tell chat GPT to act like
4:33 different personas like act like an
4:35 accountant or a master scribe or a
4:38 nutritionist or whatever and and it gets
4:41 better and better as you give it more
4:42 context and then futurepedia.io is just
4:45 a big long list of uh AI tools
4:49 and Kyle for president oh God that'd be
4:53 a [ __ ] disaster
4:56 you know my problem would be I tell the
4:58 truth and uh then things would be very
5:01 bad
5:02 oh yeah they took me in the back room
5:04 here's what's going on with UFOs I
5:06 know
5:10 all right anyway if you have questions
5:11 about AI pop them down below the other
5:13 thing I thought might be interesting is
5:15 um
5:16 if you've got a problem that you're
5:18 trying to solve and you're you know
5:19 you're trying to figure out how do I do
5:20 that with AI just pop that down below as
5:22 well see what I can see what I can do to
5:25 um
5:26 answer any of that stuff so welcome
5:29 welcome oh [ __ ] wrong shirt I'm gonna
5:32 need to go change my shirt you can't you
5:34 can't see a thing here
5:36 hope you're having a great day all right
5:38 um all right looks like you have a
5:40 popped any questions down there so feel
5:42 free to pop questions and I'm gonna go
5:43 change my shirt so we can read the text
6:03 thank you
6:09 [Music]
6:15 all right
6:17 got a little Basquiat Cassius Clay
6:22 nice
6:24 all right shirt changing commences oh
6:27 yeah I should have put on some music
6:31 [Music]
6:34 um am I doing an AI demo this evening
6:37 yes what would you like to see
6:40 greetings earthlings hello Kevin
6:43 it's loud I like it yeah it was loud but
6:46 it was not you couldn't read the text I
6:47 couldn't read the text it was it was
6:49 making my head hurt
6:51 and we don't want daddy's head hurting
6:53 that would be bad what the hair is
6:55 starting to misbehave again I got it I
6:57 got a haircut it was good for like three
6:59 days four days now it's
7:01 it's starting to head south again that's
7:04 like a two-pay tussle laughs
7:11 how about teaching us something about
7:13 prompts okay I can do that
7:17 I'll tell you what might be good
7:20 [ __ ] duck I get it
7:23 you switch those letters around and if I
7:26 wasn't careful I'd say a bad word
7:29 very clever of you [ __ ] duck huh
7:33 I'm too smart for you
7:36 like a steel trap
7:39 so are you doing an AI demo and teach us
7:42 something about prompting so what you
7:43 should do [ __ ] duck is pop down below
7:46 what's something that you're trying to
7:48 solve so rather than me just randomly
7:49 putting [ __ ] in there because I tend to
7:52 go to the same stuff just because I'm
7:54 thinking fast on the fly so I go to I go
7:56 to the same [ __ ] over and over again so
7:58 put something down there and then we'll
7:59 see if we can uh I'll see if I can come
8:02 up with a creative way to play with it
8:04 oh there's some news
8:06 so if you pay for
8:09 um open AIS API everybody has access to
8:13 gpt4 now I didn't have access to gpt4 on
8:16 the API side I do know everybody has it
8:20 and the other thing that you've been
8:21 hearing me whine about for I don't know
8:24 three weeks now four weeks is that I
8:26 don't have access to code interpreter
8:28 for chatgpt apparently everyone who has
8:32 a paid
8:33 chat GPT member is getting access to
8:36 code interpreter in fact why don't we go
8:37 check it now maybe they gave it to me
8:39 tonight but everyone is supposed to get
8:41 access to it within the week
8:44 beta features it's not there yet but
8:47 apparently within a week we all got
8:48 access to code interpreter Ethan malaku
8:51 I'm a big fan of said what does what did
8:53 he say in his tweet hang on
8:56 um
9:01 let me see code interpreter
9:04 here we go mullik
9:08 what he said he said
9:11 where's this post
9:18 everyone who subscribes oh let me flip
9:21 this around
9:22 so Ethan Malik Wharton professor
9:25 brilliant dude definitely worth
9:27 following
9:28 everyone who subscribes to chat EPT plus
9:30 is apparently going to get code
9:32 interpreter in the next week it is the
9:34 mode of AI that I find the most
9:36 promising interesting and unnerving
9:41 he
9:43 it's coming and if you don't know what
9:46 code interpreter it is you upload files
9:48 to it and it reads them and understands
9:50 them and then it can do [ __ ] for you so
9:52 like you can upload a big pile of data
9:54 and then say to it uh what's interesting
9:57 about this data and it will write you a
9:59 whole report and generate graphs and
10:01 charts
10:02 it'll write the python to generate
10:04 it's insane
10:06 it'll turn gifts into MP4s it'll pull
10:10 the colors out of a gif and create a
10:13 color palette it I I don't know what it
10:15 does because I don't have it yet
10:18 but when I have it I'll play with it
10:21 all right let's see
10:28 who's falling asleep [ __ ] up
10:33 how about teaching us something about
10:34 prompts and then you're snoring on me
10:36 now I'm not going to teach you something
10:38 about prompts
10:39 that's just rude
10:41 Kyle have you checked out threads got AI
10:44 learning Labs set up over there
10:46 um
10:47 I set it up with uh
10:50 with my Kyle Shannon account
10:53 but um
10:55 no
10:56 I mean I played with it I did one post
10:59 um let me see
11:02 can you do they have a
11:04 a web app like the Twitter threads uh
11:08 meta
11:11 -doo
11:12 I guess not
11:16 but if you go look up uh at Kyle Shannon
11:19 on threads my first post was a
11:21 mid-journey image that I created today
11:23 that was really cool you want to see it
11:25 I'll see a cool image
11:27 flop ducks like what are you gonna teach
11:29 me about prompting man
11:32 that's just how he talks dude
11:34 this is a little dark but that's the
11:36 image that I created Isn't that cool
11:39 paper cut shapes and stuff like that
11:43 that was my first post to threads I
11:46 figured I'd do a purdy image put a purdy
11:48 image up there
11:50 greetings I agree about the prompts wait
11:53 given all that's going on in the world
11:56 wait earthlings is appropriate wait
12:00 given all that is going on in the world
12:03 earthlings is appropriate oh someone
12:06 said haircut looks haircut looks rocking
12:09 that is well it does I guess it does it
12:11 looks like bad what is that late 80s
12:14 early 90s rocking
12:16 it's like it's like it's like I don't
12:20 know Wham
12:23 Wham with a receding hairline
12:28 thank you thank you very much for that
12:30 Rose that's for the uh receding hairline
12:33 yeah that's she's selling hair growth
12:36 product right yeah she's she knows
12:38 what's going down up here
12:41 20 bucks a month for
12:44 um Poe yes is poe legitimate
12:47 I would just cancel my open eye AI
12:50 subscription okay so here's the deal
12:53 well the middle one's free so you can't
12:55 pay for that I pay for open Ai and I pay
12:58 for Poe
13:00 I will cancel po as soon as open AI
13:04 expands their context window here's why
13:07 I think they're both worth
13:09 paying for and and the announcement
13:11 about code interpreter being launched
13:13 within a week
13:14 makes me tip heavily toward
13:17 um chat gbt so the top one
13:21 um so the top one if you pay for chat
13:22 GPT plus you get gpt4
13:28 um
13:28 you also
13:30 get plugins
13:33 when they when they resurrect
13:36 um browse with Bing the the gpt4
13:39 connected to the internet they just
13:40 killed it three days ago because it was
13:42 stealing
13:43 um um content behind paywalls
13:46 um I think I understand what's actually
13:47 what happened there but uh but they got
13:49 to fix some [ __ ] so that'll be back so
13:52 you'll have the default mode which is
13:53 gpt4 you'll you'll have uh it connected
13:57 to the internet you'll have plugins
13:58 which are interesting but flawed right
14:00 now and then you're gonna have this new
14:02 thing called code interpreter so paying
14:05 for chat GPT plus gives you all of those
14:08 there's other [ __ ] coming and and you'll
14:10 get access to it first but code
14:11 interpreter
14:13 seems like it's a game changer uh just
14:16 just for it just seems like a game
14:19 changer like it's it's I haven't been
14:21 excited about something in a while about
14:23 like I want the new thing to come that's
14:25 what I'm excited about okay so that's
14:27 that
14:28 now let's go to Poe pow pow pow
14:32 and here's why I think Poe is
14:35 potentially worth paying for but you
14:37 know what you're trying to accomplish
14:38 better than I do
14:40 come on focus focus focus
14:43 all right well it's not going to focus
14:45 okay so when you come to Poe
14:47 you've got these six models here Sage
14:49 gpt4 Cloud plus Cloud instant 100K Cloud
14:52 instant and chat GPT
14:54 the second the the two three and four
14:58 gpt4 Claude plus and Claude instant 100K
15:02 you have to be a subscriber for Claude
15:04 instant 100K what that means is it's a
15:07 hundred thousand token
15:08 chat chat model so translate it into
15:13 like if real people were to say that
15:15 Kyle what would they say you can put 75
15:18 000 words into your prompt
15:20 so that's like a novel you can if you've
15:23 got the Great American novel written and
15:26 you want to summarize it or you want to
15:27 reorder the chapters or you want to make
15:29 it written in the I don't know the tone
15:32 of
15:34 some favorite author of yours
15:36 um you can just paste it in there and
15:38 start interacting with it 75 000 word
15:40 context window so that's a big deal
15:43 um Poe I like like the the it's worth
15:46 using
15:48 um for the free version just because of
15:49 the Bots because I think the the little
15:51 body ecosystem is pretty cool I think
15:52 the Bots are actually probably more
15:55 useful than plugins are right now
15:58 um but that's the difference between
15:59 them recommendation
16:01 um I don't know I would lean toward if
16:03 you could only afford one I would go
16:05 toward open AI because I think code
16:07 interpreter is gonna
16:09 I won't change the game don't change the
16:12 game what is the game what game are we
16:14 playing Kyle why is your hair so funny
16:18 nice code interpreter I know
16:20 where do you get the news on open AI on
16:23 the Twitter and maybe now on the threads
16:26 did you see did you see that that tweet
16:29 that [ __ ] Zuckerberg sent to Elon
16:32 Musk Elon Musk sent him a tweet going
16:34 hey looks like you stole my you know
16:36 intellectual property for for Twitter he
16:39 goes he said something about
16:41 um Zuckerberg cheating and then and then
16:45 Zuckerberg goes well what about you
16:47 cheating with uh Sergey brin's wife it's
16:49 like oh
16:52 as muscle says the girls are fighting
16:56 oh my God
16:58 I think all of this I I here here's
17:01 here's the theory for you maybe Zuck
17:03 released threads just to Hype up the the
17:05 cage match that they're gonna go do
17:08 can mid-journey do tasteful Modern Art
17:11 more minimalist yeah can absolutely do
17:14 that
17:15 um let's go do one
17:18 oh you know what we'll do we'll go to
17:20 Pub so here we are in po
17:23 and I'm going to click the mid Journey
17:24 bot
17:26 and then I'm going to say
17:29 painting
17:32 let's see painting
17:35 of
17:37 colorful
17:39 abstract
17:42 abs
17:44 direct shapes modern
17:50 minimalist
17:53 we'll do 50s and then let's
17:56 just do that I'm going to let this thing
17:58 rewrite it so what it re rewrote was
18:00 colorful paint color painting of
18:04 colorful abstract shapes in a modern
18:06 minimalist style from the 1950s bold and
18:09 vibrant colors geometric and symmetrical
18:11 composition clean lines
18:14 sharp edges textured surface acrylic on
18:17 canvas Jackson Pollock mark rothko
18:21 well
18:22 okay that's just [ __ ] random you
18:25 can't have Pollock and Mondrian in the
18:27 same prompt so so we'll unfuck that up
18:31 that's you know one of the cool things
18:33 about playing with these these tools is
18:35 like like with mid-journey and and
18:37 understanding what it can do and what it
18:39 can understand is like you do get a bit
18:41 of an art history lesson
18:43 that was me trying to do oh that's up
18:46 with this Focus
18:50 ain't there we go
18:52 um this is me trying to do Sling Blade
18:53 it didn't quite get it it's in the
18:55 neighborhood but it wasn't quite right
18:56 all right image
19:00 all right so we're gonna do that we're
19:02 gonna pull out Jackson Pollock
19:04 because that's not minimalist at all
19:06 that is maximalist
19:08 so we'll see what this does oh and it
19:10 gave me a two by three aspect ratio
19:13 which I don't like
19:15 but we'll play with it we'll play with a
19:16 couple of cool things here there's a new
19:18 thing called
19:19 um zoom out which if you haven't played
19:21 with it is is their word for out
19:23 painting which Zoom outs much more clear
19:26 and then there's also panning you can
19:29 now pan you can so if you've got an
19:31 image here and you want to like extend
19:33 just one side of it you can pan it in
19:35 that direction and it actually makes the
19:37 image bigger
19:40 all right so that's going off and doing
19:42 its thing looks like mid Journey's being
19:44 slow so we'll come back to that let's
19:47 see awesome awesome
20:00 where's the picture it's coming
20:03 where's the picture it's just the most
20:06 remarkable piece of technology in the
20:08 history of humanity why isn't it done
20:10 yet
20:11 my microwaves faster than this with a
20:13 hot dog
20:16 here they come
20:18 here they come they're coming is it done
20:21 looks like it's done
20:25 [Music]
20:27 I don't like this let's see
20:32 let's get let's take Mondrian out of
20:34 that let's do another one slash imagine
20:38 pop that in get rid of Jackson Pollock
20:42 let's just do Mark rothko get rid of
20:44 mondrion
20:45 and then instead of bold and vibrant
20:48 Colors Let's do
20:50 um
20:51 bold
20:53 dramatic colors
20:57 um symmetrical composition clean line
20:59 sharp edges textured surface acrylic on
21:01 canvas and let's do 16x9
21:06 all right so we'll go do that and then
21:08 I'll come back and I'll show you some of
21:09 the other stuff
21:11 I just don't like those color palettes
21:12 they just look I don't know what is this
21:14 what are we what are we Easter what is
21:16 this Easter Eastern Modern Art no
21:19 no we want some devil worship Modern Art
21:23 [ __ ] some drama we want drama
21:28 all right I lost my place in a little
21:30 comment here okay here we are modern
21:32 tasteful art I I don't know what code
21:35 interpreter means really okay I'll I'll
21:37 come back to code interpreter uh I'll go
21:39 show some some things but base the basic
21:42 deal with it is
21:44 in your little chat box in chat it gives
21:47 you an upload button and then you can
21:48 upload like say so you're in fundraising
21:51 you could upload
21:53 um you know all of the uh all of the
21:56 foundations or non-profits performance
21:58 for the year let's say you got a big
22:00 spreadsheet with all the performance
22:02 data of all your non-profits and you
22:03 could say tell me what's interesting
22:05 about this data tell me the trends tell
22:07 me what's going on here and it will just
22:10 read all of the data and it'll write you
22:14 charts it'll analyze the data it you can
22:17 tell it to visualize stuff it'll write
22:19 code for you it's
22:22 it's crazy I'll go find a post on
22:25 Twitter these are actually pretty good
22:26 it's the they're they're not quite the
22:28 color palette I would go for but they're
22:30 this prompt right colorful painting
22:32 abstract with Martha Mark rothko and
22:35 then
22:36 and then um it um there's four different
22:40 images here with with these abstracts
22:42 it's hard to tell there's four different
22:43 ones
22:44 um so you got one two three four and
22:46 then you've got upscale one through four
22:49 variations one through four and you
22:50 could just re-roll it basically just say
22:52 I don't like any of them give me a bunch
22:54 of new ones but let's see I think
22:57 I think one and three are interesting to
22:59 me so let's do variations of one and
23:02 variations of three
23:04 and so now that's going to have
23:06 um uh mid Journey go do things that will
23:09 look very very similar but they'll have
23:11 slight variations and then we can blow
23:13 one of those up and then we could let's
23:15 say we wanted this thing to be square
23:16 instead of landscape we could do that as
23:19 well
23:22 what's going on here
23:41 oh there we go that's that's looking
23:43 interesting I like this one with the big
23:44 dark circle
23:50 they're not quite there yet this one's
23:52 done
23:53 wait oh no
23:56 they're still coming they're still
23:58 coming 93 percent
24:00 so you kind of watch these things if you
24:02 haven't watched these things work what
24:03 they actually start with is like a fully
24:05 noised image and then over a series of
24:07 30 or 50 or 70 steps it'll denoise it so
24:11 I like
24:12 three and I think two is pretty
24:15 interesting so what I'm going to do is
24:16 I'm going to hit upscale two
24:19 and upscale three and so now it's going
24:22 to give me big versions of those
24:31 here
24:36 by printing some website and then
24:39 printed
24:40 um
24:41 are you kidding
24:44 holy [ __ ] that was two recipes yeah I
24:47 know I'm so sorry it's all right all
24:51 right
24:52 so
24:54 so there's one of them that's actually
24:56 pretty beautiful
24:58 and then oh it's great and then there's
25:01 another one
25:03 all right so that's oh I was going to
25:06 show you the zoom thing so so let's say
25:08 um
25:09 let's say
25:12 we we like this uh this dark circle here
25:15 I like this dark circle here and let's
25:16 say I want that more in the middle of
25:18 the composition so I can now say pan
25:21 down
25:22 so if I hit this down arrow what it's
25:24 going to do is it's actually going to
25:26 add to the bottom of the painting
25:28 so it's going to give us a whole new
25:30 chunk of painting
25:33 um so this is that's something that just
25:35 came out in mid-journey like yesterday I
25:37 think or the day before
25:40 so that's pretty cool so we'll wait for
25:42 that to do that people are going to flip
25:44 when they get code interpreter I've had
25:45 it for a while now oh the silver tell me
25:47 more tell me more talk about what it is
25:49 yeah I I just like everything I've seen
25:52 in fact Let Me Go
25:54 let me go to the old Twitter
26:00 let's see
26:02 so here here's the thing Ethan Malik did
26:05 he he asked code interpreter to give him
26:09 like different kinds of how did he
26:12 describe it hang on
26:15 he goes I had no idea
26:18 chat EBT with code interpreter could do
26:20 flow charts until it did so I asked it
26:23 what other exotic python visualizations
26:25 it could do and to make samples
26:29 interactive maps pairwise plots word
26:32 clouds hierarchies machine learning
26:34 diagrams Andrews curves so like it can
26:37 just do this [ __ ]
26:39 like you upload you upload data
26:44 to this thing
26:46 and then you say like visualize this
26:49 data on a map give me you know a set of
26:52 Scatter Plots give me whatever the [ __ ]
26:54 these things are you know give me a word
26:56 cloud and it writes the python code and
26:59 Maps the data into it and provides you
27:01 the graphic
27:03 I mean good flipping Lord really
27:07 that's coming this week
27:10 can you explain the math of how neural
27:13 networks calculate gradient descent I
27:16 cannot I can tell you how large language
27:18 models do what they do and I can tell
27:21 you how
27:22 things like mid-journey do what they do
27:24 but what you're asking for is a level
27:26 more machine learning engineer than I am
27:30 unfortunately so here's the uh
27:33 here's the extended image let me I'll
27:36 just upscale one of these let's grab I
27:39 think this bottom one yeah is better so
27:42 we'll do upscale three
27:46 um so I don't understand the math of
27:48 that just just a little bit about
27:50 context about who I am I'm an
27:51 entrepreneur and a you know I'm a I'm an
27:54 innovator with technology once it sort
27:56 of hits
27:57 um
27:58 the the masses right so so when the
28:02 internet was just the internet you had
28:03 to do command line [ __ ] I used it but I
28:06 didn't use it a lot when the World Wide
28:08 Web came out I was like holy [ __ ] I
28:10 started one of the first online
28:11 magazines I started one of the first
28:13 online agencies
28:14 and same thing with chat gbt when when
28:18 this thing came out I'm like holy [ __ ]
28:19 everything's different so I'm I
28:21 understand some of the math behind this
28:23 stuff and how these things work but but
28:25 think you know from chat gbt forward
28:28 essentially
28:30 um so anyway so there's that thing that
28:31 you know the this painting ended down
28:33 here before and now it just added all
28:35 that stuff at the bottom I mean
28:38 remarkable if you ask me
28:41 you know you know just your everyday uh
28:45 every day little
28:47 making art on nothing
28:52 I haven't messed with the others yet but
28:54 I have spent some time finding the edges
28:56 of gpt's capabilities that's cool keep
28:59 digging on it push it harder
29:02 um someone asked me about prompting
29:03 before I'll do something here as soon as
29:05 I get a little inspiration about what
29:06 the prompt I will do it
29:08 retro bug the context window is my
29:10 biggest challenge so if the contacts
29:12 Windows your biggest challenge it's
29:14 probably worth paying for Poe and going
29:17 to go go play with that hundred thousand
29:20 token window or if you don't want to pay
29:22 for it
29:23 there's a open source model called MPT
29:26 7B that has a 65 000 token window which
29:30 is what 45 000 words 40 000 words
29:32 something like that 45 probably
29:34 um
29:36 if you're geeky and want to go play with
29:38 uh open source there's that or pay the
29:41 20 bucks and go play with the Claude and
29:43 hundred thousand one
29:44 where did the pick go which pick
29:49 um I made it go away code interpreter
29:51 gets around the token length issue in
29:53 some ways that's interesting
29:54 where's the picture
29:57 there's the picture
29:59 no it's gone again
30:02 uh um
30:04 when do you think Bing is coming back to
30:06 chat GPT I have no flipping idea I don't
30:09 know how bad that bug was so so here's
30:12 here's what I think okay
30:14 so when you go to
30:17 um when you go to the New York Times or
30:20 when you go to whatever site and it
30:22 shows you the article and then it like
30:23 pops up the thing that says you got to
30:25 pay you flipping loser
30:28 there's kind of two ways to do that
30:30 paywall one of them is kind of the cheap
30:33 way where if you know how to inspect the
30:35 code of a page you can actually just
30:36 sort of sneak you basically just say
30:39 show me the code of this page and you
30:40 can see the article
30:42 so so
30:45 those kind of sites that kind of had the
30:47 cheap version of the paywall for lack of
30:48 a better term I think chat GPT was just
30:52 reading the code and presenting the
30:55 paywalled content I I assume the ones
30:58 that had better versions of that better
31:00 versions of the payroll while that
31:02 didn't happen but but you know it could
31:04 have been a major player that just said
31:05 if you don't fix that we're gonna you
31:07 know make this a big deal so I I mean I
31:11 mean listen
31:12 Sam Altman and the team at open AI
31:14 they've got Microsoft as a partner and
31:17 even though Microsoft doesn't have a
31:19 controlling interest they don't want to
31:20 piss them off right Microsoft has a
31:22 bunch of advertisers on Bing they've got
31:25 massive corporate Partners so it
31:28 wouldn't surprise me if you know someone
31:30 made a stink and maybe Microsoft has
31:34 this dealt with on on their
31:36 implementation of chat jpt or on Bing
31:39 chat
31:40 um and maybe they said to open AI you
31:43 got to deal with this or you know maybe
31:44 they're just doing it I don't know but I
31:46 don't I don't have a [ __ ] clue it
31:48 seems to me that it would be something
31:50 relatively simple to fix but if they
31:53 don't know what the actual problem is if
31:55 they don't know all the places where
31:56 it's sneaking around paywalls and and
31:59 whatever other issues were raised
32:02 um it might be a while
32:04 I mean the good news is we've got Bing
32:06 good and it's not as convenient as using
32:08 gpt4 connected to the internet there but
32:11 that's the deal it's where we are yeah
32:15 I don't know why it's like this I want
32:17 it to be so much better
32:19 all of this has promise but it seems so
32:23 you know
32:28 hard forking last week in AI or great
32:31 podcasts I'm confused by the pricing is
32:34 it charged by tokens or keys I don't
32:36 know okay
32:38 so
32:40 chat GPT is flat pricing charged monthly
32:45 if you're paying for apis
32:48 um the the GPT API on the playground
32:52 and and you're doing things like zapier
32:54 and you're hooking into the API that's
32:56 charged by the token
32:58 so the the key just gives you access to
33:01 make the API calls
33:03 and then the API calls are the
33:05 combination of the prompt that you send
33:07 in and the answer that it returns that
33:09 amount of tokens is what you get charged
33:11 for each call I don't know if there's a
33:13 I don't think there's a charge per call
33:15 I think it's per chunk of tokens per
33:17 thousand tokens I think so like like uh
33:21 GPT 3.5 turbo is I don't I don't even
33:25 know what it is two cents and then GPT
33:27 for six cents like they they've they've
33:29 changed like the GPT 3.5 turbo 16k is
33:35 way cheaper and then gpt4 was slightly
33:38 cheaper so
33:40 it is confusing though because you're
33:43 paying if you're paying for chat GPT
33:45 you're essentially paying for an
33:46 application layer if you're playing if
33:48 you're paying for API access then that's
33:50 then you're paying for tokens
33:54 you can put text files into code
33:56 interpreter and greatly expand the
33:58 context in a prompt oh that's a super
34:00 cool idea
34:03 yeah because I think
34:04 I think can't isn't it like a maximum of
34:06 100 megabyte file you can put in 100
34:08 megabyte data file
34:10 so yeah so you could put in 100 megabyte
34:12 text file which that's probably a fairly
34:15 sizable amount of text
34:17 and then have that be the
34:20 the prompt
34:22 holy [ __ ] that's amazing
34:24 because you could do some some really
34:26 sophisticated prompting in that
34:30 in the uh in that document wow that's
34:32 wild is that 75 000 K of just your words
34:36 or added together questions and
34:38 responses yes it's a 70 it's a hundred
34:40 thousand token window context window so
34:44 that's that's a combination of
34:47 so you could put in a 45
34:49 000 word novel and get back a thirty
34:52 thousand word response and then you'd
34:54 still be under that and then it would
34:55 cut you off somewhere in that
34:57 neighborhood
34:58 hey I have a serious question well then
35:01 I'll give you a serious answer please
35:04 ask
35:08 all right
35:10 do you think software developers careers
35:13 are at risk due to AI well you said due
35:16 to IA definitely not at risk due to I IA
35:19 but maybe due to AI I know I'm a little
35:23 smarmy he's a little smart he's kind of
35:25 a smart ass isn't he is he trolling me I
35:28 don't think you're allowed to do that
35:34 yes you [ __ ] yes that's gone
35:38 all right do you think um so so here's
35:42 the deal software developers jobs are so
35:45 okay
35:46 so there's a guy that I'm a big fan of
35:49 named Paul rateser who started the uh
35:51 marketing AI Institute
35:53 and he's got this thing he says
35:55 ai's not going to take your job but
35:57 something somebody using AI will
36:00 um I think that AI literate programmers
36:03 AI literate Engineers are going to be
36:06 fine for some time because there's a lot
36:08 of [ __ ] to be built now how engineers
36:12 and programmers do what they do I think
36:14 is going to transform quickly and
36:16 dramatically and and you know anyone
36:18 who's using GitHub with copilot right
36:20 now you know essentially doing pair
36:22 programming with GPT tools
36:25 um knows this already right it's just a
36:26 completely different experience you're
36:28 you're sort of driving the intent of the
36:30 application and then GPT sitting there
36:33 sort of executing you know big chunks of
36:36 code and then you're sort of figuring
36:38 out is that code right is it did it do
36:40 it in a way that I like right all that
36:42 sort of stuff
36:44 so so I think that there's going to be
36:46 plenty of work there but any developer
36:49 right now that basically just sticks
36:50 their head in the sand and just says
36:52 well hand coding is the is the way I you
36:55 know I've always done it this way and I
36:58 will be coding like this for all of my
36:59 days they're going to be in trouble
37:02 because some little Punk's going to come
37:03 along and be able to 10x
37:06 um the productivity and I don't care how
37:08 tight their code is at some point
37:10 they're just going to get priced out of
37:11 the market right they're gonna they're
37:13 gonna be expected to do more than they
37:15 can do with manual coding all right
37:17 serious question from Black Swan I have
37:19 an idea I just don't know where to start
37:21 well feel free to share what you are you
37:24 feel free to share
37:25 um and I'll help
37:26 or I mean I don't know if I'll help I
37:29 will give you feedback on whatever that
37:31 might be and and tell you where you
37:33 might start
37:34 um yes they absolutely are we'll need
37:36 less developers until New Industries
37:39 expand
37:40 yeah in the end I think AI is going to
37:42 create you know massive amounts of new
37:44 jobs in the next two to three years it's
37:47 going to be a really painful transition
37:49 until those new jobs emerge right right
37:53 now all of these generative AI tools all
37:56 the stuff that's going to [ __ ] up all
37:57 the jobs is very early and it's very
38:00 kind of spit and duct tape it's it's
38:02 really janky software it's not
38:05 sophisticated the autonomous agents that
38:07 everyone got so excited about you know
38:09 Auto GPT and baby AGI all those things
38:12 are just janky pieces of [ __ ] right now
38:14 all of it is
38:16 as these tools get more and more and
38:18 more sophisticated
38:20 and
38:23 instead of being disparate features and
38:26 functions like we've got mid Journey
38:27 over here that you have to do in Discord
38:29 but we've got chat GPT over here but we
38:31 got this over here we got that over
38:32 there all this stuff's going to start to
38:34 merge where the tools that we use are
38:37 just going to be imbued with
38:40 AI capabilities
38:42 um and and and and that's when we'll
38:44 start to things will start to normalize
38:46 a bit but that's going to take a while
38:51 [Laughter]
39:00 I could listen to you all day sorry
39:03 about that I wish you were my dad
39:08 careful what you wish for yes
39:13 [Laughter]
39:17 it wears thin I wear thin
39:22 thin I wear this
39:24 I'm a lot I'm a lot
39:27 Camelot
39:32 all right I get Tick Tock just lost me
39:35 in my place of comments again so here's
39:37 the thing for those of you that have
39:39 been here before you know this about me
39:40 I got a little OCD like I want to answer
39:42 all the questions and then as I'm
39:44 scrolling to answer one question someone
39:46 will post another one and Tick Tock
39:48 jumps me to the bottom of the screen and
39:50 then I lose everything and so that makes
39:52 me cranky and so then I just sit here
39:54 scrolling
39:56 old man scrolling on a tick tock old man
39:59 tried to read words old man eyes don't
40:03 work old old man welcome to Elder talk
40:06 children
40:12 ah oh here's black Swan's serious
40:16 question all right
40:18 I'm back
40:20 Black Swan asked I want to take
40:22 every piece of Congressional Record and
40:26 feed it to an AI Okay so
40:29 Okay so good serious question
40:34 I should I should answer it seriously
40:36 and you know for talking about
40:39 Congressional records
40:43 um okay
40:45 so so there's a couple of things about
40:47 what you want to do I think what you
40:49 want to do will absolutely be done I
40:51 don't know what your background is what
40:53 access you have to that if you're just
40:54 talking about doing public records
40:57 um
40:58 I think there are so many applications
41:00 for that kind of work I was I was on a
41:02 panel
41:03 um about a month and a half ago uh with
41:06 a guy NX newspaper guy that was talking
41:08 about local journalism and the fact that
41:10 local journalism is [ __ ] and doesn't
41:12 exist and like all these you know uh
41:15 multinational companies are buying up
41:17 all the local markets and right so so
41:18 local journalism is kind of screwed but
41:21 what he was talking about was what local
41:23 journalists most of what local
41:24 journalists did was they went and got
41:26 you know public records and they
41:29 reported on them you know
41:31 here's what is happening on that corner
41:33 lot they're putting up a skyscraper and
41:35 if you want to keep the little corner
41:37 candy store you better get out there and
41:39 protest right that's what local
41:41 journalism was well that's all
41:43 discoverable public data that's
41:47 electronically discoverable
41:48 that you could
41:50 create
41:52 um you know GPT
41:53 article writers with
41:56 um automatically and so what you're
41:57 talking about is the same thing
42:00 right now it's it's it's
42:02 um in order to do that you can't just do
42:04 it with chat GPT you can do your prompt
42:06 engineering on GPT playground or on
42:09 chatgpt but ultimately what you'd need
42:11 to do with that there's a there's a
42:13 framework a project called Lang chain I
42:16 think it's short for language chain
42:18 and Lang chain allows you to essentially
42:21 Point um to certain directories full of
42:25 content and embed them and so embedding
42:28 basically means the the English uh the
42:31 the real human being phrase of embedding
42:33 is make it gptable so you basically take
42:36 all of the Congressional records you
42:38 embed them make them gptable and then
42:40 you store them in what's called a vector
42:42 database and that just basically means
42:43 they're searchable like chat GPT so it's
42:47 absolutely doable
42:48 um and depending on the access you have
42:51 um and you know you could probably find
42:53 you know an engineer or two that's got
42:55 some experience with Lang chain and do
42:57 that I would go just go hang out on
42:59 Reddit
43:00 um in the in the Lang chain subreddit
43:03 and there might even be some other
43:06 projects or extensions of Lang chain
43:08 um that you know that might make that
43:10 relatively straightforward
43:12 you can do some pretty sophisticated
43:14 stuff with not a lot of code and not a
43:16 lot of time it all boils down to how
43:18 fast people want to adopt on the scale
43:19 how do you display text on mid-journey
43:22 you don't right now
43:24 the text the text is just it's just not
43:27 there yet they'll get there with it if
43:28 you want to do text you can use adobe's
43:31 Adobe Firefly has a uh you can sort of
43:34 do
43:35 um prompting inside of a font so you can
43:38 make like fuzzy letters right the word
43:41 fuzzy all the letters are fuzzy or you
43:43 know honeycomb and they're all
43:44 Honeycombs so you can do that with an
43:46 adobe Firefly there I've I've seen some
43:49 papers in the past month that that are
43:52 talking about their methods for how to
43:54 do text my guess is that we're gonna see
43:56 coherent text on stable diffusion before
44:00 we see it on mid-journey
44:02 um just because it's open source in the
44:03 developer Community kind of attacks
44:05 everything on stable diffusion faster
44:07 than the team at mid-journey can
44:10 um so
44:11 everybody like like I feel like
44:14 mid-journey dealt with the pretzel
44:15 fingers and so those are good or getting
44:18 better now
44:20 um and and they've they've dealt with
44:22 some of the mushy faces and uh
44:25 so I would say text is probably it's
44:27 probably up there in priority whoever
44:29 gets text first is gonna people are
44:31 gonna flock to that whatever that tool
44:33 is
44:34 do I know if AI is being included for
44:36 the future of Robotics yeah there's let
44:38 me see if I can go find this one video
44:39 there's a there's a disturbingly fun
44:42 video from Boston Dynamics which I want
44:46 to like them but their robots are
44:47 [ __ ] terrified like why like if
44:50 you're gonna have a robot company why
44:52 would you make your robots look like the
44:54 ones that are in all the movies that
44:55 kill us
44:56 you know
44:58 you know what I'm saying Am I Wrong Am I
45:00 Wrong see Boston
45:03 Dynamics
45:05 dog chat GPT
45:10 oh doc dog
45:15 there it is
45:19 we use certificate to query information
45:22 and the user can ask natural questions I
45:26 was just wondering if I could dictate to
45:27 it to say something specifically what is
45:30 your battery level battery level is
45:32 currently at 53 what is the voice coming
45:35 from it's a Google Text-to-Speech we
45:38 give the Json to try GPT and explain
45:41 what the structure is and how to read
45:43 that Json Notch rgbt can answer
45:45 questions about that case all right so
45:47 so what he said there if you if you're
45:50 not into the geeky stuff that that
45:52 basically they're they're turning the
45:54 chat GPT into Json which the robot can
45:56 understand and then so they can send
45:58 stuff to the robot that the robot
45:59 understands and then the robot can
46:01 respond and chatgpt can can
46:04 um take that in so they're doing that
46:06 there what what one of the things that
46:08 large language models make possible is
46:12 um
46:13 is is putting an incredibly simple layer
46:15 on top of incredibly complicated systems
46:18 so a robot is an incredibly complicated
46:20 system of sensors and motors and
46:22 batteries and you know all that sort of
46:24 stuff
46:26 um
46:27 and and so
46:29 historically you know what you would
46:31 have to do is you would have to code it
46:33 programmatically go here go there go
46:35 there and then in the past 10 years
46:37 they've taken
46:39 um these these more specialized machine
46:41 learning
46:42 um models and put them in the robots and
46:44 that's how they do their navigation and
46:46 things like that that's how Tesla's
46:48 doing self-driving cars what the large
46:50 language models offer the potential of
46:52 is now we can just talk to them so
46:53 rather than having to have an engineer
46:55 kind of you know do all that stuff
46:57 manually if you get that large language
46:59 model interface right
47:01 um You can
47:04 um
47:05 talk to the robots and tell them to go
47:08 do things and then when you combine that
47:10 with the autonomous agents like Auto GPT
47:12 where you can just give them a goal
47:14 then you can say hey robot uh I'm low on
47:18 milk could you run down to the 7-Eleven
47:21 the robot will go sure boss
47:28 I wonder what it'll be like when will it
47:30 be each of our first moment where we're
47:32 just at a grocery store and a [ __ ]
47:34 robot comes in and starts picking up
47:36 peaches and sniffing them
47:38 yeah that's right
47:41 to happen it's not going to be that long
47:43 probably within five years
47:46 if Elon Musk has his way right he's
47:47 figured out the manufacturing [ __ ] so
47:49 he's gonna he's probably gonna pop some
47:52 pop some stuff
47:54 out into the world sooner than sooner
47:57 than the world is ready for it
48:00 uh thanks for the I don't know if that
48:03 uh coveruno I don't know if that art was
48:05 what you were looking for but I thought
48:06 that was pretty cool actually
48:09 that was pretty cool so
48:12 yeah
48:15 [Music]
48:16 any other wait any other law books like
48:20 oh Amber II
48:24 oh and other law books like
48:27 yeah so
48:29 so the stuff you want to do Black Swan
48:32 is very very doable you know I would go
48:35 if if you if you're D if you're a little
48:37 bit geeky even go read the documentation
48:41 for embedding
48:43 um at open ai go to open AI
48:47 I just go to platform.openai but let me
48:49 see what what it is here
48:53 foreign
48:54 [Applause]
49:00 yeah if you go to platform.openai.com
49:02 documentation
49:05 and then there's a guide for embeddings
49:08 go read that or go have chat GPT read
49:11 that for you and summarize it for you
49:15 actually some someone wanted to know
49:16 about that the about some prompting to
49:19 do so let's go let's go play with this
49:21 right now let me see how long this
49:22 document is oh it's not that long okay
49:25 so here's the here's the page on
49:28 embeddings in openai's documentation
49:31 okay
49:33 so I'm going to copy it we're going to
49:35 hop over to chat GPT
49:40 and this is one of these things that
49:41 these tools are really really good at
49:43 I'm going to go to gpt4 just because I
49:45 like it better
49:47 and so I'm going to say based on this
49:51 documentation
49:54 and then I'm going to hit colon shift
49:56 return does a line break so it doesn't
49:58 submit this and then I'm going to put
50:01 um
50:02 pound to pound pound
50:03 and then I'm going to put another line
50:05 break and I'm going to paste in all of
50:07 that documentation and then I'm going to
50:09 put pound pound Pound what the pound
50:11 pound does is let's chat GPT know that
50:14 that's the beginning and end of the
50:15 document you don't have to do that like
50:18 Chachi PT is incredibly good with
50:20 unstructured data with like shitty
50:22 unstructured oh [ __ ] I just got rid of
50:24 my
50:27 my double camera I somehow slid it up
50:30 can I get it back
50:31 no
50:33 ah
50:35 that what's that
50:37 what's that no
50:39 anyway okay
50:42 so so sorry 80d Panic
50:46 um
50:47 okay so I said based on this
50:49 documentation pound pound pasted the
50:51 documentation pound pound now I'm going
50:53 to hit shift return again I'm going to
50:54 say please summarize
50:59 um what
51:01 uh embeddings
51:05 mean for a non-technical
51:10 user
51:12 in
51:14 you know two sentences
51:19 and so now it's going to go read all
51:21 that for me
51:23 and it's going to tell me what's up
51:27 embeddings in simple terms are a way for
51:29 AI to convert
51:30 and represent complex information like
51:33 text into list of numbers which can be
51:35 used to find similarities or differences
51:38 okay listen
51:41 this is still too nerdy
51:46 um please use
51:49 um clear metaphors
51:52 and
51:54 um
51:55 bring the
51:58 meaning to life
52:02 with exam with real world examples real
52:05 world examples so so someone before was
52:08 asking examples someone was asking about
52:12 prompting the thing about prompting is
52:15 it's it's a lot it's very different than
52:19 playing with Google with Google you you
52:23 we've been kind of trained by Google to
52:25 create the most concise possible answer
52:27 and what happens with chat gbt is it's
52:29 more like you're having a conversation
52:31 with this thing
52:33 um so so if it gives you something you
52:35 don't like you can tell it that you
52:37 don't like it and also tell it why you
52:38 don't like it and then it will just you
52:40 know it will get better so let's see
52:42 what it says I can simplify further
52:44 imagine you're at a large party and you
52:46 want to introduce people to each other
52:47 based on their interests the process of
52:50 creating embeddings is like making
52:52 mental notes about everyone's Hobbies
52:54 background characteristics later you can
52:57 use those mental notes the equivalent of
53:00 embeddings in AI it's really cool to
53:02 introduce people to one another find the
53:05 most sports loving person or identify
53:06 someone who doesn't quite fit the crowd
53:08 yeah that's really interesting so it's
53:10 taking individual words in the context
53:12 of this Congressional you know
53:15 legislation and it's saying you know
53:17 semantically what what is that that word
53:20 is in the context of of what meaning the
53:23 meaning of the other words around it
53:24 right so like the person with their
53:26 interests
53:27 it's like giving a computer a way to
53:29 understand and compare the essence of
53:31 each text similar to how you understood
53:33 and compared the essence of each person
53:35 at the party
53:37 that's a good explanation that's not bad
53:41 what app are you using I don't know what
53:43 app you're talking about
53:45 um let's see I'm gonna I'm gonna invite
53:48 someone up here on screen I don't know
53:51 who wants to hop up here
53:54 hair growth geek you want to hop up here
53:56 because I think if I talk to you on
53:59 screen
54:01 then when I go back it'll give you back
54:03 my hello
54:06 it looks like you're processing hello
54:09 hello how are you how's it going good
54:14 another turn the camera on hold on a sec
54:17 there I am hello hello hello
54:20 how are you tonight I've been watching
54:22 you this is this is so much fun I was at
54:24 my computer trying to keep up I couldn't
54:29 yeah you know it'll it we I you know I
54:32 mean you've been here enough I cycle
54:34 through the same thing over and over
54:35 again
54:36 awesome awesome how what can I do for
54:38 you how can I
54:40 nothing can you you said you turned on
54:42 your camera I don't see you okay there
54:45 we go I made it smaller there we go hold
54:47 on
54:48 there hello you know
54:50 save it there we go okay
54:53 okay cool hello
54:56 um I don't know how are you doing are
54:57 you how how are things going are you
54:59 playing with any AI stuff uh yeah I uh
55:02 thank you I am I was actually sitting
55:03 here and I was playing with um comparing
55:06 Memento FM to get munch so we can break
55:09 down some of our YouTube content and
55:12 take some shortcuts yeah yeah the minds
55:14 were short for them
55:15 and then we post it with repurpose dot
55:19 AI but it's not the great uh greatest on
55:21 uh posting YouTube shorts so uh yeah I
55:25 know all of the tools right now are a
55:26 little janky like I I
55:28 um I've played with Vizio v i v i d y o
55:32 that's another one that does that and
55:34 it's like they're all they're all like
55:35 okay
55:37 just like but that's just the nature of
55:39 where we are in the news right now it'll
55:41 get better well I gotta tell you you
55:43 know the repurpose.io has been helpful
55:45 they're not all the way there but it's
55:47 it's it's really a big help but what I
55:50 did this weekend have you heard of
55:52 maverick.ai no
55:55 okay so this is how I spent my July 4th
55:57 so
55:58 so you know how you have your clavio CRM
56:01 flow and you have your thank you card in
56:03 the band or thank you email the
56:05 abandoned email Etc
56:08 so what I did was I recorded all those
56:11 videos
56:12 and then
56:13 they have you sit there and read names
56:17 for like 15 minutes which I know sounds
56:19 really archaic
56:20 but they can hit about eighty percent of
56:23 the names so when you abandon your car
56:26 oh yeah I saw this yeah
56:30 oh interesting yeah and um there's been
56:34 some very good
56:36 conversion rates with it
56:38 so um and I went into Ai and I said
56:41 Pretender garyvee write me an abandoned
56:44 carton thank you email and so I said act
56:47 as garyvee act as Ogilvy act as who is
56:52 the best act as Michael cialdini
56:55 um act as
56:57 um
56:58 Daniel Kahneman
57:00 act as a neuro-linguistic programmer and
57:03 so all these people and in the scripts
57:05 that came out were
57:07 actually quite similar in their
57:09 structure but they're all doing the same
57:12 thing did any did anyone show up better
57:13 than the others you know who was the
57:15 best garyvee yeah because he's got more
57:18 content out there
57:20 he's just he puts so much content out
57:22 there that makes perfect sense to me
57:24 because every all the other guys all
57:26 their shits behind a payroll wall but
57:28 garyvee's just like [ __ ] it here it is
57:30 yeah so I will share with you the
57:33 results they're processing the videos
57:34 now and we're going to set up our clavio
57:36 flows tomorrow oh that's cool and then
57:39 we'll be running our our paid app to
57:41 that as well as anybody who visits the
57:43 website because we use retention and
57:46 other services to grab the website and
57:50 add them to our clavio so even if you
57:52 don't give us your email we
57:55 yeah well of course you know and we're
57:57 spying on You by the keywords that
58:00 you're using to search and when you go
58:02 to my competitor site so I gave them a
58:04 list of my competitor sites I gave them
58:06 a list of my top keywords and we're
58:08 going to try to match and then we're
58:10 going to try to match you to your name
58:11 and the product you're looking for oh
58:14 that's cool well that's I mean all the
58:16 stuff that you're talking about I you
58:17 know I'd imagine if you're doing that
58:19 much analysis you're probably paying a
58:21 fair amount of money for that you're
58:22 going to have a blast with code
58:24 interpreter I think you just take a
58:25 bunch of your raw data and just dump it
58:28 in there and start asking it questions
58:30 you'll you'll be able to do all sorts of
58:32 custom reports that's going to be insane
58:34 all right well well so definitely if you
58:36 start playing around with code
58:37 interpreter let me know how it's going
58:38 because yeah
58:41 worthwhile yeah
58:42 yeah yeah that's super cool awesome well
58:45 thanks for popping up here
58:53 thank you
58:54 yeah
58:56 turn on all right I got my dual camera
58:58 back thank you thank you thank you hair
59:01 geek
59:03 all right
59:04 there's that I got my dual camera back
59:06 that was awesome thanks for popping up
59:08 here
59:09 um props.chat is very helpful check that
59:11 one out oh it's Shannon okay all right I
59:14 lost my place in my questions again okay
59:16 so who's here so I don't know who's here
59:17 I don't know how long you've been here
59:18 anyway my name is Kyle Shannon this the
59:20 AI learning lab as you can tell this is
59:22 a rambly sort of ADHD nightmare
59:26 hellscape of enjoyable
59:28 wacky fun with an old guy talking about
59:31 new stuff
59:33 um if you have questions about AI pop
59:35 them below I'll do what I can to answer
59:37 them
59:38 um if I can't answer them I try to be
59:40 straight about that too
59:42 um
59:42 the chat GPT isn't good at writing
59:45 people
59:47 on my last nerve yeah I know I I know
59:51 it's it's like um
59:54 like I just kind of feel like it's like
59:57 you know
59:59 I'm boring and uninspired and I asked
1:00:02 chat GPT for something and it gave me
1:00:04 something boring and uninspired like
1:00:06 well yeah
1:00:12 can you explain prompt creation yes so
1:00:16 that's what someone was saying to you
1:00:17 Shannon
1:00:18 um go to the prompts.chat
1:00:22 um this thing
1:00:23 go there
1:00:24 and that will explain to you what
1:00:27 prompting is so debate so so
1:00:32 prompting is just a geeky term for
1:00:34 talking
1:00:36 so so you're or here here's a here's
1:00:39 maybe a better way to put it what
1:00:40 prompting is is
1:00:43 finding ways to ask better questions to
1:00:46 get better answers
1:00:48 so what what chat is when you go to chat
1:00:52 GPT it looks kind of like Google
1:00:55 and in fact let me go since you're
1:00:58 you're kind of starting from the the
1:01:00 ground floor on the prompting stuff
1:01:02 I'm gonna do
1:01:04 um what I call some parlor tricks
1:01:07 because the
1:01:09 the
1:01:10 let me give you a little bit of primer
1:01:12 sort of technically some of the things
1:01:14 that are happening here it's called chat
1:01:16 GPT the GPT part of that stands for
1:01:20 generative pre-trained Transformer
1:01:24 you don't need to know much of any of
1:01:26 that but generative basically means it's
1:01:28 generating words so it's actually
1:01:29 writing original content for you based
1:01:32 on what you ask it pre-trained mean it
1:01:34 means it's been pre-trained you don't
1:01:36 need to supply it with data you can just
1:01:38 ask it questions what's it been trained
1:01:40 on it's been trained on essentially all
1:01:43 of the public internet for as much as
1:01:45 open AI could get over the past four or
1:01:48 five years
1:01:49 and
1:01:50 it's been trained on like all of
1:01:52 Wikipedia and all of the blogs and all
1:01:55 of the websites and all of the research
1:01:57 papers and all of the programming
1:01:58 languages and all of the documentation
1:02:00 and just like everything you can imagine
1:02:02 that's been dumped onto the public
1:02:04 internet for the past 35 or 40 years
1:02:07 is inside this thing right so it looks
1:02:10 like just a lovely simple little little
1:02:13 um prompt message but but you're it it's
1:02:15 you're not searching for links on the
1:02:19 internet like what Google does it's got
1:02:21 all of the content of the internet just
1:02:24 sitting in there and then the
1:02:25 Transformer part of of uh GPT is it's
1:02:30 just the technology that allows these
1:02:32 tools to be trained on all this data and
1:02:34 Google invented it and now they're at at
1:02:37 they're they're at risk because of what
1:02:40 chat gbt has done okay so some of the
1:02:43 parlor tricks so you can kind of ask it
1:02:45 anything so I could say
1:02:48 um
1:02:49 let's see
1:02:51 write me a
1:02:55 sonnet about
1:02:57 um
1:02:58 how lovely
1:03:02 thunder storms are and it will
1:03:07 write a lovely sonnet
1:03:10 and it did that actually quite well it
1:03:12 did it better than it's done it for me
1:03:13 in the past so in the skies we're
1:03:16 Tempest dance and thunder roars a
1:03:18 spectacle unfolds Enchanted sight the
1:03:21 air Ablaze with electricity's might as
1:03:23 rain Cascades lightning strikes the
1:03:26 shores right so so okay good Palm poem
1:03:30 um
1:03:31 can you make that a country song
1:03:35 boom and it will just write you a
1:03:38 country song it knows what verses are it
1:03:40 knows what choruses are it it writes
1:03:42 them it matches them
1:03:45 um what's the chorus here oh lovely
1:03:46 thunderstorms rolling on through your
1:03:48 crackling lightning and raindrops too
1:03:50 you paint the sky with your electrocues
1:03:52 in this country boy's heart I'm singing
1:03:55 the blues
1:03:58 um I could say things like
1:04:01 um please add
1:04:04 sad country
1:04:07 um chords
1:04:10 and it will now add it adds in
1:04:14 um
1:04:15 D C's and G's which if you're in a
1:04:17 country song we can do that but I could
1:04:19 say
1:04:20 um those are boring
1:04:24 chords
1:04:26 um and progressions
1:04:31 make them more interesting
1:04:35 and I misspelled a bunch of words there
1:04:37 it doesn't care about that so it threw
1:04:38 it into a minor and now it's giving me
1:04:41 some more interesting chords okay so I
1:04:44 can do that so now I can go
1:04:46 um
1:04:48 please
1:04:50 um make this a table
1:04:54 and compare the main
1:04:58 concepts with
1:05:01 Kenny Rogers The Gambler
1:05:07 why that just popped in my head hey Gen
1:05:10 X Gen X is hard man Gen X is hard so now
1:05:13 it's it's literally writing a table
1:05:15 comparing that shitty song I just wrote
1:05:17 about thunderstorms with Kenny Rogers
1:05:20 The Gambler theme Beauty and power of
1:05:22 Thunder thunderstorms theme wisdom and
1:05:25 life lessons setting Heartland wild
1:05:28 winds stormy skies setting gambling
1:05:31 table train journey
1:05:34 uh so it does that so now I could say um
1:05:38 make me a python application
1:05:44 um that
1:05:45 asks someone for a theme and then writes
1:05:52 them A Country Song
1:05:59 certainly
1:06:01 boom there's your code
1:06:03 so
1:06:04 prompting is the like as painful as it
1:06:08 is to watch me type and I I get told
1:06:11 that a fair amount you know I didn't
1:06:13 just come on this free internet channel
1:06:15 to watch you type for an hour well
1:06:17 apparently it did butthole because now
1:06:19 you're commenting on it uh
1:06:21 so but but prompting is is that
1:06:24 interaction and and it's it's
1:06:27 um I saw it referred to a couple of days
1:06:30 ago as as
1:06:32 um
1:06:34 um
1:06:35 AI psychology or GPT psychology that
1:06:39 that you're almost trying to get in the
1:06:41 head of okay if I want to get this thing
1:06:43 to respond to me as if it were William
1:06:47 Shakespeare
1:06:48 what would I have to tell it to do that
1:06:51 well I could tell it I want you to act
1:06:54 like William Shakespeare and I want you
1:06:55 to respond uh only in iambic pentameter
1:06:59 and I want you to you know make every
1:07:01 other line rhyme and whatever it might
1:07:03 be
1:07:04 um
1:07:05 and so so that's what prompting is and
1:07:07 so prompts.chat the you know that that
1:07:11 document right there will give you some
1:07:13 some hints and lessons into that
1:07:15 direction and then once you sort of get
1:07:17 it my you know I say this a lot that
1:07:19 that my My Hope for everyone on this
1:07:21 channel is that you play with chat GPT
1:07:23 long enough until you have a Kevin
1:07:25 McAllister moment which is from home
1:07:27 alone right
1:07:29 because at some point when you play with
1:07:31 this tool you'll crack through those
1:07:33 parlor tricks
1:07:35 and you'll do something that's relevant
1:07:37 to like your identity or something
1:07:40 you're good at or like you know maybe
1:07:42 you're good at whatever writing recipes
1:07:43 or gardening or copywriting or writing
1:07:48 strategies or doing marketing
1:07:50 um hair growth geek was just up here
1:07:52 she's obviously really good at the
1:07:53 marketing right so as you start to play
1:07:56 with this thing in the area that you're
1:07:58 good at at some point it's going to do
1:08:01 something where you're where you're just
1:08:03 gonna go oh
1:08:06 really should I be excited or should I
1:08:09 be scared
1:08:11 yes like the answer is yes
1:08:14 like like I'm the most excited I've been
1:08:18 in decades and and I'm totally unsettled
1:08:22 because everything's changing everything
1:08:24 in my business is changing everything in
1:08:27 marketing is changing if I owned an
1:08:29 advertising holding company right now
1:08:32 I would be
1:08:34 terrified
1:08:36 everything's changing
1:08:38 so hope that helps
1:08:45 can you demo something from prompts.chat
1:08:48 sure let's go do that that's a good idea
1:08:51 let's go to prom Star Chat
1:08:55 yeah
1:08:57 prompts.chat
1:09:00 so this is prompts.chat it's awesome GPT
1:09:03 chat GPT prompts just some dude made
1:09:05 this and then I think other people can
1:09:08 add to it I don't know if this is even
1:09:09 being updated anymore but it doesn't
1:09:11 really matter
1:09:12 so this just sort of explains what chat
1:09:14 GPT is it it gives you a free ebook
1:09:18 about writing prompts I think it sells
1:09:20 you at ebook about writing prompts I
1:09:22 wouldn't buy that
1:09:24 um
1:09:25 build your own GPT prompt app
1:09:27 that talks about the API and then
1:09:31 we get down into the the sort of meat of
1:09:34 the document so the the top thing is
1:09:36 just a primer and the meat of the
1:09:37 document is act like an English
1:09:39 translator right act as a Linux terminal
1:09:44 act as a blank position interviewer so
1:09:48 like if you want to practice for a job
1:09:49 you could say
1:09:51 um act as a project manager interviewer
1:09:53 right and you can just change out
1:09:56 um that thing so let's go find something
1:09:58 interesting here and then we'll play
1:09:59 with it
1:10:00 act as a travel guide plagiarism checked
1:10:03 or Advertiser Storyteller football
1:10:05 commentator stand-up comedian I don't
1:10:07 want to do stand-up comedian because I'm
1:10:09 a big fan of Comedy fan and when it does
1:10:12 bad comedy it's going to make me cranky
1:10:13 act as a debater that could be an
1:10:16 interesting one movie critic all right
1:10:18 let's do a movie critic
1:10:21 all right so we'll read what the prompt
1:10:23 is I want you to act as a movie critic
1:10:26 you will develop an engaging and
1:10:28 creative movie review you can cover
1:10:30 topics like plot themes and tone acting
1:10:34 and characters Direction scores
1:10:37 cinematography production design
1:10:39 special effects editing pace and
1:10:41 dialogue the most important aspect
1:10:43 though is to emphasize how the movie has
1:10:46 made you feel which is interesting
1:10:49 um what really resonated with you you
1:10:51 can also be critical of the movie please
1:10:53 avoid spoilers my first request is okay
1:10:57 so we're going to copy that we're going
1:10:58 to hop over to Chachi BT or as they say
1:11:02 on Tick Tock you can make money
1:11:06 foreign
1:11:11 if you if you listen to those people
1:11:13 then you're no longer welcome you can't
1:11:15 listen to those people they're all full
1:11:17 of [ __ ] they're just trying to take your
1:11:18 money you know how they're making money
1:11:19 with chat GPT getting you to click on a
1:11:21 stupid [ __ ] link
1:11:23 okay
1:11:26 so
1:11:29 we're gonna modify this a bit I want you
1:11:31 to act as the
1:11:35 movie critic
1:11:38 Roger
1:11:40 Ebert
1:11:43 um
1:11:45 [Music]
1:11:48 you could also be critical about the
1:11:49 movie please avoid spoilers
1:11:54 um
1:11:54 okay
1:11:56 let's see
1:11:58 um
1:12:01 um write your review
1:12:06 so that it is formatted with
1:12:12 you know bold
1:12:14 headlines and such
1:12:18 and organize
1:12:23 the review
1:12:27 into distinct
1:12:31 sections like d
1:12:34 distinct sections okay
1:12:37 now why am I doing that I don't know
1:12:39 because
1:12:40 distinct
1:12:43 um because one of the things that that
1:12:46 you can start to do is you you can
1:12:49 you you can push it to go move just
1:12:51 beyond a writing exercise and start to
1:12:53 think about how are we actually going to
1:12:55 present this right so
1:12:57 um write your review so that's formatted
1:12:59 with both headlines and such and
1:13:00 organize the reviews into distinct
1:13:02 sections each review should be
1:13:07 um I don't know 600 words
1:13:10 so so we'll go to length we'll try to
1:13:12 make this consistent my first request is
1:13:15 I need to write a movie review for the
1:13:17 movie Interstellar now we're going to do
1:13:18 it for
1:13:20 um
1:13:21 Pulp Fiction
1:13:24 all right Bang
1:13:31 so notice that it's created
1:13:35 um bold things there's a thing in chat
1:13:37 EPT called markdown
1:13:39 um which it it actually is sort of
1:13:42 coding this on the back end so that it
1:13:43 displays things as like bold and normal
1:13:46 and italic and things like that
1:13:48 so title Pulp Fiction Pulp Fiction a
1:13:50 symphony of style and substance the
1:13:52 intricate dance of narrative and theme
1:13:54 Quentin Tarantino's Pulp Fiction isn't
1:13:56 merely a film it's a paradigm shifting
1:13:58 spectacle that wriggles and dances its
1:14:00 way through a genre blender creating a
1:14:03 mesmerizing mix of high art and low
1:14:05 culture Tarantino's audacious narrative
1:14:07 structures spending time and playing
1:14:09 with right so solid writing
1:14:12 spectacular performances and authentic
1:14:15 characters right so it has now figured
1:14:17 out that because I told it I wanted to
1:14:19 break this into how did I put the prompt
1:14:22 um organize the review into distinct
1:14:24 sections
1:14:26 so it's doing that and I thought it was
1:14:27 going to give us generic things like
1:14:29 pacing or plot but no it's actually
1:14:31 writing headlines so spectacular
1:14:33 performances and authentic characters in
1:14:36 the center of the swirling cinematic
1:14:37 Vortex or Tarantino's fully realized
1:14:39 characters brought to life by Stellar
1:14:41 ensemble cast John Travolta Samuel L
1:14:43 Jackson Vincent Vega that's not right is
1:14:46 it
1:14:48 wasn't it isn't it Vincent
1:14:51 it's not Vincent Vega I think that's
1:14:52 wrong anyway
1:14:54 um
1:14:55 a Thurman's Mia Wallace is more than a
1:14:58 mafia mole she is a mirror reflecting
1:15:00 the desires and fears of men a master
1:15:03 class in Direction and writing Artistry
1:15:05 behind the camera
1:15:07 critique style over substance right so
1:15:09 brilliant okay so there's there's our
1:15:11 there's our review I say
1:15:14 um
1:15:15 how about
1:15:18 um a review of Caddyshack
1:15:25 again show my Gen X Stripes
1:15:28 [Laughter]
1:15:35 Larger than Life not so larger than
1:15:37 laughs the cast of Caddyshack is an
1:15:39 ensemble comedians who deliver coarse
1:15:41 and comedic timing Carl spackler Chevy
1:15:45 Chase
1:15:47 all right so we'll let that keep going
1:15:49 so that's writing that one so that's
1:15:51 interesting so now I'm going to say
1:15:53 um
1:15:54 give me a table
1:15:59 that compares
1:16:02 the characters
1:16:05 and plot
1:16:07 elements of these two
1:16:11 classics
1:16:16 and so here's a comparison so now it's
1:16:19 writing a it created a table
1:16:22 so here it says here's your main
1:16:24 characters here's your supporting
1:16:26 characters
1:16:29 primary plot
1:16:33 secondary plots
1:16:38 character arcs
1:16:43 the character Arc for Caddyshack is
1:16:46 generally shallow prioritizing comedy
1:16:48 over dramatic growth yeah that's the
1:16:51 point
1:16:53 um
1:16:54 so there you go so
1:16:56 so
1:16:59 the it's
1:17:01 it's
1:17:02 um
1:17:04 this idea of of telling chat to act as
1:17:08 as a specific thing it it actually why
1:17:12 this works is
1:17:14 because because these GPT models have
1:17:17 been trained on such massive amounts of
1:17:19 data
1:17:20 when you tell it to act as a movie
1:17:22 critic or in this case Roger Ebert right
1:17:24 you're kind of narrowing the world of
1:17:27 data that it's looking at to to a much
1:17:29 smaller subset so if I just said write
1:17:32 me a movie review of Pulp Fiction it'll
1:17:34 just go look at everything out on the
1:17:36 internet
1:17:38 and it'll pull up like you know Twitter
1:17:40 posts and like whatever whatever it has
1:17:42 in there just any idiot that said
1:17:43 anything about it you know from IMDb to
1:17:46 you know to some blog to Roger Ebert but
1:17:49 if you say I want you to act as Roger
1:17:51 Ebert it kind of filters out all that
1:17:53 other crap and focuses it so if if you
1:17:56 want Chachi PT to help you with your
1:17:58 quantum mechanics
1:18:00 problem solving or conceptualizing or
1:18:03 understanding
1:18:05 um point it in that direction and you'll
1:18:07 get much more focused and generally more
1:18:09 accurate less hallucinations because the
1:18:11 the way hallucinations work is
1:18:14 if if you're casting too wide a net it
1:18:17 will just kind of average out you know
1:18:19 all of the data that it has and it's a
1:18:21 prediction engine so it's just predict
1:18:23 predicting one word after the other and
1:18:25 so it will just include [ __ ] that has
1:18:27 nothing to do with reality because it
1:18:30 doesn't know any better it's not
1:18:31 actually looking at what it sends to you
1:18:35 all right
1:18:37 janky shame is my love language you can
1:18:40 make money with Chachi VT so Kyle isms
1:18:44 it's so true I do use janky a lot
1:18:48 and shame is my love language all right
1:18:51 I just got a job engineering automation
1:18:53 systems cool
1:18:55 AI is getting implemented right now yep
1:18:58 yep uh Ai and automation this generative
1:19:01 AI [ __ ] and automation oh it's gonna be
1:19:04 amazing it's gonna be like magic because
1:19:07 if you've ever done any sort of
1:19:09 automation historically or just any kind
1:19:11 of computing any kind of programming
1:19:14 like you have to think of like all of
1:19:18 the logic trees right like like you just
1:19:20 have to like
1:19:23 anticipate every possible path people
1:19:25 might go down and then you program in
1:19:27 all that logic you don't need to do that
1:19:30 with this thing it's just it like knows
1:19:33 all the [ __ ] that's that's why I'm so
1:19:35 excited about code interpreter coming
1:19:36 because
1:19:38 it's you can give it imperfect data and
1:19:41 it will turn it it will clean up your
1:19:43 data and turn it into perfect data it'll
1:19:45 find anomalies in your data and fix it
1:19:48 right so so you've got a bunch of people
1:19:50 entering
1:19:52 um you know your programmer didn't do
1:19:53 you right and didn't break up your
1:19:55 Fields enough when you're you know when
1:19:57 your customers are inputting their their
1:19:59 addresses and all their addresses are
1:20:01 all [ __ ] up
1:20:02 you'll just go fix all that
1:20:07 um I just built a slack integration with
1:20:09 chat EBT yeah I've got one of those too
1:20:11 that's it's super cool I don't my
1:20:13 problem right now is I'm I'm
1:20:15 here's something to I'm starting I'm
1:20:18 starting to get smarter about this but
1:20:20 it's taken me a while
1:20:22 um I'm falling into the Trap of being
1:20:25 intimidated by there's too much
1:20:28 possibility like this thing can do too
1:20:30 much
1:20:30 so I kind of I get like a blank screen
1:20:33 syndrome right where like blank page
1:20:35 syndrome where it's just like
1:20:38 since you could write anything like you
1:20:39 don't know where to start
1:20:41 I'm starting to find I'm doing that with
1:20:44 both AI images with with chat GPT with
1:20:48 with a lot of the tools that we're using
1:20:49 with automation tools like zapier
1:20:52 the thing that hit me today is I need to
1:20:55 flip my mental model to not think about
1:20:59 what's possible
1:21:01 but to actually start with what's a
1:21:04 problem I have right now like I don't I
1:21:06 don't care if I know how to solve it or
1:21:08 not well that's a problem that I have
1:21:10 right now that needs solving
1:21:12 and start with that and then from that
1:21:14 say okay how would I solve that with
1:21:18 chat gbt or the mid-journey with this or
1:21:20 with that or with that
1:21:22 um and that's that's that's something
1:21:23 that I'm I just feel like I have to
1:21:25 shift because it's this this stuff is
1:21:28 very seductive like once you have your
1:21:29 Kevin McAllister moment it's very
1:21:32 seductive you get you get sucked into it
1:21:34 and and then and then you'll find
1:21:37 yourself self spending four hours like
1:21:40 writing the Great American novel you
1:21:43 know
1:21:44 um in iambic in Shakespearean iambic
1:21:46 pentameter and then you'll realize I've
1:21:50 got this book now that No One's Gonna
1:21:53 read and I don't even really care about
1:21:55 but I did it because it was there right
1:21:57 that it's it's a little like that
1:21:59 um so so I think the idea of focusing on
1:22:01 a specific problem and then just you
1:22:03 know being less seduced by the
1:22:06 technology is a big one that I need to
1:22:08 work on
1:22:09 um appreciate your efforts you're a
1:22:11 great teacher thank you very much I
1:22:12 appreciate that
1:22:13 please tap it's freezing oh
1:22:17 is it still freezing sorry is there an
1:22:19 application
1:22:20 that wouldn't you stream from your
1:22:23 computer there is but it's for Windows
1:22:26 only and I'm a Macintosh guy I'm an
1:22:28 apple Fanboy from way back
1:22:30 uh I've hated Microsoft for 35 years so
1:22:33 I'm very conflicted now that they're
1:22:35 acting like apple and they're more
1:22:37 they're more Innovative than really
1:22:39 anyone out there right now because
1:22:40 because of what they're doing
1:22:43 um you know in conjunction with open AI
1:22:46 I think they're they're they're up there
1:22:49 so I pay for a chat GPT for just for
1:22:52 email stuff I'm still afraid to share
1:22:55 what I do at work yeah there's there's a
1:22:57 uh there's a really fascinating
1:22:59 um article by this guy named Ethan Malik
1:23:01 who's A Wharton Professor called
1:23:03 detecting the secret cyborgs
1:23:06 and Ampersand you are what Ethan Malik
1:23:09 would call a secret cyborg which
1:23:11 basically what he's what he says in the
1:23:13 article is these generative AI tools are
1:23:16 incredible personal productivity tools
1:23:19 but they're not really ready for prime
1:23:20 time for businesses
1:23:23 and and they're also potentially very
1:23:26 disruptive for businesses and they're
1:23:27 also potentially legally in privacy and
1:23:30 security risks for businesses so
1:23:34 businesses are like yeah we can't really
1:23:36 [ __ ] with that right now and I don't
1:23:37 know if I want to because I think that
1:23:39 can screw some things up and the
1:23:41 employees are like holy [ __ ] [ __ ] I
1:23:43 just got done in four hours what
1:23:45 normally takes me the whole week uh I'm
1:23:49 not gonna tell anyone because that's
1:23:50 [ __ ] awesome
1:23:52 so those are the secret cyborgs and what
1:23:54 Ethan Malik's saying in the piece is
1:23:56 companies need to get their [ __ ]
1:23:58 together and say hey you've got sort of
1:24:00 you know a whole class of employees here
1:24:03 that's becoming secretly AI literate
1:24:06 that's not benefit it's benefiting your
1:24:08 business but not directly right
1:24:12 um and so so the businesses need to
1:24:15 create safe zones to invite these secret
1:24:18 cyborgs Out of the Shadows
1:24:20 um and and really start innovating as as
1:24:22 organizations so I think it's absolutely
1:24:24 fascinating oh Lord he's going into
1:24:27 character it happens it's it I haven't
1:24:29 done it too much tonight a little bit
1:24:31 but as I get more tired I tend to go
1:24:33 into more characters it's also like I've
1:24:35 noticed like I don't know I don't know
1:24:37 why I'm doing self-analysis on on my
1:24:39 Tick Tock performances but um
1:24:42 but I've noticed like if I'm if I'm in a
1:24:45 bad mood or cranky or just low energy I
1:24:47 tend to be just like AI is fantastic
1:24:50 here's how you learn it and then if I'm
1:24:52 in a good mood I'm like wow
1:24:55 so
1:24:56 where am I tonight I'm somewhere in
1:24:58 between I'm I'm Midland Midland on stage
1:25:03 do you use zapier with chech EPT and if
1:25:07 so in what capacity have you found it
1:25:09 most useful so the most useful things
1:25:10 I'm using zapier for right now is just
1:25:12 the first things I used it for like you
1:25:15 know how when you first sort of crack a
1:25:16 technology
1:25:17 it's like that that use case embeds on
1:25:20 you
1:25:21 so it's like
1:25:23 um the first thing I did with chat CPT
1:25:25 was I built a Content generation engine
1:25:27 so it was a thing that asks someone
1:25:29 questions and then generated a bunch of
1:25:31 social media content content I did it
1:25:34 first in Python and then I I went to I
1:25:37 recreated it in zapier and I called it
1:25:39 the social media marketing machine
1:25:42 um and it's cool I mean it it works you
1:25:46 answer some questions about a topic and
1:25:47 it generates a bunch of actually really
1:25:49 good social media content that really
1:25:51 doesn't need editing
1:25:52 and so so I do a lot of that I did one
1:25:55 for my business so my business is called
1:25:57 storyvine and it's a video
1:26:00 um technology platform we have an app
1:26:02 that guides it it provides prompts for
1:26:04 someone we're a prompting company we
1:26:07 have been for 11 years
1:26:08 um it provides prompts for someone to
1:26:10 answer to tell a specific story so if
1:26:12 it's someone who was diagnosed with
1:26:13 cancer it'll prompt them you know when
1:26:16 were you diagnosed and
1:26:18 um you know what did it feel like when
1:26:19 you heard the news and how have you been
1:26:21 treating it how are you feeling and all
1:26:23 that sort of stuff and so writing those
1:26:25 prompts is a very particular skill that
1:26:28 we've gotten good at over the years so
1:26:29 one of the engines that I built was an
1:26:32 engine that automatically writes those
1:26:34 prompts for US based on some questions
1:26:36 that we ask a client so we get on a
1:26:38 phone with a client and an agency which
1:26:41 historically would take us three to five
1:26:44 creative meetings just to get them to
1:26:46 understand how we think about questions
1:26:48 versus how the agency does versus
1:26:51 and it was always this confusing mess
1:26:55 and so now what we do is in the course
1:26:57 of the meeting we ask them questions
1:26:58 what's the kind of story you want to
1:27:00 tell who's the audience you know what's
1:27:01 the message what are you trying to
1:27:02 accomplish
1:27:04 and then when we submit that form it
1:27:07 goes off to chat GPT it generates three
1:27:10 sets of questions a short medium and
1:27:12 long set of questions for that specific
1:27:14 story for that specific client and then
1:27:17 it sends it to our slack Channel and so
1:27:19 the account team now and when when we're
1:27:22 on one of these creative calls we ask
1:27:24 the client questions it it generates
1:27:26 these prompts they literally copy and
1:27:28 paste them into a word doc tweak them
1:27:31 kind of as we're talking you know they
1:27:32 might add one or subtract one and
1:27:35 they're good they're like we don't need
1:27:36 to tweak much and then we we literally
1:27:39 just pop them up on screen and say well
1:27:42 so based on the the answers that you've
1:27:45 given us here's a starter set of prompts
1:27:47 and so we don't have to explain how we
1:27:49 write prompts or what we do they're just
1:27:51 done and then all we're doing is
1:27:53 tweaking the prompts that are there what
1:27:56 we would have done historically is say
1:27:58 well based on this story what are the
1:28:00 kinds of questions we want to ask and
1:28:02 then the agency would go well why don't
1:28:03 we go off and we'll come back in a week
1:28:05 or two and give you that and then three
1:28:06 weeks later they would send us 27
1:28:09 questions for a two-minute video and
1:28:12 we're like you can't ask 27 questions in
1:28:14 a two-minute video and there would be
1:28:15 all this back and forth so that's gone
1:28:18 so that's what I'm doing with it
1:28:20 um we're also we've got
1:28:23 um open ai's whisper transcription tool
1:28:26 built into our system now so every video
1:28:28 in the system now has interactive
1:28:30 transcriptions
1:28:32 and I'm using zapier to prototype
1:28:36 um omnichannel content generation based
1:28:38 on a video transcript right so here's
1:28:41 what the person said in the video here's
1:28:43 the title of the video here's the
1:28:44 YouTube description here's a tweet and
1:28:47 doing all that sort of stuff so it's I'm
1:28:49 using it a lot it's pretty cool and
1:28:51 there's there's so much more you can do
1:28:53 but that's that's a fun one that I've
1:28:55 been been playing with
1:28:57 would you be willing to beta test an AI
1:28:59 app I'm working hard as hell on to get
1:29:01 launch ready I'll take a look at it yeah
1:29:03 just my my contact info is in my
1:29:06 um what's it called profile there's a
1:29:09 there's a link tree there oh you know
1:29:11 what I wanna I wanna make a request to
1:29:13 everyone on the on yield Channel
1:29:16 so a couple of days ago
1:29:20 someone had asked hey you know do you do
1:29:22 a workshop or do you have a workshop and
1:29:25 I said well no not really
1:29:27 um
1:29:28 but
1:29:29 I now have a Graphic for a workshop
1:29:33 so I'll put my little head here uh like
1:29:35 I'm like I'm actually talking to you so
1:29:38 um
1:29:38 so I'm thinking about putting together a
1:29:40 workshop but I it it it's a lot of work
1:29:43 which I'm not afraid of work I don't
1:29:44 mind it being a lot of work but what I
1:29:46 don't want to do is just
1:29:47 um create a workshop and a vacuum so if
1:29:50 you go to that URL below the link tree
1:29:52 slash real Kyle Shannon the first link
1:29:54 there is a link to a survey that's got
1:29:57 three questions in it it's like it's
1:29:59 like your name email and three questions
1:30:00 and so the questions are just like what
1:30:03 do you want this Workshop to be like
1:30:04 what are you trying where are you in in
1:30:06 your AI Adventure you know what are you
1:30:09 hoping to learn and then like where do
1:30:10 you hope you are a year from now it's
1:30:12 they're actually really simple questions
1:30:13 so if you'd be so kind hop over there
1:30:16 while I'm rambling and just if you'd be
1:30:19 interested in the workshop and then that
1:30:20 way when I put together the workshop I
1:30:22 can do it in a way that that serves as
1:30:24 many people as possible as well as
1:30:26 possible because I'm codependent I want
1:30:30 everyone to be happy if you're not happy
1:30:34 I failed as a human
1:30:37 all right
1:30:39 um so yes I'd be willing to to at least
1:30:41 talk to you about testing it I'm listen
1:30:43 I have ADD here's what I can tell you
1:30:45 about my Beta Testing if you want all of
1:30:48 the things tested I'm not your guy if
1:30:50 you want me to break your app as quickly
1:30:52 as possible I'm your guy I can usually
1:30:55 break an app within 30 seconds
1:30:58 my my developers hate me for that
1:31:02 because they're like okay we got to this
1:31:04 is good now it's solid I'm like
1:31:08 30 seconds
1:31:10 in fact I just got so we have a we just
1:31:14 got rid of one developer we added a full
1:31:16 stack developer and the full stack
1:31:18 developer is actually pretty good and
1:31:20 um it was the first time in I think four
1:31:23 years that I was given a beta version of
1:31:26 our app and I didn't break it
1:31:30 um in in two entire sessions first time
1:31:33 in in a in a long long time so that was
1:31:36 pretty cool in the beginning that was
1:31:38 the word and the Word was with God and
1:31:41 the Word was God oh my God
1:31:43 foreign
1:31:45 that's funny okay did the AI Salon
1:31:48 Meetup happened last night yes it did
1:31:50 Miss Birch and it was it was good we did
1:31:52 a thing so so normally at the AI Salon
1:31:54 let me let me give you that info as well
1:31:57 [Music]
1:31:57 um
1:31:59 so the AI Salon is it was funny we were
1:32:01 actually working on the um
1:32:04 the positioning today and we're like
1:32:06 What is the salon so the salon is this
1:32:08 we call people AI adventurers in the
1:32:10 salon so it's a community of people
1:32:13 they're all basically
1:32:16 um
1:32:17 starting the adventure of getting into
1:32:19 generative AI are there already on the
1:32:21 adventure but like the the common theme
1:32:23 is that it's people that see this as an
1:32:26 adventure there's some [ __ ] to learn
1:32:28 here it's gonna [ __ ] some stuff up like
1:32:30 a good Adventure there's reasons to be
1:32:32 excited and like a good Adventure
1:32:34 there's reasons to be terrified right
1:32:36 and that's um what the salon is so if
1:32:39 you go to the salon.ai it's a link tree
1:32:42 that's got a link to the Meetup we meet
1:32:44 every other week on on Tuesdays we met
1:32:46 last night because the fourth was on
1:32:47 Tuesday
1:32:49 um and we normally have a speaker come
1:32:51 in that is you know in the industry
1:32:53 somehow they're either an artist or
1:32:55 there's someone in business and then we
1:32:57 just talk to one another and we have the
1:32:58 speaker talk and people can ask
1:32:59 questions it's it's a really great
1:33:01 community and what we did last night was
1:33:04 we didn't have a speaker but we had
1:33:07 um what we call AI confessionals and so
1:33:10 everyone in the group who wanted to had
1:33:12 10 minutes to talk about something they
1:33:15 learned some failure they had some app
1:33:17 that they're working on some business
1:33:19 that they're starting some cool
1:33:21 experience that they had some art that
1:33:23 they want to show
1:33:25 um and it it was just great and you know
1:33:28 one guy just got on and said okay you
1:33:30 know you're all are talking about
1:33:31 prompting here's some you know prompting
1:33:33 tips I came up with so and we record
1:33:35 them they're two hours long and they go
1:33:37 up on uh YouTube so we've got recordings
1:33:40 of back back ones that that link is
1:33:42 there and then we also have a Discord so
1:33:44 the Discord where I show mid-journey
1:33:45 stuff here that's the AI Salon Discord
1:33:47 and we have a LinkedIn group as well so
1:33:49 yeah so it happened and if you want to
1:33:52 be a part of it
1:33:54 please join
1:33:58 is GitHub co-pilot useful for non-devs
1:34:02 does it have a prompt like uh GPT
1:34:07 um
1:34:08 I assume it is I have not deeply all the
1:34:11 sort of coding stuff I've done I've done
1:34:13 within chat GPT
1:34:15 um just because I'm not really an
1:34:17 engineer so I'm usually just sort of
1:34:18 [ __ ] around trying to get some little
1:34:20 chunk of code modified rather than you
1:34:22 know building a whole application uh it
1:34:25 is it is ostensibly GPT
1:34:28 um you know that's that's
1:34:30 um optimized for coding
1:34:33 um but I don't I don't know how it's
1:34:34 implemented enough and what the I think
1:34:36 it's more like a pair programmer where
1:34:38 you're kind of programming your thing
1:34:39 and it's it's sort of assisting you
1:34:41 along the way rather than a distinct
1:34:43 experience so I don't actually know
1:34:47 um you can do a lot of coding in chat
1:34:50 GPT gpt4 in particular is quite good
1:34:53 Adobe text is awesome but you can't use
1:34:56 it on a commercial site they'll change
1:34:57 that MCAT you know it's it right right
1:35:00 now it's in beta so they're basically
1:35:02 saying we don't we don't until we have
1:35:04 this thing dialed in
1:35:06 um to the point that we're happy with it
1:35:07 we don't want people using the outputs
1:35:09 commercially because you know someone
1:35:12 will produce something that they'll get
1:35:14 whatever
1:35:15 sued for someone will write a tweet
1:35:17 about and then that'll get back to Adobe
1:35:20 and Adobe shouldn't have had you know
1:35:22 boobies in their model or whatever
1:35:24 stupid [ __ ] thing will come up
1:35:27 so it'll get there um it's just not
1:35:29 there yet
1:35:31 you're not wrong they must do it on
1:35:33 purpose too smart to make that mistake
1:35:35 I'm not sure what I was talking about
1:35:36 there Kyle will you explain your setup
1:35:39 in detail I don't get how the little
1:35:41 window works yeah so when
1:35:44 when I first go live
1:35:47 there's like all these like eight little
1:35:49 icons that you can set up like goals and
1:35:53 effects and all that sort of [ __ ]
1:35:56 one of those options just says dual
1:35:59 d-u-a-l and it's dual camera so it's
1:36:02 like a little icon with two two cameras
1:36:03 stacked on top of one another
1:36:05 and then when you turn that on you can
1:36:07 either choose a square window a round
1:36:09 window or or none I don't I suppose none
1:36:12 just means you're not turning it on
1:36:15 um
1:36:15 and then all it's doing is it's using my
1:36:18 back camera
1:36:19 you know for that and my front camera
1:36:21 here
1:36:23 and then if I tap in that window so I
1:36:26 can move the window
1:36:28 like over my face which that's probably
1:36:31 better for everyone right solid although
1:36:34 wait maybe we should just cover up the
1:36:36 neck fat there
1:36:38 [Laughter]
1:36:40 so I can do that
1:36:42 and then if I tap on it it it flips them
1:36:46 and so that's all it is there's nothing
1:36:48 fancy here
1:36:50 except me I'm fancy
1:36:52 I do declare it's so fancy
1:36:56 did you notice his hair what was going
1:36:58 on with it does he do that on purpose
1:37:01 good Lord
1:37:04 all right robot dog right right my term
1:37:08 paper and do a flip exactly and the
1:37:11 robot dog will
1:37:12 robot dog stuff humping my leg robot dog
1:37:16 start humping my leg
1:37:18 robot dog What is love
1:37:21 exactly you know you joke we're not that
1:37:25 far from that we really aren't
1:37:31 oh my goodness
1:37:33 um what's going on with the sound I
1:37:35 don't know it could be my shitty voice
1:37:37 it could be my shitty internet
1:37:39 um the absolute nightmare of Robo dog
1:37:41 running into the grocery store by me
1:37:43 thanks for the nightmares I I'm telling
1:37:46 you man it's
1:37:48 I am guessing within three years someone
1:37:52 on there's 58 people on this right now I
1:37:54 assume that means including me no but
1:37:56 probably in addition there's nearly 60
1:37:58 people on here I think within three
1:38:01 years one of us is going to be in a
1:38:02 [ __ ] grocery store and we're gonna
1:38:04 hear
1:38:05 we're gonna go no no you're gonna turn
1:38:08 around and there's gonna be a [ __ ]
1:38:10 robot picking up a box of Fritos
1:38:12 Cheetos free whoa those are in bags I
1:38:16 can't even I I'm so upset about that
1:38:19 perspective meeting that I don't know
1:38:20 what's in a box or a bag anymore
1:38:24 all these robots walking around Costco
1:38:29 filling up the carts and like
1:38:31 mathematically perfect piles
1:38:33 [Laughter]
1:38:37 a little general intelligence I don't
1:38:39 know
1:38:40 um
1:38:41 Ray Kurzweil is predicting 2029 for the
1:38:44 singularity which I think is in in the
1:38:46 neighborhood of that
1:38:48 um
1:38:50 feels plausible to me I mean there's
1:38:53 there's a group of Microsoft researchers
1:38:56 put out a like a really long paper like
1:38:58 120 page paper
1:39:00 um three or four months ago
1:39:02 um where they they
1:39:04 um were doing tests on gpt4 and they
1:39:07 said gpt4 is showing signs of general
1:39:10 intelligence
1:39:12 um but but there's a couple of things
1:39:14 about it that that just aren't there
1:39:16 right it it it doesn't have memory it
1:39:19 doesn't have reasoning it doesn't
1:39:21 understand what it wrote even though the
1:39:23 things it's writing feel very human and
1:39:26 and it will respond empathetically
1:39:28 sometimes and and in ways that looks
1:39:30 like reasoning and some of the stuff it
1:39:32 does borders on reasoning it is there
1:39:35 there are you know unanticipated
1:39:37 emergent properties the technicians call
1:39:40 them the the emergent properties are
1:39:43 like holy [ __ ] I didn't know it could do
1:39:44 that we didn't train it to do why is it
1:39:46 doing that could it do that so so that's
1:39:49 happening
1:39:51 um
1:39:52 but
1:39:54 it's not true reasoning it doesn't have
1:39:56 memory right so so once you start to
1:39:58 create these autonomous agents
1:40:01 where they start behaving on their own
1:40:03 and it start has starts to have some
1:40:05 self-awareness right I think I think the
1:40:07 building blocks for general intelligence
1:40:09 there's components of we we see
1:40:12 components of them in a lot of the
1:40:14 research papers that are coming out
1:40:15 right now where they're trying to work
1:40:17 on alignment
1:40:19 alignment basically meaning we want
1:40:22 these robots to
1:40:24 um align
1:40:26 with the human values like humans are
1:40:29 valuable we want that alignment so so so
1:40:32 open AI just announced two days ago
1:40:35 um they're they're dedicating 20 of
1:40:37 their resources to
1:40:40 um what they're calling super alignment
1:40:42 which is basically don't kill us please
1:40:46 so I think I think 20 of their resources
1:40:50 to building a tool that has AGI super
1:40:54 Intel or general intelligence first
1:40:56 um and then putting 20 of your efforts
1:40:58 towards making sure that it's good for
1:41:00 Humanity
1:41:01 that's good it's better than zero
1:41:05 might want to up it to 50 but let's
1:41:07 start with 20. let's see how evil it
1:41:10 gets how how smart it gets and then how
1:41:12 evil it gets
1:41:14 um can you talk about neural networks uh
1:41:17 and and or deep learning I can talk a
1:41:20 bit more about the about how chat GPT
1:41:23 works I I don't have enough expertise on
1:41:26 the neural networks and deep learning
1:41:29 piece of it to really probably do anyone
1:41:31 any good
1:41:32 um if you want to know like how chat GPT
1:41:35 Works how the the diffusion models work
1:41:39 I'm good with that
1:41:41 um I'm just I'm just too inexperienced
1:41:43 and frankly it's it's it's not stuff
1:41:45 that I could contribute anything to
1:41:48 anyway so
1:41:50 um I'm not really learning it like I
1:41:53 don't know it was five or six years ago
1:41:55 and Google did whatever sort of nice
1:41:57 press release that that was like
1:41:58 tensorflow is now available to play with
1:42:00 and you can do machine learning and I'm
1:42:02 like oh good I'm gonna go learn machine
1:42:03 learning and I went and I started
1:42:06 reading the paper for tensorflow and I'm
1:42:07 like no I'm not
1:42:11 the words in it and I didn't have chat
1:42:14 gbt to be able to you know um distill it
1:42:17 for me explain it to me like I'm five
1:42:19 explain it to me like I didn't go to
1:42:22 Stanford explain it to me like I have a
1:42:24 degree in acting which I do
1:42:29 so yeah so it's probably not good for me
1:42:31 to talk about neural networks um I mean
1:42:34 what I can tell you is Google invented
1:42:36 this thing called the Transformer which
1:42:39 allowed these neural networks to be the
1:42:43 machine learning model sitting on the
1:42:44 neural networks to be trained on massive
1:42:46 data sets and the more compute power you
1:42:49 throw at them the better they get and
1:42:51 that's how openai did what they did with
1:42:53 gpt4
1:42:55 um I also heard who was it that was
1:42:57 saying it um
1:42:59 oh George Hots I don't know if you know
1:43:00 George hats he's from a company called
1:43:02 comma he was he he
1:43:05 um he was the first guy to jailbreak an
1:43:07 iPhone when he was 17. and he and he
1:43:09 became famous for that and then he
1:43:11 started
1:43:12 a company called comma which was a
1:43:15 self-driving car that he was designing
1:43:17 out of his garage in San Francisco
1:43:21 um and he sort of spilled the beans on
1:43:24 gpt4 is actually apparently like 16
1:43:27 different models and eight different
1:43:29 weights I don't know there's it's
1:43:32 apparently a hell of a lot more
1:43:33 sophisticated and complicated than
1:43:35 they're letting on so it's it's
1:43:38 significantly different than 3.5
1:43:40 structurally apparently according to
1:43:42 George hotz
1:43:43 but he likes he likes [ __ ] with
1:43:45 people anyway so
1:43:48 if you've not watched videos of George
1:43:50 hot's programming if you want to see the
1:43:52 human equivalent of GPT like the the
1:43:55 speed with which that so so if you
1:43:58 haven't seen GPT do programming
1:44:01 um so let me see
1:44:03 um write
1:44:04 a C sharp
1:44:07 application that
1:44:10 combines
1:44:12 Flappy Birds
1:44:16 with
1:44:18 snake
1:44:31 first the game's rules oh so so this is
1:44:34 inventing a game
1:44:36 so now it should code now
1:44:39 now we'll need several several classes
1:44:41 so here's the classes
1:44:44 so it's writing us the documentation of
1:44:46 the game it's going to build and now
1:44:47 it's building the code
1:44:49 so watching George hot's program is he's
1:44:53 faster than this
1:44:59 um
1:45:00 make it
1:45:02 a python uh script
1:45:07 that
1:45:08 um I don't know
1:45:10 uh sorts
1:45:14 um
1:45:16 ingredients from a shopping list
1:45:23 sure
1:45:24 here you go
1:45:30 so watching George shots program is like
1:45:32 he like types this fast and he talks
1:45:34 this fast like if you think I talk fast
1:45:36 George hotz is like
1:45:38 10 times faster
1:45:41 it looks like my sound is [ __ ] up
1:45:43 somehow
1:45:45 so embeddings is like an SQL code query
1:45:48 it's
1:45:50 no
1:45:51 well
1:45:54 all right so you want to see okay so
1:45:57 someone was asking about neural networks
1:45:59 I won't tell you about them I'll tell
1:46:00 you how
1:46:02 how large language models work I've
1:46:04 showed this a couple of times before if
1:46:05 you've seen this before feel free to
1:46:07 take a nap or take off
1:46:10 um
1:46:11 Keynote
1:46:15 this is how this thing works is
1:46:20 like like the more you understand how
1:46:22 how this thing works the more
1:46:24 mind-blowing it actually is
1:46:30 Okay so
1:46:33 let me get my little head out of the way
1:46:35 yeah oh no I just lost my camera again
1:46:38 God damn it all right
1:46:40 whatever
1:46:43 [Music]
1:46:45 Okay so
1:46:48 we have this thing called chat GPT GPT
1:46:50 generative pre-trained Transformer I
1:46:52 talked about that earlier generative
1:46:53 meaning meaning it's generating stuff
1:46:55 pre-trained means it's pre-trained on
1:46:57 all of the internet basically and or it
1:47:00 can be on whatever but in the case of
1:47:02 chat GPT it's trained on on the the
1:47:04 public web and then Transformers the
1:47:06 technology that lets that happen the
1:47:09 large language model is is the core
1:47:11 based model the like GPT 3.5 and gpt4
1:47:14 that's the large language model
1:47:17 um and then the latent space
1:47:19 this is this is how it works so so
1:47:23 it's
1:47:26 there are similarities to a SQL query
1:47:28 but the SQL queries for the most part
1:47:31 are explicit you have to have good data
1:47:33 well organized and then you have to
1:47:35 query on that good data good
1:47:38 well-organized data
1:47:40 what happens with these large language
1:47:42 models is
1:47:44 when when you talk about embedding
1:47:47 you're embedding a document it's taking
1:47:49 all of the words in that document and
1:47:51 it's clustering them into these
1:47:53 mathematical Vector spaces so they call
1:47:55 it a vector database so what you see on
1:47:57 screen right here is just
1:47:59 um two two Dimensions right so you've
1:48:01 got an x y coordinate and then these
1:48:02 these clusters are numbers each of these
1:48:06 dots represents a number in this x y
1:48:09 coordinate space and then they're
1:48:11 clustered by meaning and then each one
1:48:13 is tied to a specific word right so the
1:48:16 the real models are like 500 Dimensions
1:48:20 or a thousand Dimensions or even multi
1:48:23 multiple thousands of Dimensions so it
1:48:25 is impossible to comprehend how complex
1:48:29 it is but this is the simple 2D XY
1:48:32 version of it okay so that's the latent
1:48:34 space
1:48:35 the way chat GPT works is that based on
1:48:39 your prompt the large language model
1:48:41 predicts the next word mathematically
1:48:43 and then
1:48:45 repeats that over and over and over so
1:48:47 if your prompt is the quick brown fox
1:48:50 jumps over the lazy blank
1:48:52 we all know as humans well that's dog
1:48:54 because we've heard this a thousand
1:48:55 times but how did we get to that word
1:48:57 dog well how
1:48:59 um how chat gbt and how these large
1:49:00 language models do it is is in this
1:49:02 latent space when you type your prompt
1:49:05 you're creating a mathematical sort of
1:49:08 probability function that says there is
1:49:11 a word that is the most likely to be
1:49:13 there and so what it does is it kind of
1:49:16 deselects all the Clusters that are not
1:49:18 relevant and then it selects like I got
1:49:20 a bunch of light blue dots here and
1:49:22 these might all be animals right and so
1:49:24 one might be a cat one might be a
1:49:26 whatever and then one of those numbers
1:49:30 is going to have the highest probability
1:49:32 so that says 96 if you go look at what's
1:49:35 actually you know at that Vector space
1:49:37 it's just a long string of numbers that
1:49:40 are the address of that Vector space
1:49:42 and then that string of numbers is
1:49:45 attached to the word dog and then what
1:49:47 the large language model does is it
1:49:48 sticks the word dog index fence and then
1:49:52 it does that over and over and over
1:49:54 again so when it writes you a 500 word
1:49:57 response it's doing it one word at a
1:49:59 time by just doing this over and over
1:50:01 and over again and that's why
1:50:04 um that that idea of tell chat GPT to
1:50:08 act as a screenwriter it will like
1:50:11 narrow your search into into one of
1:50:13 these semantic clusters basically so why
1:50:16 you'll get better answers is because it
1:50:17 got it didn't get smarter it got less
1:50:19 smart you you made it specialize a
1:50:22 little bit right because it's designed
1:50:24 on everything it's got all the latent
1:50:27 space it's got all the words in all the
1:50:29 semantic clusters and when you narrow it
1:50:31 down you're narrowing down that semantic
1:50:33 cluster
1:50:35 trippy right isn't that insane
1:50:38 I lost my uh
1:50:40 oh that's my double camera again
1:50:44 so I'll put on I'll put on a background
1:50:45 we'll do we we've we've been in this
1:50:49 Rodeo a few times we know how to deal
1:50:50 with
1:50:51 this stupid [ __ ] yeah man
1:50:56 um so I hope that answers your questions
1:50:58 it's like a SQL query if instead of
1:51:02 writing SQL queries you just said hey
1:51:04 can I have some [ __ ] and it gives it to
1:51:06 you
1:51:07 hmm
1:51:08 and code interpreter which is coming
1:51:10 this week to a chat GPT near you
1:51:12 apparently
1:51:15 um is going to do that even better
1:51:18 even more better
1:51:22 hi professor Sonya what's happening this
1:51:25 is dope I appreciate that oh it just
1:51:28 jumped me to the bottom again you Tick
1:51:30 Tock I'm afraid
1:51:33 I know I know
1:51:37 the call is coming from inside the house
1:51:42 it's crazy this shit's crazy this shit's
1:51:46 crazy it's crazy oh Vincent Vega is
1:51:49 right from Caddyshack great that's good
1:51:51 thank you
1:51:53 yeah that's pretty legit you pay for
1:51:56 this one uh I pay for chat gbt if that's
1:51:58 what the question was prompt engineering
1:52:00 needs to be better here's the deal with
1:52:02 prompt engineering you ain't gonna need
1:52:04 to do it if you're an engineer if you if
1:52:07 you are if you are building applications
1:52:09 you're gonna have to be good at prompt
1:52:10 engineering but even then
1:52:13 um these things are going to start
1:52:14 prompting themselves and they're going
1:52:15 to start assisting us it's going to get
1:52:17 much much easier to just say make me a
1:52:19 poem about the sky and it will figure
1:52:21 out does it need to ask you more
1:52:23 questions about that does it just go
1:52:25 ahead and pick you know based on all of
1:52:28 your preferences it's going to pick a
1:52:29 poet that it knows that you like
1:52:32 um so that the prompting stuff is going
1:52:33 to drift to the background increasingly
1:52:36 um and probably pretty quickly probably
1:52:37 within a year maybe two
1:52:40 that's amazing I wonder how well the
1:52:42 code runs uh the code runs well my first
1:52:45 um the code that I was showing you
1:52:47 before
1:52:48 um my real breakthrough my Epiphany with
1:52:51 chat GPT happened in December so like
1:52:53 it's like a week or two after it came
1:52:55 out
1:52:56 I'm sitting in the living room with my
1:52:57 wife she's watching some TV show that I
1:52:59 was not interested in and I'm like I'm
1:53:01 like you know adding out on chat GPT
1:53:04 chassis with you it's exciting so I'm
1:53:07 doing that and I said oh I I I read on
1:53:12 Twitter that this chat EBT can code and
1:53:14 I've heard that I should learn Python
1:53:16 and you know I'm I just I don't have the
1:53:18 attention span to do coding but I'm like
1:53:21 maybe this can code so I said hey write
1:53:23 me write me a a python application that
1:53:26 takes an input
1:53:27 and it did it wrote this python code and
1:53:31 then I did exactly what you did uh I
1:53:34 said I realized I realized it wrote
1:53:37 python code I said I don't know how to
1:53:38 read python code I certainly don't know
1:53:40 how to run it right so I said I don't
1:53:42 know how to run python code it said sure
1:53:43 no problem just go to your terminal and
1:53:45 type in this command line and then do
1:53:47 this and install this and this library
1:53:48 but I'm like no no no I don't want to do
1:53:51 anything on the command line isn't there
1:53:52 a website I can just paste this [ __ ] in
1:53:54 and see if it works and it said oh sure
1:53:56 and it gave me replit and it described
1:53:58 okay copy this code from here paste it
1:54:01 into the left pane on replit and hit the
1:54:03 Run button I was like okay and I pasted
1:54:06 the python code into replit I hit the
1:54:08 Run button and it went error
1:54:11 so I copied the error and I went back to
1:54:13 Chachi PT and I said I did what you told
1:54:15 me and it gave me this error it said
1:54:16 well you don't have this Library
1:54:17 installed and I'm like
1:54:20 how do you install the library right
1:54:22 like working with chat GPT is like
1:54:24 working with a very simple intern
1:54:27 that's the smartest [ __ ] intern in
1:54:30 the history of interns like knows
1:54:32 everything but needs Direction on
1:54:34 everything right
1:54:36 so it said sure you go
1:54:39 click on this icon
1:54:41 a search window will pop up in the upper
1:54:43 right search for this Library name hit
1:54:44 install and then it should run and
1:54:46 within 90 minutes I had built a
1:54:50 functioning
1:54:52 content generating application in Python
1:54:55 that I was running in replit
1:54:57 that would ask you for a topic then it
1:55:01 would ask you three questions about that
1:55:02 topic and then it would write you five
1:55:04 pieces of social media content
1:55:07 it's a hashtag Cloud a blog post
1:55:11 and and uh what else did I do oh YouTube
1:55:14 script it's like
1:55:16 in 90 minutes never having written a
1:55:19 line of python fully functional so yes
1:55:21 it runs now
1:55:23 if you're a real engineer real Engineers
1:55:24 are going to look at that [ __ ] and go
1:55:26 well it's sloppy or it's too too tight
1:55:29 or not tight enough or whatever like
1:55:30 there's nuances to it so being an
1:55:32 engineer and and sort of pair
1:55:34 programming with this stuff is the way
1:55:36 to go because experienced Engineers know
1:55:39 what's good and what's not so they can
1:55:41 take the best of what these things do
1:55:43 and and you know leverage them it's also
1:55:45 really good at documentation you can
1:55:47 take code and say
1:55:49 um write me documentation for this code
1:55:50 or for this API and it'll do it or just
1:55:53 pop the code in there have it grab some
1:55:56 code out of a random bot out of some
1:55:59 stock trading site just pop the code in
1:56:02 and tell it to tell you what that code
1:56:04 does
1:56:05 it will tell you that it's a bot and
1:56:07 it'll tell you probably what model it
1:56:09 uses
1:56:13 so she's doing data analytics this is
1:56:15 dope
1:56:16 I don't know where I was before okay
1:56:18 here I was this is where I was okay good
1:56:21 um let's see regenerate geez that's
1:56:24 amazing I wonder how the code runs we
1:56:26 just did that one
1:56:29 damn because it'll hit the news and
1:56:32 three days later it's patched
1:56:35 that's pretty legit not much substance
1:56:38 he was the character Vincent Bagels
1:56:41 right pretty damn good it's just parsing
1:56:43 info it scrapes wrong Chad it's not I
1:56:46 just explained to you how it worked it's
1:56:48 well okay
1:56:50 well it's not parsing it it's it's not
1:56:53 parsing it it's not that's not how it's
1:56:55 working it's it's it's doing word
1:56:58 probability so so what chat EBT is is is
1:57:01 just like when you're in Google
1:57:03 when you're writing a Gmail and you
1:57:05 write most of a sentence and it shows
1:57:08 you the four predicted words that come
1:57:10 next that sentence completion that's
1:57:12 what it's doing
1:57:15 um so it's it's what it's not doing is
1:57:18 it's not taking chunks of like when I
1:57:21 said act as Roger Ebert it's not taking
1:57:23 chunks of his review and copying and
1:57:26 pasting them into the answer
1:57:29 it's kind of amalgamating all of the
1:57:32 possible
1:57:34 um words in that semantic region and
1:57:36 saying of all these words which is the
1:57:38 next most probable single word
1:57:41 and then it takes that entire prompt and
1:57:44 that word and says okay based on all
1:57:45 that what's the next most probable word
1:57:47 and then what's the next most probable
1:57:49 word so it's literally the generative
1:57:51 part of generative pre-trained
1:57:53 Transformer is it's generating original
1:57:55 content it's mind-blowing that it does
1:57:58 what it does like how anyone figured
1:58:01 this [ __ ] out and figured out latent
1:58:03 space and Vector databases and you know
1:58:06 thousand Dimension mathematical like
1:58:09 like just the level of math in this [ __ ]
1:58:12 is just mind-numbing but this goes back
1:58:15 to my favorite there's an Arthur C
1:58:17 Clarke saying any significantly Advanced
1:58:22 um technology is indistinguishable from
1:58:24 Magic this shit's magic it's magic
1:58:28 if you understand how it works I think
1:58:30 it gets more magical
1:58:33 question chat cpt4 includes all of the
1:58:35 visible internet
1:58:37 probably not all of it
1:58:39 but basically all of the public internet
1:58:41 so anything that's behind firewalls like
1:58:43 private corporate data like Bloomberg
1:58:46 just introduced a thing a few months ago
1:58:48 called
1:58:49 Bloomberg GPT where they took all of
1:58:52 their proprietary financial data and
1:58:54 they created a GPT you know sitting they
1:58:56 did all the embeddings of that and and
1:58:59 did a private version of Bloomberg data
1:59:03 that you can interact with with GPT
1:59:06 um so that stuff's not in the model but
1:59:08 basically as much data as they could
1:59:11 find as open AI you know could find and
1:59:14 tag and and do what they needed to do
1:59:16 for the embedding
1:59:18 um that was on the public internet
1:59:19 that's what it's been trained on so so
1:59:22 yeah it's essentially all of the visible
1:59:24 internet like
1:59:26 realistically it's probably some
1:59:29 percentage like 70 I don't know like
1:59:31 they might know like there might be
1:59:33 there might be a percentage that they
1:59:35 know it is and and I assume gpt5 will be
1:59:40 you know if if GPT 3 was 50 and 4 was 70
1:59:45 I'm sure five will be like 95 plus a
1:59:49 bunch of proprietary data sets where
1:59:51 they cut deals with those companies
1:59:55 um how it resonated from Wikipedia so
1:59:58 Wikipedia's reception from Wikipedia
2:00:00 Wikipedia is a general opinion site
2:00:02 he definitely paid for this I don't know
2:00:05 what I definitely paid for
2:00:07 but possibly
2:00:09 Caddyshack sucks Chad you're so dark if
2:00:12 you're gonna troll please be creative
2:00:13 boring trolls not interested
2:00:17 um it parses prompt engineering Are You
2:00:20 The Prompt engineer no I'm a dude I'm an
2:00:23 entrepreneur
2:00:25 um but I'm a writer being good with
2:00:27 words helps with the prompting I sent an
2:00:29 email yesterday had to edit it to not
2:00:31 sound so friendly grumpy old man
2:00:34 obviously you can if it if it makes [ __ ]
2:00:36 that's too
2:00:38 um flowery for you you can tell it to
2:00:39 limit Pros it'll make it more sparse you
2:00:42 can tell it to make it more serious you
2:00:44 can tell it to make it more firm
2:00:48 um I saw I saw a tick tock today I was I
2:00:50 was driving and I was flipping through
2:00:52 tick tocks and this guy this guy goes uh
2:00:56 what do I say rewrite this resignation
2:00:58 letter I [ __ ] hate this company I
2:01:01 [ __ ] hate the way you treat me the
2:01:03 stupid [ __ ] dinners you forced me to
2:01:05 go to every other night or a waste of my
2:01:07 [ __ ] time you don't pay me what I'm
2:01:10 worth I mean he just [ __ ] went on
2:01:11 this you know [ __ ] three minute rant
2:01:13 into his phone and then it rewrote this
2:01:16 professional resignation letter you know
2:01:18 my experience with the company well
2:01:21 positive has been you know
2:01:25 it was pretty good
2:01:27 so yeah you can you can make it go the
2:01:29 other direction
2:01:30 all right I have I have 15 more minutes
2:01:32 and then I gotta I gotta leave
2:01:35 it looks like we got some new folks here
2:01:36 so my name is Kyle Shannon I normally
2:01:38 have a dual camera going but that went
2:01:40 away whatever tick tocks special
2:01:43 um
2:01:44 I'm an entrepreneur I'm the CEO of a
2:01:46 company called storyvine This is the AI
2:01:48 learning lab all these URLs behind my
2:01:50 head why they're here let me get out of
2:01:53 the way
2:01:54 um
2:01:55 if you haven't played with chat GPT I'm
2:01:59 obviously a little passionate about it
2:02:00 you know enough so that I spend like two
2:02:02 or three hours a night talking about
2:02:03 this [ __ ]
2:02:05 um that's the official chat GPD website
2:02:07 so go there right now
2:02:09 start playing with it if you haven't
2:02:10 bing.com is also chat gbt but that's
2:02:14 Microsoft search engine
2:02:16 and if you click on the chat button it's
2:02:18 gpt4 and it's connected to the internet
2:02:20 and you can make free images there so
2:02:23 this is these are the guys that invented
2:02:24 it these are the guys that gave them 10
2:02:27 billion dollars and are completely
2:02:28 incorporating it in everything they're
2:02:30 doing right they're going to roll it
2:02:32 into office 365. it's in GitHub it's
2:02:35 going to be in Windows it's in all the
2:02:37 Azure cloud services it's going to be
2:02:38 everywhere
2:02:39 so that's that and then poe.com is quora
2:02:43 if you know this the question and answer
2:02:44 site quora that's their search engine or
2:02:47 that's their chat bot but you can do
2:02:49 there's two interesting things here one
2:02:51 is you can play with six different large
2:02:53 language models so you can kind of
2:02:54 compare them one of them has a hundred
2:02:56 thousand token context windows so you
2:02:59 can put a lot of words in it and then
2:03:01 they also have a bot maker and so
2:03:02 there's a lot of bots here you can play
2:03:04 with which are just think of them as
2:03:05 like specialized chat gbts that one will
2:03:09 rewrite your mid Journey prompt for you
2:03:11 one will help you write emails there's
2:03:14 like thousands of them now and you can
2:03:15 make your own bot so you can practice
2:03:17 your prompt writing uh by making a bot
2:03:19 and then prompts.chat
2:03:21 will teach you how to prompt these
2:03:24 things how to talk to these things and
2:03:25 how to make them act in ways that are
2:03:28 valuable to you and then the final Thing
2:03:29 Future pedia is a directory of like 3
2:03:33 700 uh AI tools so
2:03:37 can you move the blue window it's
2:03:39 covering your face
2:03:41 uh that was because I think I moved it
2:03:43 on my face audio's fine Milk Bones
2:03:45 there's already a restaurant with a
2:03:47 robot weight staff yep and I think I
2:03:50 think Wendy's is testing a uh a
2:03:53 synthesized voice that talks to you at
2:03:56 the when you pull up to the
2:03:57 drive-through it's not a real person I
2:04:00 think Wendy's is testing that right now
2:04:02 Kurzweil predicted AGI and aging are
2:04:05 going to be overcome by the 2030s Yeah
2:04:06 Yeah his a recent interview with
2:04:08 Kurzweil said um 2029
2:04:11 is is that was the year he predicted and
2:04:14 he said he's sticking by it he feels
2:04:15 like that's on track
2:04:17 open the pod bay doors hell I'm sorry
2:04:20 Dave I'm afraid I can't do that
2:04:26 spelling that's correct h-o-t-z
2:04:29 um wild intelligent crazy as [ __ ] dude
2:04:33 but brilliant it like whenever whenever
2:04:37 I want to be humbled there's there's two
2:04:39 guys right now that that are in that
2:04:41 space for me one is George hotz although
2:04:44 I haven't watched him in a while he's
2:04:45 sort of resurfacing because he's doing
2:04:47 some of this AI stuff and the other one
2:04:49 is this guy named David Shapiro um who's
2:04:51 neuro spicy as he describes it
2:04:55 um you know on this on the Spectrum and
2:04:58 um
2:04:59 and uh he's just brilliant and he's
2:05:02 written uh David Shapiro has written
2:05:04 five books on on artificial intelligence
2:05:07 and he's got a number of Open Source
2:05:10 projects to to create AGI he wants to
2:05:12 create an open source artificial general
2:05:15 intelligence so he wants to do it in
2:05:18 parallel with what open AI is doing but
2:05:20 do it with open source you know
2:05:24 um large language models and with tools
2:05:26 he's writing and he's building a whole
2:05:27 Community around it but he's [ __ ]
2:05:29 Wicked smart and he does like these
2:05:31 these long like PowerPoint presentations
2:05:34 on theories of like economics and like
2:05:38 uh uh uh Universal basic income and and
2:05:41 like what are the implications like if
2:05:43 these things are smarter than us and the
2:05:45 alignment issue of like we want to make
2:05:47 sure that these robots are aligned with
2:05:49 our values so they don't de-prioritize
2:05:52 us
2:05:53 um really really smart guy I appreciate
2:05:55 your willingness to share and respond to
2:05:56 questions will you remind me of the a uh
2:06:00 the AI Salon URL yes I will
2:06:05 that's it
2:06:07 the salon the AI and then this thing
2:06:10 down here is a newsletter that I did
2:06:12 with a guy named Greg mushin so that's
2:06:15 the Twitter handle everyday AI news
2:06:19 um but yeah that's the salon so if you
2:06:21 want to join the salon we've got a
2:06:22 Discord we have meetups we've got a
2:06:25 LinkedIn group and it's just it's like a
2:06:28 group that are just kind of on this
2:06:30 adventure together trying to figure it
2:06:32 out and there's people everyone from
2:06:34 like people have been working in machine
2:06:36 learning for like you know 20 years like
2:06:39 you know baller kind of Engineers to
2:06:42 like people that are like I heard this
2:06:44 say I think it's something I should pay
2:06:46 attention to I'm scared is it gonna kill
2:06:48 us
2:06:50 everyone in between those two ranges
2:06:53 it's it's a cool group
2:06:56 do you think in the future AI can create
2:06:58 new jobs I think AI is going to create
2:07:00 lots and lots of new jobs like like jobs
2:07:02 we can't even imagine right now
2:07:04 um I think it's going to create a lot of
2:07:06 them I think it's going to be a net
2:07:07 dramatic positive
2:07:10 um assuming we can get the alignment
2:07:11 thing right and um
2:07:14 but
2:07:16 short term it's gonna [ __ ] up a lot it's
2:07:18 gonna disrupt a lot of jobs it's going
2:07:20 to change a lot of jobs
2:07:22 any job that's kind of a tactical
2:07:24 execution anything where your your
2:07:28 some form of a cog in the in the wheel
2:07:31 of this big complicated machine where
2:07:33 you're doing repetitive
2:07:36 like Soul crushing the same kind of work
2:07:40 over and over and over again whether
2:07:42 it's cold calling as a salesperson or
2:07:44 doing data entry or cranking out you
2:07:48 know resizing images for social media in
2:07:50 a social media department or cranking
2:07:52 out you know copy for 57 variations for
2:07:56 a b testing anything where you're a cog
2:07:59 in a wheel
2:08:00 or you're a cog in the in the machine
2:08:03 of your company or or Society
2:08:07 those jobs are going to be automated out
2:08:10 now are those people going to be
2:08:12 automated out not necessarily because
2:08:14 you know
2:08:16 someone's got to operate the machine
2:08:18 that the cogs are in so if if the cogs
2:08:21 get automated and you can make more
2:08:23 things faster and more efficiently well
2:08:25 what what does that mean it potentially
2:08:27 means you don't need as many people but
2:08:29 it also potentially means you could use
2:08:31 the same amount of people to generate
2:08:33 you know more output right so you could
2:08:36 give lower prices to your customers or
2:08:38 you could give higher quality or you
2:08:40 could turn things around faster with the
2:08:42 same amount of work some companies are
2:08:44 just going to jettison employees others
2:08:47 are going to find ways to keep employees
2:08:49 and basically upskill them to improve
2:08:51 the company
2:08:53 um but it's gonna be a it's going to be
2:08:55 pretty ugly for my guess is two to three
2:08:57 years and it's already starting like
2:08:59 we're already starting to see I forget I
2:09:00 just saw one today the 40 4 500 jobs
2:09:04 were being eliminated from a big company
2:09:06 direct result of of generative AI
2:09:10 um which AI platform generate graphics
2:09:12 for free so bing.com so let me pull up
2:09:15 that graphic again
2:09:18 bing.com is
2:09:20 um if if you're in you go to bing.com
2:09:23 and then you click on the chat button in
2:09:25 the chat window just say make me an
2:09:26 image of and describe whatever you want
2:09:29 and then if you want go to poe.com
2:09:32 and there's a there's a bot for Dolly
2:09:35 d-a-l-l-e and and you can type in a
2:09:38 little prompt for an image and it will
2:09:40 rewrite it to to give you a better image
2:09:43 so you can go here find a dolly bot
2:09:46 and then go to bing.com and paste The
2:09:49 Prompt in to bing.com and they'll make
2:09:52 you one so that's free
2:09:54 um if you're if you have Adobe Creative
2:09:57 license Creative Cloud it's called
2:10:00 um the you you can get access to Adobe
2:10:03 Firefly which is generative AI so it's
2:10:06 not free but if you if you've already
2:10:07 got a Photoshop license you can do
2:10:09 generative fill within that
2:10:12 um
2:10:13 sign up for a there's a site out there
2:10:16 called leonardo.ai like Leonardo da
2:10:19 Vinci leonardo.ai
2:10:22 and that requires you to sign up for a
2:10:25 wait list but it's very worth it because
2:10:27 they give you like 150 credits a day
2:10:29 which is I think creating an image
2:10:31 there's like eight credits so you can do
2:10:34 you can do a fair amount a day uh and
2:10:36 then I don't think it's super expensive
2:10:38 to do like their lowest tier
2:10:40 um so that's that one
2:10:42 um
2:10:43 stable diffusion
2:10:45 um there are all sorts of free options
2:10:47 but you got to be geeky you gotta like
2:10:49 set up a virtual machine or you gotta
2:10:51 have some decent iron or you got there's
2:10:55 like there's just a fair amount you got
2:10:56 to do to get it running but stable
2:10:58 diffusion is open source and there's an
2:11:00 incredible developer Community like
2:11:01 there's if you want flexibility and
2:11:04 power stable diffusion is the way to go
2:11:06 if you want quality
2:11:09 um Mid journey is the way to go and that
2:11:11 that you you have to pay like I pay 10
2:11:13 bucks a month for Mid journey and that
2:11:15 gives me enough I don't I haven't really
2:11:17 hit my boundaries on that I don't I
2:11:18 don't make a ton but it gives me enough
2:11:20 but it's from a quality standpoint
2:11:22 there's nothing like mid-journey right
2:11:23 now
2:11:25 um if you're in a business and you want
2:11:27 it to be commercially safe and viable
2:11:29 Adobe Firefly and generative fill
2:11:33 because it's trained on non-copyritten
2:11:35 images
2:11:36 uh the crappiest one is Dolly the thing
2:11:39 that sits underneath Bing but that'll
2:11:41 get better at some point um but it's
2:11:43 free
2:11:44 and then there's another one called Blue
2:11:46 Willow that's kind of like mid-journey
2:11:49 you log into a Discord server for Blue
2:11:51 Willow
2:11:52 and then you type uh commands to make
2:11:55 your images I think it's built on top of
2:11:57 stable diffusion but that's free and I
2:12:00 have no idea how they're doing that for
2:12:01 free I don't understand who the [ __ ] is
2:12:03 paying for that but maybe it's maybe
2:12:05 it's one of those models where they
2:12:07 raised enough money where they're like
2:12:08 let's go get customers first and then
2:12:09 we'll figure out how to monetize it
2:12:14 that's too stressful for me I couldn't
2:12:16 start that business I started a service
2:12:18 business and then my current business
2:12:19 storyvine we sold it first like we
2:12:22 basically didn't build storyvine until
2:12:24 we had someone to pay for it
2:12:26 uh
2:12:28 uh the salon.ai you're helping each
2:12:29 other that's awesome
2:12:32 not that trippy up it just jumped me to
2:12:34 the bottom again I lost my pace and
2:12:36 that's my place in the questions people
2:12:38 I have seven minutes left until I turn
2:12:40 into a pumpkin and then I must leave I
2:12:43 must go to the ball Papa I must
2:12:46 oh papa
2:12:48 why do you treat me so
2:12:55 do you need to be great at math to learn
2:12:57 Ai No
2:12:59 well
2:13:01 depends on which side of the inflection
2:13:04 point you are so here's here's the
2:13:05 here's the deal with this
2:13:09 pre-november 30th 2022 the answer was
2:13:13 essentially yes like you need to be you
2:13:15 don't need to be good at math but you
2:13:16 need to be pretty geeky you need to be
2:13:18 an engineer you need to understand
2:13:19 programming you need to understand
2:13:20 python you need to understand
2:13:23 yeah so the answer is [ __ ] yes
2:13:27 November 30th 2022
2:13:31 that site launches
2:13:34 so that
2:13:37 the application that you play with there
2:13:39 chat GPT
2:13:42 the original version of it that they
2:13:44 launched was written in two weeks
2:13:47 two weeks
2:13:50 how did they do it that quickly because
2:13:53 the documentation for chat GPT was
2:13:56 actually uh the the instructions for how
2:13:59 to build chat GPT we're in the
2:14:00 documentation of gpt3 and nobody built
2:14:03 it and so Sam Altman Microsoft's getting
2:14:06 ready to make this billion dollar
2:14:07 investment announcement and he's like we
2:14:09 gotta have something to show because we
2:14:10 can't just show them the the dev tools
2:14:12 so go build chat GPT so they built it in
2:14:15 two weeks and they launched it on
2:14:16 November 30th 2022.
2:14:20 what that tool does or what that tool
2:14:23 did was it basically turned
2:14:26 the English language or your native
2:14:28 language but let's just say for the
2:14:30 purposes of the words I'm saying here
2:14:33 it turned the English language into a
2:14:35 programming language
2:14:36 so you just go there and you just tell
2:14:39 it words and it does [ __ ] for you you
2:14:42 can say make me an application that
2:14:45 calculates the average speed between
2:14:48 this object and that object and it will
2:14:50 go right that application
2:14:52 and it will probably work and if it
2:14:55 doesn't work it will help you debug it
2:14:57 um you can also just ask it to help you
2:15:00 write a paper or a proposal or a
2:15:02 marketing plan or copy or whatever or
2:15:05 create a table so
2:15:07 um no you don't the other amazing thing
2:15:10 about that date is from the day this
2:15:14 launched until they got to a hundred
2:15:16 million users was six weeks
2:15:18 it was the fastest adoption of consumer
2:15:21 technology in history the World Wide Web
2:15:24 when the World Wide Web launched it took
2:15:26 six years to get to 100 million users
2:15:29 this took six weeks so you've got at
2:15:32 this point probably two or three hundred
2:15:34 million people that are using this thing
2:15:36 where the English language now can
2:15:40 deliver like magical results right
2:15:43 you've got all of the knowledge of the
2:15:45 public internet at your fingertips and
2:15:48 knowledge not just information not like
2:15:50 I've got to go read an article about
2:15:52 Python Programming no no it will write
2:15:55 the python code for you
2:15:58 so
2:16:00 um knowing math
2:16:01 probably doesn't hurt but I have a
2:16:04 degree in acting
2:16:06 like
2:16:07 I get
2:16:09 like I've got a I I will occasionally
2:16:12 stumble upon a thing I'm like oh I could
2:16:15 use algebra to solve this and I'm like
2:16:17 wait a squared plus b oh [ __ ] like I
2:16:19 gotta go back to like eighth grade like
2:16:21 all the time so so like no you don't you
2:16:24 don't need to be good at math at all
2:16:31 Reflection Point if you want to if you
2:16:33 want to play on the machine learning and
2:16:35 build models and do things like Lang
2:16:36 chain and do your own embeddings and all
2:16:38 that sort of [ __ ] yeah you should be
2:16:39 good at math you want to use these tools
2:16:41 nope
2:16:43 what you do need to be good at is
2:16:45 critical thinking
2:16:46 that's the new thing like if you're just
2:16:48 sort of a cog in the wheel I just show
2:16:50 up at work I do my thing
2:16:53 that shit's gonna get automated out
2:16:56 the new skill is being able to to
2:16:59 understand what you want to accomplish
2:17:01 and how to
2:17:04 figure out which of these tools can
2:17:06 accomplish that and then when you when
2:17:08 it starts generating stuff go oh that's
2:17:10 good that's bad that's good okay if I
2:17:12 took that and put that here that
2:17:13 critical thinking [ __ ] the value of that
2:17:15 goes up dramatically
2:17:20 all right it learned to auto complete
2:17:22 programming languages by modeling
2:17:24 predictive tests text on flawed code oh
2:17:26 that's cool I didn't know that
2:17:28 big game I love the name that's awesome
2:17:31 the input has lots of flaws so the
2:17:34 generated output will also end up with
2:17:36 similar flaws yeah exactly that's I mean
2:17:38 the hallucination thing they'll get that
2:17:40 worked out it's not magic it no no I
2:17:42 didn't say well it is it is perceived as
2:17:45 magic right any significantly Advanced
2:17:48 Technologies indistinguishable from
2:17:50 Magic that's that's what I was saying
2:17:52 when when that thing [ __ ]
2:17:55 writes you something that you know would
2:17:57 have taken you two hours and it does it
2:17:59 better than you in 30 seconds that's
2:18:01 [ __ ] magic how it does it is not
2:18:03 magic how it does it it's a [ __ ]
2:18:05 calculator
2:18:06 but it's it's a it's a sophisticated
2:18:09 enough calculator that it's like that's
2:18:11 [ __ ] Magic
2:18:13 when when you watch it write code and
2:18:15 then the code [ __ ] runs I did this
2:18:17 with a 10 year old
2:18:18 he he had a a clone of Flappy Birds and
2:18:22 it had a bug in it he couldn't figure it
2:18:24 out and we copied and pasted it into
2:18:26 chatgpt it found the bug you know of
2:18:29 course like that it fixed it he copied
2:18:31 it back into
2:18:33 um
2:18:34 into unity and hit play and it ran
2:18:39 and he did this he was like
2:18:41 and he looked at me and he goes this
2:18:44 changes everything
2:18:47 that's [ __ ] magic I'm sorry
2:18:50 it's magic
2:18:53 I mean Magic's not magic right you know
2:18:55 magic tricks aren't magic it's sleight
2:18:57 of hand and it's like 30 years of
2:18:59 practice
2:19:01 but if you don't know how to do that
2:19:02 it's magic right same thing with these
2:19:05 tools it's 30 years of neural networks
2:19:08 and Mathematics and Technology and [ __ ]
2:19:11 ups and you know Jeffrey hinton's
2:19:14 original you know neural network and
2:19:17 large language model was whatever it was
2:19:19 like 18 words that you know
2:19:21 the original prediction was this in
2:19:24 incredibly small thing so all of that's
2:19:27 in this right all of that history all
2:19:29 that Legacy is there the 30 years of a
2:19:31 magician practicing sleight of hand so
2:19:33 it shows up as magic so chat GPT shows
2:19:35 up on November 30th it's [ __ ] Magic
2:19:39 because the 100 million people that
2:19:42 signed up for that weren't building
2:19:45 neural networks weren't engineering this
2:19:48 [ __ ] they were just like
2:19:50 cut it right could it could it write a
2:19:53 blog post oh [ __ ]
2:19:55 but it but it can't write a tweet good
2:19:57 right oh
2:19:59 but but it could it it doesn't do
2:20:01 hashtags oh
2:20:03 oh my is my job in trouble
2:20:14 corner of the internet their little
2:20:16 corner of knowledge worker of creative
2:20:17 work they're realizing isn't quite so
2:20:21 special after all
2:20:24 and so what do you do with that
2:20:26 that's existential right this is a lot
2:20:29 of people's identity is in the work they
2:20:31 do
2:20:32 and Along Comes A Tool
2:20:34 that can do it like that
2:20:38 it's intense
2:20:42 it's just a load of Matrix manipulations
2:20:45 or multiplications it is
2:20:47 and it's magic
2:20:51 I'm a prompt employee to see now that's
2:20:54 solid Polonius that's good
2:20:56 that's a solid see that's somewhere
2:20:58 between a troll and a compliment this is
2:21:00 this is the kind of trolling we like at
2:21:02 the AI learning lab smart smart trolling
2:21:05 love it
2:21:06 and that's not even trolling that's just
2:21:08 like a good joke
2:21:09 I'm a prompt employee you get it
2:21:13 great session lot learned thank you very
2:21:15 much appreciate it love this thank you
2:21:16 what's the salon url url it's the
2:21:19 salon.ai
2:21:21 can you give some examples of where you
2:21:23 expect big opportunities
2:21:25 will be with autonomous agents thank you
2:21:28 for the rose
2:21:29 um and I gotta go
2:21:31 um
2:21:33 yeah I I think autonomous agents are
2:21:36 ultimately where this goes so right now
2:21:38 we've kind of we kind of have to go to
2:21:40 the to the AI right we have to go to
2:21:43 chat GPT right we have to go to Bing
2:21:46 when when
2:21:48 um these guys add code interpreter
2:21:50 you're gonna have to go there and upload
2:21:52 your spreadsheet right
2:21:56 the autonomous agents are going to be
2:21:59 like so so imagine
2:22:02 that we were all like a big cigar
2:22:04 smoking CEO in the corner office and you
2:22:07 know we got we got the the the the you
2:22:10 know the the uh like the 50s secretary
2:22:12 March get your ass in here you know go
2:22:15 get me lunch
2:22:16 so imagine if everyone had an agent that
2:22:21 you could ask it to do anything and
2:22:23 it'll just go do it
2:22:25 that's where it heads right so so I
2:22:28 think initially it's going to start out
2:22:29 with relatively mundane stuff like
2:22:32 um I wanna I want to
2:22:35 um you know plan you know I want to plan
2:22:39 a nice vacation for the family right so
2:22:42 right now how would you do that well
2:22:43 you'd go you'd you'd go talk to the
2:22:45 family and you'd say you know who wants
2:22:47 to go where and maybe we'll go Google
2:22:49 and look at destinations or I've always
2:22:50 wanted to go here right and you get in
2:22:52 the fight you get in the fight about
2:22:54 where we're going and then and then
2:22:56 everyone's in a bad mood and then you go
2:22:57 to work the next day and then you go
2:22:59 Google you know the your least favorite
2:23:02 place because it's like the it's the one
2:23:04 that you know people won't kill each
2:23:05 other at right so so you go there and
2:23:08 then you start googling right and what
2:23:10 does Google do it presents you with like
2:23:13 8 000 [ __ ] options and
2:23:16 um
2:23:18 60 of them are paid SEO injected
2:23:22 [ __ ] sites just trying to sell you
2:23:25 some rip-off thing right so most of the
2:23:28 time trying to like plan your vacation
2:23:30 is sifting through horrific
2:23:34 SEO just [ __ ] [ __ ] right like like
2:23:38 the usefulness of Google has just
2:23:40 plummeted right as as their Monopoly
2:23:43 increased and everyone started following
2:23:46 their rules for SEO and as SEO agencies
2:23:48 got better and better in what they did
2:23:50 and writing all this fake content to
2:23:52 generate the views to get the people to
2:23:54 click that that right it became useless
2:23:58 so with an autonomous agent I might just
2:24:01 give it a go
2:24:03 you know
2:24:05 um my wife wants this kind of vacation I
2:24:07 want this kind of vacation my kids want
2:24:09 this kind of vacation can you suggest
2:24:12 five places that will make us each the
2:24:17 least pissed off
2:24:20 here's five
2:24:22 I forgot about whatever you know
2:24:24 Gatlinburg Tennessee that's where
2:24:27 Dollywood is okay that could work never
2:24:30 would have thought of it good
2:24:32 can you go now book me
2:24:34 um
2:24:35 you know we want to go in September and
2:24:38 and uh we don't want to spend more than
2:24:40 this much on travel and accommodations
2:24:42 sure Boom the agent will just go off it
2:24:45 will automatically find the best deals
2:24:47 book all the [ __ ] because you'll have
2:24:48 given it permissions for whatever
2:24:50 accounts it will go book all that stuff
2:24:52 and then while it's doing that and it'll
2:24:54 probably do that in 10 minutes maybe
2:24:57 five like depending on how quick those
2:24:59 apis are it'll just go do that [ __ ] and
2:25:01 it'll come back and it'll list it'll
2:25:03 provide you with a list here's some
2:25:04 options do these work for you oh no I
2:25:06 forgot we hate flying in the morning we
2:25:08 want to fly in the afternoon so we just
2:25:10 get to the hotel and go to sleep okay
2:25:12 great it'll just go rebook everything
2:25:14 right
2:25:15 and then while it's off doing that you
2:25:17 say well what are we going to do during
2:25:19 the day because we'll one day we'll be
2:25:21 Dollywood and but we're gonna we're
2:25:22 gonna be there for five days so we want
2:25:25 one day where we do some shopping stuff
2:25:27 that the wife likes and one day we want
2:25:29 to do outdoorsy stuff that I like and
2:25:30 one day we want to do something that the
2:25:32 kids like right so you give those
2:25:34 instructions to your agent you just say
2:25:35 here's here's my goal I want to make
2:25:37 these people happy we have four days to
2:25:39 do it we don't want to spend more than
2:25:40 500 a day on activities and food go book
2:25:44 us daily itineraries
2:25:46 foreign
2:25:50 so right now you can do that you can go
2:25:52 to chat GPT and with plugins you can use
2:25:56 some of the plugins and things like that
2:25:57 so
2:26:01 all right
2:26:03 I gotta go
2:26:04 let's see
2:26:06 so anyway so so that's where they start
2:26:09 I think is with little personal projects
2:26:11 shopping travel
2:26:14 probably a lot of work stuff project
2:26:16 management autonomous agents would be
2:26:19 great at project management why because
2:26:21 all the documents you've already got all
2:26:23 the all of the project management tools
2:26:25 are out there
2:26:26 we also
2:26:28 know how complete all the documents are
2:26:31 like we know that that document wasn't
2:26:33 finished
2:26:34 and we know that Sally was working on
2:26:36 that document
2:26:37 so the agent can just go hey Sally how's
2:26:39 that thing coming along would you like
2:26:41 me to help you finish that document sure
2:26:43 that would be great
2:26:45 here's thirsty can you just give that
2:26:48 the thumbs up Sally and then we'll move
2:26:49 that on to the next step and stay on
2:26:51 track with this so you'll see it and
2:26:53 work and then I think they start to get
2:26:55 more sophisticated and where they
2:26:57 ultimately go is you know Elon Musk
2:27:00 starts to turn his robots into
2:27:03 um
2:27:04 a purchase that you would consider
2:27:06 buying just like you buy a car right so
2:27:09 you're going to have a personal robot
2:27:10 for the family just like you have the
2:27:11 family car you'll have the family robot
2:27:14 and then you can just say you know
2:27:16 you're you're out back working on the I
2:27:19 don't know the the porch the board
2:27:21 popped up on the porch and you realize
2:27:23 you got the wrong kind of nails and
2:27:25 you're just like oh [ __ ] uh Jimmy you
2:27:28 named your robot Jimmy because why
2:27:29 wouldn't you you like Jimmy and Jimmy
2:27:32 comes over like could you run to the
2:27:33 Home Depot and take these nails back and
2:27:36 get me the nail the right Nails sure
2:27:38 it'll go off and I'll hop in your car
2:27:40 I'll drive to Home Depot it'll walk in
2:27:42 it'll freak the [ __ ] out of the people
2:27:43 at Home Depot for the first year
2:27:46 and it'll get your nails and they'll
2:27:48 come back right so so It ultimately
2:27:49 plays out there like it plays this all
2:27:52 this [ __ ] plays out to The Jetsons
2:27:56 it really it really does it plays out
2:27:59 like like we're living in Jetsons and by
2:28:02 the way three days ago the FAA approved
2:28:04 the first fly-in car
2:28:07 um design
2:28:11 three days ago the FAA said yes this is
2:28:14 we we will consider this you know you
2:28:17 know we're going to allow you to put
2:28:19 this design into testing and if they
2:28:22 approve it then we're going to have a
2:28:23 you know an FAA approved flying car so
2:28:26 we're going to literally have the
2:28:27 Jetsons it's coming like
2:28:30 we're not that far away
2:28:34 sorry Dave I'm afraid I can't do that
2:28:39 all right everybody listen I know I
2:28:41 didn't get to all the questions uh we'll
2:28:42 do one more
2:28:44 what AI company should you invest in for
2:28:47 a risk reward perspective you're asking
2:28:49 an actor what company too
2:28:51 invest in I don't know
2:28:53 um I like what Microsoft is doing I
2:28:56 think Google's in trouble but they're
2:28:58 Google and they get a bunch of smart
2:29:00 [ __ ] people there and they invented
2:29:01 this [ __ ] so
2:29:03 um
2:29:04 I would probably be equally surprised if
2:29:07 Google didn't come out with something
2:29:09 remarkable
2:29:10 um
2:29:11 as as if they came up with something
2:29:13 that like like they're probably going to
2:29:16 come out with something quite remarkable
2:29:19 um
2:29:21 Nvidia like like AMD is trying to get
2:29:24 into the Nvidia game but nvidia's got it
2:29:27 kind of locked up right now but they
2:29:28 have a trillion dollar market cap so I
2:29:30 you know do they go to 2 trillion I
2:29:32 don't know but right now they're kind of
2:29:34 the only game in town for this GPU [ __ ]
2:29:36 for AI
2:29:39 um I think anthropic's really
2:29:40 interesting but I don't think they're
2:29:42 public yet
2:29:43 um
2:29:51 I think Apple's gonna [ __ ] kill it I
2:29:54 think apple with the autonomous agents
2:29:56 back to the autonomous agent thing what
2:29:58 Apple's got that no one else has
2:30:01 is
2:30:02 they haven't caved on privacy like they
2:30:05 will not give the government backdoors
2:30:08 to their technology for privacy so
2:30:09 they've got they've got some real
2:30:11 credibility with privacy
2:30:13 um and and that's going to get
2:30:14 increasingly important and and the other
2:30:17 thing that Apple's doing is they're
2:30:19 building their own silicon and they're
2:30:21 they're able to do a lot of machine
2:30:22 learning processing locally so instead
2:30:24 of having to send [ __ ] out to
2:30:28 um big Microsoft servers like when you
2:30:30 do an open Ai call that's going to
2:30:32 Microsoft service so you're sending your
2:30:33 data to the to the machine to the cloud
2:30:36 Apple's going to be able to do all that
2:30:38 processing locally so so I think Apple
2:30:41 their combination of user interface like
2:30:44 Kick-Ass technology and and the Privacy
2:30:47 thing they're going to be able to do
2:30:50 um autonomous agents really well so I
2:30:53 think they're going to be amazing
2:30:55 um and if you want to bet on the Vision
2:30:57 Pro goggles
2:30:59 um I mean you know Apple's always
2:31:01 interesting
2:31:02 um I don't know those are the those are
2:31:04 the ones I'm paying attention to right
2:31:05 now it's it's a fairly small list
2:31:08 um for for me the one that that is
2:31:12 I guess I guess the two that are most
2:31:13 interesting to me right now are are
2:31:16 anthropic and
2:31:19 um open AI although what's the one
2:31:21 there's there's a guy that found a deep
2:31:24 mind
2:31:26 inflection he just he just got funded
2:31:29 for 1.3 billion dollars for their series
2:31:32 B
2:31:33 1.3 billion dollars for a series B
2:31:37 um so that's probably an interesting
2:31:39 company to look at
2:31:41 but anyway all right everybody
2:31:44 um I gotta run I do appreciate you
2:31:46 hanging out with me it's uh it's always
2:31:48 humbling to know this many people can
2:31:50 hang out and I hope this helps I know it
2:31:53 was kind of all over the place tonight
2:31:54 but that's what you get you get add some
2:31:56 Focus sometimes it's not
2:31:58 peace everybody have a good night