aiLL 6/1/23 - Learn About ChatGPT, AI Ethics, Job Loss, & More!

Description

Chapters

Transcript

0:04 foreign
0:10 [Music]
0:16 learning lab I'm Kyle Shannon and my
0:18 Wi-Fi sucks so I was live and then I was
0:21 not live and then I was sort of Live And
0:23 then everyone disappeared and then they
0:25 started to come back and then it went
0:26 away but uh this is the AI learning lab
0:29 if you got questions about AI pop them
0:31 in the comments uh if you're new to AI
0:33 That's the official chat gbt website
0:35 right there and a couple other resources
0:37 I am not affiliated with them in any way
0:40 other than I'm trying to get people to
0:44 get curious about AI
0:47 um it doesn't want you to talk about
0:49 Boston apparently not uh yeah so if
0:51 you're asking about Boston before
0:54 um this document
0:57 the city of Boston it is an incredibly
1:00 well thought out
1:02 well organized
1:04 thoughtful
1:06 concise
1:08 useful
1:11 um
1:12 set of guidelines from a city from
1:15 Municipal
1:17 municipality I think my computer's dying
1:21 um
1:22 so I think that's absolutely fascinating
1:25 and someone was asking me about schools
1:27 and I thought it was the school system
1:29 in Boston but it wasn't it was the city
1:32 of Boston and then it's the New York
1:34 City
1:36 school district is reversing their ban
1:40 of chat GPT sorry I'm changing my I'm
1:44 plugging in my laptop because otherwise
1:46 she is gonna die
1:49 so
1:51 um yeah welcome back uh there's 18 of
1:55 you here now welcome all 18 of you
1:58 hopefully that'll go up a bit but
2:01 nice to see everybody uh this is the AI
2:03 learning lab if you've got questions
2:05 about AI pop them in the comments below
2:08 I'll do it again about them
2:11 um now it looks like some folks are
2:12 coming in uh feel free to share this uh
2:15 chat if you
2:17 um think people might enjoy learning
2:19 about AI chatting about it
2:21 um I'm happy to do that
2:24 so greetings from coyote California
2:28 greetings greetings indeed
2:31 I appreciate it welcome welcome I'm from
2:34 Denver Colorado here I'm uh I'm an
2:36 ex-new Yorker I'm a New Yorker at heart
2:38 I miss New York real bad
2:41 but
2:42 we make our choices so I chose to come
2:45 out to Denver Denver's nice but it ain't
2:47 New York
2:51 I like the sauciness of New York well
2:53 now we're up to 120 folks hey everybody
2:55 this is the AI learning lab Lab
2:59 my name is Kyle Shannon
3:02 um I've got I'm an entrepreneur multiple
3:04 time entrepreneur I'm a CEO of a company
3:07 called storyvine which is a an automated
3:10 video production platform a video
3:12 storytelling platform and uh and but
3:16 it's it's not I'm adding AI to it but
3:19 it's not an AI driven company yet it's
3:22 coming uh we're doing AI generated
3:24 transcriptions so we're we're starting
3:26 to incorporate some things into it
3:29 um so that's pretty exciting but this
3:30 channel is all about getting people up
3:32 to speed on chat GPT and generative Ai
3:35 and answering questions so if you have
3:37 questions about AI ask them in the
3:38 comments below I will try to keep up
3:40 with them
3:41 um I either get yelled at for skipping
3:43 questions or for taking for not keeping
3:47 up with questions so I will screw up
3:49 somehow I've also have been booted off
3:52 live twice now for Community guidelines
3:54 because I've never read them and I don't
3:55 really give a [ __ ] about them but but uh
3:59 I don't know why I've gotten kicked off
4:01 but I just try to I just try to talk the
4:03 truth about AI here all right what do
4:05 you think about the AI drone that took
4:07 out its operator did it actually take
4:09 out the operator did it actually shoot
4:12 them down like you know or shoot shoot
4:14 down the thing or was it a test I don't
4:16 know I I don't know enough about that
4:18 but you know
4:21 these tools are like the generative AI
4:24 tools the large language model tools
4:25 they've been out for a while they've
4:26 been out for for a bunch of years the
4:29 the really sophisticated ones like gpt4
4:31 are relatively new but you gotta think
4:34 that the militaries had their hands on
4:36 some fancy s AI ship for a while
4:40 um
4:41 I would assume the Air Force thing was
4:43 testing but if they're doing live fire
4:45 exercises with AI driven drones
4:48 and the drones take on a mind of their
4:51 own well that's just [ __ ] stupid so
4:53 they should not do that until I got this
4:56 [ __ ] worked out uh I think they need to
4:59 uh put some safety measures in place
5:01 okay uh your live is how I learned about
5:05 chat gbt and now I'm obsessed thank you
5:07 you're welcome wait what's your name
5:09 Holly Holly you're welcome Holly I
5:11 appreciate the uh I appreciate the kind
5:12 words yeah that's why I'm here the whole
5:14 the whole point here is just you know
5:17 um just talk about this stuff get it out
5:19 there it's a little late tonight I was
5:21 out at a CEO Soiree hanging out with a
5:23 bunch of CEOs here in Denver uh which
5:26 was kind of fun but uh this is more fun
5:29 I like this better than hanging out with
5:32 boring CEO types
5:35 um what's the longest length prompt that
5:37 you have used
5:38 is it worth a super long prompt that's
5:41 actually a really good question
5:43 um
5:43 so
5:46 if
5:47 if you want to have chat EPT analyze a
5:51 big long chunk of text
5:54 um so so so one thing I've done one
5:56 thing I did that would live in the
5:58 neighborhood of a long prompt
6:00 was based on this text and then I
6:03 literally copied and pasted my entire
6:05 LinkedIn profile and then I pasted in
6:08 three different bios that I've written
6:11 about me over the years and then I
6:13 pasted in like an article I wrote so I
6:15 pasted in three really long things but
6:17 the prompt was only based on all this
6:20 crap
6:21 and then I said write me a whatever
6:24 three year four paragraph bio
6:27 um
6:27 about me uh for speaking at an AI
6:33 conference basically so add AI to my bio
6:35 right and then so so the prompt itself
6:38 was actually short but what I pasted in
6:40 was quite long
6:42 um the the current limit to chat EPT is
6:46 four thousand tokens or three thousand
6:47 words gpt4 will ultimately go to thirty
6:50 two thousand and probably pretty quickly
6:52 exceed that
6:54 there's a model out there right now
6:55 called clot Claude 100K by a company
6:59 called the anthropic and if you go to
7:01 poe.com which I'll go to right now it's
7:04 I think I actually have it up here
7:06 somewhere
7:10 come on come on internet is my internet
7:13 sucking okay so if you go to poe.com
7:17 what Poe is is it's it's a Chachi gpt4
7:21 is gpt4 Claude Claude instant 100K and
7:24 Claude instant are from anthropic which
7:27 is a company I'm a big fan of and
7:29 there's Chachi BT but Claude instant
7:32 100K means a hundred thousand tokens you
7:35 can paste in a 75 000 Word document into
7:39 Claude 100K and it will parse it I I
7:43 took a transcript from one of the AI
7:46 salons that I run every every other week
7:49 we record them and and transcribe them
7:52 and I pasted a transcription in it was
7:54 like 56 pages and then just started
7:56 started asking it to summarize things
7:58 about the salon so you can do that
8:01 I assume what you're asking here though
8:03 I know I answered a lot of uh things
8:06 there
8:07 um but what I assume you're asking is
8:11 does it make sense to do long
8:13 complicated prompts like all in one shot
8:16 and then try to get the perfect answer
8:18 out of it
8:19 my my experience is that it's not good
8:24 to do that what's good to do is actually
8:25 break it up into discreet little chunks
8:28 so if you want to write the Great
8:29 American novel for example
8:31 maybe what you start with is prompted to
8:33 say teach me about different kinds of
8:35 story Frameworks and it will tell you
8:38 about save The Cat and hero's journey
8:41 and whatever different Frameworks there
8:43 are and and then you can say Okay based
8:45 on you know why don't why don't you
8:47 combine the save the cat story framework
8:50 with Aristotle's
8:52 theory of story I don't know sorry not
8:55 educated enough there why don't you
8:56 think these two-story Frameworks and
8:58 give me a hybrid of those because I kind
8:59 of like both of them and then
9:02 um
9:03 ask me a series of questions about my
9:06 story
9:07 until you have enough information to
9:09 write me an outline and then go back and
9:11 forth with it and then say okay write me
9:14 that outline and then if you like the
9:15 outline great if you don't like the
9:17 outline say hey around the second act
9:19 you know it took a dark turn you know I
9:22 was thinking of this more like a comedy
9:24 and so rewrite the outline is that and
9:27 then you know go in and have do a little
9:29 prompt about the you know first scene or
9:32 a middle scene or whatever it is so I've
9:35 found that it's better to kind of break
9:37 your prompts up
9:38 think of chat GPT less like a Google
9:41 search where you're trying to put in the
9:42 perfect prompt to get back the perfect
9:44 answer and more like you've got a
9:46 collaborative partner that's smarter
9:48 than anyone you've ever [ __ ] met
9:50 because that's what it's like
9:52 20 years of AI don't quite know what
9:55 that means but yes it's been around for
9:57 a while
9:58 um but November 30th the day November
10:00 30th 2022 six months ago five months ago
10:03 is the day that everything changed
10:05 that's the day that chat GPT came out
10:08 here's my contention
10:09 um back in the mid 90s when Tim
10:12 berners-lee launched the World Wide Web
10:14 the internet had been around for decades
10:16 but it was largely the territory of
10:20 scientists and researchers and things
10:21 like that and when the World Wide Web
10:23 came out it kind of made this thing
10:24 available to the rest of us and that was
10:27 when I kind of pounced on it and said
10:29 I'm going to start a magazine and I'm
10:31 going to start an agency and I'm going
10:32 to do everything I can to get my head
10:34 around this thing because I think it's
10:35 going to change everything and it did
10:38 um and I think this is another one so uh
10:41 ai's been around for you know a couple
10:43 of decades now three four five decades
10:45 depending on where you define where it
10:47 began
10:48 and but chat GPT is the thing that said
10:52 hey everyone who's not a data scientist
10:55 how would you like to play with
10:57 something that's [ __ ] Magic
11:00 and and 100 million people did in six
11:03 weeks in six weeks chat GPT went from
11:05 zero users to 100 million users that's
11:08 staggering
11:10 uh context is everything context is
11:12 absolutely everything that's in in
11:14 regard to prompting I assume
11:17 and I agree with you
11:19 [Laughter]
11:22 yeah oh it was a simulated test but AI
11:24 determined that the operator is
11:26 preventing its objective yeah I went
11:27 back and shot him down that's okay good
11:30 I'm glad it was a simulated test I can't
11:32 imagine that the Air Force was was doing
11:34 a lot live fire with large language
11:36 models especially if they're you know
11:38 basing it on you know gpt4 that came out
11:41 five months ago from a private company
11:44 yeah they got some they had some uh they
11:47 got some what do they call it some
11:48 testing to do they got some testing to
11:50 do there okay
11:52 um played with vicuna uh oh have I
11:55 played with fukuna or other local llms I
11:58 haven't
11:59 uh well that's not true I played with
12:02 what did I play with there's one called
12:04 GPT for all is that right
12:07 that you don't need to do any sort of
12:09 technical installation let me see if
12:11 that's right GPT for all
12:16 yeah a free to use locally running
12:19 privately aware chat bot no GPT or GPU
12:23 or Internet required a little bit
12:27 um I think open source large language
12:29 models are going to be huge
12:33 um
12:34 there was an interesting
12:37 um
12:37 document that was leaked from Google a
12:39 few weeks back about that was titled uh
12:43 we have no moat meaning we have no
12:47 intellectual property defensive defenses
12:50 that that open source models for a
12:53 hundred dollars with an open source
12:55 model that can fit on a thumb drive
12:58 um are essentially almost at the same
13:01 performance level as what Google spent
13:03 10 million dollars training in open eye
13:05 AI spent five million dollars training
13:08 um so I I think they're going to be
13:09 quite a big deal also the open source
13:12 large language models don't have the the
13:14 safety guard rails that these other
13:16 things do which may or may not be a good
13:18 thing
13:18 all right
13:21 um context is everything okay let's see
13:25 can you talk about how chat GPT
13:29 wait can you talk about how to I guess
13:32 put chat gbt into learning mode well
13:37 so
13:38 so every time we use chat GPT they're
13:41 they're using it to train and refine the
13:43 core model or not the core model they're
13:45 they're using it to re to to fine-tune
13:49 the the base large language model that
13:51 the training for the base model
13:53 stopped at September 2021 all of the
13:56 stuff that we've been doing since is
13:57 kind of adding to a layer on top of it
14:00 for lack of a better term and
14:03 um
14:05 so so it's constantly learning if what
14:08 you're referring to is how do you train
14:09 it on your data there's a couple of
14:11 couple of issues there if you want to
14:13 train a large language model on a lot of
14:16 your data then you need to use something
14:18 like Lang chain l-a-n-g chain
14:21 which is a series of tools that allows
14:24 you to embed your own data put it into
14:26 what's called a vector database good
14:29 geeky terms here and then be able to
14:31 search on those and be able to connect
14:33 it to Google and things like that and
14:34 then you can have your own sort of
14:36 version of chat GPT trained up on your
14:38 own data that's called Lang chain it's
14:41 relatively straightforward to install
14:42 but it's geeky enough that like I
14:44 haven't done it yet because it's just
14:46 like you need to know some command line
14:48 stuff and it's a bit of a pain in the
14:50 ass
14:51 um
14:52 you can also do what's called I think
14:55 it's called few shot learning which is
14:57 basically this let's say
14:59 let's say you're a writer and you've got
15:01 a particular style when you blog you
15:04 blog with I don't know wit and charm
15:07 rather than saying chat EBT write me a
15:10 Blog with wit and charm what you can say
15:12 is based on this blog post and then you
15:16 paste your blog post into chat gbt
15:19 write me a Blog about topic A B C or D
15:22 with the following facts and figures in
15:24 it and use that earlier example for tone
15:29 match the tone of that earlier piece so
15:31 you can train in a single session you
15:34 can train chat GPT to to mimic you
15:38 um so it's pretty good at that again
15:40 it's got this limited what's called
15:42 context window you have kind of this
15:44 think of it like a 3000 word sliding
15:46 window where anything within 3000 words
15:50 as as you generate more and more stuff
15:53 the stuff at the top is going to get
15:55 forgotten so sometimes you're working
15:56 with chat GPT and it's working great and
15:59 then all of a sudden it's like it forgot
16:01 what you were talking about that's
16:03 because it literally forgot what you
16:04 were talking about so that's some of the
16:06 limitations right now
16:08 all right let's see
16:12 can I explain sorry I'm trying to figure
16:14 out how to read this can I explain how
16:17 to add wait how add-ons work what the
16:20 add-ons uh do certain things
16:23 can't
16:24 the can't chat can't do without okay so
16:28 let me show you plugins
16:31 plugins are interesting I tried to show
16:33 them last night and they they were it
16:35 was not behaving GPT has been
16:38 I think it's been overloaded for the
16:40 past couple of days because
16:42 um my zapier automations are so if
16:46 you're at chat GPT
16:48 in order to access plugins you have to
16:50 be a chat GPT plus member so that the
16:53 little plus sign here
16:56 um means that I this is my little badge
16:59 that says I pay you 20 bucks a month but
17:02 what you get is you get access to gpt4
17:04 then to turn on these fancy new settings
17:08 you go to your settings
17:11 and then in your settings you go to Beta
17:15 features
17:16 and you turn on browse with Bing
17:20 which is chat GPT connected to the
17:22 internet and you turn on plugins
17:24 so that's how you turn the stuff on now
17:26 when you go to chat GPT you have three
17:29 options hang on let me see if I can get
17:32 this to focus that's a little better
17:34 um you have three options you've got
17:35 default which is basically the old gpt4
17:38 not connected to the internet so that's
17:39 the one that says I don't know anything
17:41 past September 2021 you've got browse
17:44 with Bing beta
17:47 um so let me let me con you know do a
17:50 chat with with the one connected to the
17:52 internet so now I can say
17:54 um what's the weather for the next
17:59 three days in Denver question mark so
18:04 obviously you couldn't do that if it
18:06 weren't connected to the internet but
18:08 now it says browsing the web
18:10 it says Denver weather forecast for the
18:13 next three days clicked on a link you
18:15 can open this up and see what it's doing
18:17 so it clicked on weather.com It's
18:19 thinking
18:21 it's thinking
18:23 it's reading content right so so it is
18:27 off performing functions and now it's
18:29 saying the weather forecast is Friday
18:31 the temperature is expected to be
18:33 between 52 and 77 Saturday 50 and 63
18:37 with rainstorms and thunderstorms it's
18:40 been raining and thundering so much here
18:41 in in Denver it's bizarre it's like East
18:43 Coast kind of kind of thunderstorms for
18:46 the past two or three weeks been very
18:48 very strange
18:51 um so anyway that's how you use that one
18:53 okay so now if we want to use plugins
18:59 um you click on the little plugins Tab
19:02 and then
19:03 there's a there's this pop-up menu at
19:06 the top
19:08 um where you've got these are your
19:10 installed plugins so I went to the
19:11 plug-in store and I found things that
19:13 looked interesting and I installed a
19:14 bunch of them and then you can activate
19:16 up to three of them at a time
19:19 so the two that I've got activated right
19:21 now are ask your PDF and Link reader and
19:26 so
19:27 um so what it will do is I can just pop
19:29 in a URL and it will go to that URL and
19:32 read it so let me go find
19:36 is this a PDF yeah this is a PDF so this
19:38 is the Boston
19:41 um
19:42 guidelines for using generative AI so I
19:45 copied that URL so I'm going to come
19:47 back over here to chat EVT and I'm going
19:49 to pop in
19:51 um based on this colon and then I'm
19:55 holding down the shift key to make some
19:57 line breaks without submitting it and
19:59 then I popped in my URL there
20:02 and then I'm going to say
20:04 what are the top three
20:10 recommendations spelled very wrong
20:14 um from this article
20:17 boom
20:18 and so now it should figure out that it
20:22 needs to go use one of its plugins so it
20:24 says using ask your PDF
20:28 and if I open this up here's a little
20:32 chunk of code that it gave
20:34 it generated a summary of the piece
20:39 and I said based on the document the top
20:43 three recommendations are
20:49 be specific about prompt summarizing
20:52 text coding and programming
20:56 oh I see so it's got It's kind of
20:57 categorized them from page four drafting
21:00 documents or letters okay so so the the
21:02 top three recommendations are you can
21:04 use it to draft documents or letters you
21:06 can use it to summarize text and you can
21:07 use it for coding and programming and
21:09 then it gives me some some bullets from
21:10 it so what the plugins do is they give
21:14 you additional little superpowers that
21:18 you you can have these combinations of
21:21 them installed like for example you
21:22 might have
21:23 some food related things like OpenTable
21:26 for for making reservations and whatever
21:30 the grocery one is for delivering
21:32 groceries and you could get calories
21:34 from wool from alpha so you can make
21:35 your own little applications by
21:37 combining the uh the plug-in so I hope
21:40 that answers
21:41 how those work and what they are it's
21:43 taken me a while to get my head around
21:44 them it's it's a bizarre kind of concept
21:46 that you've got this thing that you can
21:48 sort of ask anything in any order
21:51 and then it kind of figures out which of
21:53 those plugins it needs to use based on
21:55 what you're asking and then it kind of
21:57 puts them together sometimes they work
21:59 sometimes they don't all this stuff is
22:01 very very new but here's the other thing
22:03 about plugins Microsoft just announced
22:06 that Bing and Office 365 and windows 11
22:11 and everything that's going to have gpt4
22:14 built into it
22:15 is all going to use the open plug-in
22:18 architecture so any plugins that you use
22:21 in chat GPT are also going to work in
22:23 all the Microsoft products so this is
22:24 going to be this is like the App Store
22:26 on crack
22:28 because all of these things are going to
22:30 be intermingled and intermixed and
22:32 dynamically called it's going to be
22:34 absolutely wild so if you want to start
22:36 getting your heads around it I think
22:37 it's worth playing with
22:39 okay this is the AI learning lab if you
22:42 are new here hello welcome I'm Kyle
22:44 Shannon I like talking about AI crap and
22:46 I'm a pain in the ass and I'm a Gen X
22:48 curmudgeon and
22:49 I swear
22:51 um that URL chat.openai.com that's the
22:55 official chat GPT website there's a lot
22:59 of scam sites and scam apps out there so
23:01 that's the official one they have an
23:03 official app now too
23:04 um but basically I'd say start there
23:06 um start playing with this stuff start
23:08 using it it's it's incredibly powerful
23:10 and so what I talk about here is just
23:12 sort of answering questions if you have
23:14 any questions about AI pop them into the
23:16 comments I try to keep up with them I'm
23:18 old my eyes are going you know all that
23:21 sort of stuff
23:23 um Jim Carrey I don't know what that
23:25 means
23:27 I don't think I look like Jim Carrey but
23:29 maybe I'm making faces like Jim Carrey
23:33 I don't think that was Jim Carrey at all
23:35 but it was something it was some bad
23:37 cartoon of a character wasn't it oh wow
23:39 I'm really behind on my comments so
23:41 here's what I'm doing now I'm scrolling
23:42 back up through the comments so if you
23:44 think I'm ignoring you or I'm not paying
23:45 attention to the latest things that's
23:47 because I'm way back up at the top all
23:49 right but I am trying to answer
23:51 everyone's okay can I explain the
23:52 add-ons I just did that okay tree of
23:54 thoughts and other uh
23:59 I don't know what that word is
24:03 discretized I don't know if that means
24:04 prompting methods have been shown to
24:06 increase accuracy yeah there's there's a
24:08 lot of different methods where
24:11 where the
24:13 answers that a large language model
24:15 gives you are going to be run through
24:17 other checks and balances basically like
24:20 anthropic is one they've got a thing
24:21 called the Constitutional learning model
24:23 where they write a set of rules and then
24:25 when it produces an answer it checks
24:27 that answer against the rules and if it
24:29 violates any of those rules it rewrites
24:31 it until it gets it so so it follows
24:34 them I think that whole line of sort of
24:36 multi-step
24:38 um answer review before it gives you the
24:42 answer I think that's going to be an
24:43 incredibly common and and really
24:46 important thing for things like the Air
24:47 Force where where the plane went back
24:49 and attacked the operator
24:51 in the simulation okay cool that helps
24:54 I'm getting better with these prompts
24:56 yeah keep playing with the prompts keep
24:57 again with the prompts don't I I think
25:00 prompt engineering is the wrong term it
25:03 it is kind of like programming with
25:05 natural language
25:07 um but I prefer the term prompt crafting
25:09 think of it like a conversation
25:12 that you're having with a really smart
25:14 intern or a really smart creative
25:16 collaborator that's that's kind of what
25:18 it's like and if if you start to
25:19 mentally shift
25:21 from I've got to write the perfect
25:23 prompt to let me just start talking to
25:26 this thing and see if if together we can
25:29 solve a problem
25:30 it's like it's like collaboration it's
25:32 like a Saturday afternoon movie from
25:35 when we were kids when I was a kid
25:39 we had Saturday morning TV I'm new to
25:42 this love the topic awesome welcome
25:44 um if if you're new to this please go to
25:46 that website that's the
25:48 chat.openai.com is the official chat gbt
25:50 website I don't get anything from that I
25:53 just have it because that's the official
25:54 one and I think it's important for
25:55 people to play with it prompts.chat is a
25:58 pretty cool site that'll teach you how
25:59 to do prompting it gives you kind of
26:01 this never-ending list of prompts that
26:03 you can just copy and paste and play
26:04 with and start to explore how [ __ ]
26:07 powerful this thing is and then
26:09 futurepedia.io is just a list of lots
26:12 and lots of AI tools there I mean there
26:15 are so many tools being released in the
26:16 wild right now it's impossible to keep
26:18 up with it do I think open AI will go
26:21 public soon I do not I think they're
26:23 going to keep it private for as long as
26:25 they can
26:26 um
26:27 I I think I think being public puts them
26:30 in a whole different place and Sam
26:31 Walton's got it as a capped prophet
26:35 for-profit meaning that the investors I
26:38 don't think can make more than 100 x on
26:40 their investment or it might be 10x but
26:42 I think it's 100x might be a thousand X
26:44 I don't know whatever there's some cap
26:46 on the on the limit so it's not quite a
26:49 non-profit which pissed Elon Musk off
26:52 enough that he went and started a
26:53 competitor uh because because he's Elon
26:59 um and he did that a week after saying
27:01 we should take a six-month pause on
27:03 developing uh AI technology started a
27:06 competitor open I AI a week later I
27:08 thought that was pretty pretty funny
27:11 um all right it was just scrolling and
27:13 found this great to see you at dinner
27:15 tonight oh hey
27:17 nice nice
27:18 Danny we were at a CEO Soiree we were
27:22 just a bunch of CEOs hanging out
27:25 eating sort of Mexican food Talking Shop
27:28 you know what you do
27:31 all right uh do you see the FDA approved
27:34 elon's neural link for human trials I
27:36 know that's crazy if you don't know what
27:37 neurolink is basically you Elon musk's
27:41 sort of contention is that
27:44 these these fanciful phones we have in
27:47 our pockets
27:49 um are are you know we're cyborgs
27:52 already but we have a really shitty
27:53 interface and the shitty interfaces are
27:56 eyeballs and how long it takes the
27:58 signal to get to our brain and how long
28:00 it takes our brain to process all that
28:02 crap uh and he said if you could just
28:05 connect your you know basically the the
28:07 data in your smartphone directly into
28:09 your brain uh then you can do it so it's
28:12 it's a remarkable piece of technology if
28:14 you haven't watched videos on it it's
28:16 well worth it he basically had to invent
28:18 a robot company to do the surgery to
28:22 install these little you know thinner
28:25 than hair like electrodes into your
28:27 brain and it literally sort of punches
28:29 it in and figures out where the blood
28:30 vessels are and punches the electrode in
28:33 between the blood vessels it's insane
28:36 um so anyway yeah so they're doing human
28:37 trials on that and there was also there
28:39 was a guy in was it France or Greece
28:41 somewhere in Europe
28:42 um a guy been paralyzed for 12 years uh
28:45 that they they used large language
28:47 models to read his brain scans and had
28:50 him think about moving his legs then
28:52 they did a Bluetooth Bridge down to an
28:55 um
28:56 and a set of electrodes on his spine and
28:59 now he's walking so they basically use
29:02 AI to bridge a motorcycle injury after
29:06 12 years it's amazing
29:08 this is so hard but I'm trying tell me
29:11 what's hard about it I'll do what I can
29:12 to help um just just keep trying just
29:14 keep playing here's here's one of the
29:16 tricks of if if you're if you're a
29:17 little flummoxed about what to do with
29:19 chat GPT it's very overwhelming when you
29:21 have a tool that can essentially do
29:23 anything from like let's talk philosophy
29:26 of Plato to you know tell me about
29:28 quantum physics and anything in between
29:30 um
29:31 it can be very overwhelming so one thing
29:33 that I found that helps is pick a really
29:35 small project in your life like a
29:37 specific project I want to eat better
29:39 okay have it create a new nutrition plan
29:42 for you and then have it generate
29:45 um the shopping list and then have it
29:46 organize the shopping list by section of
29:49 the grocery store
29:51 um to your specific grocery store and it
29:53 will probably do it uh but but pick
29:56 small little discrete projects that you
29:58 have to solve a problem and then use the
30:00 tools to solve that and you'll get
30:01 better at using it because otherwise it
30:04 just gets completely overwhelming
30:06 is the model's capacity to remember and
30:09 build conversations limited by the max
30:11 amount of tokens in a single session yes
30:13 if you use something like
30:16 um Lang chain
30:17 which is a series of tools strung
30:19 together so Lang chain you've got the
30:21 large language model right so think chat
30:24 GPT GPT three or four
30:27 you've got an embedding engine where
30:30 you're taking some amount of data and
30:32 you're bet embedding it you're making it
30:34 sort of large language model friendly
30:36 you've got a uh what's it called a
30:39 vector database that stores that the
30:41 vector databases where you store the
30:43 memory um and then you've got like tools
30:46 like connecting to Google and things
30:47 like that so Lang chain strings these
30:49 things together and it can have longer
30:52 term memory but if you're just talking
30:54 about chat GPT it's limited to the
30:56 context window which is is three
30:59 thousand words four thousand tokens so
31:00 yes that's if you're talking within a
31:03 single session within chat gbt however
31:05 long your context window is that's kind
31:07 of how long your memory is which can be
31:10 a pain in the ass if you're if you're
31:11 working on on something long generating
31:14 lots of content you can forget the
31:15 earlier [ __ ] and you have to kind of go
31:17 back and grab stuff and keep moving it
31:20 down into the conversation
31:23 um let's see llamas freak
31:26 oh llamas leak from meta is showing you
31:29 can get 99 similar performance with
31:32 local use yeah exactly this is so so if
31:34 you don't know So Meta Facebook
31:37 their large language models like chat
31:39 GPT the core models actually leaked and
31:42 then the developer Community was like
31:44 cool we'll take those and what they
31:48 actually did this was brilliant and this
31:49 happened within like two weeks
31:52 um they took this leaked large language
31:54 model and they said hey why don't we
31:55 take gpt4 and have it train the large
31:59 language model to do what it does and
32:02 they figured out a way to basically have
32:03 chat CPT generate all the training that
32:06 trained the smaller large language model
32:08 and so they're sort of knocking these
32:10 things out like once a week they'll come
32:13 out with ones that like that are now
32:15 approaching gpt4 level quality passing
32:19 similar kinds of tests
32:21 um where they train it on like a hundred
32:23 dollars of compute time I mean it's
32:25 staggering and and they're they're
32:26 they're optimizing them one of the
32:28 things they figured out they call it
32:30 quantit quantization I think is that
32:32 what they call is that the right number
32:33 of syllables
32:35 quantization I think yeah
32:38 um basically where they're they're
32:40 stripping out like 90 of the data out of
32:42 these large language models and they
32:44 still work like they still work
32:45 remarkably well so what what they're
32:48 doing is like pulling out 90 of the
32:50 volume of data compressing them and and
32:53 making like five ten you know 50
32:56 megabyte installs that you can install
32:58 in a phone and run on a phone
33:01 um and and they're incredibly high level
33:02 of of performance it's amazing
33:06 um okay check out text generation web UI
33:11 web UI similar things exist for image
33:14 gen but it makes it easy okay I will do
33:17 that what types of jobs will be
33:18 displaced
33:22 I think almost
33:25 every knowledge working job creative job
33:29 consulting job
33:31 anything where you're using knowledge is
33:34 going to be disrupted massively like
33:38 massively and then I think any job
33:41 that's impacted by complex systems so
33:45 think you know
33:47 I.T security
33:50 um
33:51 those are all knowledge workers
33:53 Logistics
33:55 um all sorts of things like that where
33:57 you might have people who've got jobs
33:58 where they're doing data entry or data
34:01 analysis or things like that where kind
34:03 of the low level execution of of all of
34:05 the really complicated stuff are
34:08 probably going to be automated out and
34:09 and and just disrupted away but you're
34:12 gonna you're gonna see disruptions I I
34:14 mean I'm already starting to see it in
34:16 the advertising world where
34:18 um the just the rules are changing kind
34:21 of in real time for what you can charge
34:23 for creative and how long how long you
34:26 can take to turn it around like clients
34:28 are already expecting higher quality
34:31 results faster and cheaper there used to
34:34 be this old adage in the ad game called
34:36 you you know good fast and cheap you can
34:38 have any two of the three right it can
34:41 be fast and good but not cheap or it can
34:44 be fast and cheap but not good well with
34:46 generative AI you can have all three
34:50 um and I think clients are starting to
34:51 know that so I think I don't think that
34:53 any jobs are safe safe from disruption
34:56 um largely speaking
35:00 um I almost want to let AI save Earth at
35:03 this point I you know what I'm with you
35:05 like I feel like who's that Alexander
35:07 yeah I I feel like we've we've certainly
35:09 here in America but but it seems like
35:12 we're we're kind of tanking this thing
35:14 globally as well it just feels like
35:16 we've gotten so divisive and weird in in
35:18 this country that um
35:21 as painful as a reset is going to be I
35:24 think we need a reset and I think kovid
35:26 was the beginning of that reset and I
35:27 think generative AI is the second half
35:29 of that reset the good news about
35:31 generative AI is that while it's going
35:35 to displace a lot of people and disrupt
35:37 a lot of jobs it's also putting these
35:39 superpowers in the hands of the people
35:42 that are being displaced so the
35:44 implication of that the way I think
35:45 about it is
35:46 okay
35:48 um someone decided not to learn Ai and
35:50 they got laid off from their job or they
35:52 just you know were doing their job and
35:54 all of a sudden AI you know the boss
35:56 said hey we're gonna have ai do that now
35:57 good luck
35:59 um
36:00 a lot of those people are gonna be like
36:02 wait a minute and and uh you know I
36:05 didn't ask for this and they're gonna
36:06 find that job doesn't exist for them
36:09 anymore
36:10 and some people are going to be you know
36:12 upset and pissed off and sad and kind of
36:14 stay in that place right they'll go into
36:16 sort of
36:18 um just kind of Despair I think some
36:20 people will be
36:23 um
36:24 you know displaced and they're like okay
36:27 well I'm Gonna Learn these tools and I'm
36:28 gonna go back and I'm going to compete
36:30 against my ex you know my ex-employer so
36:33 I think we'll see some Innovation like
36:34 that and then I think you're going to
36:36 see a whole other wave of people that
36:37 are like they get laid off from the job
36:39 it's going to be initially really
36:40 painful and then what they're going to
36:43 realize is I hated that job anyway I'm
36:47 glad I don't have that job you know what
36:48 I've always wanted to do blank and then
36:51 they'll use generative AI to to
36:53 completely reinvent themselves I think
36:55 we're going to just see an absolute
36:57 explosion of people Reinventing
36:59 themselves inventing things creative
37:02 expression I I think I think we are
37:05 entering what I'm calling the great
37:06 Renaissance that's going to make the
37:08 traditional Renaissance look look quaint
37:10 by comparison
37:13 um I Heard it can access hyperlinks
37:14 posted yeah that's I just showed that
37:17 before so now I've got two plugins
37:19 installed one's called ask your PDF
37:22 and another one I have turned on is
37:24 called link reader and you can basically
37:25 just go get any website
37:27 so let me pop over here to the Twitter
37:33 let's go find a
37:39 an article let's see
37:42 oh that's just my
37:45 profile
37:49 um
37:51 let's see let's type in
37:55 white paper by the way I'm in Denver and
37:59 the Denver Nuggets won their game
38:01 against the Miami Heat tonight we're
38:03 going to see a a uh a Denver
38:05 championship
38:07 well I mean we are seeing the
38:09 championship are we going to see them as
38:10 Champions I can't get it to focus
38:11 anymore come on
38:14 God dang it
38:17 there we go
38:19 um let me go grab a white paper
38:21 advancing the next generation of
38:24 download it now okay this is a white
38:26 paper
38:28 is there anything here oh download white
38:30 paper
38:32 it's gonna make me fill out a form no
38:34 okay
38:35 so that's PDF so I can just grab that
38:36 URL
38:37 and I can pop over here to chat GPT and
38:41 go based on
38:44 this
38:46 and there's the URL please
38:50 tell me
38:53 what this is about as if I were a
38:59 five-year-old
39:02 Bang
39:03 and then assuming it works it should
39:06 fire up one of the plugins okay using
39:09 ask your PDF it's doing it
39:12 and let's see what it thinks a 50 year
39:15 old should know about that article
39:18 I don't I didn't even look at what that
39:19 article was called so
39:22 yeah I got nothing so let's see what it
39:23 comes up with
39:27 I'm sorry but there was an issue
39:28 processing the PDF all right so that one
39:30 didn't work
39:32 uh
39:33 [Music]
39:35 upload your document oh that's
39:37 interesting I have a feeling click here
39:39 to upload your document I've never seen
39:41 this before
39:43 I don't feel like doing it anyway so so
39:45 yeah so what the plugins do is is they
39:48 kind of extend what what chat GPT can do
39:55 uh
39:58 um it says chat GPT you can't refer
40:01 it says chat EBT can't refer online
40:03 website yeah so if you're not paying for
40:08 chat GPT plus you don't have access to
40:11 gpt4 if you do have access to to plus
40:14 then you can turn on web browsing the
40:16 other thing you can do is you can use
40:18 bing.com so if you go to bing.com
40:25 it says ask me anything there's TPT and
40:29 there's sort of a more creative version
40:31 a more balanced version and a more
40:33 precise version this is gpt4 and it's
40:35 connected to the internet
40:37 so
40:38 um so yeah you can use all that and that
40:41 and Bing is going to support plugins so
40:44 if you're new here that's the official
40:46 chat GPT URL so go there and start
40:49 playing and then ask me any questions
40:51 you have about AI it repeats itself
40:53 sometimes too yeah it does it also lies
40:56 a lot I like to think of chat GPT as
40:58 mansplaining as a service
41:00 because it will confidently tell you
41:03 what it thinks even if it's wrong
41:05 jump scare you mentioned zapier do I
41:09 like it how do you use it so I've been
41:11 using zapier to create I'll show you a
41:14 couple of things I've done with it
41:16 one organization that I'm actually the
41:19 chief generative officer for
41:21 I don't know that there are any other
41:23 Chief generative officers in the world I
41:25 think I might be the first
41:27 but we created this thing called The
41:29 Challenge engine which you can give it a
41:31 business challenge so this is a little
41:33 Google form here that's embedded on the
41:35 website you put in your email your
41:38 business category you give it a
41:40 challenge that you're facing answer a
41:42 couple of questions about it and then
41:44 within a few minutes
41:47 let me pop over here to Gmail
41:51 within a few minutes of asking the
41:54 challenge engine about your challenge
42:05 so I gotta find one of the answers here
42:09 challenge engine response
42:12 it generates an email that talks about
42:15 you know it restates your
42:19 hang on
42:24 there we go it restates uh what your
42:26 challenge was and then it provides you
42:28 with three questions in six different
42:31 categories and so zapier takes the input
42:36 of the Google form
42:37 it sends the the data that it collected
42:41 in a prompt that that I created with the
42:43 team at content Evolution it sends that
42:46 to open AI
42:48 um it gets back a bunch of raw content
42:50 and then I basically send that raw
42:53 content back to open AI to do all the
42:55 the HTML formatting so so all of the
42:58 HTML formatting in this email was done
43:01 automatically by
43:03 um
43:04 buy some just some decently clever
43:06 coding and buy a a nice prompt a
43:09 well-constructed prompt to open Ai and
43:12 then it takes the output of that HTML
43:16 pops it in an email and sends that via
43:18 Gmail so that's one of the things I've
43:21 built so I've been building these these
43:22 content generation engines
43:24 um with zapier do I like it
43:26 it's a janky piece of [ __ ] but they've
43:30 got like 5 000 Services connected to it
43:33 um it's the the challenge with all of
43:36 the no code
43:37 um tools is that they are they
43:42 to make them no code they have to put a
43:44 lot of interface in front of the code
43:46 which makes it
43:47 unnecessarily complicated so you kind of
43:50 have to learn all of the intricacies of
43:53 how their no code stuff isn't coding
43:55 which is kind of like coding like it's
43:57 it's probably about as easy to learn
43:59 coding as to learn zapier but not really
44:02 once you kind of learn the quirks of
44:04 zapier you can knock stuff out pretty
44:05 quick the problem with it is just
44:08 you're using what zapier effectively
44:11 does is it is it links together multiple
44:13 API services so you're taking multiple
44:16 cloud services and just kind of stacking
44:17 them on top of one another to make these
44:20 automations
44:21 if any one of those cloud services goes
44:23 down your whole application kind of
44:25 fails I mean but that's that's cloud
44:26 computing in general but when you're
44:28 using something like zapier if if they
44:30 change some code or the third party
44:33 changes code it messes stuff up but I
44:35 like it it's pretty good there's one out
44:37 there called make m-a-k-e that looks
44:40 good looks really pretty I think it's
44:41 newer so but it's probably got some some
44:44 uh you know some some quirks to it as
44:47 well but it looks prettier than zapier
44:50 um
44:51 others need to download to use I don't
44:54 know what you're talking about there
44:55 um maybe you need a very
44:58 maybe you need a new provider you
45:00 unfortunately lagging very often yeah my
45:03 um Wi-Fi here is
45:06 horseshit it's Xfinity and it's bad if
45:10 anyone knows anyone at Xfinity tell them
45:11 to get their [ __ ] together in North
45:13 Denver it's really bad it's really bad
45:15 can you upload files uh you can't right
45:19 now on chat GPT but there's a there's a
45:21 a mode coming in chat GPT called code
45:24 interpreter where you can upload files
45:26 and it is remarkable
45:30 um zapier last I used IRT I don't know
45:35 what that means IRT in in real time
45:38 [Music]
45:40 ascots analysis for the Ukrainian
45:43 Russian conflict nah I've I've been I've
45:47 been booted off of Live Twice now for
45:50 seven days each for Community guidelines
45:52 I have no [ __ ] idea what I said but
45:55 the second time I was booted off I was
45:56 talking about
45:58 um someone asked me about bias in the
45:59 models and I was explaining how the bias
46:01 might get in there and and just in
46:03 talking about it they kicked me off so I
46:05 don't know I'm not going to get into
46:06 anything politics you can explore that
46:08 all you want
46:10 uh uh uh uh uh map or
46:14 esri integration ideas I don't know what
46:19 those are but you know what I can do
46:26 what oh wait I'm just gonna go to normal
46:30 I'll go to I'll go to uh chatgpt with
46:34 Bing what our map what is it ersi e s r
46:39 i n
46:42 Integrations I have a degree in acting
46:44 so I don't know a lot of the super
46:45 technical stuff here
46:47 environmental systems Research Unit
46:50 GIS software okay
46:53 oh map map okay so basically mapping
46:56 software
46:57 um
47:00 boy like a lot of ideas I think I think
47:03 I think you're gonna see Google come out
47:05 with a [ __ ] ton of AI tools sitting on
47:08 top of Google Maps
47:11 um but you but you know so so let me let
47:14 me see
47:14 um what are some scripting or coding
47:20 languages uh for mapping
47:24 which are the most commonly used
47:31 so let's see what it's going to tell us
47:33 here so here's some so python okay
47:36 Python's a scripting language sure of
47:38 course python does everything JavaScript
47:43 um
47:46 sequel yep
47:51 um
47:51 can you make me code
47:57 in uh let's see yeah I'll do python
48:01 um
48:02 mapping census
48:08 data to
48:11 major U.S cities that's a really crappy
48:15 prompt but let's see what it gives me
48:18 to map Census Data you'll need a few
48:21 different components the Census Data
48:23 itself the list of cities
48:25 you can do geo you're going to need
48:28 libraries
48:30 geopandas
48:32 matplotlib and pandas libraries and then
48:35 it wrote the code and then it explains
48:38 what the code does this is a very basic
48:40 example
48:42 now it's talking about creating a census
48:44 data CSV and how you would use that so
48:47 so yeah I think it gives you
48:50 explanations of what to do so so yeah I
48:52 think you can do all sorts of crazy
48:53 stuff with it and and with actual
48:56 programming
48:57 um where you're doing some coding and
48:59 connecting data sets together I think
49:00 you can do remarkable things but I
49:02 Google's going to come out with some
49:03 crazy stuff
49:05 um also Apple has their worldwide
49:07 developer conference announcement on
49:10 Monday Monday at
49:13 10 a.m PST
49:15 ific time I think it's 11 A.M Denver
49:18 and
49:20 um
49:20 they're going to be announcing their
49:22 fancy AR goggles
49:24 um that are probably going to kick the
49:25 [ __ ] out of meta's uh whatever the hell
49:28 those are called
49:29 their VR goggles that
49:32 um are kind of big and bulky but I think
49:34 they're also going to be
49:36 announcing some AI stuff
49:39 uh maybe things like agents like one of
49:42 the things they've already announced is
49:43 that in the next IOS operating system
49:46 you're going to be able to clone your
49:48 voice and I would be very surprised if
49:50 they didn't connect that with their
49:51 memoji technology where you'll be able
49:53 to take you know some version of your
49:55 face some version of your voice and
49:57 create a generative clone like like an
50:00 avatar that's that sounds like you looks
50:03 like you
50:04 um
50:05 and they're going to do some interesting
50:07 stuff with generative agents I do have a
50:09 feeling all right
50:11 a resume plug-in let me upload a file to
50:14 chat GPT that's cool yeah there's so
50:17 many plugins right now I'm very behind
50:18 on I think there's 160 or some some
50:21 crazy amount of plugins let me go see
50:24 how many are there
50:25 um
50:27 plugins plugins plug plugins go to the
50:30 plugin store
50:35 um
50:36 all
50:38 so there are 56 times four pages so
50:44 224 224 plugins it's a lot of plugins
50:48 it's a lot of plugins
50:50 it's a lot of plugins I don't know if
50:52 you knew this there's a lot of plugins
50:53 why is he talking like that I don't know
50:55 it's a lot of plugins a lot of plugins
50:58 all right Quest bought us West okay
51:02 bias is inherent I agree with that
51:03 although I think anthropics
51:05 constitutional learning model
51:09 um it's not going to get rid of bias but
51:11 I think what it will allow you to do is
51:13 they're going to be different models
51:15 where you can opt in to the bias of your
51:17 choice if you want one that leans more
51:19 this direction or that direction you can
51:22 do that or if you want something that's
51:23 particularly neutral you can do that
51:25 don't ask it for code why not it's it's
51:28 brilliant at code
51:30 all right what else do we got here other
51:32 questions sorry I lost uh Tick Tock if
51:36 I'm scrolling when new questions come in
51:38 it jumps me all the way to the bottom
51:40 and then I lost where I was in my little
51:41 line it's it's annoying
51:44 it's annoying this Tick Tock thing
51:51 Universal basic income I think that was
51:54 probably in response to me talking about
51:56 Job disruption
51:58 um
51:59 I think Universal basic income is going
52:01 to be a tool that's that's important I
52:04 think it's going to be a short-term
52:05 important thing
52:07 um I finally understand it Andrew Yang
52:09 was talking about it in the last
52:10 presidential election and I I was like
52:13 conceptually I'm like yeah okay that
52:15 that sounds right but I didn't quite get
52:16 it and uh now that I've been digging
52:19 deep into this AI stuff I'm like oh I
52:21 get it is chat GPT plagiarism so it's
52:24 well okay that's an it's a really
52:27 interesting question
52:29 um
52:31 the model itself is trained on kind of
52:34 all of the internet all of the public
52:36 internet there's a bunch of stuff out
52:38 there that's copy written so it's
52:39 trained on some copywritten data and
52:41 then it's trained on lots and lots of
52:43 non-copy written data So within the
52:45 model within so think of it kind of like
52:48 the large language model as a brain the
52:50 brain went to lots of universities and
52:52 learned lots of [ __ ] right and a lot of
52:54 the [ __ ] that it learned is copywritten
52:56 [ __ ] and so if you ask it to give you
52:58 some specific you know copywritten thing
53:01 it will give you things that look like
53:03 that it might quote it accurately it
53:05 might quote it inaccurately
53:08 um but you can get it to write in the
53:09 style of a certain person and and
53:11 there's a very much a gray area there
53:13 but the one thing that these large
53:15 language models are not doing
53:17 they're not copying and pasting chunks
53:21 of copy written text and and merging
53:23 them together that's not how it works
53:26 they're literally typing or sort of
53:29 retrieving word by word the next most
53:32 probable word in a whole string of words
53:34 so you type in a prompt and and when you
53:37 submit that prompt it figures out
53:39 mathematically based on all of the words
53:42 that you said there what is the next
53:44 most likely word in all of the words
53:46 that I know which is like all of the
53:48 words across the internet is the next
53:50 most likely word and then what's the
53:52 next most likely one after that so it's
53:54 literally generating original content
53:57 out of this massive set of knowledge
54:01 um
54:02 the other thing about
54:04 the the way
54:06 you actually use these tools versus how
54:10 they're written about in The Press a lot
54:12 a lot of how how these tools are
54:14 described as like the robots are
54:17 stealing you know are stealing from the
54:19 humans and the robots are doing this
54:22 that's not quite right the humans are
54:26 using this tool called a large language
54:28 model which is a bot like thing but it's
54:32 a it's a large language model it's a
54:33 prediction model actually
54:35 humans are saying hey all of this vast
54:39 knowledge can you help me write a
54:41 paragraph that sounds like Emily
54:43 Dickinson and out it will come
54:45 is that plagiarism no because it's
54:47 writing original content is it is it a
54:51 very much a gray area that this the the
54:53 base model is is trained on all this
54:55 stuff I think it's a gray area because
54:57 it's kind of like this if I went to
54:59 we'll take images for example if I went
55:02 to art school for for four years and I
55:05 studied pop art and I love Andy Warhol
55:07 and I learn everything I can about Andy
55:09 Warhol
55:10 and I get out of school and and you know
55:13 I'm totally inspired by Andy Warhol and
55:16 I create pieces that look like Andy
55:17 Warhol
55:19 um is that stealing um well you know
55:22 depending on I I think is it the Keith
55:25 Herring it's it's one of the one of the
55:27 the foundations that for one of the
55:29 artists that have passed sues the [ __ ]
55:32 out of people if if pieces look too much
55:34 like that artist
55:36 um others don't really give a [ __ ] um so
55:38 you know I think it's very much a gray
55:40 area I think the IP attorneys are gonna
55:42 make a fortune in the next 20 years
55:45 um I'm I'm not worrying too much about
55:48 any of that stuff I'm just trying to
55:50 keep my head down and do the work
55:53 um can I AI help bury a body I mean it's
55:56 actually an interesting question there
55:58 are all sorts of safety
56:00 um guard rails built into chat GPT that
56:02 if you ask it what it thinks is
56:04 dangerous stuff it will avoid doing that
56:07 but then there's a whole hacker
56:08 community that works around what they're
56:11 called jailbreaking chat GPT to do all
56:13 the stuff that it's not supposed to do
56:15 uh if you want to
56:17 um if you want to explore that go to
56:19 Reddit and and search for uh chatgpt
56:22 damn d-a-n do anything now so there's
56:25 sort of this whole battle going back and
56:27 forth between the open AI team and this
56:30 group of hackers on Reddit writing these
56:31 Dan Bots which are basically just super
56:33 long sophisticated prompts that trick
56:36 chat GPT into ignoring its safety
56:39 guidance because why not because this is
56:42 how we play in the internet worlds all
56:45 right I've been working on AI work I was
56:48 wondering what was up with the tokens
56:50 yeah tokens are a pain in the ass right
56:52 now
56:53 um I saw a thing today if so Nvidia just
56:56 announced a new chip that some crazy AI
57:01 you know happy thing
57:04 um we're probably in not too long an
57:08 amount of time we'll have context
57:10 lengths of like a million tokens like it
57:12 will be it's going to be kind of like
57:14 you know RAM on computers like when I
57:17 was growing up in the 80s like you could
57:19 you could you know you could dial up to
57:20 the Internet and literally watch the
57:22 characters type across screen like
57:23 everything was so slow and you know
57:25 every time it got better and better and
57:27 better and better and now it's just good
57:30 enough and it'll be the same thing with
57:31 these context Windows right now they're
57:33 they're a pain in the ass limitation but
57:35 but still like even with the limitations
57:37 what what this stuff can do is
57:39 remarkable loving this convo thank you
57:42 very much I appreciate that will AI be
57:45 the end of the world I don't think it
57:46 will like here's the deal
57:49 you know are there a lot of dangers
57:52 there and could it absolutely
57:55 be you know put in the hands of Bad
57:57 actors to do bad stuff with it yeah but
58:00 these tools are also in the hands of the
58:02 good actors
58:03 and I also think that we've got 50 years
58:05 of Hollywood that's told us pretty
58:08 convincingly that the robots will kill
58:10 us
58:11 and so I think I think at least in
58:14 America I don't know how it is globally
58:17 I think there's a default fear of AI
58:20 just because of Hollywood that they're
58:23 gonna [ __ ] kill us
58:24 what's a
58:27 what I think some a more interesting
58:30 um
58:32 um sort of way of looking at the future
58:35 is well what if they dramatically
58:37 improve things what if they amplify our
58:39 humanity and and
58:42 we channel the power of them in a good
58:45 way so there there's a non-zero chance
58:47 that it goes incredibly good as well
58:51 right I think we all know the incredibly
58:53 bad it can go I think it can also go
58:55 incredibly good and that's things like
58:57 solving disease solving energy
59:01 um you know cancer genetic diseases
59:04 neurological diseases
59:07 um
59:07 just scientific breakthroughs you know
59:11 happening on a weekly basis that
59:14 historically would have taken decades
59:17 I you know I if you look at the
59:20 advancement in what's happening in AI
59:21 right now that's what's happening like
59:23 every week I do a newsletter called
59:25 everyday Ai and every week is dozens of
59:30 remarkable things that are happening
59:34 um so you know
59:36 it could it could kill us all but I'm
59:39 I'm trying to live on the other side of
59:40 that because otherwise it'd be pretty
59:41 [ __ ] depressing what's the cost for
59:44 me to walk again I don't know if you're
59:46 not walking I'm sorry about that but you
59:48 know I
59:49 they just I mean they just successfully
59:52 did the first Bluetooth bridge between
59:55 the brain they they jumped
59:57 um with Bluetooth they jumped a spinal
59:59 cord injury and and that guy's walking
1:00:01 after 12 years of not walking I I assume
1:00:04 it was done at a university
1:00:06 um there's limitations to it obviously
1:00:08 but but it's the first time that they've
1:00:10 done that so
1:00:12 um so I think breakthroughs are coming
1:00:13 in that which is really really amazing
1:00:15 what's the base AI tool that can predict
1:00:18 outcomes of football matches
1:00:20 um
1:00:21 you know start using
1:00:24 um when code interpreter comes out dump
1:00:27 all the all the football match data into
1:00:30 code interpreter and go crazy
1:00:33 um tell me again how llm is used to
1:00:35 create another llm I don't quite
1:00:38 understand it exactly but
1:00:42 so how they train these models is they
1:00:45 they use what's called human
1:00:46 reinforcement learning or something like
1:00:48 that human something reinforcement
1:00:50 learning
1:00:52 so they they build this base model and
1:00:55 then they start asking it questions and
1:00:57 and and there's a bunch of tag data in
1:01:00 in the set that they train and so
1:01:03 there's part of it is
1:01:05 getting good data into the base model
1:01:08 that's when Sam Altman talks about open
1:01:10 ai's Advantage he said they spend a
1:01:12 tremendous amount of time and money
1:01:14 getting the right data into the model so
1:01:16 the the first thing is getting the the
1:01:18 base data right
1:01:20 and then they use humans to kind of
1:01:23 write questions get answers back and
1:01:26 then say yes that was good or no that
1:01:28 wasn't good
1:01:29 and then I assume they also put in like
1:01:31 here's what might be better
1:01:33 and so
1:01:35 after tens of thousands of this human
1:01:38 reinforcement learning that was good
1:01:40 that was bad that was good that was bad
1:01:41 that was good that was bad the the large
1:01:43 language the fine tuning of what's
1:01:45 called you got the base model down here
1:01:47 you got the fine tuning up here
1:01:49 at a certain point there's this Tipping
1:01:51 Point where it just gets really good
1:01:54 like like remarkably good
1:01:57 so
1:01:58 so what the what the open source
1:02:01 community did was they basically said
1:02:03 okay we've got the base model that meta
1:02:05 built and it leaked so why don't we use
1:02:07 gpt4 rather than having humans do all
1:02:10 that typing and we have to pay them 5 or
1:02:12 10 or 15 bucks an hour why don't we just
1:02:15 have chat GPT simulate humans and have
1:02:18 chat GPT write the questions and answer
1:02:20 if they're good or not and so they're
1:02:22 basically using a fine-tuned trained
1:02:25 model gpt4 to train an unfine-tuned you
1:02:29 know open source model with no
1:02:30 restrictions on it so that's how they're
1:02:32 doing it it's remarkable Geeks are
1:02:34 amazing when when Geeks get their mind
1:02:37 put their mind on something you know
1:02:38 they just you know the hacker mentality
1:02:41 is amazing what do I think about
1:02:43 companies using their own version of
1:02:45 chat EPT to help their their software
1:02:47 um engine yeah or engagement or whatever
1:02:49 oh Engineers um yeah I I think that I
1:02:53 think that
1:02:54 personal data sets or company data sets
1:02:57 internal data sets are going to be
1:02:58 massively powerful I think uh Microsoft
1:03:02 has announced that that large language
1:03:05 models are going to operate at the
1:03:07 operating system level so Windows 11 is
1:03:09 going to have you're going to be able to
1:03:10 just ask it you know fix my you know
1:03:13 Windows like configure this printer
1:03:15 driver whatever it might be and it'll
1:03:17 just go do it so these language models
1:03:20 are going to be throughout the Microsoft
1:03:21 suite and I think they're going to make
1:03:22 it very very easy for companies to train
1:03:25 on their own data I think Google will do
1:03:27 that as well
1:03:29 um so I think that's I think that's
1:03:30 going to be huge
1:03:31 um
1:03:32 how do you break into AI as a business
1:03:34 analyst start learning AI start using so
1:03:38 what I would do if if you're a business
1:03:40 analyst is I would start using chat GPT
1:03:43 to support you in your work
1:03:47 um have it help you do research
1:03:50 um
1:03:52 triple check the research
1:03:54 there was a lawyer that was just busted
1:03:56 he wrote a legal brief that cited
1:03:59 citations of of you know cases and the
1:04:03 judge busted him because he just did it
1:04:05 on chat GPT and it made up some cases
1:04:08 that didn't exist and the judge was like
1:04:10 uh you referred me to these cases and
1:04:13 these are all [ __ ] what's going on
1:04:16 um but but I would start using chat gbt
1:04:18 to support you in your business
1:04:21 um and and just start learning about it
1:04:22 and start talking about it because your
1:04:25 customers are going to start asking you
1:04:27 about it
1:04:28 medical coding I don't quite know what
1:04:30 the question is there but yeah
1:04:32 um any job that has a routine task I
1:04:34 agree with that
1:04:35 um it's already massively affected my
1:04:38 job as a software engineer yeah I'm
1:04:40 hearing what I'm hearing uh I heard this
1:04:42 tonight I was at a CEO Soiree and uh the
1:04:47 there were some investors talking on a
1:04:49 round table and they said that the their
1:04:51 portfolio companies are seeing 30 to 40
1:04:54 percent increases in efficiency uh
1:04:57 because of chat GPT
1:04:59 um set grad student here how can I have
1:05:01 chat GPT do statistical analysis how
1:05:05 would I share a spreadsheet okay so
1:05:07 that's Chad B so Chad B
1:05:10 go look for I'll I'll just go hands on
1:05:14 it but go look for
1:05:17 let's see we'll hop over here to the
1:05:20 YouTube
1:05:23 and you're going to want to go search
1:05:26 for code interpreter
1:05:33 chat GPT
1:05:35 and then there's going to be a whole
1:05:36 bunch of videos talking about code
1:05:39 interpreter
1:05:40 and what code interpreter is is mind
1:05:45 melting so
1:05:47 with code interpreter what you'll do is
1:05:49 you'll go to your chat GPT window like
1:05:51 you normally do and in your little box
1:05:53 where you type in your query there's
1:05:55 there's an icon where you can upload a
1:05:57 spreadsheet or a document
1:05:59 and when you upload that document
1:06:02 um chatgpt learns it
1:06:05 and I think you can put in like 100
1:06:07 megabyte which isn't huge from a data
1:06:09 perspective but pretty
1:06:11 beefy piece of data
1:06:14 and it just learns it and then you can
1:06:16 start asking it data analysis questions
1:06:18 like what's interesting about this data
1:06:20 what are the trends
1:06:22 um okay how do you compare this and that
1:06:24 then you can ask it things like could
1:06:26 you make me a map of the United States
1:06:28 where you change the height of the state
1:06:31 based on the Census Data or whatever
1:06:33 data you want
1:06:35 and they will then write the python code
1:06:38 to generate a map of the United States
1:06:39 with the states at different heights
1:06:41 generate the image and hand that back to
1:06:43 you like it's mind-blowing so go watch
1:06:45 videos on code interpreter uh and and
1:06:48 just be patient grasshopper it will come
1:06:51 when it comes pounce on that [ __ ] there
1:06:54 are probably some third-party companies
1:06:56 out there that are using things like
1:06:57 Lang chain to allow you to upload
1:07:00 spreadsheets and and play with them also
1:07:02 if you haven't played with there there
1:07:05 are now plugins for like Google Sheets
1:07:07 there's one called
1:07:09 GPT for Sheets there's probably a ton of
1:07:12 them now where basically every cell in
1:07:14 your Excel spreadsheet can be a call to
1:07:16 open AI so so you can have like put one
1:07:19 word in a spreadsheet and then the next
1:07:21 cell over says based on this word
1:07:24 generate 10 other words and then you
1:07:26 know the cells to the right of that can
1:07:27 can generate their own things based on
1:07:29 that so you can have a spreadsheet that
1:07:31 self-populates original content so if
1:07:33 you want to make synthetic data you
1:07:36 could set up a spreadsheet in probably
1:07:38 15 or 20 minutes that generates you know
1:07:41 massive amounts of synthetic data that's
1:07:43 exactly what you need
1:07:45 it's amazing everything that's coming is
1:07:48 so amazing if you know what I mean it's
1:07:49 amazing it's amazing the what's coming
1:07:53 on the Internet is the Mad thing
1:07:57 all right can you tell it's late here
1:07:59 and I'm a little Punchy and I had a
1:08:01 drink or two at the CEO Soiree all right
1:08:04 why is AI not working in my language I
1:08:07 would love to help
1:08:09 I would love to help to improve it how
1:08:12 can I do it okay I don't know what
1:08:14 language where where you are but that
1:08:17 thing
1:08:19 chat.openai.com understands all
1:08:21 languages it should work in your
1:08:23 language
1:08:25 um it's remarkable like it you can just
1:08:27 tell it to translate this thing into
1:08:29 that language and it does it I had a guy
1:08:31 I told I told the guy I said it's
1:08:34 probably good enough at languages that
1:08:35 it can invent a language any any those
1:08:38 were my tick tock on on my recorded
1:08:41 things any and he sent me back a comment
1:08:43 he said he said yeah I I uh had an
1:08:45 interaction with it were we talking
1:08:47 about Star Wars and I had to invent a
1:08:49 language for me so it should be able to
1:08:52 understand your language my guess is
1:08:54 that you're using
1:08:56 a non-official
1:08:58 um app that someone put together that's
1:09:00 just a piece of [ __ ] so
1:09:02 go there chat.openai.com and you should
1:09:05 be unless it's blocked in your country
1:09:07 you should be able to use it how do you
1:09:09 break into okay that was the business
1:09:11 analyst one uh uh
1:09:14 um
1:09:15 why is it called generative AI okay so
1:09:17 GPT those three words chat GPT where
1:09:21 does that come from generative
1:09:22 pre-trained Transformer
1:09:25 for Transformers
1:09:28 I'll get back to generate in a second
1:09:30 pre-trained means so historically with
1:09:32 with large language models or with with
1:09:35 machine learning you had to bring your
1:09:37 own data so there were these tools out
1:09:39 there and you could bring a massive data
1:09:41 set and then you could put it into these
1:09:43 large language models and then you could
1:09:44 search on it
1:09:46 pre-trained means it's pre-trained
1:09:48 meaning it's trained on what in the case
1:09:51 of gpt4 kind of all of the internet all
1:09:54 of the public internet
1:09:55 um Transformer what's that about
1:09:58 that's a piece of technology that Google
1:10:01 invented in 2017 that allowed these
1:10:04 large language models to be trained on
1:10:06 massive data sets so so basically before
1:10:09 you had to have small data sets the
1:10:12 Transformer allow you to basically stack
1:10:14 Transformers and and it was kind of like
1:10:16 the more data you threw at it the better
1:10:18 it got and that's what openai got really
1:10:20 good at once they once openai sort of
1:10:22 cracked the nut of the more we throw at
1:10:25 this thing the better it gets they said
1:10:27 let's throw as much as we can at it and
1:10:29 it just kept getting better and that's
1:10:30 kind of where why we are where we are so
1:10:33 so generative means it's literally
1:10:35 generating word by word
1:10:38 the response to your query
1:10:41 using a pre-trained data set that's
1:10:44 sitting on top of these this Transformer
1:10:45 architecture so that's what those three
1:10:47 things mean generative means it's
1:10:49 generating based on your input original
1:10:52 content or original images essentially
1:10:55 from nothing
1:10:57 um one word at a time or one pixel at a
1:10:59 time in the case of the images it's it's
1:11:01 pretty remarkable stuff the more you
1:11:02 learn about how it works it's just like
1:11:04 like your face just goes you're like
1:11:07 holy crap who invented this [ __ ]
1:11:11 back to school training entrepreneurial
1:11:13 Innovation Universal basic income
1:11:17 um kind of I'm weight loss you know what
1:11:19 I'm going to do I'm going all the way to
1:11:20 the bottom here and I'm just going to
1:11:22 grab a couple of the questions most
1:11:23 recent questions thank you for that look
1:11:25 at that cool little heart hat thank you
1:11:27 whoever sent me that all right how do I
1:11:30 use AI to expedite social sciences
1:11:33 research and article finding
1:11:36 um I would do something like krita
1:11:38 zapier
1:11:40 automation to
1:11:44 you know anytime
1:11:46 you know some something is tweeted or
1:11:48 posted on LinkedIn that's got certain
1:11:50 keywords in it I would sort of use
1:11:52 zapier to flag articles and then
1:11:56 um pull them in and
1:11:59 um
1:12:01 pull them into openai to summarize them
1:12:04 and then like I would use zapier to
1:12:07 create automations because I think you
1:12:09 could do lots of stuff with that you
1:12:11 could dump all the all the stuff it
1:12:13 finds into a spreadsheet and then you
1:12:15 could do analysis on the spreadsheet I I
1:12:16 think there's lots you could do there
1:12:18 GitHub copilot has statistical and
1:12:20 program assistance too oh oh yeah that's
1:12:22 that's that's true so there was a
1:12:24 question someone was asking about
1:12:26 statistical stuff uh depending how geeky
1:12:28 you are go over to GitHub because
1:12:30 co-pilots got some pretty powerful
1:12:34 um pretty powerful tools on that front
1:12:37 um what happens when the AI gets the
1:12:39 wrong info
1:12:40 um so this is an important things thing
1:12:42 to know it's called hallucination in the
1:12:45 Craft um they call it uh these models
1:12:48 hallucinate hang on a sec
1:12:51 um
1:12:54 I'm apparently not allowed to drink
1:12:55 liquid
1:12:57 on camera because that violates
1:12:59 Community guideline violations it's the
1:13:02 dumbest [ __ ] thing I've ever heard
1:13:04 but apparently I can call the guidelines
1:13:06 the dumbest [ __ ] thing I've ever
1:13:08 heard but not drink water on camera
1:13:10 whatever
1:13:15 okay so what happens when AI gets it
1:13:17 wrong it gets it wrong a lot
1:13:21 the way the models work they're
1:13:23 literally prediction engines of what
1:13:25 word is the most likely to come next in
1:13:27 a series so it's not actually conscious
1:13:30 of what it's writing it's literally
1:13:32 saying the next word should be this and
1:13:34 the next word should be that and the
1:13:35 next word should be that so it doesn't
1:13:37 have awareness of what it actually wrote
1:13:39 I think there will be systems in the
1:13:41 future that analyze what it wrote and
1:13:43 somehow keep track of that [ __ ]
1:13:46 right now it doesn't know
1:13:48 so if you say
1:13:51 give me your opinion on this court case
1:13:53 and cite case law that supports your
1:13:56 argument it will do that
1:13:59 sometimes it cites
1:14:01 case law and it'll put hyperlinks in
1:14:04 there too and sometimes it cites case
1:14:07 law and those hyperlinks just go to
1:14:08 nothing there's nothing there it just
1:14:11 makes [ __ ] up so what I would say is
1:14:13 this if you're doing creative writing it
1:14:15 doesn't matter if it hallucinates if
1:14:17 you're doing research um it matters a
1:14:19 lot if it hallucinates so if you're
1:14:21 doing anything where data Integrity or
1:14:23 accuracy data accuracy is important then
1:14:27 you need to double and triple check your
1:14:30 work because you don't want to be like
1:14:32 the lawyer that just got busted by the
1:14:33 judge that said did you write this on
1:14:35 chat GPT This legal opinion This legal
1:14:38 brief and he did and he got busted
1:14:44 yes Ben can use chat GPT plugins
1:14:47 there's an Innovative browser extension
1:14:49 called chat GPT for Bing chat
1:14:53 you don't need chat GPT for Bing I don't
1:14:56 think
1:14:57 because Bing is chat gbt so but whatever
1:15:01 I I am I am not a fan of extensions
1:15:04 right now because I think most of the
1:15:06 extensions are just kind of these
1:15:08 thin layers sitting on top of
1:15:11 um simple open AI calls most of them are
1:15:15 just going to become features within the
1:15:17 browsers or within operating systems
1:15:20 anyway so I'm just I'm just waiting I'm
1:15:22 not doing anything with extensions
1:15:24 what would you say is currently the best
1:15:26 usage case for AI there are so many like
1:15:29 it's it's almost like I mean that's why
1:15:31 it's so mind-blowing I mean
1:15:34 here's this will give you a sense one of
1:15:36 the things I have and show you this
1:15:40 chat
1:15:46 so so this is awesome chat GPT prompt so
1:15:49 the beginning of this document is
1:15:50 basically just teaching you how to do
1:15:53 prompting
1:15:54 but the bottom of this document is
1:15:58 a a seemingly never-ending list of
1:16:02 prompts of how you can have chat GPT Act
1:16:06 so act as a travel guide
1:16:09 act as a character from any movie book
1:16:12 or anything
1:16:14 act as an Advertiser act as a
1:16:16 Storyteller act as a football
1:16:17 commentator act as a stand-up comic a
1:16:20 movie critic a relationship coach a poet
1:16:22 a rapper a motivational speaker right
1:16:24 this just goes on and on and on every
1:16:28 one of these is a rabbit hole you can
1:16:30 run down with chat GPT
1:16:32 so the answer to the question of what's
1:16:34 currently the best usage like I don't I
1:16:37 don't have enough time in the day to
1:16:39 explore enough to know what what it's
1:16:42 best at
1:16:43 it's really [ __ ] good at diagnosing
1:16:47 diseases according to doctors but it's
1:16:49 really bad at citing the cases that sit
1:16:51 underneath its decisions it's really
1:16:54 good at law it's really good at code
1:16:57 it's really good at creative writing
1:17:00 um it's really good at at
1:17:02 um at things like reading complicated
1:17:05 documents and and you know giving you
1:17:09 interpretations of it act as a florist
1:17:12 act as a self-help book
1:17:16 um act as a statistician act as a prompt
1:17:19 engineer act as a password generator a
1:17:23 Morse code translator an instructor in a
1:17:25 school I think education if you go look
1:17:27 at what um
1:17:29 what Khan Academy has done with with
1:17:32 Khan Migo so conmigo is a GPT for chat
1:17:37 bot that so
1:17:39 um Khan Academy was given access to GPT
1:17:41 for last August August of 22 and they
1:17:45 took until I think it was early may this
1:17:47 came out they launched this thing called
1:17:49 conmigo which is this
1:17:52 um
1:17:52 it's it's like an instruction assistant
1:17:56 like a t it's like a teacher and it will
1:17:59 teach you [ __ ] and it and if you just
1:18:00 say give me the answer to this it won't
1:18:02 just give you answers it gives you like
1:18:04 this Socratic back and forth thing to
1:18:06 understand if you actually understand
1:18:08 the concepts it's brilliant so
1:18:11 I think that
1:18:13 these large language models are capable
1:18:15 of being good at almost anything in the
1:18:19 right hands so I don't have a better
1:18:21 answer for you there but it's good at a
1:18:23 lot
1:18:24 it's good at a lot a lot of things it's
1:18:27 so many things I don't know if you know
1:18:29 that's because a lot of things it's a
1:18:31 lot of things all right companies rely
1:18:34 on AI and it gets it wrong could end the
1:18:36 company absolutely
1:18:37 companies that use powerful tools
1:18:40 stupidly will be punished we will see
1:18:44 many headlines of idiotic things like
1:18:47 the the group in Samsung that
1:18:50 put a bunch of sensitive company data on
1:18:53 what's effectively an open
1:18:56 and open at the time chat GPT and you
1:19:00 know and and leaked their secrets
1:19:03 ah
1:19:06 [Music]
1:19:08 all right what other questions do I see
1:19:11 here oops
1:19:13 hold please
1:19:15 I got lost in my questions damn
1:19:26 cash flow Queen oh there's a couple of
1:19:28 people sharing this live thank you for
1:19:29 sharing the live I appreciate that
1:19:31 there's a bunch of you in here so if
1:19:33 you're new here the the reason I have
1:19:35 those things there
1:19:37 is that's the official chat GPT website
1:19:41 chat.openai.com they do have an official
1:19:44 app but there's a lot of Faker apps out
1:19:46 there so you don't need to pay for apps
1:19:48 you don't need to pay for chat GPT you
1:19:51 can you can pay for What's called chat
1:19:53 GPT plus that gives you access to
1:19:55 plugins web browsing and uh gpt4 uh
1:20:01 natively but you can also use GPT 3.5
1:20:03 which is still incredibly powerful and
1:20:05 very fast
1:20:07 all right if this Tech is so powerful
1:20:09 and potentially dangerous why would they
1:20:11 open source it well they didn't so
1:20:14 okay so a couple of things so here's
1:20:16 here's some of the scoop of what's going
1:20:18 on
1:20:21 um
1:20:22 so
1:20:23 Google invents this thing called the
1:20:26 transformer in 2017. I think
1:20:30 as they start exploring what this stuff
1:20:32 can do they realize that
1:20:36 a tool that just gives you answers is a
1:20:40 threat to Google's business model
1:20:43 because their business model does not
1:20:45 want to give you answers they want to
1:20:46 give you a list of 40 possible links and
1:20:49 have you go click on as many of them as
1:20:51 possible because every time you click
1:20:53 they make money so they kind of
1:20:56 mothballed this technology they kind of
1:20:58 you know just said Ah that's good we'll
1:21:00 just keep this back here but they
1:21:02 released this paper to the world this
1:21:03 thing on Transformers
1:21:05 and so the rest of the development
1:21:08 Community kind of adopted this and over
1:21:10 five or so years people were developing
1:21:12 things open AI was created as by Elon
1:21:17 Musk as a counter to Google all the
1:21:20 stuff Google was doing with AI
1:21:23 Elon Musk saw that as a threat so he
1:21:27 founded
1:21:28 um openai
1:21:30 as a non-profit
1:21:32 to be an open source
1:21:35 company
1:21:37 to counter the proprietary [ __ ] that
1:21:40 Google was doing so that that was kind
1:21:42 of how those two became rivals
1:21:45 and then
1:21:47 as Sam Altman realized the more compute
1:21:51 power we toss at these large language
1:21:54 models the better they get
1:21:56 we need way more compute power than we
1:21:59 have than money we could plausibly raise
1:22:01 as a non-profit so we're going to take
1:22:03 it for profit Elon Musk huffed out of
1:22:06 there
1:22:08 um didn't like that at all Sam Altman
1:22:10 said we're going to try to control this
1:22:12 so so open AI is not open source it's
1:22:15 not non-profit anymore it's a capped
1:22:17 for-profit
1:22:18 so that happened when chat GPT was
1:22:21 released
1:22:23 it went from zero to 100 million users
1:22:25 in six weeks it was the quickest
1:22:27 adoption of software in history I mean
1:22:31 the World Wide Web took six years to get
1:22:33 to 100 million users chat CPT took six
1:22:36 weeks
1:22:38 because it took six weeks everyone kind
1:22:41 of panicked Google said oh [ __ ] we got
1:22:43 to get our AI stuff out there meta said
1:22:45 oh [ __ ] we've got AI stuff too Amazon
1:22:48 said we've got a whole AI Division
1:22:50 coming like everyone's coming out of the
1:22:52 Woodworks Apple's announcing on Monday
1:22:54 I'm sure there's going to be some AI
1:22:55 stuff in there when meta announced that
1:22:58 they had these large language models
1:23:00 like chat GPT they they inadvertently
1:23:03 kind of released
1:23:05 um the core large language model out to
1:23:08 the developer Community kind of kind of
1:23:10 in a controlled but not controlled kind
1:23:12 of way and the developer Community went
1:23:14 thank you very much they took their core
1:23:16 models the Llama models and they just
1:23:18 sort of blew them out into the open
1:23:20 source world and so the developer
1:23:23 Community has been pouncing on these
1:23:25 open source uh or these these leaked
1:23:29 models fine-tuning them creating their
1:23:31 own and all sorts of stuff like that so
1:23:34 um
1:23:36 so it wasn't a mistake that they're open
1:23:40 sourced but it I think it was just in
1:23:43 the Panic of people of companies trying
1:23:47 to stay relevant in the game as openai
1:23:49 went from zero to like a million
1:23:51 overnight
1:23:53 um I think that's what caused some of
1:23:55 the open source stuff so I think that's
1:23:56 that's what's going on there I think
1:23:59 what stocks would be good to invest in
1:24:02 for the AI boom I don't know like listen
1:24:04 I'm you know I've got a degree in acting
1:24:05 I'm really shitty with money
1:24:07 um you know I think that what micro
1:24:10 Microsoft is acting like a startup like
1:24:12 like nothing I've seen since they
1:24:15 decided to adopt the internet
1:24:18 um they're kind they're kind of
1:24:19 gleefully
1:24:20 um kicking the teeth of of Google
1:24:23 um
1:24:24 I think open AI is really interesting
1:24:27 they're not public I think
1:24:29 um stability AI is very interesting
1:24:33 um I don't know I like Nvidia they just
1:24:35 did a trillion dollar market cap
1:24:37 um and
1:24:39 show no signs of slowing so you know but
1:24:42 it's a trillion dollar market cap how do
1:24:44 you spell zapier uh
1:24:47 z-a-p-i-e-r z-a-p-i-e-r
1:24:53 um
1:24:53 [Music]
1:24:55 they're trying rewarding the process now
1:24:58 which should reduce hallucinations quite
1:25:00 a bit
1:25:03 oh yeah they're they're basically
1:25:05 putting weightings on if if it gets it
1:25:07 right it
1:25:09 yeah I there's gonna be whole there's
1:25:12 gonna be whole multi-layer systems that
1:25:14 that deal with accuracy like anthropic a
1:25:17 big part of why they were started was to
1:25:19 to deal with some accuracy stuff so
1:25:22 mm-hmm
1:25:24 why say hallucinate versus lying it's
1:25:27 called marketing if you say that AI lies
1:25:30 no one would use it if you say it
1:25:31 hallucinates it's like that's kind of
1:25:33 trippy oh it's like it's on acid in the
1:25:35 sticks these
1:25:38 [Music]
1:25:42 all right
1:25:44 it hallucinates when writing code too
1:25:46 yeah it does
1:25:48 listen I I mean here's the deal with
1:25:50 this thing I my you know my experience
1:25:53 with using these large language models
1:25:54 in a work environment is that
1:25:57 it gets you to 80 percent almost
1:26:00 instantly and then the last 20 is up to
1:26:02 you and you can use chat GPD to help you
1:26:05 get from you know 80 to 100 but it's
1:26:08 really up to you to take the steaming
1:26:11 pile of [ __ ] that chat apt put in front
1:26:13 of you and turn it into something you
1:26:15 can sculpt it into something
1:26:17 um some use cases
1:26:19 it's you get to like 99 really quickly
1:26:23 some use cases you get to 50 and then
1:26:25 but I would say on average it kind of
1:26:27 gets you to 80 very very quickly guys
1:26:30 it's past your bedtime good night
1:26:33 it is past my bedtime all right it's 12
1:26:35 20. yeah I got to go to bed
1:26:37 um I started this late I need an AI wrap
1:26:40 male voice using text input which app is
1:26:43 free for this purpose
1:26:45 there's no free ones well I'm sure
1:26:47 there's some open source ones if you go
1:26:49 to GitHub there's probably [ __ ] that you
1:26:51 could you could install on Google collab
1:26:53 or whatever
1:26:55 um I use
1:26:57 um 11 labs and did
1:27:01 um I think did's got some free like you
1:27:05 can do some amount of free stuff with
1:27:06 did so it's d-id.com or eleven labs
1:27:11 e-l-e-v-e-n-l-a-b-s
1:27:15 um for some of the voice synthesis stuff
1:27:18 let me see the title I don't know what
1:27:20 title you're talking about if you tell
1:27:22 me what title I will why not include an
1:27:25 act as oh dang it dang it tick tock
1:27:30 yeah let's see if I can find that
1:27:32 question again I was in the middle of a
1:27:34 question
1:27:35 and I was scrolling
1:27:37 please let me see the title that's where
1:27:39 it was okay
1:27:39 why include the act as prompt it gives
1:27:43 similar results without so
1:27:46 there's something about the way that
1:27:50 um
1:27:51 hang on there's something about the way
1:27:54 that
1:27:56 um
1:27:59 these large language models work that
1:28:02 the act as thing
1:28:05 makes a difference
1:28:08 creative mornings here we go I've got
1:28:11 I'm downloading damn it
1:28:13 my my hard drive on my laptop is full so
1:28:17 everything's in the cloud so anytime I
1:28:19 want to open something it's not there
1:28:21 do I have that do I have it in there
1:28:31 yeah well it's not there anyway so so
1:28:34 the reason they act as kind of works is
1:28:38 there's this thing called the latent
1:28:41 space or vector space the latent space
1:28:43 is
1:28:47 when when these large language models go
1:28:50 out and they learn all of the words
1:28:51 they're not just learning words as
1:28:54 discrete entities they're learning words
1:28:56 in semantic forests right so if you're
1:29:00 learning about
1:29:02 you know I don't know dogs and and
1:29:04 you're talking about pack animals you
1:29:07 know that's one thing if you learn about
1:29:10 dogs as you know men who philander
1:29:13 that's another kind of semantic ocean So
1:29:16 when you say act as
1:29:18 um
1:29:19 you know a Freudian therapist it it kind
1:29:23 of points the large language model into
1:29:25 a semantic direction that all of your
1:29:29 prompts from that point forward sort of
1:29:31 live in that semantic area region so I
1:29:33 think that's why it works better
1:29:35 um but you you don't have to to tell it
1:29:38 to act as a certain thing it can improve
1:29:40 things
1:29:41 um I don't always do it I think one of
1:29:43 the things that that works pretty
1:29:45 universally well is when it gives you an
1:29:47 answer you have it you say review review
1:29:50 your answer and tell me how you would
1:29:52 improve it
1:29:54 um and then have it improve it that
1:29:56 apparently works incredibly well
1:29:59 is there automation data an information
1:30:01 database that AI language models base
1:30:04 their answers on or do they actively
1:30:07 look for info so
1:30:12 so the core model
1:30:14 they stopped they basically stopped
1:30:18 trading the core model in September of
1:30:19 2021 for open AI
1:30:22 um Bard might be different but I think
1:30:25 it's it's around the same time so they
1:30:27 basically were ingesting all of this
1:30:29 data and tagging all of this data and
1:30:32 whatever they do to build the base large
1:30:33 language model this is very expensive
1:30:35 time-consuming thing
1:30:38 so that stopped in September of 2021 but
1:30:41 they've been fine-tuning it and now that
1:30:42 it's connected to the internet
1:30:45 um so so when I go this this option that
1:30:49 says browse with Bing so that's now
1:30:51 connected to the internet so so I'm now
1:30:54 querying this massive you know data Lake
1:30:58 that's that includes up to September
1:31:01 2021 but I'm now augmenting that with
1:31:04 with um
1:31:06 with current data so
1:31:09 um who won the nuggets
1:31:12 game tonight question mark
1:31:18 so that says browsing the web and now
1:31:21 it's searching Bing Denver Nuggets game
1:31:23 result June 2nd 2023
1:31:25 and the Denver Nuggets won their game
1:31:27 104.93 and it's got a citation and that
1:31:30 citation takes me too probably some
1:31:32 Denver
1:31:33 paper odds whatever
1:31:36 so that's how that works hope that helps
1:31:40 hope that helps all right you're a
1:31:44 fascinating person I'm staying tuned and
1:31:46 look forward to learning more from you
1:31:47 thank you very much I appreciate that
1:31:49 [Music]
1:31:51 thank you for explaining absolutely open
1:31:53 AI was created by a bunch of people to
1:31:56 be honest not just Elon Musk I know and
1:31:59 he takes he takes all the credit for it
1:32:00 yeah it was a group of whatever seven or
1:32:03 eight people but the point was was that
1:32:05 if you hear him tell the story he was
1:32:07 like you know Larry Page is really
1:32:09 pissed off at me because I started open
1:32:11 AI to you know because he was doing it
1:32:14 wrong
1:32:15 um but yeah there were a lot of people
1:32:17 that that got that thing off the ground
1:32:19 so yes I'm sorry if I just gave all the
1:32:22 credit to Elon he's got plenty of plenty
1:32:24 of credit going on right now so I'm
1:32:27 sorry I missed where you were sharing
1:32:28 about how to train chat GPT
1:32:31 um yeah I'll talk about that again I I I
1:32:33 I go over stuff
1:32:35 a fair amount on this channel so just
1:32:38 hang out on my lives I've been trying to
1:32:40 go live daily except for when I get
1:32:43 um banned and put in detention I fear AI
1:32:46 I believe it will be used for pure evil
1:32:48 and was will destroy Humanity one day
1:32:50 you've seen too many movies yeah I think
1:32:52 Hollywood has absolutely trained us that
1:32:54 uh AI is going to destroy us
1:32:57 um
1:32:58 I personally think that's kind of a
1:33:00 boring future
1:33:02 um if that happens that happens
1:33:04 I'm of the belief that we're we're early
1:33:07 enough right now and because we have
1:33:09 access to these tools that anything
1:33:11 positive individuals do right now with
1:33:14 AI is going to have a disproportionately
1:33:15 positive effect on the future
1:33:19 um so yes Bad actors can play with this
1:33:21 not play with Bad actors can leverage
1:33:23 these tools to do really bad things but
1:33:26 good actors can as well to counter
1:33:27 balance that
1:33:30 um so go do some good [ __ ] with it so
1:33:32 rather than just sitting on the sideline
1:33:34 start learning it and start doing good
1:33:35 things with it would be my suggestion
1:33:38 I'm sure uh
1:33:41 I'm sorry sir you shared something I
1:33:43 really wanted to take note okay sorry
1:33:46 ask me again I'll see if I can say it
1:33:49 again I don't know what I said maybe an
1:33:50 AI fund a blended AI fund zapier yes I
1:33:53 use zapier to automate
1:33:55 all wait onboarding all my new customers
1:33:59 and automating estimates yeah zapier is
1:34:01 really powerful I mean you have to
1:34:03 really
1:34:03 the thing about zapier is it's kind of
1:34:05 like chat CPT it can do so much that
1:34:07 it's hard to get your head around like
1:34:08 where do I start
1:34:10 so I've just been using it to create
1:34:11 these content generation engines which
1:34:13 it's a cool they're powerful
1:34:17 thoughts on Air Force Air Force drone
1:34:20 incident I'm glad it was a simulation
1:34:23 um if I asked chat GPT who I am it makes
1:34:25 things up a lot yes that it does it it
1:34:28 often gives me really powerful
1:34:30 credentials I kind of like what it makes
1:34:32 up about me I think I might keep some of
1:34:34 them
1:34:35 sounds pretty accurate yeah it gets you
1:34:37 to 80 pretty quickly I've been using
1:34:38 chat CPT to write copy for classic cars
1:34:40 and they've been selling like crazy yeah
1:34:43 because because that's the thing like
1:34:45 you know the large language models have
1:34:47 been trained on all the internet all the
1:34:49 internet all the public internet and if
1:34:52 you think about
1:34:53 all of the car Geeks that have been
1:34:55 reading writing about every obscure
1:34:58 you know detail of every classic car for
1:35:02 the past 30 or 40 years all of that's in
1:35:05 there so if you say you know talk about
1:35:08 you know a 69 GTO Judge in you know in
1:35:13 the classic orange color that's my
1:35:15 favorite muscle car from the 70 or from
1:35:17 the 60s
1:35:19 um it'll know a lot of details about it
1:35:21 you can have it really geek out I think
1:35:23 that's brilliant it's a brilliant use of
1:35:24 it and it's fun like that's a that's a
1:35:26 fun one and it's it's you know I can
1:35:28 imagine it's incredibly good at that
1:35:31 do you uh let's see
1:35:34 do you have the most wait do you have
1:35:36 the most amazing camera lens I've ever
1:35:38 seen no really it's just an iPhone 11.
1:35:42 yeah it's an 11 it's an 11. it's an old
1:35:44 iPhone with a cracked with a cracked uh
1:35:47 front front
1:35:50 um when will quantum computers combine
1:35:52 with chat gbt we're all dead most likely
1:35:55 I don't think we're all dead we're all
1:35:56 dead or maybe it's curious cancer
1:35:59 when quantum computers when when they
1:36:01 can start training these things on
1:36:03 quantum computers I think you'll get
1:36:04 real-time training so the large language
1:36:06 models won't be cut off at a certain
1:36:08 date you'll you'll just get things that
1:36:10 are trained in real time and I think
1:36:12 you'll start we'll start solving the
1:36:14 world's problems Sam Altman talked about
1:36:16 he thinks In Our Lifetime there's going
1:36:17 to be
1:36:18 a single year where we have as much
1:36:21 scientific advancement as we've had in
1:36:23 the last 500 years so go to chat GPT and
1:36:27 have it give you a list of the top 20
1:36:29 scientific advancements in the past 500
1:36:32 years just look at that list or past top
1:36:34 50 whatever some long list
1:36:37 look at that list and imagine a single
1:36:40 year where we have that level of
1:36:42 scientific advancement compressed into a
1:36:44 single year like that's the kind of
1:36:46 thing where when you when you start
1:36:47 saying quantum computers get overlaid
1:36:49 with with
1:36:51 um with AI that's possible now could it
1:36:54 just go ahead and [ __ ] exterminate us
1:36:56 all sure it could I don't like that
1:36:59 future so I'm not giving that any power
1:37:02 I give you no power
1:37:03 negative future I give you no power
1:37:07 uh concession stop gerrymandering and
1:37:10 draw Fair boundaries well it's not chat
1:37:12 well AI is probably generating them but
1:37:16 who's asking the AI to generate them
1:37:20 it ain't at the a the AI ain't prompting
1:37:23 itself that's human beings being
1:37:26 [ __ ]
1:37:27 all right AI is going to make us build
1:37:29 stuff faster faster better I yes I think
1:37:33 we're entering a great Renaissance
1:37:35 we are at the beginning of a great
1:37:36 Renaissance we're going to look back at
1:37:38 these times and go holy crap remember
1:37:39 when chat GPT just did text oh that was
1:37:42 so stupid you could only put 3 000 words
1:37:45 in it wow it didn't make movies back
1:37:48 then no we had to watch movies on this
1:37:51 thing called streaming services what's
1:37:53 that oh it doesn't matter
1:37:55 wait you only had one movie for all the
1:37:58 people
1:37:59 yeah Gone With the Wind was just like
1:38:01 Gone With the Wind and it just had that
1:38:03 one actor in it yeah it was just Clark
1:38:06 Gable
1:38:07 you couldn't put Superman in it yeah
1:38:10 that's what it's gonna be like hey I
1:38:12 made up legal cases and citations yes it
1:38:15 did and the lawyer didn't check it and
1:38:17 he's an idiot
1:38:20 I like the Will Smith world the first
1:38:22 half yeah I know the first half of all
1:38:24 the AI movies is great it's the second
1:38:26 half that we want to avoid the best
1:38:28 camera image ever well thank you I don't
1:38:30 know what's going on with my camera
1:38:31 image but yeah maybe it's just a when
1:38:35 you got this going on maybe it's not the
1:38:38 camera you know what I mean you know
1:38:39 what I mean I just felt like I sounded
1:38:41 like Ted lasso there for a minute did
1:38:43 you see the last episode of Ted lasso is
1:38:45 bawling my eyes out God damn that was
1:38:47 good
1:38:48 all right I used AI to help with my
1:38:50 non-profit kitten rescue cool save the
1:38:53 kittens you say so what is AI good for
1:38:55 saving kittens see you heard it here
1:38:57 first
1:38:59 that's awesome
1:39:00 I like that
1:39:01 all right I lost my place again I'm
1:39:03 going back I'm going back I'm doing it
1:39:06 uh-oh all right
1:39:09 did you see the yes I saw the AI
1:39:13 drone that went rogue but it was a
1:39:15 simulation so we're all good
1:39:19 I went back too far I've lost my place
1:39:22 okay I just got to the kitten rescue
1:39:24 thing okay
1:39:25 I think it's about I think is about how
1:39:30 to train AI prompts you know folks put
1:39:33 complete sentences in and that won't
1:39:36 sound like such an Egypt
1:39:39 um
1:39:40 all right can I use AI to write my
1:39:42 resume and make me look like a rock star
1:39:44 yes you can
1:39:45 um just so go get a resume format that
1:39:49 you like or have it recommend a resume
1:39:51 format that you like
1:39:53 pop that into chat EBT then go copy and
1:39:56 paste all your LinkedIn stuff paste that
1:39:58 into chat EBT and say write me a resume
1:40:00 that follows the format of the resume up
1:40:03 top and populates it with the details of
1:40:06 what's from my LinkedIn profile
1:40:08 for a job in whatever it might be
1:40:12 what's the Air Force incident so so the
1:40:15 Air Force
1:40:16 had a simulation going of an armed drone
1:40:20 that was set to take out enemy positions
1:40:23 and the the the person sending
1:40:27 instructions to the Drone basically
1:40:29 called it back and said you can't go
1:40:31 kill that thing and it didn't like that
1:40:33 so it turned around and attacked the
1:40:36 operator now it was a simulation so so
1:40:39 no shots were fired but uh yeah assuming
1:40:42 they're a missile attached to that their
1:40:44 drone
1:40:45 um that wouldn't have been good so that
1:40:47 was the Air Force incident so
1:40:49 um good that they're testing good that
1:40:51 they're doing it in a simulated way and
1:40:54 they got some Kinks to work out you know
1:40:56 Elon Musk said we'd have full driving
1:40:58 cars by what four years ago and it's
1:41:01 still not quite there but it's close
1:41:03 well
1:41:04 imagine flying things with bombs on them
1:41:08 they might take a while to get it right
1:41:11 uh why did you say 69 GTO Judge that's
1:41:14 creepy I just sold one oh dude I I don't
1:41:16 know there's something about a 69 GTO
1:41:18 Judge that's just the most beautiful car
1:41:20 car in the world
1:41:24 [Music]
1:41:26 uh
1:41:28 do you have a website for more info wait
1:41:30 prompts for code Generation
1:41:33 Um I don't but that's the the
1:41:36 SE prompts right here so prompts.chat
1:41:40 that'll take you to the prompting Thing
1:41:41 Future PD is a bunch of tools and then
1:41:44 chat.openai.com is uh the official chat
1:41:47 GPT website
1:41:48 um if you're asking about information
1:41:50 about me I've got a link Tree in my bio
1:41:52 that that's got a bunch of my [ __ ] there
1:41:54 like my company and I think I put a link
1:41:56 to my email and things like that so
1:42:00 um
1:42:01 where can you get your own model to run
1:42:03 locally on a server cluster just go
1:42:06 search for uh just go to Twitter and
1:42:09 search for open source llm
1:42:13 um and and go to go look up trending
1:42:16 projects on GitHub and hugging face
1:42:18 hugging face is a place to get all those
1:42:21 things there are so many of them out
1:42:23 there that
1:42:24 you know probably the place to go is
1:42:26 Reddit and and and just find the large
1:42:29 line I'm sure there's an open source
1:42:31 large language model subreddit
1:42:33 um and just go start hanging out with
1:42:36 those Geeks and you'll find some some
1:42:38 amazing stuff I'm a fantastic teacher
1:42:41 thank you very much I appreciate the
1:42:42 kind words
1:42:43 and I gotta go to bed I gotta go to bed
1:42:45 people what what are we doing here what
1:42:48 are we doing why are you watching me
1:42:50 stop
1:42:52 stop it's not that interesting
1:42:55 thing at all
1:42:57 who are your people
1:42:59 um let's give AI the goal of making the
1:43:03 Star Trek Society yeah I think it could
1:43:05 I think we're gonna
1:43:06 I think we're gonna see some really
1:43:09 really amazing things I think there's
1:43:11 going to be some headline making
1:43:12 catastrophes
1:43:15 um but I think we're also going to see
1:43:16 some remarkable remarkable stuff
1:43:19 coming out of this don't forget we're so
1:43:22 early with this and and you know
1:43:25 ai's been around for a long long time
1:43:27 so the foundation of all this is has
1:43:30 been around for a while but what's new
1:43:32 is
1:43:33 what what started on November 30th last
1:43:36 year 2022 is when Chachi BT came out it
1:43:40 gave regular people access to these
1:43:43 tools that were largely the the you know
1:43:47 the territory of data scientists and
1:43:49 researchers and and basically Sam Altman
1:43:53 said here's a gift
1:43:54 do you know that they wrote chat GPT in
1:43:57 two weeks
1:43:59 do you know that in the documentation
1:44:02 for gpt3 which came out like a year and
1:44:04 a quarter before chat GPT
1:44:08 the instructions for how to build chat
1:44:10 GPT were in the documentation for gpt3
1:44:14 for like a year
1:44:15 and Sam Altman said we assumed someone
1:44:18 would build it and no one built it
1:44:21 and then last November uh I think it was
1:44:24 because Microsoft was going to announce
1:44:26 their billion dollar investment Sam
1:44:28 Altman said hey uh engineering team can
1:44:31 you whip up that chat GPT thing they
1:44:33 built it in two weeks
1:44:35 and six weeks later it hit 100 million
1:44:38 users I mean
1:44:39 that's remarkable
1:44:41 fastest adoption of software in history
1:44:44 all right I want to be a part of this
1:44:47 Renaissance what's a good field or thing
1:44:50 to study with AI
1:44:55 ubiquitous so so I would sort of push
1:44:57 that back on you and say what is the
1:44:59 thing you're most interested in and then
1:45:01 I would dig deep on
1:45:03 How can I do what I'm interested in
1:45:06 better
1:45:07 faster more powerfully with AI so if it
1:45:11 involves images start learning
1:45:12 mid-journey or stable diffusion or
1:45:16 um the new Adobe generative fill if if
1:45:19 you look at what's happening with
1:45:21 General Phil within Photoshop it's
1:45:23 remarkable
1:45:25 um
1:45:26 if it's around words or business or
1:45:28 analysis or programming or things like
1:45:31 that then dig into chat EBT but I like I
1:45:33 I just think it can't hurt anyone right
1:45:36 now to go to that address like why I
1:45:38 have it on screen that's the official
1:45:39 chat EBT website just go there start
1:45:42 using chat GPT to solve problems in your
1:45:45 life
1:45:46 um
1:45:47 I would say lean into what you're
1:45:50 passionate about what you're good at
1:45:53 um what you have a point of view on and
1:45:56 just figure out which tools can support
1:45:58 you in doing that
1:45:59 because again it's so [ __ ] capable
1:46:02 that it can do anything if I ask chat
1:46:04 CPT to write a book about vampires and
1:46:06 you ask the same question the same way
1:46:08 will we get the same story we will not
1:46:11 um in fact you can do that yourself
1:46:15 um
1:46:15 create a new chat you know put in a
1:46:18 decently sophisticated
1:46:20 um prompt about your vampire book have
1:46:22 it writes a chapter or whatever or an
1:46:24 outline and then start a new chat and
1:46:27 use the exact same prompt and generate
1:46:29 it and you'll you'll see it'll give you
1:46:31 it'll give give you different things
1:46:34 um
1:46:36 and that's largely because they've got
1:46:37 there's a thing called temperature if
1:46:39 you if you use the the back end of these
1:46:42 GPT tools I think chat GPT is set for
1:46:45 0.4 it goes from zero to one
1:46:48 zero is the most factual answer so if
1:46:51 it's zero you you'll likely get things
1:46:53 that are all the same but if you dial
1:46:55 that up to one you'll get things that
1:46:56 are much more creative so I think
1:46:59 default chat GPT is somewhere near the
1:47:01 middle
1:47:02 everything using chat GPT apis are going
1:47:05 to be paid Services right because
1:47:06 developers using apis need to pay for
1:47:09 its subscriptions yeah but they're
1:47:10 relatively cheap I mean unless you're
1:47:13 doing unless you're doing training and
1:47:14 embedding which can be expensive
1:47:18 um
1:47:18 just the calls to open AI like I set a
1:47:22 Max monthly spend of 20 bucks I don't
1:47:24 think I've ever exceeded 10 and I like I
1:47:27 do a decent amount with this I don't do
1:47:29 a ton 65 Riviera is more beautiful I
1:47:33 don't know
1:47:35 I don't know 69 judge the the Riviera's
1:47:39 it's a bit more of a boat it's I don't
1:47:41 know it's a little rounder I see where
1:47:43 you're going it's it's maybe more unique
1:47:46 I like the I like the muscly aspect of
1:47:48 the GTO
1:47:49 which job sector will it be affected
1:47:51 negatively from a AI from Full stack
1:47:54 cloud and devops
1:47:57 I don't know like all of them
1:48:00 I don't know if they'll be negatively
1:48:02 impacted they're just going to be
1:48:04 dramatically impacted
1:48:07 the negative part is
1:48:10 some jobs are going to go away some jobs
1:48:12 are going to get very very different and
1:48:14 people might not like them there's a
1:48:16 really sad Reddit post about a guy that
1:48:17 did 3D modeling
1:48:20 and he said I hate I'm quitting my job
1:48:22 he was lamenting his job because he's no
1:48:24 longer a 3D modeler he now just writes
1:48:27 prompts and then curates images that
1:48:29 this thing generates and he and he hates
1:48:31 his life and he's he's leaving the
1:48:33 profession so I think you'll see some of
1:48:34 that but I think I think every single
1:48:37 sector is going to be impacted by this
1:48:39 especially for things like devops and a
1:48:41 lot of that stuff the coding and the all
1:48:43 the tech stack
1:48:45 um just every all that complexity is
1:48:48 just gonna get subsumed by the
1:48:50 Simplicity of the large language model
1:48:52 sitting on top of it
1:48:54 so so this is really interesting so
1:48:57 right now you've got this incredibly
1:48:58 complicated technical world right with
1:49:01 cloud services and you know devops and
1:49:04 you've got all these machines and
1:49:06 platforms and services
1:49:08 and then you have entire companies that
1:49:11 build software to manage and glue all
1:49:13 this [ __ ] together right and you have
1:49:15 entire Tech divisions to do that so some
1:49:19 combination of services and Technologies
1:49:21 and human beings
1:49:23 make sense of this madness down here
1:49:27 well what large language models are
1:49:29 going to do is they're going to put a
1:49:30 layer up here that just does all this
1:49:32 [ __ ] automatically
1:49:34 and so companies that serve that that
1:49:38 thrive on the complexity of the ship
1:49:41 below it
1:49:42 because they figured out how to simplify
1:49:44 this stuff they're going to get
1:49:46 displaced and people who spend their
1:49:48 days doing menial
1:49:51 I don't know data analysis
1:49:53 they're going to get displaced and so
1:49:56 that that whole middle layer sitting on
1:49:58 top of all that complex stuff down
1:49:59 beneath I think that's where you're
1:50:01 gonna see the biggest
1:50:03 the biggest thing the algorithm has me I
1:50:06 don't know what that means
1:50:09 invest in stocks and cryptos and invest
1:50:12 in metaverse as well I would
1:50:16 not recommend investing in crypto or
1:50:18 metaverse right now let AI take take off
1:50:21 so I spent about a year deep diving in
1:50:23 blockchain and nfts and what was
1:50:26 interesting was the deeper I dug into
1:50:27 that the less I believed in it and with
1:50:30 AI it's the complete opposite where the
1:50:31 deeper I dig dig the more my mind is
1:50:34 blown I think AI is going to drive
1:50:36 adoption of blockchain but we're going
1:50:39 to come up with a use case first and
1:50:41 then the technology as opposed to right
1:50:42 now where crypto and and blockchain and
1:50:46 nfts are kind of leading with technology
1:50:48 and and so many of those companies are
1:50:50 just money grips they're just it's just
1:50:53 bad
1:50:55 um is that GPT how to build guides still
1:50:57 available
1:50:58 GPT how to build the prompts.chat yeah
1:51:05 prompts.chat is I think that's the thing
1:51:08 you're talking about
1:51:11 you should be if be an affiliate for
1:51:14 them I know I talk about them all the
1:51:16 time they should give me they should
1:51:17 give me access to code interpreter the
1:51:19 [ __ ]
1:51:22 0.7 over here it was preset to 0.7 oh
1:51:25 chat GPD was preset to 0.7 GT all day
1:51:28 let's go GTO people truck drivers are
1:51:31 the worst well the Riviera is not a
1:51:33 truck it's a boat boat drivers are the
1:51:35 worst Maybe
1:51:37 um platform is preset to 0.7 yeah so
1:51:39 it's decently creative I thought it was
1:51:41 lower than that for some reason
1:51:45 how do professionals in fields using AI
1:51:48 set themselves apart from the folks
1:51:51 who can now fake the skills that's an
1:51:54 interesting question
1:51:57 um
1:51:59 the part that's still important is the
1:52:02 critical thinking and the strategy
1:52:06 so
1:52:09 how you decide what what you want these
1:52:11 tools to do and then how you're going to
1:52:13 use it and how you deploy it that's
1:52:15 still humans
1:52:17 and so if you have someone that's just
1:52:20 using chatgpt to generate documents that
1:52:24 look like that
1:52:25 that's just a different thing than
1:52:27 someone who can look at a bunch of
1:52:29 output from chat GPT and go that I like
1:52:32 that I don't like based on their 30
1:52:34 years of experience so I think how you
1:52:36 combat
1:52:38 charlatans is how you always because
1:52:39 here's what's going to happen
1:52:41 everyone has access to these tools so
1:52:46 the perceived
1:52:50 level of idiocy in the workplace is kind
1:52:53 of is gonna kind of rise like this where
1:52:55 oh you know maybe Janie isn't such an
1:52:59 idiot because her her writing quality
1:53:01 went up because she's using a tool but
1:53:03 then what you're going to see is people
1:53:05 with actual experience and brains and
1:53:08 strategy are going to learn how to use
1:53:09 these tools just at a diff a slightly
1:53:11 different level so you're going to start
1:53:13 to see some work do this so all work's
1:53:16 gonna go like this and then the good
1:53:18 people are going to rise to the top as
1:53:19 they always do
1:53:21 because these are just tools in the end
1:53:23 and the people that are smartest at
1:53:25 using and deploying the tools are going
1:53:27 to be the ones that rise above the noise
1:53:29 and so
1:53:31 be smart one and start learning AI start
1:53:34 getting good with this stuff so that
1:53:36 you're one of those ones that rises
1:53:37 above the charlatans because there will
1:53:39 be lots of them good thing about chat
1:53:42 CPT is it will force a human to see that
1:53:44 their mind
1:53:45 is not what are
1:53:48 what they are
1:53:50 is that their mind is not what they are
1:53:53 um
1:53:54 yeah I I think that one of the Silver
1:53:56 Linings of this the disruption that's
1:53:59 coming is
1:54:02 is that
1:54:05 I think it's going to allow people to be
1:54:07 a bit more free to say what do I really
1:54:09 want
1:54:12 because like if what I really wanted was
1:54:14 to make beautiful images
1:54:17 I've got photography and I'm good at
1:54:19 that but like when I see a good
1:54:21 illustrator it like makes my heart ache
1:54:23 because I want to be able to do that but
1:54:25 I suck at drawing I suck at it
1:54:29 um and then
1:54:31 mid-journey and stable diffusion come
1:54:33 along and I can take the idea that I
1:54:35 have in my head of something I want to
1:54:37 visualize and now I can just visualize
1:54:39 it and so I think you're going to see a
1:54:41 lot of people discovering creative sides
1:54:44 of themselves or creative expression of
1:54:46 who they are
1:54:48 um that they might not have otherwise
1:54:49 ever gotten to I think it's going to be
1:54:52 beautiful I I think people I think the
1:54:54 amount of reinvention
1:54:56 human reinvention that's coming is just
1:54:59 going to be amazing
1:55:03 preach go chat GPT I'll tell you I tell
1:55:08 you it's gonna save you the AI is Gonna
1:55:10 Save You It's Gonna Save You Brethren
1:55:12 It's Gonna Save You
1:55:15 all right I can do a few more questions
1:55:17 and then I gotta go it's one in the
1:55:19 morning here what are you people doing
1:55:20 to me what are you doing with your lives
1:55:23 I hope you're drinking
1:55:25 all right I use chat tvt to mimic
1:55:29 a human AI Persona with the experience
1:55:33 and skills I use it as a CFO yeah
1:55:35 there's a cool project there's a guy on
1:55:37 uh Tick Tock called McKenzie Bose who's
1:55:40 got a project called GPT boss
1:55:43 and he's it's basically a wrapper app
1:55:45 that sits on top of uh gpt4 but it's all
1:55:49 these different characters and
1:55:50 generators and so there's like a CFO and
1:55:53 there's a coder an elite engineer and a
1:55:56 copywriter and a marketing expert and an
1:55:58 SEO expert and so he's built all these
1:56:00 sort of pre-prompted tools that you can
1:56:02 use to run your business it's pretty
1:56:04 it's pretty slick
1:56:06 all right what happens when the tools
1:56:08 become so easy anyone can use them
1:56:10 that's what we're in the middle of
1:56:11 finding out I mean
1:56:16 yeah like
1:56:19 writing uh a uh like like you can do
1:56:24 anything right now you can do anything
1:56:25 so so
1:56:27 the tools are easy enough that anyone
1:56:30 can use them right now but you have to
1:56:31 have a little bit of skill to sort of
1:56:33 piece part the different things together
1:56:36 but
1:56:37 within two or three years there's going
1:56:39 to be slick interfaces on top of all
1:56:41 this stuff it's going to be [ __ ]
1:56:42 miraculous so what what that means is
1:56:45 the people that start learning it now
1:56:47 are going to be way better at this [ __ ]
1:56:49 than when it gets easy to use two years
1:56:51 from now so all the opportunity is
1:56:54 happening right now let's call this the
1:56:56 summer of opportunity get curious about
1:56:58 ai go to that website
1:57:01 chat.openai.com that's the official chat
1:57:03 GPT website
1:57:05 go to prompts.chat and just start
1:57:07 playing with the prompts there and just
1:57:08 learn about prompting
1:57:11 start learning it's just a tool I'm a
1:57:14 senior Oracle DBA and I use it to help
1:57:15 you write documentation yeah it's really
1:57:17 good at documentation and and parsing
1:57:19 things I got chat cheap
1:57:23 knowledge up to year 2021
1:57:26 guess who sings the song anti-hero
1:57:29 accurately
1:57:31 well you can I mean
1:57:34 you might be able to get it to I I mean
1:57:36 it can't guess anything unless it's been
1:57:38 fine-tuned to things
1:57:40 after 2021 it can't know that so maybe
1:57:44 you had a trained model but gpt4 is
1:57:46 connected to to the internet now via
1:57:49 Bing so
1:57:50 that is all water under the bridge the
1:57:53 it only goes to September 2021 thing
1:57:55 that's over
1:57:57 that's over because it's all connected
1:57:59 to the internet now with the Bing the
1:58:02 Bing from the Microsoft Microsoft
1:58:04 company out of Redmond Washington they
1:58:07 made software they made shitty software
1:58:09 for years and then they discovered this
1:58:12 AI stuff and all of a sudden they're
1:58:13 like [ __ ] Apple what the hell what
1:58:16 the hell is going on in the world I used
1:58:18 to hate Microsoft for years I'm an apple
1:58:20 fanboy
1:58:23 can I just say this Apple's got their
1:58:26 worldwide developer conference on Monday
1:58:28 if they do not blow it out of the
1:58:30 [ __ ] water and make themselves
1:58:32 relevant again what is going on in the
1:58:35 world
1:58:38 uh that was an interesting comment I
1:58:40 think it will make real artwork even
1:58:41 more valuable
1:58:43 absolutely agree with that
1:58:45 I think that authenticity things like
1:58:47 in-person music
1:58:49 real art
1:58:51 um
1:58:51 you know video of actual human being
1:58:54 sharing their stories authentically I
1:58:56 think the value of authenticity is going
1:58:59 to go up dramatically so so we're going
1:59:02 to see the sea of synthesized content
1:59:04 that just sort of boils up
1:59:06 and I think there's going to be a hunger
1:59:08 and and I the the 2024 election in the
1:59:12 U.S there's going to be so much
1:59:13 disinformation and misinformation that's
1:59:16 going to be AI Amplified it's going to
1:59:18 be insane
1:59:20 I think the backlash of that so think
1:59:23 you know 2025 will be an incredible
1:59:27 demand for authenticity and I think the
1:59:31 value of authenticity is going to go up
1:59:33 I think you're going to be able to build
1:59:34 businesses around it I think authentic
1:59:37 Arts authentic experiences I think value
1:59:41 goes up dramatically so I totally agree
1:59:43 with that Holly
1:59:45 what's the best way to use this little
1:59:48 wait what's the best way to use this
1:59:51 with little technology I like again I
1:59:54 would just go to
1:59:56 um just start learning chat gbt like you
1:59:59 use it as as an Arts major as a
2:00:02 Humanities major
2:00:03 um have it right song lyrics have it
2:00:05 write poems have it help you
2:00:08 um
2:00:08 create outlines for a book write
2:00:10 chapters of a book have it help you
2:00:12 rewrite you know if if you wrote wrote
2:00:15 like I wrote the
2:00:17 shitty beginning of a screenplay and I
2:00:19 had to help me rewrite the first couple
2:00:21 of scenes of it just to see how it did
2:00:23 it did great like
2:00:25 um use it for things that you're
2:00:28 passionate about don't worry about the
2:00:29 tech piece of it if you want to do
2:00:32 technical stuff have chat GPT start to
2:00:34 teach you that like you can start to
2:00:36 geek out and run down rabbit holes if
2:00:37 you want to but leave it leave it up at
2:00:41 the tech leave it up at the at the
2:00:43 humanities level
2:00:45 um it's it's remarkable at that [ __ ]
2:00:47 like gpt4 in particular like it's
2:00:50 writing is really good you have to you
2:00:52 have to get good at prompting it if if
2:00:54 you give it shitty prompts it will
2:00:55 comply and give you shitty answers
2:00:58 all right what's the best way to use
2:01:00 okay that was that that's fascinating
2:01:02 all right have you tried prompt
2:01:06 AI PRM I have not it's really good okay
2:01:09 I'll check it out
2:01:11 um I've been doing some Consulting for
2:01:12 ad agencies interested in texted video I
2:01:15 think my
2:01:17 summer will be short-lived well I I
2:01:21 think listen I think companies are going
2:01:23 to be hungry to figure out what the hell
2:01:25 these tools are so I'd say Consulting
2:01:27 with this stuff even though I mean here
2:01:30 would be my only advice if you're
2:01:31 Consulting with AI don't tell them
2:01:33 you're an expert no one's a [ __ ]
2:01:35 expert right now everyone's trying to
2:01:36 figure this out and everything's
2:01:38 changing week to week
2:01:39 um so I would just say be humble about
2:01:41 it just say hey I'm learning as much as
2:01:43 I can about this AI stuff you know can I
2:01:45 come in and tell you what I know
2:01:47 um and just you know try not to put
2:01:49 yourself on too much of a pedestal
2:01:51 because
2:01:51 this [ __ ] is changing so fast it's it's
2:01:54 very humbling like I I kind of pride
2:01:56 myself in being able to quickly grock
2:01:58 [ __ ] and keep ahead of stuff I'm not
2:02:00 even close to keeping ahead of this
2:02:02 stuff it's moving so fast
2:02:05 [Music]
2:02:09 National digital banking starts in
2:02:14 America on July 17 2023 are you ready no
2:02:19 I'm not that's going to be interesting
2:02:21 to see what they do with it we'll see
2:02:23 I'm not all that worried about it right
2:02:25 now
2:02:25 thanks for sharing your time and
2:02:27 knowledge you're welcome you're funny
2:02:28 thank you agreed but looks aren't
2:02:31 everything
2:02:31 [Laughter]
2:02:39 all right do you think there will be
2:02:41 a regression toward the mean AI content
2:02:44 of AI content of AI content
2:02:47 that's a really interesting question
2:02:48 Krista I don't know I mean
2:02:52 I suppose if you well
2:02:55 yeah it's really interesting
2:03:01 I suppose that the point at which you
2:03:03 pull humans out of the contribution
2:03:06 Matrix right if you just if you say
2:03:09 right now humans are still very actively
2:03:13 prompting and and refining
2:03:17 and crafting content in conjunction with
2:03:20 AI if the AI gets so good that that
2:03:24 humans increasingly kind of back out of
2:03:26 the process and just let the machines do
2:03:28 the bulk of the work I think at that
2:03:30 point you might start to see kind of
2:03:32 diminishing
2:03:34 quality
2:03:36 but
2:03:37 but at the same time as as the tools get
2:03:40 more sophisticated I think we're going
2:03:42 to see the tools get trained on
2:03:44 increasingly large data sets like like
2:03:47 at some point the large language models
2:03:50 are essentially going to be the entire
2:03:53 output
2:03:55 of the history of humanity
2:03:57 and and it will be updated in real time
2:04:00 so at that point
2:04:02 I don't know I think I think we'll just
2:04:04 be able to do magical magical [ __ ] with
2:04:07 it but I don't know it's possible I
2:04:09 don't know I don't know enough about
2:04:12 large data degradation to
2:04:16 to talk to that at all so sorry about
2:04:20 that I'll just I just think I don't have
2:04:23 the book smarts
2:04:28 all right I gotta go to [ __ ] bad
2:04:31 people I gotta go to bed eight to see
2:04:33 what the military is going to do with it
2:04:34 they just shot their own well in a
2:04:37 simulation a drone using AI went back
2:04:40 and attacked its Commander because it
2:04:42 didn't like the the uh
2:04:45 the command it gave the Drone so went
2:04:48 back and and shot the operator so
2:04:51 they're gonna have to they're gonna have
2:04:52 to tweak that little algorithm all right
2:04:55 all right I got a few more here for sure
2:04:56 I mean if I welcome for anyone who's new
2:05:00 how do you know this stuff I just
2:05:02 started so so here's how I know this
2:05:03 stuff I I went through
2:05:07 I I discovered the World Wide Web
2:05:09 personally in
2:05:12 um 1994 when when I when I discovered
2:05:14 you know what what had been out for a
2:05:16 few years
2:05:18 um I had a massive Epiphany that holy
2:05:20 [ __ ] this is going to change the world
2:05:21 but like the world has changed it's just
2:05:23 no one knows it yet and I started one of
2:05:24 the first digital agencies
2:05:27 so
2:05:28 I've been through this before and and
2:05:30 this AI thing had the parallels to what
2:05:33 I was going through with the world wide
2:05:35 web in the mid 90s and what's going on
2:05:37 with this stuff now are insane
2:05:40 um and so the minute I had this Epiphany
2:05:42 with chat GPT like I've been playing
2:05:44 with language models before that and
2:05:46 then I had just a breakthrough with chat
2:05:48 GPT like holy [ __ ] everything's
2:05:50 different
2:05:51 um I just dove in I started this Tick
2:05:53 Tock Channel I started an AI Salon I
2:05:56 started in a newsletter
2:05:58 um I I basically started saying yes to
2:06:00 every speaking gig I could get every
2:06:04 Workshop someone said hey can you do a
2:06:06 workshop I'd go do that so I've just
2:06:09 been
2:06:11 insufferably diving deeper and deeper
2:06:14 into learning everything I can about it
2:06:15 and it's impossible to learn because
2:06:17 there's so much of it
2:06:18 um but but just got curious about it
2:06:20 because I think it's incredibly
2:06:21 important
2:06:23 just enjoying the ride thank you
2:06:24 stranger in a stranger land we are in
2:06:27 surreal times I feel like we're in post
2:06:31 we're entering a post sci-fi world I
2:06:33 feel like there's just all these Sci-Fi
2:06:35 movies it's like check yeah we can do
2:06:37 all that [ __ ] now check we can do most
2:06:39 of that [ __ ] now check oh we don't have
2:06:41 time travel yet but you know we'll
2:06:43 probably figure that out in the next
2:06:44 decade
2:06:47 um excellent perspective thank you very
2:06:48 much very cool thanks so much for
2:06:49 sharing that okay listen I gotta go it
2:06:52 is one in the morning here peace out
2:06:55 everybody I appreciate it thumbs up to
2:06:57 all of you I will check back in tomorrow
2:06:59 probably and uh I hope you're well go
2:07:03 there if you haven't go to
2:07:06 chat.openai.com start using chat GPT
2:07:09 do it good night