8/12/2025 - ChatGPT 5's Secrets: AI's Evolution and The FUEL Framework: Mastering AI Prompting

Description

Chapters

Transcript

0:00 live. We'll do it live.
0:03 We'll do it live.
0:07 [Music]
0:09 [Applause]
0:11 [Music]
0:16 [Music]
0:36 freedom came my way that night
0:39 [Music]
0:42 just like a jet plate in and out of
0:46 sight.
0:48 I was hauling ass at a million miles an
0:52 hour, wondering how hard I'd hit
0:59 when they came into the station.
1:04 [Music]
1:06 They said I was bad beyond repair,
1:11 but I got no qualms with my situation.
1:18 say here I am.
1:21 [Music]
1:25 So
1:27 say Sheree Sheree Sheree won't you dare
1:31 do
1:33 say Sheree Sheree Sheree won't you dare
1:37 do say Sheree Sheree Shereehu
1:46 yeah leave a message and your number
1:50 please
1:53 take your time to want to satis satisfy
1:56 me.
1:59 Take all these old fantasies and send
2:02 them care of me.
2:04 [Music]
2:09 Oh yes, yes, yes, yes, yes. What day is
2:13 today? Tuesday. It's Tuesday, people.
2:16 It's Tuesday. Tuesday.
2:19 What do we do on Tuesdays?
2:22 What do we do on Tuesday? Well, we often
2:24 have AI salons. That's not tonight.
2:27 What else do we do on Tuesdays?
2:31 We reveal secrets.
2:34 I'm going to reveal secrets tonight
2:38 that no one else knows
2:40 about Chat GPT. I'm going to be like a
2:43 magician like
2:48 [Music]
2:51 in a westerly direction.
2:56 This car is my train.
3:01 I've been driving. I've been wondering
3:06 what it is I'm running from again.
3:11 Feel like an old man
3:15 holding on to 29
3:19 and up ahead on that horizon
3:23 [Music]
3:24 is California life
3:28 [Applause]
3:30 [Music]
3:34 of a head of trucks carrying a wide
3:36 load.
3:38 Preab house cut in half.
3:43 Cute little front door and two windows.
3:45 My love
3:47 ain't sure whether to cry or should
3:49 laugh.
3:52 You see, I broke a home of myself once.
3:56 As I stumbled to that door,
4:00 I read a note by the dawn light
4:04 said, "Don't you come around here
4:09 anymore.
4:14 I had enough
4:18 of this freedom of
4:22 never was good with decision.
4:25 At
4:26 least that's what I've been told.
4:31 [Applause]
4:35 >> Hi, Kyle.
4:38 [Music]
4:40 [Applause]
4:42 [Music]
4:57 I do hate gender up champ. But we do
4:59 have an actual question on Tik Tok for
5:01 you.
5:02 >> Yes. Yes.
5:02 >> Which I which I can kind of read uh
5:04 through your shirt. So if you'd be so
5:07 kind to switch your cams.
5:09 >> Flip flip my little cams. I can do that.
5:12 >> Beautiful.
5:15 >> Huh? That's
5:16 >> I would say while we're at it, I just
5:17 want to make sure you have Miro up.
5:20 Oh, I do have Miro up. I don't have the
5:27 um the Post-it notes in my view, but I
5:30 now do. And I'm going to put the purple
5:33 below the blue so because I can see them
5:35 both vertically off to the right here.
5:37 Perfect. There you go. Champy sings the
5:41 blues. He does. Okay. Um, hey Pate M,
5:44 what's happening? Side Hustle Mimi Joy
5:46 Party in the house. got some some good
5:48 folks. Professor Oscar G.
5:52 is GPT5 or is it misunderstood?
5:57 Um, we're going to talk about that
5:59 tonight. That's the secret of chat GPT5.
6:02 Professor,
6:04 Professor Oscar,
6:06 you're jumping the gun.
6:09 Can't tell you the secret of Chat GPT5.
6:13 Right off the gate. Right. Right off the
6:15 gate. Out of the gate. Off the bat. Off
6:18 the bat. Out of the gate.
6:22 Yeah, exactly. How dare you ask a
6:25 question that I was going to give the
6:26 answer to?
6:29 [Music]
6:32 But I think you hit it. I think you hit
6:35 it. Here's the secret of chat GPT5.
6:39 [Music]
6:44 I think everyone including
6:48 the people at OpenAI don't have a clue.
6:54 I think that's the secret of ChatgBT5.
6:57 We don't know yet.
6:59 It's too early. We don't know how to use
7:01 it. They don't know how to present it.
7:05 An hour before we came on here, they
7:08 changed the interface again.
7:11 We are back to a model picker with eight
7:13 different models in it as of like an
7:16 hour ago.
7:18 So the secret to chat GPT5 is we don't
7:21 know now. I think I think
7:27 it's better. It's way better, but it's
7:29 not a thing.
7:32 Okay, let's see. What should we do
7:34 tonight? Image benchmarking video
7:36 avatars research tools. Music night. Oh,
7:40 no, no, no, no. We've got an agenda for
7:42 tonight. I was going to uh somebody
7:43 asked for a poll of whether it should
7:45 whether GPT5 was bad or just
7:47 misunderstood. And it it pulled a wrong
7:50 poll.
7:50 >> Oh, it pulled an old poll. Okay, good.
7:53 >> But there is a poll coming. So,
7:54 >> okay, there's a poll on Tik Tok.
7:57 >> On Tik Tok, there's a poll coming. So,
7:59 ignore that. Um, if you're over there
8:02 >> on YouTube, uh, and
8:06 >> what's that?
8:07 >> We're doing it on TikTok. We're doing
8:09 the poll on Tik Tok.
8:10 >> Yeah, doing the poll on Tik Tok. If
8:12 you're on YouTube, um feel free to just
8:14 comment if you have comments.
8:18 Um and if anyone has any questions, feel
8:22 free to uh to pop them in either place,
8:24 Tik Tok or YouTube. Um and we we'll put
8:27 a poll up in a second here. Um better
8:30 than a nose picker.
8:32 You can pick your friends. You can pick
8:34 your nose, but you can't pick your
8:35 friend's nose.
8:39 [Music]
8:51 [Music]
9:01 Good morning.
9:03 [Music]
9:21 She came on him like slow moving cold
9:25 fronts.
9:30 Well, his beard was warmer than a look
9:32 in her eyes.
9:35 [Music]
9:38 Okay. Up in the uh up in the upper
9:41 leftand corner. Oh, wait. Yeah, up in
9:43 the upper leftand corner over there,
9:46 there's a there's a a blue box on an
9:49 angle
9:51 and pull that up. GPT5 is bad,
9:53 misunderstood math.
9:56 Give me my 40 girlfriend back. Um, so
9:59 yeah, go do a poll over there. All
10:01 right.
10:03 [Music]
10:07 Alultman once said, "We will see a
10:09 soloreneur hit a billion dollar
10:10 valuation with AI soon." He just said
10:12 that on a a podcast the other day. He
10:15 talked about that like a year and a half
10:17 ago. That that's a standing bet um
10:20 amongst uh founders in Silicon Valley.
10:23 It has been for a while. Um
10:28 that's funny, Vicki.
10:30 [Music]
10:35 [Applause]
10:38 [Music]
10:45 Thank you. Sad hustle. Mimi.
10:50 [Music]
11:00 Standing between
11:03 you and a hard place is insane.
11:09 Standing too near
11:12 you and a fire makes it clear.
11:17 [Music]
11:19 You're trouble to me.
11:24 Real trouble. Can't you say
11:27 [Music]
11:30 leaving in close?
11:33 Smell of your perfume scares me most.
11:39 Leaving away
11:42 feels stronger every day.
11:47 [Music]
11:48 You're trouble to me.
11:54 Real trouble. Can't you see?
12:01 All right, let's get this show started.
12:04 Let's get this show on the road, people.
12:07 Um, that last song was actually a song I
12:10 wrote. Most songs I sing I did not
12:11 write. That last one I did. Uh, the poll
12:14 uh, five people took it. Chat GBT is
12:17 misunderstood. Was 100% of the answer.
12:20 Um, that's the big secret to GPT5 right
12:23 now. We just don't know. We just don't
12:25 know. We don't know how to prompt it. We
12:28 don't know how to use a model that's
12:29 going to route our prompts to different
12:33 models. Um, and now I want to let me
12:36 show you something they did here which I
12:38 find astounding.
12:42 Um,
12:44 [Music]
12:52 [Music]
12:55 okay.
12:59 There we go. All right. You can all see
13:02 that, right? You can see that there.
13:04 Fant Tik Tok pen. Two block tom. I
13:07 actually got some good historical
13:09 fiction writing with thinking mode. I
13:11 was impressed. Yeah. So, a couple of
13:13 things you should notice with five. One,
13:16 it shouldn't hallucinate as much.
13:19 They've got a what do they call it? A
13:24 I don't know. It looks at the answer
13:25 before it gives it to you. Should
13:27 hallucinate less. Um, look what they
13:30 did. I can't believe they [ __ ] did
13:32 this.
13:34 We are now back to
13:38 one, two, three, four, five, six, seven,
13:42 eight, nine different model choices.
13:48 Nine.
13:50 Nine.
13:52 Nine of them.
13:56 The whole point of chat GPT5 was we're
13:59 going to hide all these models from you
14:00 so you don't have to think about it.
14:03 Then they launched chat GPD5 and
14:06 everyone who liked chat GPT4 because
14:09 they were using it as their girlfriend
14:11 freaked out on Reddit and so they added
14:14 back in GPT4 and they're like we don't
14:17 know what it's doing. How are we
14:19 supposed to know what it's doing?
14:23 So now all this crap is back. It's
14:25 unbelievable to me.
14:28 You don't have all those legacy models
14:30 really.
14:33 I don't want to revert to four because
14:35 it will go away. And then the whiny
14:38 babies came out and Sam caved. I agree.
14:40 Like I I feel like I feel like what I'm
14:42 facing right here is So Vicki, do did
14:47 you So in order to see the legacy
14:49 models, you have to go into your
14:50 settings. You might not have been here
14:51 because I think you were traveling or or
14:53 working on work stuff. If you go into
14:55 settings and scroll down to the bottom
14:57 of the general page, you have to say oh
15:01 and show additional models. Now they
15:03 even changed that. That said show legacy
15:05 models as of yesterday. Now it says show
15:08 additional models. So you have to turn
15:10 that on.
15:11 You have you have auto fast thinking and
15:14 legacy 4 model. Yeah, I've got
15:17 four, 4.1, 03, and 04 mini.
15:22 But what's interesting here is, okay,
15:25 you've got auto mode. So, most people
15:27 are just going to leave it in auto mode,
15:30 right?
15:32 I have mine set to robot. No fluff, no
15:34 BS, just direct, straight to the point.
15:37 Yeah, that's I mean, if if you're not
15:40 going to be in voice mode ever, I think
15:42 that's a fine way to go. And if that's
15:44 listen, if that's the way your brain
15:45 works, then yeah, do your custom
15:47 instructions to do that. So, let's walk
15:50 through these. Um, now we kind of know
15:52 what's in here, right? So, they've got
15:55 fast,
15:57 which this is probably GPT5 mini.
16:05 This is probably or or no, this is
16:07 probably
16:09 GPT50.
16:11 If if we're going to follow their old
16:13 conventions, fast is probably their
16:17 like 40, but it's version five and it's
16:20 fast and it is. It's really fast. I was
16:22 just using it. It's super quick. Then
16:24 they've got Thinking Mini, which is
16:27 GPT05
16:29 mini. So, reasoning mini. And then
16:31 they've got reasoning. And then they've
16:34 got reasoning pro, which is like that'll
16:36 run deep for you. So you've got 40 is
16:40 like fast.
16:43 41 is just an anomaly. I don't know why
16:45 they have 41 in here because it's not
16:47 omnimodal. It's weird. 03 is like
16:52 thinking and then 04 mini is like
16:55 thinking mini. But this is this is 05
16:58 mini. This is 05 reasoning. This is 50
17:01 omni channel is my guess. That's what
17:03 those three are. And then pro just
17:05 thinks longer. Pro is this one, but it
17:07 thinks longer. It gives it more compute.
17:10 Okay.
17:12 Okay. I see them now. Turning them on
17:14 and back off again. Okay. I had to
17:16 reload the page. Okay. Hello, Ocean. All
17:19 right.
17:25 Yeah. And here's the deal.
17:27 Unless you're hankering to go back to
17:30 40,
17:33 I I would just pop this into auto mode
17:36 and forget it. Or if you want something
17:39 really fast, flip into fast mode. Or if
17:41 you want to force thinking, you know,
17:44 jump into thinking mode. I can't imagine
17:46 ever wanting to flip into thinking mini.
17:49 It's just like if I want thinking, I
17:51 want thinking.
17:53 And if I want fast, I want fast. So, I
17:55 could see flipping between fast and
17:56 thinking, but but the their router
18:00 probably works pretty well. Like, if I
18:02 just say, um, if I just say, uh,
18:08 let me see. Um,
18:12 why is the sky
18:14 blue? That probably doesn't require
18:16 thinking.
18:18 There you go. Because sunlight, see how
18:20 fast that is? So that probably used fast
18:24 because sunlight is made up of many
18:25 colors and our atmosphere is like a very
18:27 picky bouncer.
18:29 Scatter sunlight more than others. Now
18:32 if I say um
18:36 share the calculation wait uh explain
18:41 it to me scientifically
18:46 with
18:49 formulas
18:52 now it should go
18:56 into some well it's doing some math here
18:59 it's getting mathy on us. I don't know
19:01 that that went into thinking. So, that
19:03 just says to me it's pretty [ __ ]
19:04 smart. Now, I'm gonna say um
19:09 tell me
19:12 what would have to be different
19:17 and Oh, no. It probably knows this one.
19:24 Uh ask it to
19:28 Yes. Go on.
19:30 [Laughter]
19:33 Um,
19:39 get much more creative
19:47 and deeper
19:49 scientifically.
19:53 um change
19:56 my life with your answer.
20:00 See if it it now we can't get it to
20:02 think. What am I I'm I'm in auto mode.
20:06 All right, Kyle. So So it's fast. It's
20:08 doing some mathy math here. Last night
20:11 we were doing we did um
20:14 someone said to do is
20:18 8.88888 88888.
20:22 The same as
20:24 exactly the same as
20:31 nine.
20:37 So, this got it right and it's not doing
20:39 thinking. This
20:41 honest to God, it feels like this thing
20:42 got better since last night.
20:47 Chat GBD5 did.
20:50 Let's flip over to thinking. Well, let's
20:53 go to thinking mini and say explain
21:00 wavelengths
21:02 in math.
21:15 All right, that's pretty good. And
21:18 that's pretty fast. Tik Tok pin a shoe.
21:23 Hey, what's happening? I'm here for the
21:25 secret to chat GBT. Here's the secret to
21:27 chat GBT. If you if you came and and you
21:29 saw my Tik Tok video and you want to
21:30 know the secret, here's the secret. The
21:32 secret is nobody knows. Nobody knows
21:35 right now. It's too new.
21:38 From a product marketing standpoint,
21:42 they're kind of screwing the pooch a bit
21:44 here because
21:47 what they launched was, hey, we're going
21:50 to do this thing that's going to do all
21:51 the picking for you, all the model
21:53 picking for you. You're not going to
21:54 have to deal with that anymore. And they
21:56 have fully capitulated back to all of
22:00 the
22:02 now,
22:03 let's call them chat GPT lites.
22:07 So the chat GPT lites fell in love with
22:09 chat GPT40 and the minute they took it
22:13 away they went on to Reddit and whed
22:15 like a bunch of whiny babies.
22:19 Now, to be fair,
22:22 um
22:26 what what Alman has said in in some
22:28 interviews since they did that is that
22:33 the relationships people have with these
22:37 generative AI models, with the these
22:39 large language models,
22:42 are of a fundamentally different nature
22:45 than the relationships they have with
22:47 say a feature in Gmail. email that gets
22:49 deprecated, right? If Gmail had some I I
22:53 think at some point they had some kind
22:54 of chat thing that they got rid of or
22:56 they had groups or I don't know what it
22:57 was. And some people loved it, but they
23:01 just it didn't work for them. It wasn't
23:03 working, whatever. And they just got rid
23:04 of it. They just deprecated. They said,
23:06 "On this date, this thing's going to go
23:08 away." Blah, and it went away. Um
23:11 the reaction to getting rid of GPT40 was
23:14 like, "You killed my girlfriend."
23:17 It was like it was it was not good. Um
23:20 people have relationships with these
23:22 things. A lot of people are using these
23:23 as companions and um you know therapists
23:27 and things like that. And so if you if
23:30 you just swap out the model, it's going
23:32 to change the personality by default
23:34 because it's trained on different stuff.
23:36 Um, but they have now put back nine
23:40 different choices for us to all go be
23:42 confused by.
23:45 Just [ __ ] ridiculous. Now, to be
23:47 fair, you have to turn on legacy models,
23:50 right? These aren't just there. But
23:52 yesterday, this was one. Yesterday, this
23:53 was GPT40.
23:55 I didn't I don't care if we get 4.1
23:58 back. I don't frankly care if we get 03
24:00 back because we have five thinking. I'm
24:03 gonna assume
24:05 five is better than three.
24:10 04 mini is some stepping stone to what
24:12 we have.
24:15 It's clearly not this thinking mini.
24:18 This is05 mini I or 50 mini or05 mini
24:23 I'm guessing.
24:25 So anyway, so the the the thing that I
24:28 thought would be good to do tonight is
24:30 just is just play with it a little bit.
24:33 And what I'm noticing immediately
24:35 tonight
24:36 is I don't know if they fired up a new
24:38 bank of servers or if they had if it was
24:42 choosing some older models, but it seems
24:45 incredibly fast tonight and
24:48 and good.
24:52 So,
24:54 oh, here's another thing I learned.
24:55 Here's another This was This was
24:57 actually what inspired my video of like
24:59 I've got the secret to chat GBT5.
25:04 So, one of the nicknames of this channel
25:06 is Chat TMZ, right? Where I'll just
25:09 speculate on [ __ ] you know? I'll see
25:11 some rumors and I'll see a couple of
25:13 things on Twitter and I'll just
25:14 speculate on [ __ ] and then we'll make
25:17 predictions and some come true and some
25:18 don't. But it's just sort of [ __ ]
25:21 speculation.
25:23 Twitter right now is full full of GPT5
25:29 TMZ speculation [ __ ]
25:32 And they they uh so like one of the one
25:36 of the chat TMZ kind of rumors is that
25:41 chat GPT5 is only 32,000 tokens. So to
25:46 put it in context, Gemini has a model
25:49 that's a million tokens. They have
25:51 another one that's two million tokens,
25:55 but their their pro model tok their pro
25:57 model on Gemini is a million tokens.
26:00 Enthropic is 200,000.
26:02 So chat GPTBN at 32,000 tokens. And
26:06 basically what what tokens are, if you
26:08 don't know that term, the token length,
26:12 like 32,000 is basically the the window
26:15 of think of it like short-term memory,
26:18 right? You start having a conversation
26:20 and at about 25,000 words, the beginning
26:24 of your conversation falls out of
26:26 memory.
26:27 Well, apparently
26:30 um someone from OpenAI had to step into
26:32 the Twitter conversation and say, "No,
26:34 no, the GPT5 thinking model,
26:38 like the big nutso
26:42 thinking model has 192,000 or 196,000. I
26:47 think it's 192. Um but 192,000 tokens,
26:51 not 32,000."
26:53 So, so like we don't know. Nobody knows.
26:55 Nobody knows.
26:58 Don't worry about all of the chat TMZ
27:01 [ __ ] Don't worry about the
27:03 speculation saying it's the worst model
27:05 ever. Don't worry about the speculation
27:07 saying it's the best model ever.
27:11 Just use it. Just use it. Just use it.
27:16 Um I saw that too. Yeah. You know,
27:19 another another funny thing is uh
27:22 Perplexity wants to buy Chrome from
27:25 Google, which here's why I would support
27:28 that.
27:29 Fix the memory links for tab hoarders.
27:32 If you're a tab hoarder and you use
27:34 Google Chrome, it's [ __ ] unusable at
27:37 this point. Um I'm I'm a full-on tab
27:40 hoarder. I'm not going to apologize for
27:41 it for it. I love my tabs. It's how I
27:44 get through my day. 27 tabs open, not
27:47 enough. 80. That's more like it.
27:51 Is it going to [ __ ] up my browser? Oh,
27:53 yeah. Oh, yeah.
27:56 All right.
28:00 Ah, I don't think you should be teaching
28:02 this content. I think you're confusing
28:04 people.
28:06 Which content? Which one?
28:13 AI. I shouldn't be teaching AI. I could
28:17 teach people how to get their dog to
28:18 sing. Step one, find a dog that sings
28:21 when you play guitar. You're done.
28:26 You get a singing dog and you get a
28:28 singing dog and you get a singing dog.
28:31 Um,
28:32 another thing struck me today that get
28:34 that that again gets back to we just
28:37 don't know how good these things are
28:39 yet.
28:42 What was it three weeks ago? Open AAI
28:45 launched. No, it wasn't even three weeks
28:47 ago. Two weeks ago, OpenAI launched
28:50 agent,
28:51 this autonomous agent,
28:54 right? So, we've now got Manis, we've
28:56 got Gen Spark. Now, we have Chat GBT.
28:58 There's probably some other ones in
28:59 there. We've got the Perplexity Comet
29:01 Browser.
29:04 Um, so we have these agentic tools.
29:10 Well, no one's talking about agents
29:12 anymore. They're all They're all talking
29:14 about, "Did you take away my girlfriend?
29:16 Is Chat GPT5 the worst model ever?" It's
29:20 there's just noise right now. So, I
29:22 think I think
29:26 what we need to do on this channel for
29:28 the next two or three weeks is just
29:29 play. Nobody knows what to do with an
29:31 agent. Exactly.
29:34 Everyone knows what to do with with a
29:36 girlfriend.
29:38 That's pretty good. producer Brandon
29:40 coming in hot with the with the zingers.
29:45 Um, but it's possible that chat GPT
29:48 agents are going to get much better when
29:51 they have access to these models. So,
29:53 you put it in
29:55 you put it in auto mode and then you
29:58 click on the plus down here
30:01 and you can put it into agent mode. What
30:04 are the sources here?
30:14 Can that really go into my Gmail?
30:27 Huh?
30:29 I don't know. It really can. Is it Is it
30:32 working?
30:39 Let me turn off agent. Turn on sources.
30:47 Oh, that's Google Drive. But I don't
30:48 have Gmail here.
30:51 So that's not a connector. I don't I
30:53 don't get it.
30:59 I've got 22 deep researches left. Oh,
31:02 then there's deep research.
31:05 It only works with agent and deep
31:07 research. Ah, okay. Um,
31:10 that might be something to play with,
31:12 but I I find that terrifying.
31:16 Like like
31:17 OpenAI right now is is a privacy
31:21 cesspool.
31:24 like like like one of the things Alman
31:26 said in this recent interview was
31:29 everyone's putting all their medical
31:31 information in this thing and they're
31:33 having these psychological conversations
31:35 with this thing and at at this point
31:38 there is no expectation of privacy. The
31:40 government has said to open AI you keep
31:43 [ __ ] everything and we can subpoena
31:46 anything we want to at any time.
31:49 There there is no expectation of
31:51 privacy. Like when you go to your
31:53 therapist or your doctor or your lawyer,
31:57 there is a there is there is a framework
32:00 to keep your conversations private.
32:04 That doesn't exist with chat GPT. And
32:06 Altman's like, uh, it's the law. Wish we
32:10 had better ones.
32:11 Um, there's no privacy with ChatgBT.
32:14 Yeah, exactly. So, I'm I'm a little
32:16 skittery about using Gmail with it. But
32:18 anyway, there's so there is so much more
32:21 for us to explore, even if we don't play
32:23 with that stuff.
32:25 Um,
32:40 so wait, so wait, we have nine models
32:43 now.
32:44 Well, no. Let's we we'll get rid of auto
32:46 because auto is just going to choose one
32:48 for us. So, you got three here, four if
32:50 you have pro, and you have four here.
32:52 So, we have eight models. And then we
32:54 have
32:57 1 2 3 4 5 6 7
33:02 modes
33:04 plus the ability to prompt from an image
33:06 or from files from your drive
33:12 plus all the settings plus custom GPTs
33:16 plus codecs plus Sora
33:20 and voice chat.
33:23 So we're at
33:31 to just have a basic understanding of
33:34 chat GBT
33:35 [Laughter]
33:40 like like a solid baseline understanding
33:43 like if someone asked you hey what's the
33:46 big deal about chat GPT for you to be
33:48 able to answer it
33:51 about all the different things you could
33:52 Do you you right now need to master
33:54 about 15 different tools?
34:04 Whatever happened to you can have any
34:05 color you want as long as it's black.
34:07 Yeah, I don't know. I I don't know.
34:09 Well, I do know what what's happening
34:11 with chat GPT and and frankly with all
34:14 of the models. Um Gemini suffers from
34:17 this enthropic to a degree, perplexity
34:20 certainly.
34:22 um is is just model creep. Uh you know
34:26 uh uh co-pilot with Microsoft is totally
34:31 bloatware at this point, right? People
34:33 won't even use it.
34:35 Um
34:41 all right.
34:43 So I think what I'm going to do is this.
34:54 Ah, for biz consulting. Ah, cool. Okay.
34:57 So, let's try something here.
35:00 So, so Nora, welcome. Um,
35:04 you want to use chat GPT5 for biz
35:06 consulting.
35:08 So, let's let's leave this. We're gonna
35:10 leave this in auto mode right now. I'm
35:12 going to do a new chat.
35:15 Um,
35:17 what kind of what kind of industry
35:19 sectors, Nora?
35:28 [Music]
35:34 I broke co-pilot today asking it to
35:36 write a simple job description.
35:47 Um, how can I use AI in uh chat GBT and
35:51 business consulting? Um, okay, let's uh
35:56 let's do this. Let's say
36:00 um
36:04 let's let's say a company that you're
36:07 you're working with what what you say to
36:09 people is I'm going to help you drive AI
36:13 adoption
36:14 and what if one of the first things that
36:17 they say is let's see all industries
36:20 mainly works work with subject matter
36:22 experts. Okay. Okay. Okay. Great. This
36:26 is perfect.
36:29 So, I'm going to say
36:33 um
36:41 I've got a
36:44 consulting
36:46 business
36:48 and I mostly work withmemes
36:54 and as AI gets
36:58 better
37:00 and better.
37:02 It seems
37:05 It seems the role
37:09 ofmemes
37:12 is going to
37:17 disappear
37:19 or transform
37:21 at a minimum.
37:28 I'm just gonna say this. What do you
37:30 think?
37:32 [Laughter]
37:35 Oh, I've got a fun idea for this one.
37:37 Okay, I think you're right. RO is
37:38 already changing and will continue. The
37:41 wrote knowledge part ofmemes is getting
37:44 automated. Right. So, the fact that
37:46 you've got 30 years in import export
37:50 regulation,
37:52 chat GBT doesn't care. It'll just go
37:53 find all the rules and synthesize them
37:56 for you. This means that I I know the
37:59 rules, let me tell you what they are
38:01 portion ofme value is shrinking fast.
38:04 Like like that insight right there,
38:07 the wrote knowledge part ofmemes is
38:09 getting automated. Like that's that's
38:11 actually pretty pretty valuable in and
38:13 of itself right there. Interpretation,
38:15 nuance and judgment will remain
38:18 critical. This is one of the things I
38:19 deeply believe that
38:23 it's not the knowledge inme, it's your
38:25 trust of them. Right? So relationships
38:29 go up in value in this new world. Just
38:32 wrote knowledge goes down in value.
38:35 Understanding why certain rules are best
38:38 practices. Yep. Making judgment calls.
38:41 Knowing people um knowing real world
38:43 constraints as opposed to what chat GPT
38:46 thinks they are. Spotting when an AI's
38:48 answer is technically right but
38:49 practically wrong. Yep. Um the SME role
38:54 may morph into AI augmented decision
38:57 maker.
38:59 Okay. The competitive edge will belong
39:01 tomemes who speak AI. That's good.
39:04 That's big, right?
39:07 Your consulting business can evolve to
39:10 lead that transformation. Okay. So,
39:12 we're going to do a couple of things
39:13 here. one is
39:16 I'm going to say
39:20 let's flip this bad boy into thinking
39:23 not thinking mini thinking.
39:26 So we're going to put it in thinking
39:28 mode. So this is the the the Mac Daddy
39:30 thing and then we're going to do
39:34 deep research.
39:37 Do I want to do deep research or do I
39:39 want to do an agent?
39:42 Ah, I'm going to do an agent.
39:48 I'm going to do an agent.
39:51 And what I'm going to say,
39:55 we'll leave web search on. I'm going to
39:57 turn off Gmail.
39:59 I'm going to say,
40:03 great. I want you to make me a slide
40:08 deck
40:12 that takes these basic concepts
40:20 and builds
40:22 a one-day
40:26 workshop
40:28 for 20memes
40:32 and their
40:34 executive
40:36 management.
40:42 The deck should have
40:47 relevant research
40:51 exercises
40:53 to identifymemes
41:01 in the room.
41:07 with high AI aptitude
41:12 even if they aren't AI trained.
41:21 eg
41:22 things like curiosity
41:27 um
41:29 um
41:30 adaptability
41:34 etc. Right?
41:37 Um
41:41 so the deck should have relevant
41:42 research exercises uh to do it. Um
41:47 and
41:49 the team
41:51 should leave
41:54 the day
41:57 with a massively
42:01 actionable
42:03 um plan
42:05 for the next
42:10 six months.
42:13 Okay.
42:15 So, if you haven't used agent before, so
42:17 what I'm about to kick off, the reason I
42:19 put a lot of [ __ ] in there
42:22 is
42:25 base chat GPT is you put in a prompt, it
42:28 gives you an answer.
42:30 thinking chat GPT the reasoning models
42:34 you put in a prompt and it has a
42:36 conversation with itself
42:39 and sometimes even uses tools and goes
42:41 and searches the webs web and things
42:43 like that and then gives you an answer
42:46 and then agent is a a kind of standalone
42:52 sort of little mini computer instance
42:55 that it spins up where it's it's
42:57 basically its own little computer
42:59 environment and It's going to go off and
43:00 search things on the web and it's going
43:02 to create things and it's going to give
43:04 itself a plan and it's going to follow
43:06 that plan and it's going to update that
43:07 plan. It's going to take a while.
43:10 So, so we're going to fire this off and
43:13 we'll take a look at what it looks like
43:14 when it starts.
43:20 Most projects have been less tech
43:22 oriented and more traditional like a
43:25 gallery for furniture. Most projects
43:27 have been less tech
43:29 Oh, I see. Like like a gallery for
43:32 furniture. Okay. You you could still do
43:34 a similar sort of thing here. So So you
43:36 could say, you know, go do research on
43:38 this thing. So So what this is doing now
43:42 is it's spun up a little
43:45 a little window.
43:49 So it says I'll start I'll open with a
43:51 second. So now it's doing research. So
43:54 So this is now off. It's doing its
43:56 thing. We can we can step back through
43:58 this and see what it what it's been
44:00 doing.
44:01 Um, I need to create a slide deck. I'm
44:05 using container.exec to open files
44:08 basically.
44:10 Uh, I'm going to open a second search
44:12 result. It went to McKenzie and found
44:14 some [ __ ] So now it's off working. All
44:16 right. So it's going to do it's going to
44:18 give us an I don't know plan or our
44:21 workshop our workshop slide deck. So
44:23 that's going to go do that. But now I'm
44:25 gonna flip over to this tab. Oh, no.
44:28 This one I want to use. Hang on. Let me
44:30 flip over to
44:33 this tab.
44:36 All right. Quick Tik Tok question. James
44:39 Hell, AI
44:41 can't become super intelligent without
44:43 being trained by a super intelligent
44:48 individual thoughts.
44:51 Um,
44:53 no. I fundamentally disagree with that.
44:58 Although I mean you could certainly
45:00 argue that the researchers in these top
45:02 research labs
45:05 are are I I don't know if they're super
45:08 intelligent, but they've certainly
45:10 stumbled upon things like recursive
45:12 self-improvement
45:14 um that are that are making the models
45:18 get better and better and better faster
45:20 and faster and faster. So you could
45:22 argue that that the super intelligence,
45:26 you know, is in the mathematicians that
45:28 are, you know, executing these models.
45:30 But if if your if your statement is you
45:34 need a super intelligent, you need a
45:36 person more intelligent than the AI to
45:39 make it as as intelligent as them.
45:41 That's absolutely not the case. That's
45:44 absolutely not the case.
45:47 A lot of the researchers in these labs,
45:50 they don't quite know how these things
45:51 do what they do. They they're finding
45:54 things that they call emergent
45:56 behaviors. I don't know how much you
45:57 know about this [ __ ] You may you may
45:58 know a lot, but
46:01 they will train AI as as they're doing
46:04 these generalized capabilities. What's
46:06 happening is this. The more that they
46:07 train these things,
46:10 the better that they get. And they're
46:11 finding that they get better at things
46:13 they didn't train them for. They call
46:15 those emergent properties or emergent
46:17 behaviors. So, for example,
46:20 big well-trained large language models
46:23 are excellent at language translation.
46:26 They're excellent at cultural
46:28 interpretation, you know, based on who
46:30 who's who's doing the talking, who's
46:32 doing the prompting. Um, they weren't
46:35 trained for that. It's just that when it
46:37 gets enough, you know, access to
46:39 language and knowledge, it's just good
46:42 at that. um they see that again and
46:44 again and again and and the the the
46:47 milestone that we seem to have have
46:50 passed in the past I don't know three
46:53 months
46:56 about three months ago all of the
46:58 frontier model companies stopped talking
47:00 about AGI
47:05 um
47:07 and they started talking about ASI
47:09 artificial super intelligence and what's
47:11 come out is that what they've what
47:13 they've gotten to is basically recursive
47:17 reinforcement learning, recursive
47:18 self-improvement, where the
47:22 model gets trained and it starts putting
47:24 out outputs and the model itself or some
47:27 other models are judging the quality of
47:30 the outputs and it says, "No, that one's
47:32 not good, but that one is good and that
47:34 one is good and they're getting better
47:35 and better and better." Um, and it looks
47:37 like on YouTube you're talking about
47:39 synthetic data. That's one of the other
47:40 things that that they've um cracked is
47:45 up to this point to build these models
47:48 you had to go out on the internet, find
47:50 a bunch of data, categorize that data,
47:53 embed that data into the model, right?
47:56 Which was massive human involvement.
48:00 Those models got strong enough now that
48:03 they can basically mine mine extract
48:06 from the latent space from this
48:08 mathematical knowledge base. They can
48:11 pull out imperfect information and
48:13 basically turn it into perfect
48:16 information synthetically, right? So
48:18 they just pull out lots and lots of
48:20 crap. They say, "What if this is wrong?"
48:22 And then they make it right. And now
48:23 they have a whole new set of data that
48:25 they can train new models on. They call
48:27 it synthetic data. And so a lot of the
48:30 new models are being trained on
48:31 synthetic data produced by the old
48:33 models and combined with recursive
48:36 self-improvement, these things are
48:37 getting smarter exponentially.
48:40 Um,
48:42 so no, you don't need a person smarter
48:44 than the model for the model to
48:46 dramatically exceed
48:50 all of us. And quite frankly,
48:54 for me personally, the GPT4 class and
48:58 and the 03 reasoning models, 01 and 03
49:01 reasoning models, they were already
49:03 significantly beyond my capacity to know
49:05 what they do to do with them. GPT5 is
49:08 some level above that. We're about to
49:10 get Gemini 3 is going to be
49:13 significantly above that. We're about to
49:14 get to new a new thing from Grock in
49:16 probably a month or two. It's going to
49:18 be significantly above that. So, I think
49:21 right now today, and and this is why I'm
49:23 saying nobody knows anything about GPT5
49:26 right now. It just hasn't been out long
49:27 enough. Um unrelated pin. Um
49:32 the AI has has to have a path.
49:38 The AI has to have a path and the path
49:41 is only as good as who taught it.
49:48 I don't know what you mean by it has to
49:49 have a path.
49:52 These things evolve,
49:57 but it's not
50:00 Oh. Oh, yeah. If if you're thinking
50:02 about AI as the way computers used to
50:05 work where you're programming logic in
50:08 if if if that's the question, James, if
50:11 the question is
50:14 are these artificial intelligence these
50:16 these generalized artificial
50:18 intelligence these machine learning
50:20 models that we're playing with
50:21 generative AI,
50:24 they are not um predictive or
50:28 deterministic
50:29 logic machines, right? They're they're
50:31 not they're not you program in some
50:33 logic and it outputs the result of that
50:36 logic. That's not how they work at all.
50:39 They work in a weirdly similar way. And
50:42 and it's not all that weird because
50:45 neural networks were basically designed
50:47 to mimic neurons in our brain from the
50:50 1950s on. So the way these
50:54 computer chips are connected together
50:56 are very similar to how our brains are.
50:59 And what happens is this. It's a it's a
51:02 relatively simple mathematical well it's
51:05 not simple math. It's very complicated
51:07 math but it's simple conceptually.
51:12 What generative AI is is a probability
51:14 calculator. So you take all this data of
51:16 the world and you do what's called
51:18 embedding. And what happens when you
51:21 embed something? If I take the
51:23 Declaration of Independence,
51:26 when I embed it into an AI model, the
51:29 Declaration of Independence as a
51:31 document ceases to exist. It gets
51:34 shattered into what are called tokens,
51:36 which are fragments of words and
51:38 punctuation and things like that. And
51:40 those fragments of words get blasted
51:44 into, and this is going to sound like
51:46 [ __ ] science fiction, but this is
51:48 actually what happens.
51:51 It gets categorized into
51:52 thousanddimensional mathematical space
51:55 in semantic clusters. So you might have
51:58 a token, this token might be the word
52:01 dog. And in this thousand dimensional
52:05 space, the word dog can live in lots of
52:07 different semantic clusters. One of the
52:10 semantic clusters dog can live in is the
52:14 furry thing that sings with you when you
52:16 play guitar. My doggy champ. Right. And
52:19 so all dog related tokens will be in
52:21 that semantic cluster in this massively
52:24 complicated space. And then there's
52:26 another word dog which means the guy did
52:29 his girlfriend bad and he's a dog that
52:31 lives in the dating cluster, right?
52:33 Whatever that might be. The original
52:36 documents cease to exist. The only thing
52:38 that exists are these little tokens that
52:40 get blasted into this crazy mathematical
52:42 space. And when you write a prompt,
52:46 it creates a probability.
52:48 And the probability is that somewhere in
52:51 that mathematical thousand-dimensional
52:53 space is a token that has the highest
52:56 probability of being the next
52:58 appropriate token on the page.
53:01 And so when you see it write a whole
53:03 paragraph or or sort of recreate
53:06 something from the Declaration of
53:07 Independence, what it seems like is it's
53:10 copying and pasting whole documents.
53:12 It's not. It's generating original
53:15 passages every single time.
53:18 It's insane.
53:20 And so what happens is they just keep
53:22 putting more and more data into
53:25 increasingly,
53:27 you know, um,
53:30 deep compute cycles that makes all of
53:34 these semantic clusters more and more
53:37 sophisticated. And so when you type in a
53:39 thing like the chances of it getting
53:41 better are better. And there there's all
53:44 sorts of math on top of that. But
53:47 computers right now as of um November
53:51 30th, 2022, which is the launch of Chat
53:53 GPT,
53:54 large language models do not behave the
53:58 same way computers have behaved in the
53:59 past. In the past, you would put in an
54:03 input, it would process it, and it would
54:05 predictably give you an output. That's
54:07 based on the logic. Large language
54:10 models, you put it in input and
54:13 depending on how the programmer
54:14 programmed it, it's going to give you
54:16 some creative crazy ass answer. They can
54:19 they can muck with a thing called
54:20 temperature which makes it more or less
54:22 creative effectively,
54:26 but it's generating original content
54:29 every single time.
54:31 Every image that's generated is an
54:33 original piece of content every single
54:36 time. Even if you prompt it to make it
54:38 look like something that came before,
54:40 it's still the way it's being generated.
54:42 It's it's it's generative. The G in GPT
54:46 generative AP uh you know uh pre-trained
54:49 transformer is critical.
54:53 We've never had computers be able to
54:56 generate from nothing effectively. And
54:59 that's what they're doing now. That's
55:00 why it's so weird.
55:03 I agree. that that's been happening
55:04 since the Gutenberg press. Uh yeah, but
55:07 not this not in this concentrated away.
55:12 And and quite frankly, the large
55:14 language models, what the Gutenberg
55:16 press did was accelerate and democratize
55:19 how knowledge got into our large
55:21 language models.
55:24 We are large language models.
55:27 If you want to think about it that way,
55:29 the way our brain works, we go out in
55:32 the world, we see things and stepdaddy
55:34 said something to us when we were little
55:35 and we went to that museum over there
55:37 and we read that book over there and
55:39 mama gave us some reinforcement about
55:42 that kind of behavior there. All of that
55:45 information
55:46 gets embedded
55:48 into little fragments of knowledge in
55:50 our brain
55:52 and then at some point you're at a party
55:54 and someone says, "Hey Kyle, what do you
55:57 do?"
55:59 And then your body decides that that
56:01 question is a fundamental threat to your
56:04 existence and you have a panic attack,
56:06 right? it it puts together a bunch of
56:09 little fragments and it says the next
56:12 most predictable token is [ __ ] get
56:14 out of there,
56:16 right? Or you give them some coherent
56:20 answer, right? And it's just stitching
56:22 together little pieces of things you've
56:24 learned along the way. That's the way
56:26 these things work.
56:29 You put in a thing and it's got some
56:31 probability and it squirts out some sort
56:33 of answer. Sometimes it's right,
56:34 sometimes it's wrong.
56:36 I've offended Pate. Of course I have.
56:43 That's a few things that are just wrong.
56:44 What did I get just wrong, Pate?
56:47 Pate actually knows this stuff. I just
56:50 have a degree in acting.
56:54 Neural networks are tensors, more or
56:55 less. I don't actually Pate is the one
56:58 to answer that question.
57:01 Pate works on making uh interacting with
57:06 tensors uh uh more and more efficient.
57:10 That's his job. We've had
57:12 non-deterministic for years. It's not
57:15 guaranteed to be unique or original.
57:17 Yeah, I I agree with that. I
57:21 I'm I'm kind of broadly generalizing it,
57:23 Pate, because chat GPT, you know, if you
57:27 turn the temperature down to zero, you
57:28 can get it to be very deterministic, but
57:32 the general tools that we're using right
57:34 now are not that. So, so yeah, I I
57:37 agree. I again, it's, you know, I'm just
57:40 trying to generalize what's going on,
57:43 but but it's
57:45 it's not how computers used to behave.
57:49 Um, paid is technical and we try to
57:51 generalize things. I know. Exactly.
57:53 Exactly.
57:57 Oh, I know you are. I'm just keeping you
58:00 honest. Aka not fun. No, it's okay.
58:02 Listen, listen. My my uh my co-founder
58:05 at StoryVine uh we call her the crusher
58:07 of dreams. It's like I think it's it's
58:10 it's really important
58:15 particularly
58:17 particularly as these models are getting
58:19 more sophisticated, Pate,
58:23 how I'm understanding what they're doing
58:26 may may not be on base. So So I
58:29 appreciate you keeping me honest, but
58:30 it's actually really important to have
58:32 that balance. Hey Kyle, acting is more
58:34 than a degree. If you can act on stage,
58:37 you can do most anything in marketing.
58:39 Absolutely.
58:42 I can act like a CEO. I can go into a
58:45 boardroom and act like, you know, I know
58:47 what I'm talking about with a bunch of
58:48 executives.
58:52 Been doing that my whole career. Um,
58:55 okay, cool. Um, so I hope that I hope
58:57 that explains it. Um yes probably more
59:00 broadly generalized than is than is
59:03 responsible but but in general that's
59:05 what's going on is so so let me get let
59:09 me give you an let me give you an
59:10 example um from from the very first time
59:13 I used chat GPT this was this was my big
59:15 epiphany with it
59:21 for years people have said to me oh Kyle
59:25 you have a tech company you should learn
59:27 Python
59:28 Python's the next big thing and it's
59:29 easy and
59:32 I'm old and I have add like learning
59:35 something new at this point is like I
59:37 don't [ __ ] give a [ __ ] right?
59:41 But if I wanted to learn Python in the
59:44 before time
59:46 before chat GPT,
59:50 what would I do? I'd either go take a
59:52 class
59:53 and then someone with more knowledge
59:55 than me would put it into some sort of
59:57 organizing fashion and make me practice
59:59 it so I could learn it or I'd go to
1:00:01 Google
1:00:03 and Google would give me a list of, you
1:00:06 know, 5,628
1:00:10 Python
1:00:11 resources
1:00:13 of which 4,595
1:00:16 are some [ __ ] horrific sales funnel
1:00:20 to just try to get me to buy something.
1:00:22 So, it seems like I'm learning something
1:00:24 and then about the time I push the
1:00:26 button to say, "Okay, actually teach me
1:00:27 this." It's like, "You can buy this
1:00:29 now." Right? And so, just to be able to
1:00:36 understand what you needed to do to
1:00:40 learn Python was it was probably I don't
1:00:44 know four, five, 8, 10, 12 hours just to
1:00:48 get to the start line.
1:00:50 And then you have to like read books or
1:00:54 read sites or go into little tutorials
1:00:57 that break it up for you. And you would
1:00:59 learn little bits of Python at a time
1:01:01 and you would do hello world and you'd
1:01:03 be like okay I've done that before and
1:01:05 then you would do like the little hello
1:01:07 world now blinks and now hello world I
1:01:10 don't know is in a form field right Tik
1:01:13 Tok pin. It's so difficult to stay a
1:01:16 breast of everything, especially in my
1:01:18 work of how fast AI is evolving.
1:01:21 Tinkerbell,
1:01:23 give it up. Don't give up AI. Give up.
1:01:27 Forgive yourself. Let yourself off the
1:01:29 hook for trying to keep up with it. You
1:01:31 You literally can't. You just can't.
1:01:33 It's like like It's kind of the the
1:01:36 reason I wanted to do tonight the secret
1:01:38 of GPT5 is no one knows what the [ __ ] is
1:01:40 going on with it. No one knows what it
1:01:42 is or how to use it because they don't.
1:01:45 And and it is probably going to take us
1:01:49 on this channel talking and sharing
1:01:51 things probably a month to even get our
1:01:53 heads around what changed and what's
1:01:55 real and what's not and things like
1:01:57 that. So just give yourself some grace.
1:01:59 Let let it go. Um let let go of the
1:02:02 expectation that you have to keep up
1:02:03 with it because you can't. And and quite
1:02:06 frankly the fact that you're here just
1:02:07 in this conversation means that you're
1:02:09 way ahead way ahead of where most people
1:02:12 are.
1:02:13 So anyway, back to the Python thing. So
1:02:15 in the olden timey world, what Google
1:02:18 did really well
1:02:20 was index information. So I could go
1:02:24 find all the resources I need. But then
1:02:26 when it came to knowledge,
1:02:28 I actually had to use my gray matter,
1:02:31 right? I actually had to
1:02:34 figure out one of those pieces of
1:02:35 information that was organized in a way
1:02:37 that worked with my ADHD crazy straw of
1:02:40 a brain.
1:02:42 And then I would have to have the
1:02:44 attention span and the passion enough to
1:02:48 get in little bits of information. And
1:02:50 then over three or four months I would
1:02:52 understand enough about Python that I
1:02:54 could actually be a little bit nimble
1:02:56 with it. And then after a year or two I
1:02:58 could be good. And after five years, I
1:03:00 could be really good. Right?
1:03:04 The week after chat GPT came out,
1:03:08 the week after
1:03:10 I typed into ChatGpt,
1:03:13 um, write me some Python code that takes
1:03:16 an input and modifies it.
1:03:19 Out came Python code. It's like, oh
1:03:21 [ __ ]
1:03:22 And and apparently that Python code did
1:03:26 what I asked it to do. I didn't know
1:03:28 because I don't know Python, but I could
1:03:30 look at it and I could sort of see the
1:03:32 logic in it. I'm like, I think that sort
1:03:34 of does what I think it does.
1:03:36 But then I realized, oh [ __ ] I don't
1:03:39 know how to run Python code. Like it was
1:03:41 it was this really embarrassing moment
1:03:43 where like, I don't know what to do
1:03:44 here.
1:03:46 And normally you'd have to call a friend
1:03:49 and be embarrassed and be like, I don't
1:03:50 know how to run Python code. Can you
1:03:52 teach me how to run Py? Or you'd go to
1:03:53 Google, how do you run Python code? And
1:03:56 you'd have to go to 8,000 sites to
1:03:58 figure it out. So I just said to chat
1:04:01 GPT, how do you run Python code? Here's
1:04:03 how you run it. Oh no, I don't want to
1:04:05 run it in terminal. Can I can I just do
1:04:06 it on a website? Can I just copy and
1:04:08 paste code into a website? Sure. Here's
1:04:11 five. So the first one was replet, which
1:04:13 I had heard of.
1:04:15 And I took my code and I pasted it in
1:04:17 there and I hit run and it gave me an
1:04:19 error. And I said, "Hey, chatbt." It
1:04:22 gave me an error. It said, "Oh, that's
1:04:23 because you don't have this library
1:04:24 installed, you dumb dumb." I was like,
1:04:26 "Oh,
1:04:28 how do you install that library?" And
1:04:31 this is where my head [ __ ] exploded.
1:04:34 Chat GPD, this is the week after it came
1:04:36 out in 2022.
1:04:38 Chat GBD said, "Well, on your replet
1:04:42 page in the lower leftand corner, click
1:04:44 on this button and then you're going to
1:04:46 see a choice of libraries in the upper
1:04:48 right hand corner and type into the
1:04:51 search box this name and then say
1:04:53 install library and then hit run and it
1:04:56 should run." I was like, "Oh shit."
1:04:59 Okay. So, I went to Replet. I clicked on
1:05:01 the button in the lower left. I looked
1:05:03 for the library in the upper right. I
1:05:05 hit install. I hit run and the code ran.
1:05:10 Within 90 minutes, I had a fully
1:05:13 functional Python application that was
1:05:16 making API calls to OpenAI and
1:05:19 generating content based on a user's
1:05:21 input. Within 90 minutes, a fully
1:05:24 functional Python application. That is
1:05:27 not how computers have ever behaved
1:05:29 before.
1:05:32 That's knowledge on demand. It's not
1:05:34 information on demand because I didn't I
1:05:38 still don't have the knowledge to be
1:05:41 able to write that program. But I wrote
1:05:43 that program. Well, Chat GBT wrote that
1:05:46 program at my behest.
1:05:49 That's not how computers have ever
1:05:52 behaved before. So, we are in a
1:05:54 fundamentally different place than we've
1:05:57 ever been. Um, Shiobani, thank you.
1:06:01 Kyle, you uh you you helped me start
1:06:03 vibe coding with lovable and now I'm
1:06:06 lying uh with uh matching code patterns.
1:06:10 That's awesome. That's vying with That's
1:06:13 awesome.
1:06:14 That's really cool. Um
1:06:18 yeah. Okay. So, what was I going to do?
1:06:21 We were gonna we we we have chat GPT
1:06:24 agent off doing something. I was going
1:06:26 to do something else. This is also chat
1:06:28 add. We Let's go back to uh let's go
1:06:31 back to uh the agent. So the agent
1:06:36 What's going on? Am I
1:06:43 It looks like my
1:06:46 computer is all behind, but maybe not.
1:06:49 Okay,
1:06:52 AI and SMMES.
1:06:56 So it worked for 18 minutes. So, if we
1:06:59 want to see what it did, here's the the
1:07:02 replay
1:07:04 of what it did for 18 minutes. So, it
1:07:07 went off and did research. It wrote us a
1:07:09 bunch of code and then it made these
1:07:11 slides.
1:07:13 Here's your PowerPoint presentation for
1:07:15 the one-day workshop complete with
1:07:18 research insights, exercises to identify
1:07:21 AI readymemes, and a six-month
1:07:23 actionable plan for leveraging AI while
1:07:26 valuing domain expertise.
1:07:28 The narrative flow flows from current AI
1:07:31 adoption statistics to a collaborative
1:07:34 action plan with visually engaging
1:07:36 charts, tables, and diagrams based on
1:07:39 credible sources. McKenzie and Stanford.
1:07:42 You can download it here. Well, let's
1:07:44 just go play with it.
1:07:50 All right, you can all see that, right?
1:07:52 Tik Tok question.
1:07:54 Have you tried meal planning with chat
1:07:57 GPT and have it great write grocery
1:07:59 lists? It's [ __ ] awesome at that.
1:08:03 I'll tell you what, if especially turn
1:08:05 on thinking and and research, it's crazy
1:08:09 at that because especially if you've got
1:08:12 a complicated family situation like this
1:08:15 one's vegan and this one's lactose
1:08:17 intolerant and this this one only likes
1:08:19 pickles this week, you can just vom. One
1:08:23 of the most impressive things about
1:08:26 large language models to me
1:08:30 is that
1:08:33 they can synthesize anything.
1:08:37 You can give it bad data. You can give
1:08:40 it just literally vomit. just turn on
1:08:44 voice mode like where where you're
1:08:46 typing into it with your voice or put it
1:08:48 into conversation mode
1:08:51 and just literally talk into it. Okay, I
1:08:54 got my daughter right now only likes
1:08:56 pickles.
1:08:58 Bobb's lactose intolerant. My husband's
1:09:01 on a vegan kick and I like burgers,
1:09:04 right? You know, you can say all we do
1:09:07 is fight about food. So, I want you to
1:09:10 come up with a week-long meal plan
1:09:12 that's going to make everyone happy that
1:09:14 I only have to cook one meal. And then I
1:09:17 want you to give me a grocery list. And
1:09:19 I want you to organize that grocery list
1:09:22 for Kroger's,
1:09:26 maybe even in your neighborhood,
1:09:28 and it will just [ __ ] do it. It'll
1:09:31 just pull it in and it'll figure it out.
1:09:34 I'll tell you another thing. If you
1:09:35 don't know this about about chat GPT,
1:09:38 you can take a picture of ingredients in
1:09:40 your fridge or or in your cupboard in
1:09:42 your pantry and say, "Come up with a
1:09:45 recipe for tonight based on what I
1:09:47 have." Um, it's it's bonkers for that
1:09:50 stuff. Bonkers. Okay, let's see if this
1:09:52 presentation's any good. One day
1:09:54 workshop formemes. Um, all right,
1:09:57 there's the agenda. AI adoption and
1:09:59 readiness. n 92% of companies plan to
1:10:02 invest.
1:10:04 78 have reported using AI in 2024, up
1:10:08 from 55% from the year before. 77%
1:10:13 um have adopted AI. 38%.
1:10:18 So the the chart here doesn't seem to
1:10:21 match these things, but okay, that's
1:10:24 probably fixable. AI excels at instant
1:10:27 access to public data and facts. Check
1:10:29 the sources.
1:10:33 Yeah. Well, you would check the sources
1:10:35 like down here in the bottom left are
1:10:37 are sources.
1:10:39 Um,
1:10:41 automating routine and repetitive
1:10:43 synthesizing large val volumes of
1:10:45 information, providing general insights
1:10:46 and recommendations. SMEs excel at
1:10:49 proprietary and tacet knowledge, guiding
1:10:52 high stakes and ethical decisions,
1:10:54 customizing solutions, applying
1:10:56 creativity and empathy. Okay. Um,
1:10:59 assessing AI aptitude, curiosity,
1:11:01 adaptability, digital fluency, problem
1:11:03 solving, collaboration, and empathy.
1:11:05 That's a pretty good list.
1:11:09 Group exercise, AI opportunity mapping.
1:11:12 Form cross functional teams. List
1:11:15 processes and tasks in your workflow.
1:11:17 Identify suitable tasks suitable for AI
1:11:20 augmentation.
1:11:21 Um, there's your six-month action plan,
1:11:24 conclusion, and next steps. So, it's I I
1:11:27 would call this a little thin,
1:11:30 but um
1:11:33 you know what might be interesting? Let
1:11:35 me go grab this.
1:11:38 I'm going to grab this prompt.
1:11:42 I'm going to pull a a chat add moment.
1:11:45 We're going to go over to Gen Spark. So,
1:11:47 I did that within um Chat GPT agent,
1:11:50 which chat GPT agent's not all that
1:11:52 good. Genpark's really good. So, I'm
1:11:55 going to do the same prompt.
1:11:59 I want to make a slide deck that takes
1:12:02 these basic concepts. Let's go get the
1:12:04 basic concepts. Oh, I know. I'm not in
1:12:05 there. Hang on.
1:12:16 Um, I have these insights
1:12:21 aboutmemes
1:12:23 today.
1:12:25 So, I'm going to put in all those
1:12:26 insights we had, and then I'm going to
1:12:27 say, I want you to So, now we're going
1:12:29 to have GenSpark do the exact same thing
1:12:32 as we just had ChatBT do, and we'll
1:12:34 we'll see if this one gets any better.
1:12:36 Um, okay, it's 9:13. I want to walk you
1:12:40 through
1:12:42 um I created a new framework, a new
1:12:45 prompting framework.
1:12:48 So, let me change how I'm sharing. We'll
1:12:50 come back to this. Don't let Don't let
1:12:51 me forget to come back to this, Brandon,
1:12:53 because you know me. I'll get all
1:12:55 excited about um what I'm about to
1:12:57 share.
1:13:02 Okay. So, uh so I'm working on a book
1:13:06 right now
1:13:08 and a new brand called Feed Your Prompt.
1:13:10 I'm really excited about it. And the
1:13:11 basic idea here is
1:13:14 AI is not the genius. You're the genius.
1:13:17 And if you learn how to feed your
1:13:19 prompt, you
1:13:22 AI amplifies your genius. Right? If you
1:13:25 treat AI like this genius off to the
1:13:28 side where you expect to give it a
1:13:30 prompt and it's going to give you a
1:13:31 great answer, it's horrible. That's why
1:13:34 most people are disappointed with it.
1:13:36 They don't know how to use it.
1:13:38 Um, and so that's that's the core
1:13:40 premise of the book. Um,
1:13:45 in putting together the book proposal,
1:13:47 um,
1:13:49 I wanted to create
1:13:52 a framework for
1:13:55 how to how to prompt well.
1:13:59 And
1:14:01 one of the things that struck me
1:14:02 recently is that
1:14:06 people who don't know anything about AI
1:14:08 assume that AI is like this single
1:14:10 transaction. You push a button
1:14:15 >> leg
1:14:16 >> and out comes an answer, right? That's
1:14:18 what most people think AI is. Oh. Oh,
1:14:21 you put it in AI. AI can't be creative.
1:14:23 No, no, it can't be creative. That's Oh.
1:14:25 Oh, you you did the lazy button. You hit
1:14:28 the lazy button. Yeah, I understand how
1:14:29 that works. Yeah. Yeah. You push the
1:14:31 button and out comes some AI slop,
1:14:33 right? World's greatest plagiarism
1:14:35 machine, I like to call it.
1:14:39 That's what most people think it is.
1:14:41 That's not what it is.
1:14:43 When you use AI effectively, there is a
1:14:46 chain of craft. There's a there's a
1:14:48 whole lot of interactions
1:14:51 that get you to a usable result.
1:14:55 Um, and you only discover that after
1:14:58 like stumbling through this for a while.
1:15:00 And so I kind of wanted to shortcut
1:15:02 that. So before we get to that prompting
1:15:05 framework, I want to do a little ad
1:15:07 here. I want you to go check out if
1:15:09 you're not if you're new here. If you're
1:15:11 new to AI, it looks like we've got some
1:15:12 folks that are new to AI.
1:15:15 Go to community.thesal
1:15:17 salon.ai.
1:15:20 So, this is the AI salon. So, this is a
1:15:22 community that I started the week after
1:15:24 chat GPT came out, right around when I
1:15:26 had that epiphany with Python. Um, I
1:15:29 started this group with a photographer
1:15:30 out of Boston named Leah Faston. And
1:15:34 this is a remarkable community. And it's
1:15:36 a community of people that are trying to
1:15:37 figure this AI stuff out day after day
1:15:40 after day. They're incredibly generous.
1:15:42 They're incredibly welcoming. If you go
1:15:44 in there and you say, "Hey, I'm totally
1:15:45 clueless. I don't know where to start."
1:15:47 This group is very open, very welcoming.
1:15:50 Showing up nightly to these things is
1:15:53 also good. But the the salon is is
1:15:55 really good. Next Tuesday, we have a
1:15:57 meet and greet. So, you should join
1:15:59 online, introduce yourself, and then
1:16:01 come next Tuesday to the meet and greet.
1:16:03 Okay. Beautiful. All right.
1:16:07 I think Kyle Shannon is is new to AI.
1:16:10 I'm I'm I'm to Listen, the the one thing
1:16:13 that I can promise you is that there are
1:16:16 no experts in generative AI right now.
1:16:18 Anyone that tells you they're an expert
1:16:19 is full of [ __ ] So, I will absolutely
1:16:22 cop to being uh
1:16:27 being clueless. Uh what did I do?
1:16:30 Something I had to reload. Oh, reload
1:16:34 required to apply changes. Oh, it was a
1:16:36 network error. Thank you.
1:16:39 Good question.
1:16:41 How do you teach people if you don't
1:16:42 know how it works? I do know how it
1:16:44 works.
1:16:47 How I teach people?
1:16:50 That's not true. There are experts. I've
1:16:52 worked with them at Apple and Facebook.
1:16:53 Oh, Jeffrey. So, listen. Let me let me
1:16:56 talk about my my context for when I say
1:16:59 there are no experts. Um,
1:17:02 AI has been around for decades, right?
1:17:04 AI's been around for
1:17:07 really since the 50s, but you know,
1:17:11 specialized AI systems have been around,
1:17:14 you know, 80s, 90s, 2000s, like it's
1:17:17 it's embedded in everything we do. There
1:17:19 are absolutely experts in creating AI
1:17:23 systems. What I'm talking about is post
1:17:26 chat GPT. Here, here's the here's the
1:17:29 thing.
1:17:30 In the mid 90s,
1:17:34 the internet had been around for decades
1:17:36 and it was largely the purview of
1:17:38 researchers and um scientists and
1:17:43 librarians and things like that and you
1:17:45 had to use command line and you had to
1:17:47 understand what IP addresses were and
1:17:49 you had to you know jump from this
1:17:52 server to that server and you had to
1:17:53 know how to do it. And then Tim Berners
1:17:56 Lee comes up with this thing called the
1:17:57 worldwide web where you can now just
1:17:59 click on
1:18:01 sites.
1:18:03 The experts that knew how to use the
1:18:06 internet for seven decades
1:18:12 didn't know what this worldwide web
1:18:14 thing was all about. It was a completely
1:18:16 different paradigm. But what that did
1:18:18 was it allowed the 98.5%
1:18:21 of people that are not engineers and
1:18:23 researchers
1:18:25 to be able to click on the internet.
1:18:28 And when I got involved in the internet
1:18:30 was post worldwide web,
1:18:34 right? What do we do? What do you do
1:18:36 with this thing now that everyone can
1:18:39 use it? And so I started one of the
1:18:42 first online magazines and I started one
1:18:43 of the first digital agencies, an agency
1:18:45 called agency.com.
1:18:47 And we built a lot of the first websites
1:18:49 for the Fortune 500. And we were there
1:18:52 figuring out how do you take this
1:18:55 technology that is now democratized
1:18:58 and how do we put that in the world?
1:19:01 That's exactly what chat GPT did for
1:19:04 machine learning and AI. Like I've been
1:19:07 following machine learning and AI for
1:19:10 years and for years I I would see a
1:19:14 press release from Google and like oh
1:19:15 there's a there's a new um version of
1:19:18 TensorFlow that's a lot more user
1:19:20 friendly and I'd be like great I can
1:19:21 finally go do AI and I'd go in there and
1:19:24 I'd start looking at it and I'm like oh
1:19:26 [ __ ] I still need a [ __ ] math degree
1:19:30 and I just didn't do it. And so I just I
1:19:33 said at some point something is going to
1:19:36 happen where the rest of the world is
1:19:39 going to have access to these tools and
1:19:41 that day was the day that chat GPT
1:19:44 launched November 30th 2022.
1:19:47 So my expertise is in okay now that the
1:19:51 technology has the possibility of being
1:19:54 democratized what do we do with it? How
1:19:57 do we use it for work? How do we use it
1:19:59 for play? How do we use it at home? How
1:20:01 do we use it to amplify what we can do?
1:20:05 How do we use it to completely reimagine
1:20:08 what we can do? So that's what this
1:20:10 channel is about. This is not about
1:20:12 mathematics and weights and and and
1:20:15 building AI. This is about now that it's
1:20:18 been democratized, how do we use it? So,
1:20:21 and in that context, there are no
1:20:23 experts.
1:20:25 ChatGpt 5 just launched. Nobody knows
1:20:27 how to [ __ ] use that. There's no
1:20:29 manual for it.
1:20:31 Open AI doesn't even know how to [ __ ]
1:20:33 you put it out in the world. They keep
1:20:35 changing it. It's been out for seven
1:20:38 days and we've had seven different
1:20:39 interfaces.
1:20:43 It's like
1:20:46 how do you use that? I don't know. Do
1:20:48 you?
1:20:50 So, so generative AI is a fundamentally
1:20:53 different flavor and it's a
1:20:55 fundamentally different audience and
1:20:56 it's a fundamentally different set of
1:20:58 challenges and opportunities.
1:21:01 Um, it's not better or worse. It's just
1:21:05 there's there's a lot more people who
1:21:07 are not engineers than our engineers and
1:21:09 that's my interest. So, all right. Um,
1:21:12 okay. So, so let me go check out the AI
1:21:16 salon if you haven't and then let me
1:21:18 talk to you about this new prompting
1:21:20 methodology. So, um, one of the guys
1:21:23 that I I I really like, uh, and and I
1:21:27 think is someone to to aspire to be like
1:21:30 is this guy named Rick Rubin, uh, who is
1:21:33 who is a uh, a music producer who is not
1:21:37 a musician. Doesn't know anything about
1:21:39 music. Well, no, he knows a lot about
1:21:41 music. He doesn't know anything about
1:21:43 actually producing it. He can't work a
1:21:46 recording console. He doesn't write
1:21:48 songs. He doesn't write lyrics. He
1:21:49 doesn't know music theory. But he knows
1:21:52 what he likes.
1:21:54 And we are entering a world where we get
1:21:58 to be
1:21:59 like Rick Rubin where we can just sit
1:22:01 back and let the tools do the work. In
1:22:04 his case, it's musicians and engineers
1:22:06 and things like that, but he's brought
1:22:08 in for his point of view. And so part of
1:22:11 the idea with feed your prompt is that
1:22:13 you don't just learn prompting
1:22:16 frameworks. Uh you know you put these
1:22:19 words in in this order.
1:22:22 The whole idea here is you figure out
1:22:24 who you are and what your taste is and
1:22:26 then all the prompting comes out of
1:22:28 that. And so within that idea is what
1:22:33 I've called what I'm calling the fuel
1:22:35 prompting methodology. And I I think
1:22:37 methodology I'll probably use system. It
1:22:40 might be I don't know. Um but there's
1:22:42 four steps to it.
1:22:44 Two of the steps
1:22:47 likely don't even involve AI.
1:22:50 So there's two steps that involve AI and
1:22:53 there's two steps that don't. So the F
1:22:56 in fuel stands for frame. Frame your
1:22:59 intent. Like what are you trying to
1:23:01 accomplish? Um CJ Fletcher, who's a
1:23:04 who's a an AI artist that's part of the
1:23:06 AI salon, he gave a talk at AI Festivus
1:23:08 last year, and he said he he was
1:23:10 teaching midjourney, and he said the
1:23:13 first thing that you should do when you
1:23:15 when you sit down at midjourney
1:23:18 is breathe.
1:23:22 We're like, wait, what? He's like, you
1:23:25 should breathe. You should take a
1:23:26 breath. You should close your eyes. You
1:23:29 should think about
1:23:31 what am I trying to do?
1:23:33 Do do I want to create funny images or
1:23:35 inspiring images or do I want to make
1:23:37 art or do I want to make do I just want
1:23:40 to entertain myself?
1:23:42 He said frame your intent, right? Just
1:23:44 set it up.
1:23:46 So before you prompt, get clear about
1:23:48 what's your goal, who's the audience,
1:23:51 what are you trying to do to them, why
1:23:53 are they the audience, right? How, you
1:23:55 know, how are you trying to impact them?
1:23:57 So, no matter what you're doing, whether
1:23:59 it's images or words or music or or
1:24:01 anything, step one, frame your intent.
1:24:03 Take a moment. Understand what success
1:24:06 looks like,
1:24:09 who you're doing it for.
1:24:12 Step two
1:24:14 is use your voice. So, this is
1:24:16 absolutely you're going to prompt
1:24:19 you're going to prompt the AI tool.
1:24:23 But there's a big difference between,
1:24:26 you know, give me an article for
1:24:28 LinkedIn on soloreneurs
1:24:32 and
1:24:33 I love soloreneurs and I want to write
1:24:36 an article for LinkedIn that celebrates
1:24:39 them and lets people know how magic they
1:24:43 are and how much guts it takes to be
1:24:46 one, right? Like that that comes from a
1:24:49 place of of point of view, right? of of
1:24:52 having a point of view. Um, and I call
1:24:55 that I call that writing a you prompt.
1:24:58 So you can write a prompt. Make me an
1:25:01 article for LinkedIn. It will [ __ ]
1:25:03 write you an article for LinkedIn
1:25:06 and it will suck.
1:25:09 It'll be boring and hackneed and it'll
1:25:11 be like everything else you've ever read
1:25:12 and it'll be AI slop.
1:25:16 But if you write a you prompt, if you
1:25:18 take that moment and say, "What's my
1:25:19 intent here?" And then the prompt that
1:25:22 you write is laced with your point of
1:25:24 view, the result you get back will be
1:25:28 marketkedly better. And you could stop
1:25:30 right there. Right? And again, I think
1:25:32 this is where most people think AI is,
1:25:35 is you figure out the perfect prompt,
1:25:38 you put it in, and you get back
1:25:39 brilliance.
1:25:41 AI is not the genius.
1:25:45 You're the genius. AI is there to
1:25:47 amplify your ideas.
1:25:49 So use your voice. Frame your intent.
1:25:52 Use your voice.
1:25:59 All right. He uses uh stable diffusion.
1:26:01 Sorry. CJ Fletcher uses stable
1:26:02 diffusion, not mid journey. Similar
1:26:05 idea.
1:26:07 Okay.
1:26:09 Step three. Edit, iterate, and refine.
1:26:12 Now this is where the magic happens in
1:26:14 AI.
1:26:16 Um,
1:26:24 I'm trying to think
1:26:27 maybe
1:26:29 3% of the time when I'm using AI, I'll
1:26:32 put in something, get an answer out, and
1:26:34 use that answer. Maybe 3% of the time
1:26:38 most times. And and what's funny is
1:26:43 when I talked about talk talk to people
1:26:45 about using AI,
1:26:48 one of the questions I'll get is, well,
1:26:50 what happens if it gives me a shitty
1:26:52 answer?
1:26:53 The answer is you tell it it gave you a
1:26:56 shitty answer.
1:26:58 So step three is about the back and
1:27:00 forth. So So in step two in use your
1:27:03 voice, you're you're lacing the prompt
1:27:06 with with your intention, your
1:27:08 personality, your ideas.
1:27:11 And rather than treating it like a
1:27:13 transaction, I'm going to put in this
1:27:15 prompt and get out an answer. Think of
1:27:17 it more like starting a collaboration,
1:27:19 starting a conversation,
1:27:21 starting a a creative back and forth.
1:27:24 And step three is where you do all that
1:27:26 back and forth. This is where the magic
1:27:28 happens. AI's first answer is usually
1:27:31 crap. You shaping it is what is what
1:27:34 gets it there. So, I want to show you
1:27:36 something. So
1:27:38 um this
1:27:43 this is something I worked on with chat
1:27:45 GPT. This is the extended version of the
1:27:48 fuel prompting methodology. Right? And
1:27:50 so this is in a canvas here
1:27:53 and um you know I've got frame your
1:27:56 intent and I've got you know examples
1:27:58 and then I've got use your voice and
1:28:00 I've got examples and prompt prompting
1:28:03 examples for business creative and
1:28:04 storytelling.
1:28:08 I didn't just say, "I want to come up
1:28:10 with a a four-step framework with an
1:28:13 acronym, write it, and it generated this
1:28:15 and I'm done." No. Let me show you
1:28:17 something.
1:28:21 The back and forth to get to this simple
1:28:23 framework that I'm I'm walking you
1:28:25 through right now
1:28:28 probably took me
1:28:30 off and on about six hours.
1:28:34 I'd go in and I'd ask it to do something
1:28:36 and then it would do it and then it
1:28:38 would be sort of right and then I'd do
1:28:39 something else and I I think the
1:28:41 original acronym I picked was fire
1:28:46 and then at some point just fire didn't
1:28:48 feel right and I went back and I looked
1:28:49 at well what other ones did it give me
1:28:51 and fuel was there and I was like yeah
1:28:52 fuel's better and I liked something
1:28:54 about it but to create that short little
1:28:58 document that like one or two page
1:29:01 document explaining this framework
1:29:03 was like six hours of back and forth and
1:29:05 creative collaboration. I think most
1:29:07 people don't know that
1:29:12 this is the work of AI. It's what I call
1:29:15 chain of craft. That that when you get
1:29:18 good with AI, when you become a
1:29:20 professional is when you understand that
1:29:22 you've got it's this back and forth and
1:29:24 it's you who are the one that's saying,
1:29:27 "No, that's not good enough. No, that's
1:29:29 not good enough. No, you [ __ ] that up.
1:29:32 I want it better. I want it better. Oh,
1:29:34 that's really good. Amplify that. And
1:29:36 then it does. It's really good at
1:29:38 responding to what you give it, but it's
1:29:40 not really good at doing it in an
1:29:41 instant.
1:29:43 When people try to give or sell you
1:29:46 prompts, um, is that an annoyance or you
1:29:50 mean for me personally,
1:29:53 I just don't Here's my philosophy on on
1:29:56 on
1:29:58 prompt libraries and prompt recipes.
1:30:01 Um, anyone can give you a prompt that
1:30:04 will work. Any prompt will generate
1:30:08 something in midjourney single word word
1:30:11 prompts can generate really amazing
1:30:13 results like really amazing.
1:30:17 You can use incredibly complicated
1:30:19 prompting systems that include JSON
1:30:22 structuring and and things like that.
1:30:25 I've used those before. You could also
1:30:28 just talk to it. I I mean, one of my
1:30:30 favorite things to do with chat GPT is
1:30:33 just throw it into
1:30:36 um into voice mode and just talk to it.
1:30:38 Just talk to it like a person. And it's
1:30:40 it's writing. It's transcribing your
1:30:42 whole conversation. Um so, no, I mean
1:30:45 those those prompt libraries don't
1:30:49 don't bug me. What what bugs me is how
1:30:52 they're positioned.
1:30:53 How they're positioned is I'm going to
1:30:56 give you this library and it's going to
1:30:58 solve all your problems. No, all they
1:31:00 did was basically give you a new thing
1:31:01 to Google. Right now, I've got a if if
1:31:05 it's a comprehensive prompt library,
1:31:08 they just use chat GPT to write all
1:31:10 those prompts. They came up with a
1:31:12 framework and said, "Write a bunch of
1:31:13 prompts for all these different use
1:31:14 cases, threw it in a PDF, and sold it."
1:31:18 And so now you're you're digging through
1:31:20 that thing thinking the answer's in
1:31:22 there. And what I'm here to say is the
1:31:24 answer is not in there because the tools
1:31:27 keep changing. That prompt library may
1:31:29 or may not work well with GPT5. It might
1:31:32 have worked well with GPT3, maybe GPT4,
1:31:36 maybe not with GPT5. We don't know.
1:31:38 Nobody knows.
1:31:40 What I'm trying to do with the this fuel
1:31:43 prompting methodology is make it tool
1:31:45 independent and prompt independent.
1:31:48 Um, so so pop the slide back up again if
1:31:51 if I'm still sharing my screen. Am I?
1:31:54 Yeah.
1:31:57 Um, so the third one is iterate. So
1:32:00 that's it's the back and forth. It gives
1:32:01 you an answer
1:32:03 and then at some point you're going to
1:32:05 get it to a point that it's good enough.
1:32:07 And and the fourth phase of the fuel
1:32:11 prompting methodology
1:32:13 is launch.
1:32:15 And again,
1:32:17 this may or may not involve AI. It
1:32:20 likely doesn't. Launch is like you
1:32:23 taking the output and going
1:32:26 Okay. Now, where where do I want this to
1:32:28 live? Is this going to go into a word
1:32:30 document? Am I going to use Grammarly to
1:32:32 go, you know, fix up the grammar? Am I
1:32:35 going to read through it again and and
1:32:38 fix any little noobly things? You're
1:32:41 going to polish the output
1:32:43 because this is something you're going
1:32:45 to put in the world with your name on
1:32:46 it. And back to the the the consulting
1:32:49 question earlier with subject matter
1:32:50 experts,
1:32:52 um
1:32:56 those subject matter experts, even if
1:32:58 they're using AI, they need to be
1:33:00 trusted. And so what they put back in
1:33:03 front of a client, shouldn't just be
1:33:05 whatever gets vomited out of an a AI
1:33:07 session. They should actually look at
1:33:09 it, package it, put it together, refine
1:33:11 it, things like that. You might use AI
1:33:13 to help you do that. Maybe add some
1:33:15 images, things like that.
1:33:17 But you might also just
1:33:20 use your brain and a pencil.
1:33:23 That works too. So anyway, so that's the
1:33:26 new the new methodology. So frame your
1:33:29 intent,
1:33:31 use your voice when you're doing your
1:33:33 prompting.
1:33:36 Iterate,
1:33:38 edit, refine it, back and forth, back
1:33:41 and forth, back and forth. And then once
1:33:44 you get something usable, then polish it
1:33:46 before you share it with the world. All
1:33:48 right.
1:33:50 Should I learn Python for AI? It
1:33:52 depends, Carl. If you want to be,
1:33:56 it probably doesn't hurt you to know it.
1:34:00 I don't think you necessarily need to
1:34:02 learn it to execute it. If you want to
1:34:05 work behind the scenes in AI, if you
1:34:08 want to work on the IT side and the
1:34:10 implementation side and the operations
1:34:12 side, probably not a bad idea to know
1:34:15 it. But what I would say is you damn
1:34:19 well better be good at vibe coding,
1:34:21 right? using cursor or or replet agent
1:34:26 um or depending on where you are in the
1:34:28 tech stack maybe lovable if you're if
1:34:30 you're closer to the front end um you
1:34:34 should understand software development
1:34:36 and you should understand the process um
1:34:39 I think learning Python's not a bad
1:34:41 thing but I think learning to program
1:34:44 with vibe coding tools and being um
1:34:48 nimble with that I think is going to
1:34:50 serve you better YouTube question. Do
1:34:53 you tend to start at 30,000 foot level
1:34:56 and refine or break such a project into
1:34:59 smaller chunks and work on them
1:35:00 individually? That's a great question.
1:35:03 Um,
1:35:07 I tend to personally, if I'm not using
1:35:10 AI,
1:35:12 I am very much a start with the three
1:35:15 acts and then break the act into scenes
1:35:18 and then break the scenes into beats and
1:35:20 then break the beats into moments.
1:35:22 Um, I I am very much I start at 50,000
1:35:26 feet and come down.
1:35:29 Um,
1:35:31 AI is weird. I've like one of the most
1:35:34 successful things I did
1:35:38 was I took
1:35:42 I took an incomplete outline where I
1:35:45 where I had written
1:35:47 most of act one in the outline, most
1:35:50 most of the bullets of act one in the
1:35:52 outline and just a handful of bullets of
1:35:54 act two and act three. This was for a
1:35:55 screenplay.
1:35:57 And then I I had gotten excited and I
1:35:59 wrote like the first
1:36:02 six or seven scenes of act one.
1:36:06 So I had these two documents. I had this
1:36:08 incomplete 30,000 ft view and then I had
1:36:13 some scenes that that like down at
1:36:15 tactical level, right? I put both of
1:36:18 those into chat GPT and into it was 40.
1:36:22 I don't know when it was. This was
1:36:23 probably November of last year.
1:36:27 And
1:36:29 I told it to
1:36:32 Oh, wait. This was 01. This is right
1:36:34 when 01 came out. This was the thing,
1:36:36 the reasoning model. I had it um
1:36:41 I said, "I want you to complete the
1:36:43 outline
1:36:45 first." So, I I uploaded both. Here's my
1:36:47 scenes. Here's my incomplete outline.
1:36:50 And I said, "Finish the outline." And
1:36:53 one of the things it recognized was that
1:36:55 I had started the outline in a in a
1:36:58 particular
1:36:59 it was called um it was from the
1:37:03 software company called Mariner
1:37:06 and it was oh [ __ ] was it called? It
1:37:08 began with a C. Anyway, it was a
1:37:11 particular kind of outlining software
1:37:13 and it recognized it and it finished the
1:37:15 outline in the structure of of that
1:37:18 proprietary thing. It was called Contour
1:37:21 I think. Um,
1:37:23 so that was really impressive. So it
1:37:26 finished the outline and then I said,
1:37:29 "Take the most dramatic scene that you
1:37:31 created new in the outline and based on
1:37:35 the script that I gave you, write write
1:37:38 that scene." And so it picked this
1:37:41 pivotal turning point scene in at toward
1:37:44 the end of act two and it added
1:37:47 characters and it maintained characters
1:37:49 from the the ones I had written and it
1:37:52 was good. It wrote like an eightpage
1:37:54 scene. So I kind of feel like with AI,
1:38:00 one of the things we talk about in the
1:38:01 AI salon is you should play first.
1:38:04 One of the things AI is really good at
1:38:06 is just throw [ __ ] at it. Just throw
1:38:08 [ __ ] at it. Just throw [ __ ] at it. See
1:38:09 what it does. It It will likely surprise
1:38:12 you how good it is. Um, Tik Tok
1:38:14 question.
1:38:17 Lol. He's an actor.
1:38:24 I do have a degree in acting. Um,
1:38:26 anyway, so that's the fuel prompting
1:38:29 methodology, right? Frame your intent,
1:38:31 use your voice, iterate, iterate,
1:38:34 iterate the edit phase, and then launch
1:38:36 it. Give it a polish. Um,
1:38:39 so anyway,
1:38:41 hope you had fun tonight. Oh, let's go
1:38:43 look at We've got one more thing to look
1:38:44 at. Let's go back to um, Gen Spark and
1:38:48 see how it did.
1:38:57 So this was we we did this with um with
1:39:01 chat GPT
1:39:04 and this this is the theme AI
1:39:07 transformation workshop a strategic
1:39:09 one-day workshop for subject matter
1:39:10 experts. Okay.
1:39:14 Um agenda is more nicely formatted. One
1:39:18 of the things I noticed about GenSpark
1:39:20 is it's it's just more nicely presented.
1:39:24 This looks better. A window for SM
1:39:27 transformation is closing rapidly as AI
1:39:29 adoption accelerates. 70% of knowledge
1:39:32 work activities potentially automatable.
1:39:35 30% of theme digital transformations
1:39:38 succeed.
1:39:40 AI adoption has le leapt from 33% to
1:39:43 72%. That 72% was in the the open AI
1:39:47 one. Generative AI can automate 60 to
1:39:50 70% of knowledge work. That's a good
1:39:52 slide. It was from McKenzie. Five core
1:39:55 insights exploring the changing theme
1:39:57 landscape.
1:39:59 Wrote knowledge automation. What's at
1:40:01 stake? So, GenSpark, you'll notice here
1:40:04 it took the five ideas that I gave it,
1:40:06 it incorporated them into this, which
1:40:08 chat GPT didn't.
1:40:12 Test at knowledge, the last human
1:40:13 frontier.
1:40:18 I don't know what's supposed to be
1:40:19 there. Oh, I guess these are not done.
1:40:21 Maybe it crapped out. There's 15 slides,
1:40:24 though.
1:40:26 What happened to those slides? Oh,
1:40:28 they're in here. Okay. It just It just
1:40:30 the the display didn't show.
1:40:36 Interactive AI, aptitude, self
1:40:38 assessment. Where are you in curiosity
1:40:41 and learning? Adaptability and new
1:40:42 tools, system thinking, comfort with
1:40:44 ambiguity.
1:40:46 Really good. Build your six-month
1:40:48 transformation template.
1:40:53 Yeah, I would say GenSpark did did
1:40:56 better. Now, it didn't do perfect
1:40:58 because it didn't
1:41:04 it says five core insights. Insight one,
1:41:07 insight two,
1:41:10 and then it moves on to the aptitude
1:41:11 test. So, it missed insights three,
1:41:13 four, and five. So, again, not perfect
1:41:16 again. And and listen, this is in in
1:41:19 that fuel framework. This is the edit
1:41:21 phase, right? It it gave us a really
1:41:23 good start here. It's missing three
1:41:25 slides, right? So, you got to go ahead
1:41:28 and tell it, "Hey, you [ __ ] this up.
1:41:30 You're missing three slides." Anyway,
1:41:33 all right. I'm getting out of here. So,
1:41:36 tomorrow at 400 p.m. Mountain time, 3:00
1:41:39 p.m. Pacific, 6:00 p.m.
1:41:43 Um,
1:41:46 wait, is it 400 PM or is it 3 p.m.? It's
1:41:49 4 p.m. Yes, 4. Uh, 6 6 PM Eastern. Um,
1:41:54 we've got the AI Readiness Project
1:41:56 podcast with Ann Murphy and myself. Um,
1:41:58 we have a really good guest who's a Tik
1:42:00 Tok AI influencer. I'm really excited
1:42:03 about it. Um, so you can check that out
1:42:05 at a readiness project.com
1:42:09 and you can see where we stream. We
1:42:10 stream on YouTube and a bunch of places.
1:42:12 Um, our podcast is now also available,
1:42:16 the AI readiness project on Apple
1:42:19 podcast and Amazon and Spotify and
1:42:23 wherever you find those things. So, go
1:42:25 check that stuff out. And, uh, yeah, I'm
1:42:28 gonna get on out of here. Hope you had
1:42:30 fun tonight.
1:42:32 Hope you people had fun tonight.
1:42:36 [Music]
1:42:42 Final thoughts,
1:42:44 questions,
1:42:46 questions?
1:42:49 Yeah. Yeah. Come back tomorrow. We'll
1:42:51 see how many new versions of Chat GBT
1:42:54 are available. I I just here here's what
1:42:59 here's my thought is you should keep
1:43:02 coming back here at least for the next
1:43:04 two weeks. I'm just going to keep
1:43:06 playing with chat GPT and and trying to
1:43:08 figure out what it's good at. Like the
1:43:10 agent experiment we did tonight. It's
1:43:12 kind of shitty at slides. We know that
1:43:14 now. All right,
1:43:17 but let's go find out. Maybe there's
1:43:19 things it's good at. Um, it looks
1:43:22 tonight like it got a lot better than it
1:43:24 was for the past three days. Why? I
1:43:28 don't know.
1:43:30 We're back to having eight models,
1:43:32 though.
1:43:34 It's so annoying. It's so annoying. Oh,
1:43:38 man. All right.
1:43:41 [Music]
1:43:47 All right.
1:43:48 I know.
1:43:50 you aren't into open stuff stuff, but
1:43:53 Olama Turbo is pretty cool. Elong Ma,
1:43:57 it's not that I'm not into open source
1:43:59 stuff. It's that I don't have the
1:44:01 bandwidth to or the hardware, frankly,
1:44:03 to install it all and and just it it's
1:44:09 Yeah, I'm just at this point I'm a
1:44:11 prepackaged guy.
1:44:14 I'm in the prepared food section of the
1:44:17 grocery store, not the raw ingredients
1:44:20 where the chef shop,
1:44:24 but the open source stuff is getting
1:44:25 really really amazing. All right. Um,
1:44:32 it case it makes fake case law and would
1:44:35 not correct it. Yes. Yes, it will
1:44:37 hallucinate. Now, GPT5 should be better
1:44:40 at hallucinations than GPT4 would. Um,
1:44:42 but yeah, this is the thing. AI at this
1:44:45 point is not good enough to be left to
1:44:47 its own devices. You have to be there.
1:44:49 That's the whole reason I put that fuel
1:44:51 framework together. That third section,
1:44:54 edit and iterate, that's really
1:44:56 important. If you're trying to do legal
1:44:58 cases, you're going to have to check all
1:45:00 of what it produces because it will
1:45:02 [ __ ] hallucinate. And it, to your
1:45:05 point, you you tell it to fix it, it'll
1:45:07 go, "Oh, you're absolutely right. I did
1:45:09 that wrong." And then it'll do it wrong
1:45:11 again. It's maddening. But yeah, that's
1:45:14 that's just the state-of-the-art right
1:45:15 now. That's where we are. All right.
1:45:17 Hope everyone has a good night and uh I
1:45:19 will see you come to the AI readiness
1:45:21 project tomorrow at 4 PM Mountain.
1:45:24 AI readiness project.com. And uh I will
1:45:27 see you tomorrow night 8:00 PM. Same bat
1:45:29 time, same bat place, same shitty home
1:45:33 office.
1:45:34 Peace.