1 00:00:00,000 --> 00:00:04,720 This is a little bit of a salacious document, but this document goes into a 2 00:00:04,720 --> 00:00:10,440 whole bunch of different details on what likely went into training GPT-4, a 3 00:00:10,440 --> 00:00:34,280 little thing called speculative decoding of all things. Ladies and gentlemen, boys 4 00:00:34,280 --> 00:00:37,160 and girls, children of all ages, dogs, cats, robots, and everybody in between, 5 00:00:37,160 --> 00:00:46,280 especially you slow and lazy chat GPT-4 responses. This is HTTTA, how to talk to 6 00:00:46,280 --> 00:00:51,080 AI. I am your host, Wes the SynthMine, SynthMine Wes, and as always, in a world 7 00:00:51,080 --> 00:00:56,120 where data reigns and algorithms dictate our future, one voice stands out to 8 00:00:56,120 --> 00:01:00,640 decode the mystery, simplify the complexities, and navigate uncharted 9 00:01:00,640 --> 00:01:05,600 pathways of artificial intelligence. Who, might you ask, is pioneering the 10 00:01:05,600 --> 00:01:09,640 frontier between humanity and technology? Your guide, your mentor, your friend, 11 00:01:09,640 --> 00:01:13,680 with her radiant knowledge, relentless curiosity, she dives into the realms 12 00:01:13,680 --> 00:01:18,840 where machines learn and grow. She is your tech guru. It's go to go everybody. 13 00:01:18,840 --> 00:01:25,040 Gee, how are you? Hi, I am great. For a special occasion, I think I should say 14 00:01:25,040 --> 00:01:33,080 everyone, let's meet wistful, wise, and wonderful Wes the SynthMine. Oh, because 15 00:01:33,080 --> 00:01:38,560 we are having a little celebration. Yeah. If you're not going to bring this up, 16 00:01:38,560 --> 00:01:43,280 I'm bringing this up because I'm so incredibly proud of you. And thank you. 17 00:01:43,280 --> 00:01:50,320 The Upwork, we need to talk about this. You are killing and making history. You 18 00:01:50,320 --> 00:01:56,080 flatter me, Miss GoToGo. So yes, we got some super news in conjunction with 19 00:01:56,080 --> 00:02:00,720 Upwork's AI Hub release. The last few weeks they've been cultivating and 20 00:02:00,720 --> 00:02:04,560 interviewing and screening a whole variety of freelancers and freelance 21 00:02:04,560 --> 00:02:09,360 agencies on the platform to add to this new AI Hub. This is a culmination of 22 00:02:09,360 --> 00:02:13,760 your one-stop shop. If you're looking for someone that can consult, can develop, 23 00:02:13,760 --> 00:02:18,240 can build, can prompt engineer, can make you some AI artwork. And if you happen to 24 00:02:18,240 --> 00:02:22,440 scroll down on the AI Hub, we'll put the link up. There's your friendly 25 00:02:22,440 --> 00:02:29,360 neighborhood SynthMine who happens to be one of the now Upwork certified top 1% AI 26 00:02:29,360 --> 00:02:36,760 consultants, I guess. We need some sort of claps. I can do claps. There we go. Claps around for 27 00:02:36,760 --> 00:02:42,800 everybody. Okay, let's take a step back. I remember when we started talking 10 28 00:02:42,800 --> 00:02:48,400 years back in AI years, which is what? Maybe four months ago. Wow. Fourteen weeks 29 00:02:48,400 --> 00:02:53,640 ago. Yeah. I remember when you told me that, hey, I listed on Upwork to start 30 00:02:53,640 --> 00:02:57,720 doing some prompt engineering jobs. Let's see how it goes. And I remember when you 31 00:02:57,720 --> 00:03:02,880 were like, oh, I just finished a job and like got rating. You're five out of five 32 00:03:02,880 --> 00:03:08,360 stars now straight. So fast forward. This is just, you have the testament how 33 00:03:08,360 --> 00:03:12,920 curiosity in prompt engineering and understanding how to talk to AI and 34 00:03:12,920 --> 00:03:19,600 going, following basically that route, can grow into complexity, into 35 00:03:19,600 --> 00:03:24,880 opportunities, into SynthMine's collective, coming together with everyone 36 00:03:24,880 --> 00:03:31,560 else. I am grateful for a team that includes you, G and a bunch of other 37 00:03:31,560 --> 00:03:36,520 folks that I've been lucky to link up with and meet that share a deep 38 00:03:36,520 --> 00:03:41,640 passion and excitement for everything AI. And because of some of those folks who 39 00:03:41,640 --> 00:03:47,080 also happen to be on Upwork, we have our SynthMine's agency on there. And in 40 00:03:47,080 --> 00:03:52,040 conjunction with this release and us getting this notoriety and this badge on 41 00:03:52,040 --> 00:03:56,240 our profile that we are an Upwork certified AI expert, they mentioned that 42 00:03:56,240 --> 00:04:00,400 we are the fifth most, I don't know if it's popular is the right word for it, 43 00:04:00,400 --> 00:04:05,440 but fifth most globally AI consultants in terms of like jobs that we've done, 44 00:04:05,440 --> 00:04:10,800 finished and rating. That's the fifth largest grossing between me and some of 45 00:04:10,800 --> 00:04:18,640 my team on SynthMine. So yeah, it's been a meteoric rise, I suppose. But hopefully 46 00:04:18,640 --> 00:04:22,720 this is just the beginning. And you know, I remember there's all these discussions 47 00:04:22,720 --> 00:04:27,160 like, oh, it's prompt engineering, who is actually really making money? If you're 48 00:04:27,160 --> 00:04:31,920 creative and you're curious, there's so many opportunities, especially in the 49 00:04:31,920 --> 00:04:36,560 field that people don't know what it is. And just knowing extra, going and 50 00:04:36,560 --> 00:04:40,920 figuring out and helping people and being ahead of the time in that way. 51 00:04:40,920 --> 00:04:45,880 Everything what I'm seeing, more and more stories coming out and just the 52 00:04:45,880 --> 00:04:51,080 SynthMine's collective, everyone involved. It's just amazing. It's a testament that 53 00:04:51,080 --> 00:04:56,080 yes, it is a real thing. It is early days. Maybe it's not going to be around for 54 00:04:56,080 --> 00:05:01,440 five years, but it's going to change its form and shape. But the fundamentals 55 00:05:01,440 --> 00:05:07,360 being first, being first listed on Upwork, that is huge. 56 00:05:07,360 --> 00:05:11,320 And just for some context too, we've got 19 completed jobs. We've got a bunch of 57 00:05:11,320 --> 00:05:17,000 other ones that were underway. And then there's my agency too, that has other 58 00:05:17,000 --> 00:05:23,040 work. Like this is something that was started in March or July now, right? My 59 00:05:23,040 --> 00:05:28,760 first job, I got $32 an hour to do prompt engineering on here. Not big bucks, not 60 00:05:29,080 --> 00:05:34,480 paid consultancy meetings. Someone taking a chance on me based on the profile, 61 00:05:34,480 --> 00:05:37,560 based on what I was putting out there and saying, all right, well, let's see 62 00:05:37,560 --> 00:05:42,280 what you got. And grateful for that opportunity because on platforms like 63 00:05:42,280 --> 00:05:46,680 this, like Upwork, like Fiverr, and we've talked about it before, we as humans 64 00:05:46,680 --> 00:05:53,240 now, internet existing beings that do commerce, interact and talk and just live 65 00:05:53,240 --> 00:05:58,520 online a lot of the time. What's our nature? Who's the best rated? Who's the 66 00:05:58,520 --> 00:06:02,440 most reliable? Who's done the most jobs? Who's earned the most? It's tough to 67 00:06:02,440 --> 00:06:07,520 break out or break onto the scene sometimes. So I'm grateful for that first 68 00:06:07,520 --> 00:06:11,400 opportunity where it all started. Because to come to this point three, four months 69 00:06:11,400 --> 00:06:17,680 later, and now being able to be in a class where Upwork has looked through 70 00:06:17,680 --> 00:06:24,240 our offerings, evaluated how we interact with clients and our outputs, the 71 00:06:24,240 --> 00:06:28,080 outcomes that we create for them. They've done all that and said, all right, you've 72 00:06:28,080 --> 00:06:32,320 got what it takes to handle our enterprise tier of clients. So all right, 73 00:06:32,320 --> 00:06:33,520 let's welcome to the big leagues. 74 00:06:33,520 --> 00:06:39,440 And meaning that you had a physical like call with Upwork that they've been 75 00:06:39,440 --> 00:06:43,840 going through asking questions. So I really like this approach that they're 76 00:06:43,840 --> 00:06:49,280 taking this very seriously and actually doing the background check who is behind 77 00:06:49,280 --> 00:06:50,920 how legit things are. 78 00:06:50,920 --> 00:06:55,000 And let me tell you too, when they had me fill out and update some of the things 79 00:06:55,000 --> 00:06:58,600 that I offer and the agency offers, even some of the things that they were saying, 80 00:06:58,600 --> 00:07:03,440 like, hey, what models do you have experience in? What different coding 81 00:07:03,440 --> 00:07:07,360 languages, what libraries you know how to use. It was clearly very well thought 82 00:07:07,360 --> 00:07:11,880 out. And as a result of that call, you know, I got a little bit of insight into 83 00:07:11,880 --> 00:07:18,360 their long term goals. The gig economy is definitely live and well in, I think 84 00:07:18,360 --> 00:07:21,960 since obviously Uber is kind of the first thing that put it on the scene over a 85 00:07:21,960 --> 00:07:26,640 decade ago, but just since the pandemic, this is I think a very real, you know, 86 00:07:26,640 --> 00:07:31,400 people make their full time incomes from stuff like this, have entire businesses 87 00:07:31,720 --> 00:07:35,640 on these type of platforms that they're hiring people around the world. 88 00:07:36,280 --> 00:07:41,160 So just to, you know, have something on here where they're where they're forward 89 00:07:41,160 --> 00:07:44,440 thinking enough to go, all right, let's try to have have that for everybody, 90 00:07:44,840 --> 00:07:49,520 but also be able to serve big businesses in a way that they can feel a little bit 91 00:07:49,520 --> 00:07:54,880 better about shopping or their choices and who to use for, you know, their AI 92 00:07:54,880 --> 00:07:57,880 objectives, I think is very, very with the times. 93 00:07:58,200 --> 00:08:04,360 And that kind of reminds me of what the ability I found her Mohammed Ahmad. He 94 00:08:04,360 --> 00:08:10,320 was talking about who is going to be that consultants are running AI 95 00:08:10,320 --> 00:08:16,240 departments at AI companies or any kind of institution. And he had very strong 96 00:08:16,240 --> 00:08:20,720 standard, it's not going to be a conventional MBA or conventional 97 00:08:20,720 --> 00:08:25,960 developer. And he was talking about incredible power of communities and also 98 00:08:25,960 --> 00:08:29,800 open source communities. And he was saying that in his company, there is 99 00:08:29,800 --> 00:08:34,760 like 19 year old PhD student, there is, I think he even mentioned someone from 100 00:08:34,760 --> 00:08:41,640 Amazon warehouse who just literally self-taught how to program. So he was 101 00:08:41,640 --> 00:08:47,480 pointing out that you have to be incredibly passionate and take part in a 102 00:08:47,480 --> 00:08:52,560 way leading the way with everyone else involved. And this is how you build up 103 00:08:52,560 --> 00:08:57,880 real expertise, not necessarily going to school and learning what worked in the 104 00:08:57,880 --> 00:09:02,440 past. So I think, yeah, testament to gig economy and communities and 105 00:09:02,440 --> 00:09:05,560 communities coming together to form this collective. 106 00:09:05,640 --> 00:09:11,400 Yes, I am a grateful recipient of those benefits, both from our learn prompting 107 00:09:11,400 --> 00:09:16,640 community that brought us together, G and also my team at SynthBinds. And we 108 00:09:16,640 --> 00:09:21,400 are a complete reflection of what Mo over at Stability, their CEO, you know, 109 00:09:21,400 --> 00:09:26,520 his approach as well. We have got 18 year old freshmen in college that are 110 00:09:26,520 --> 00:09:31,480 just, you know, don't have the education or anything about coding or prompting or 111 00:09:31,480 --> 00:09:35,080 anything like that, but just are super excited and passionate about it. All the 112 00:09:35,080 --> 00:09:41,840 way up to my my old timers that have 25 years of development, leading dev teams 113 00:09:41,840 --> 00:09:45,400 and at all levels of enterprise. We welcome their contributions. 114 00:09:45,400 --> 00:09:51,160 Professors, architects, everyone. And I think another very important aspect is 115 00:09:51,200 --> 00:09:56,920 people bringing their very specific expertise and knowledge. Yes, when you 116 00:09:56,920 --> 00:10:00,480 are in a meeting and you talk about AI and you go like, okay, this is a company, 117 00:10:00,800 --> 00:10:05,920 they want to integrate this chatbot. What other things we can put on top or 118 00:10:05,920 --> 00:10:10,080 ideas? So if there is someone from marketing is if there is someone from 119 00:10:10,080 --> 00:10:14,920 customer service, imagine just those conversations happening in the moment. 120 00:10:14,920 --> 00:10:20,760 And then you throw, yeah, architect on top or I don't know, like someone from 121 00:10:21,320 --> 00:10:27,560 bioengineering. These kind of such a diverse different perspectives, you'll 122 00:10:27,560 --> 00:10:29,280 such innovative idea. 123 00:10:29,320 --> 00:10:34,240 That is the ticket here. The diversity of a team is what really raises the bar, 124 00:10:34,480 --> 00:10:37,560 especially in this because everyone's different viewpoint, different 125 00:10:37,560 --> 00:10:41,840 background can contribute. And that's what all this AI is. It's the great, 126 00:10:41,840 --> 00:10:48,600 great equalizer. And in my mind, so thank you for calling out our little happy 127 00:10:48,600 --> 00:10:53,600 news. Feel free to come find us over on Upwork or elsewhere if you have some AI 128 00:10:53,600 --> 00:10:54,080 goals. 129 00:10:54,400 --> 00:10:59,720 What is the other place where everyone can meet us and engage with us and learn 130 00:10:59,720 --> 00:11:00,720 from us, please? 131 00:11:00,840 --> 00:11:06,840 Our Learn from thing Discord and synthminds.ai, howtotalk2.ai, all those 132 00:11:06,840 --> 00:11:09,160 good places. Did you have something else in mind? 133 00:11:09,160 --> 00:11:14,360 I'm getting more and more messages, people asking where we can learn and 134 00:11:14,360 --> 00:11:19,760 meaning not just self paced learning, like learn prompting is great. Like 135 00:11:19,760 --> 00:11:25,880 this is just to set your curiosity on fire what's possible. But learning from 136 00:11:25,920 --> 00:11:30,400 experts, learning from people who actually do it, I have one place in mind, 137 00:11:30,480 --> 00:11:31,840 which is Core-Eyes. 138 00:11:31,880 --> 00:11:37,400 Yes, indeed. Where we'll be launching our first course offering with AI and 139 00:11:37,400 --> 00:11:43,160 chat GPT for everyone. That will be the 21st of August is convening. The link 140 00:11:43,160 --> 00:11:46,560 has been in the newsletter and in the show notes. We would love to see you 141 00:11:46,560 --> 00:11:50,440 down there. We're going to be expanding on everything from every aspect in 142 00:11:50,440 --> 00:11:55,000 learn prompting, tons of real world use cases, really just going to be the 143 00:11:55,000 --> 00:11:58,760 foundation for hopefully a lot more offerings on there in the coming months 144 00:11:58,760 --> 00:12:04,600 that maybe even we'll see data architect and renowned YouTube celebrity go to go 145 00:12:04,600 --> 00:12:06,560 teaching something on Code Rise. 146 00:12:06,560 --> 00:12:12,240 Definitely going to be supporting you in this and everyone else. Talking about 147 00:12:12,400 --> 00:12:19,760 gig economy and work and expert also up work. I keep getting notifications from 148 00:12:19,760 --> 00:12:25,400 Reddit about you know chat GPT being down. And then I went to read comments 149 00:12:25,520 --> 00:12:29,840 because it's really feels like people are freaking out about it. And the 150 00:12:29,840 --> 00:12:38,080 comment section is incredible. How many people became so dependent that there 151 00:12:38,080 --> 00:12:42,640 was someone saying like, okay, so everybody throw ideas. How do I tell 152 00:12:42,640 --> 00:12:47,800 that I'm not going to meet any of my deadlines? Because that's it. My 153 00:12:47,800 --> 00:12:53,200 superpower is gone. Or someone was saying like, also, I need to now remember 154 00:12:53,200 --> 00:12:57,320 how to code and go to this stack overflow. Like what there is all the 155 00:12:57,320 --> 00:13:04,040 speculations about that GPT four is not as smart as it used to or not about 156 00:13:04,040 --> 00:13:10,560 what's not as good as it should be. And I know we discovered something. So do you 157 00:13:10,560 --> 00:13:12,920 want to just dive into that? 158 00:13:13,040 --> 00:13:17,320 I mean, we definitely could. So this is, I mean, a little bit of a salacious 159 00:13:17,320 --> 00:13:21,160 document, so to speak, but hey, I'm just we're commenting on it. We didn't do 160 00:13:21,160 --> 00:13:26,000 anything with posting, you know, do not claim to be authors or necessarily 161 00:13:26,000 --> 00:13:30,200 endorse or abide by any of the things put out in there. But to just reiterate, 162 00:13:30,200 --> 00:13:35,120 this is leaked by individual by the name of Yam Peleg, scientist, entrepreneur, 163 00:13:35,280 --> 00:13:41,040 founder and CEO of deep trading. This reason this is kind of a salacious post 164 00:13:41,040 --> 00:13:44,720 and on a random pay spin site right now. And we'll put the link in the 165 00:13:44,880 --> 00:13:50,080 description and in the show notes. But it very well may get taken down is because 166 00:13:50,080 --> 00:13:53,840 this was part of a research study that was put behind a paywall that was 167 00:13:53,840 --> 00:13:58,160 1000 bucks to access. And I think the the publisher, the author here just put 168 00:13:58,160 --> 00:14:01,440 it up there. And he's already had a couple of takedown complaints. But it's 169 00:14:01,440 --> 00:14:06,960 the internet. It was something from a study titled GPT for architecture 170 00:14:06,960 --> 00:14:11,800 infrastructure that was behind a paywall that was 1000 bucks to even access the 171 00:14:11,800 --> 00:14:17,320 article. So thank you, Yam for potentially taking the bullet for all of us so we 172 00:14:17,320 --> 00:14:21,400 can learn a little bit. Yeah, I mean, that's what it is. But this document goes 173 00:14:21,400 --> 00:14:26,520 into a whole bunch of different details on what likely this is not confirmed 174 00:14:26,520 --> 00:14:33,280 information, but what likely went into training GPT for and specific to your 175 00:14:33,280 --> 00:14:38,120 comment, go to where people are saying it's down or the quality is deteriorated, 176 00:14:38,320 --> 00:14:44,000 you know, and it's it's no good anymore. There's something to that here in this 177 00:14:44,000 --> 00:14:49,160 document, which is a little thing called speculative decoding. Of all things, they 178 00:14:49,160 --> 00:14:53,320 actually talk about that in here. It's actually a function of the architecture 179 00:14:53,360 --> 00:15:00,040 of GPT for why some of the quality might has reduced or people thinks that the 180 00:15:00,040 --> 00:15:04,240 quality has gotten a little worse. This is my favorite insight from it. There's 181 00:15:04,240 --> 00:15:09,840 this calculation chinchilla is optimal, which basically goes, hey, for this 182 00:15:09,840 --> 00:15:13,480 number of parameters, you're going to need this number of tokens to train on. 183 00:15:13,480 --> 00:15:20,440 And for GPT for if you use this ratio, it should have trained on twice the 184 00:15:20,440 --> 00:15:29,560 amount of tokens that it actually ended up training on, which was 1.8128 trillion 185 00:15:29,800 --> 00:15:36,400 parameters, 120 layers trained on 13 trillion tokens. So what that speaks to 186 00:15:36,400 --> 00:15:42,120 is they are having trouble finding free quality data. So if you've already 187 00:15:42,120 --> 00:15:46,880 scraped all all the Internet, all these resources, I think here's a couple of 188 00:15:46,880 --> 00:15:53,600 things where they confirmed what was in their libgen, SciHub, all of GitHub, you 189 00:15:53,600 --> 00:15:58,320 know, all the Wikipedia, these type of things. But that still is not enough. 190 00:15:58,440 --> 00:16:03,080 Common crawl and refined web were both 5 trillion tokens. So that means that even 191 00:16:03,080 --> 00:16:09,280 all of this all of GitHub, all of SciHub and libgen was 3 trillion tokens, and it 192 00:16:09,280 --> 00:16:11,160 technically still should have had more. 193 00:16:11,160 --> 00:16:16,640 This is something what I wanted to talk with you about AI data pollution and the 194 00:16:16,640 --> 00:16:21,680 comparison to nuclear impact from 1945. 195 00:16:21,840 --> 00:16:27,120 So you mentioned quality, right? It actually has the perceived reductions in 196 00:16:27,120 --> 00:16:32,040 quality actually are speculatively due to this thing called speculative decoding 197 00:16:32,120 --> 00:16:37,360 in GPT-4's inference. Again, author not 100% sure here. But basically, since 198 00:16:37,360 --> 00:16:42,440 there's faster models underneath to decode the tokens in advance, that then 199 00:16:42,440 --> 00:16:48,840 gets feed into the big GPT-4 large model, right? The quality of some posts might 200 00:16:48,840 --> 00:16:54,400 have deteriorated because to save on costs, they're letting the smaller models 201 00:16:54,640 --> 00:16:59,760 like answer the questions in some instances for like lower probability kind 202 00:16:59,760 --> 00:17:06,000 of inputs. So that would mean you have these smaller models that are mainly just 203 00:17:06,000 --> 00:17:11,880 meant to parse through a input prompt and then feed it to the right, you know, 204 00:17:11,920 --> 00:17:15,240 because there's nine models that apparently make this up, make up GPT-4, 205 00:17:15,240 --> 00:17:18,880 feed it to the right places to get the response in instances where it's like, 206 00:17:18,920 --> 00:17:22,920 oh, well, this is a simple question, simple input. I'm only going to have the 207 00:17:22,920 --> 00:17:28,560 lower quality speculative decoding models, the smaller faster ones respond, 208 00:17:28,600 --> 00:17:33,040 not big GPT-4. So there's another case for prompt engineering right there. 209 00:17:33,040 --> 00:17:37,960 Simple zero-shot prompt, not as good right now because to save costs, hey, 210 00:17:37,960 --> 00:17:43,040 I'm going to have the smaller, more narrow model that's main job is just 211 00:17:43,080 --> 00:17:48,240 translating and organizing for big GPT-4. I'm just going to have it answer it 212 00:17:48,280 --> 00:17:52,440 because it saves us on costs and speed, which is the big limitation. 213 00:17:52,720 --> 00:17:59,040 But it makes so much sense in a way because if you look at like population 214 00:17:59,040 --> 00:18:06,000 who use charge GPT, a percentage of people asking absolutely basic questions. 215 00:18:06,240 --> 00:18:11,760 So why would they let GPT-4, like why would they pay so much money for this 216 00:18:11,760 --> 00:18:18,080 compute? So optimizing it and fine tuning and letting, as I said, lower quality 217 00:18:18,080 --> 00:18:24,560 models do the job because that's just enough. And yeah, that's absolutely right 218 00:18:24,560 --> 00:18:29,640 with it. If you're smart with your prompt and you know how to prompt, well, 219 00:18:29,640 --> 00:18:33,440 you can kind of unlock the actual GPT-4. 220 00:18:33,480 --> 00:18:39,000 Yeah. So here's the no kidding hard costs that they have to pay in power 221 00:18:39,000 --> 00:18:43,800 and usage, you know, for a thousand tokens on GPT-4, the smaller GPT-4 222 00:18:43,800 --> 00:18:50,400 inference model is 0.0049 cents per a thousand tokens to basically process 223 00:18:50,400 --> 00:18:55,640 the data. Right. And since this is getting used so, so much at a time, 224 00:18:55,920 --> 00:19:00,800 Chad, you know, even OpenAI is processing like our individual queries one at a 225 00:19:00,800 --> 00:19:04,440 time, like, Hey, I'm going to write something in. It's going to go to, you 226 00:19:04,440 --> 00:19:08,800 know, to some server farm somewhere that is running GPT-4 and answer my 227 00:19:08,800 --> 00:19:13,280 question. It's literally all batched together second by second, everybody's 228 00:19:13,280 --> 00:19:18,720 inquiry, right. And sent to it at once. And these things are bulk decoded by 229 00:19:18,720 --> 00:19:25,160 these smaller inference models and executed by the big GPT-4 models. So 230 00:19:25,480 --> 00:19:30,640 that saves on their utilization that by keeping these batch sizes high, but I 231 00:19:30,640 --> 00:19:36,360 think when people, you know, use simpler prompts, when they're not asking more 232 00:19:36,360 --> 00:19:41,400 complex questions, it's way more easy to just offload those responses in simple 233 00:19:41,400 --> 00:19:46,840 ways, then continually push them in. Especially if that, you know, that model 234 00:19:46,840 --> 00:19:50,720 determines there's probably some threshold because every time the 235 00:19:50,920 --> 00:19:55,000 language model is responding where the transformer architecture works, it's 236 00:19:55,000 --> 00:19:59,720 calculating this like singularity value. And if it, it seeds this little value, 237 00:19:59,720 --> 00:20:04,680 which is the probably best probable response, if it probably doesn't meet 238 00:20:04,680 --> 00:20:09,440 certain threshold, GPT-4 is just having these smaller organization models 239 00:20:09,800 --> 00:20:13,720 inference them and respond. Why some people think, Hey, the quality has gone 240 00:20:13,720 --> 00:20:13,960 down. 241 00:20:13,960 --> 00:20:19,480 So if I go on Reddit and I see every post which says, Oh, quality is down. 242 00:20:19,600 --> 00:20:25,240 I'm not achieving the results I used to. I can assume that prompting in not 243 00:20:25,240 --> 00:20:26,800 exactly well made way. 244 00:20:26,920 --> 00:20:29,080 Their prompting probably sucks. 245 00:20:29,240 --> 00:20:31,120 This is like the vetting process. 246 00:20:31,120 --> 00:20:34,560 You know, it's just one of those things where, Hey, we have a podcast so we can 247 00:20:34,560 --> 00:20:37,440 hopefully get people to think about these things a little different and try 248 00:20:37,440 --> 00:20:42,440 to get a little more juice out of the GPT-4 orange by some good prompting 249 00:20:42,440 --> 00:20:46,880 engineering and some good little tips and tricks on how to talk to AI. 250 00:20:47,360 --> 00:20:54,160 It's kind of built this way that if you don't try and shape and put some 251 00:20:54,160 --> 00:20:57,560 thought and care and effort and maybe some technique into your things, 252 00:20:57,560 --> 00:21:02,840 you're passing into GPT-4, it's just going to do what it was designed to do 253 00:21:02,880 --> 00:21:05,640 and answer your question in the simplest, most likely way. 254 00:21:05,680 --> 00:21:11,480 I think it was from Albert Einstein saying that if I have a problem, I 255 00:21:11,480 --> 00:21:18,560 have a problem and I have one hour to solve it, I would spend 59 minutes 256 00:21:18,880 --> 00:21:23,640 thinking about how to solve it or how to frame the question and then actually 257 00:21:23,640 --> 00:21:26,080 do it, something between those lines. 258 00:21:26,600 --> 00:21:29,720 So this is like being very intentional. 259 00:21:29,720 --> 00:21:31,240 What are you trying to achieve? 260 00:21:31,680 --> 00:21:36,880 How do you communicate that to actually get you the results is important. 261 00:21:36,880 --> 00:21:42,440 But this is like, again, the subset of a population who really want to 262 00:21:42,440 --> 00:21:45,440 use AI in an extraordinary manner. 263 00:21:46,320 --> 00:21:52,120 However, the other 99% of people are just going to be happy and satisfied 264 00:21:52,200 --> 00:21:56,880 with results that they get, even with simple prompting or already 265 00:21:56,880 --> 00:22:00,680 pre-optimized prompting from their chatbot or organization, don't you 266 00:22:00,680 --> 00:22:04,560 think, but not everyone needs to have access to superpowers. 267 00:22:04,640 --> 00:22:05,000 Yeah. 268 00:22:05,040 --> 00:22:05,800 And with what? 269 00:22:05,800 --> 00:22:08,480 What's the line with great power comes great responsibility. 270 00:22:08,640 --> 00:22:12,400 If you're going to be able to reach things, well, you know, it's probably 271 00:22:12,400 --> 00:22:14,960 in your best interest to educate you and the best and the safest and 272 00:22:14,960 --> 00:22:16,880 the most ethical ways to do so. 273 00:22:16,920 --> 00:22:22,240 Now that we are back on a quote, writing that train, this is another 274 00:22:22,240 --> 00:22:28,040 part is that, you know, if you give to the child, super advanced computer 275 00:22:28,040 --> 00:22:33,280 and programming, and we don't even know how to use it, there is no point in that. 276 00:22:33,280 --> 00:22:33,640 Right. 277 00:22:33,640 --> 00:22:38,720 So this kind of like optimizing that, okay, for the lower quality prompts 278 00:22:38,720 --> 00:22:43,680 or tasks, this model is good enough and AI can determine that, of course. 279 00:22:44,320 --> 00:22:49,800 And for anything which is more advanced, okay, then we kind of, this is like 280 00:22:49,800 --> 00:22:50,960 qualification, right? 281 00:22:51,120 --> 00:22:56,920 Like where you pass the level to be able to actually communicate with GPT-4. 282 00:22:57,280 --> 00:22:59,960 This is also interesting looking in the future, right? 283 00:23:00,480 --> 00:23:03,280 There is going to be a group of people who are going to be 284 00:23:03,280 --> 00:23:08,160 people who have access to the best of the best or know how to access 285 00:23:08,280 --> 00:23:11,440 versus everyone else who have good enough. 286 00:23:11,480 --> 00:23:15,560 It could be the great unifier or the great divider, at least when the way I 287 00:23:15,560 --> 00:23:19,200 kind of think about it is it's all out there for you to kind of determine 288 00:23:19,200 --> 00:23:20,640 your own destiny, your own future. 289 00:23:21,200 --> 00:23:24,520 I wanted to talk about just, I wanted to mention the training cost. 290 00:23:24,840 --> 00:23:26,760 What are those numbers? 291 00:23:26,840 --> 00:23:31,920 That's the amount of processing power that it took to train GPT-4. 292 00:23:31,920 --> 00:23:34,880 So that's 2.15 times E to the 25th. 293 00:23:34,880 --> 00:23:38,720 So that means you take this little decimal point and move it 25 spaces to 294 00:23:38,720 --> 00:23:40,800 the right and then put zeros in the spaces. 295 00:23:40,880 --> 00:23:45,880 That's how many, many flops here was needed to train it. 296 00:23:46,360 --> 00:23:49,560 So they trained on 13 trillion tokens of data. 297 00:23:49,760 --> 00:23:51,840 They did it in six goes, right? 298 00:23:51,960 --> 00:23:55,440 Two goes did all the text data and then four were required for all of 299 00:23:55,440 --> 00:23:57,120 the code data trained on. 300 00:23:57,400 --> 00:23:57,760 All right. 301 00:23:57,760 --> 00:24:04,480 I think the most, I mean, just talk about a ordeal, 25,000 A100. 302 00:24:04,640 --> 00:24:10,080 These are the top end machine learning GPUs for 90 to a hundred days, but they're 303 00:24:10,080 --> 00:24:14,800 only able to use them at about 32 to 36% utilization. 304 00:24:15,080 --> 00:24:18,800 And that's because at all these checkpoints, they kept having different 305 00:24:18,800 --> 00:24:20,800 failures that needed to be restarted from. 306 00:24:21,120 --> 00:24:24,800 So that's where that million lines of code comes from. 307 00:24:24,800 --> 00:24:29,520 It's they're having to tune in different parameters and tweaks to even ingest 308 00:24:29,520 --> 00:24:32,600 all this data into a model because there's so many different conflicts. 309 00:24:33,040 --> 00:24:36,840 They said that speculating on technology that was available back 310 00:24:36,840 --> 00:24:41,640 when GPT-4 was training, it took 90 to a hundred days, probably cost 311 00:24:41,640 --> 00:24:46,840 about $63 million to train and that's just in server costs. 312 00:24:47,040 --> 00:24:51,280 So if you now put in all your architect, your, all your architects, all your 313 00:24:51,920 --> 00:24:53,760 designers, your coders, all your overhead. 314 00:24:53,760 --> 00:24:56,840 I mean, like that's, that was a hundred million bucks to train a model 315 00:24:56,840 --> 00:24:57,840 that can perform like that. 316 00:24:58,280 --> 00:25:01,560 Today, some of that could have probably been done a little more efficiently 317 00:25:01,560 --> 00:25:06,000 because they have these H100 GPUs now that Nvidia puts out probably going to 318 00:25:06,000 --> 00:25:08,560 even be less could have done in half the time, third of the budget. 319 00:25:09,040 --> 00:25:13,480 To me, that's fascinating to be like, wow, just the amount of, I mean, 320 00:25:13,480 --> 00:25:14,480 like where do you even begin? 321 00:25:14,480 --> 00:25:18,240 We like, oh, this is going to be a million lines of code to, to do this. 322 00:25:18,240 --> 00:25:22,080 Like imagine starting at line one, like, all right, let's keep going. 323 00:25:22,080 --> 00:25:26,400 It's insane. Like, you know what I would like to see someone or maybe we can even 324 00:25:26,400 --> 00:25:29,120 do a translating this in the energy. 325 00:25:29,200 --> 00:25:31,200 How much energy did this consume? 326 00:25:31,240 --> 00:25:36,800 For 90 to a hundred days, 63 million, like investment in training. 327 00:25:37,600 --> 00:25:40,440 How much that was also electricity bill. 328 00:25:41,160 --> 00:25:46,320 I would love to see visualization just like powering mega cities. 329 00:25:46,440 --> 00:25:49,040 That is the limitation right now. 330 00:25:49,040 --> 00:25:52,760 And I think it's also why, and we think we've mentioned it on here, Sam Altman, 331 00:25:52,760 --> 00:25:57,960 and there's a lot of AI, you know, experts and leaders that are very 332 00:25:57,960 --> 00:26:04,480 invested into renewable energies, fusion technology, because if it cost that 333 00:26:04,480 --> 00:26:12,680 much in just compute time, 63 million to train GPT-4 for a 1.8 trillion parameter 334 00:26:12,680 --> 00:26:16,920 model, if their true mission is AGI, artificial general intelligence, 335 00:26:16,920 --> 00:26:21,480 that thinks in reasons like a human does, well, our brain has a hundred trillion 336 00:26:21,560 --> 00:26:25,080 parameters, basically connections between neurons. 337 00:26:25,760 --> 00:26:29,200 So if we're just going to go apples to apples, if you've got to get a model 338 00:26:29,200 --> 00:26:37,440 that's 50, 60 times bigger than GPT-4 to maybe achieve AGI, a one, how do you 339 00:26:37,440 --> 00:26:41,360 think data to train that considering that they already identified that they 340 00:26:41,360 --> 00:26:45,800 were 13 trillion tokens short in training this, but they're not going to 341 00:26:45,800 --> 00:26:47,040 short in training this? 342 00:26:47,400 --> 00:26:51,520 One AI.com is kind of floating around. 343 00:26:51,560 --> 00:26:53,480 That's something interesting to look into. 344 00:26:53,800 --> 00:26:59,080 They claim that it's AI, which is trained on your internal documents. 345 00:26:59,080 --> 00:27:01,800 So it inherits the knowledge from your business. 346 00:27:02,040 --> 00:27:07,040 And basically they provide a bunch of different use cases, products, like 347 00:27:07,080 --> 00:27:13,760 custom skill development, biz, GPT, language analytics, multilingual AI, and so on. 348 00:27:13,760 --> 00:27:16,360 I'm very curious to try it. 349 00:27:16,360 --> 00:27:20,640 And I was thinking to book a demo definitely would be interesting to report on that. 350 00:27:20,720 --> 00:27:24,400 What I love about something like this, just looking at it with my little like, 351 00:27:24,400 --> 00:27:29,440 you know, AI hat from someone that builds things like this, this is essentially top 352 00:27:29,440 --> 00:27:35,280 layer fine tuning with a lot, a lot of spit and polish on every, every aspect. 353 00:27:35,640 --> 00:27:41,040 Folks can go out there with a handful of GitHub repos that are free as can be, do 354 00:27:41,040 --> 00:27:47,280 this themselves, but I think it's a very real thing when you have big businesses 355 00:27:47,280 --> 00:27:54,000 that are not AI experts or tech centric or whatever, they are more than happy to 356 00:27:54,000 --> 00:27:57,200 just pay big dollars to go you there. 357 00:27:57,400 --> 00:28:03,400 You have a shiny AI website, make me the thing where it can be chat GPT for my 358 00:28:03,400 --> 00:28:03,800 info. 359 00:28:04,120 --> 00:28:06,320 So kudos to them for taking advantage of it. 360 00:28:06,360 --> 00:28:09,680 I mean, that's, that's what we're trying to do on a different or smaller scale. 361 00:28:10,000 --> 00:28:10,520 Exactly. 362 00:28:10,520 --> 00:28:14,280 It's you take everything what's available and you package it in a beautiful 363 00:28:14,280 --> 00:28:16,640 website and then you go and sell it. 364 00:28:16,920 --> 00:28:21,200 What's interesting to me as an outcome, because I worked in big data company 365 00:28:21,680 --> 00:28:29,040 that you can bring anything like this, but if your data is a mess, if your 366 00:28:29,040 --> 00:28:33,880 systems is a mess, nothing speaking with each other, the next natural step is 367 00:28:34,080 --> 00:28:37,160 same company is going to be like, yeah, we used it. 368 00:28:37,360 --> 00:28:38,800 It's not that good. 369 00:28:38,800 --> 00:28:42,400 The response from this company should be like, oh yeah, of course, but we can 370 00:28:42,400 --> 00:28:46,400 help you with, you know, introducing machine learning, looking through your 371 00:28:46,400 --> 00:28:52,320 whole processes, optimization, and then we prepare your businesses for actually 372 00:28:52,320 --> 00:28:52,840 using it. 373 00:28:52,840 --> 00:28:58,400 So this is kind of double upselling and I think a lot of businesses is going to be 374 00:28:58,400 --> 00:28:59,400 experiencing that. 375 00:28:59,560 --> 00:29:05,200 And this is what I told you when I had conversation with the lady sitting to 376 00:29:05,200 --> 00:29:07,760 seats below Sasha from Microsoft. 377 00:29:07,760 --> 00:29:14,840 She will exactly talked about enterprise level that enterprises buy anything AI. 378 00:29:15,360 --> 00:29:20,840 90% of them don't have infrastructure to actually fully use it. 379 00:29:20,880 --> 00:29:27,560 And there's so much opportunity for not just consultancy, but how to integrate, 380 00:29:27,560 --> 00:29:32,200 how to make processes run because it's like putting a lipstick on a pig. 381 00:29:32,280 --> 00:29:35,880 What is the saying or garbage in garbage out? 382 00:29:35,880 --> 00:29:38,720 What I understood is also what she was saying. 383 00:29:38,960 --> 00:29:43,280 You have AI on your business data and that's basically the future. 384 00:29:43,280 --> 00:29:47,400 And another interesting thing when we talked about hallucinations and like, I 385 00:29:47,400 --> 00:29:49,520 was like importance of prompting. 386 00:29:49,800 --> 00:29:54,520 And she was like, these things were never meant for factual data. 387 00:29:54,520 --> 00:29:56,240 That's what you have Google for. 388 00:29:56,480 --> 00:30:01,440 And I was like, okay, for Microsoft, they've been, you know, on it for a while. 389 00:30:01,440 --> 00:30:07,080 And it was meant for internal how to optimize processes, PowerPoint things, 390 00:30:07,080 --> 00:30:13,560 summarizing meetings, all if you think about it, really like these LLMs solve 391 00:30:13,560 --> 00:30:19,640 those problems very much white collar day to day job problems, knowledge centric 392 00:30:19,640 --> 00:30:20,640 jobs, exactly. 393 00:30:20,640 --> 00:30:24,680 So, but those problems center around specific knowledge. 394 00:30:24,920 --> 00:30:31,080 So now of course, there is a part of, you know, interesting things that are 395 00:30:31,080 --> 00:30:32,080 happening in the world. 396 00:30:32,080 --> 00:30:36,920 So, you know, internet and training as a knowledge and testing and improving and 397 00:30:36,920 --> 00:30:37,440 all that. 398 00:30:37,680 --> 00:30:43,360 The point is that for her, this prompt engineering absolute drop in the ocean. 399 00:30:43,400 --> 00:30:48,480 She is on completely different level and scale of thinking. 400 00:30:48,800 --> 00:30:50,400 For her is enterprise. 401 00:30:50,440 --> 00:30:55,960 For her is like every enterprise, the clients of Microsoft enterprise, this is 402 00:30:55,960 --> 00:30:57,200 where real money is. 403 00:30:57,200 --> 00:31:02,000 For everything AI, every enterprise project, everything, what is not hardware. 404 00:31:02,000 --> 00:31:05,880 She is the head of globally selling these things. 405 00:31:06,240 --> 00:31:08,160 So she said everyone is buying it. 406 00:31:08,200 --> 00:31:09,080 Everyone wants it. 407 00:31:09,280 --> 00:31:14,080 90% don't know how to deal with this, what to do with this. 408 00:31:14,360 --> 00:31:16,960 And it's the same problem as always. 409 00:31:17,000 --> 00:31:22,400 If you have shitty processes inside your business, if already departments don't 410 00:31:22,400 --> 00:31:27,520 speak, data is completely unstructured in random places. 411 00:31:27,880 --> 00:31:30,920 You have to rethink the whole business infrastructure. 412 00:31:31,120 --> 00:31:36,880 And she was saying that implementing AI into enterprises, this is where money is. 413 00:31:37,080 --> 00:31:38,480 Oh, no doubt, without a doubt. 414 00:31:38,560 --> 00:31:38,880 Yeah. 415 00:31:39,000 --> 00:31:43,920 But you have to make some sacrifices to make things simpler, to make them bigger 416 00:31:43,920 --> 00:31:44,520 and broader. 417 00:31:44,680 --> 00:31:47,360 So I think that's why she probably doesn't think too highly about prompt 418 00:31:47,360 --> 00:31:47,920 engineering. 419 00:31:47,920 --> 00:31:52,200 And as a matter of fact, like I kind of, you know, the statement to say where 420 00:31:52,200 --> 00:31:55,600 these things were never meant for factual data. 421 00:31:55,920 --> 00:31:59,120 Well, no, they are, but you need alignment. 422 00:31:59,120 --> 00:32:00,240 You need super alignment. 423 00:32:00,240 --> 00:32:04,640 That's why opening AI is hiring a super alignment team to work through these 424 00:32:04,640 --> 00:32:05,040 things. 425 00:32:05,360 --> 00:32:09,680 It's a perfect example of, in this little article here on Pastebin that talks 426 00:32:09,680 --> 00:32:13,120 about everything that goes into GPT-4, right? 427 00:32:13,120 --> 00:32:18,640 There is over a million lines of code that is just for alignment related 428 00:32:18,640 --> 00:32:22,240 purposes to get it to reason the way it should. 429 00:32:22,280 --> 00:32:25,400 A million lines of code, you know, just for alignment. 430 00:32:25,440 --> 00:32:31,840 And that's for a model that has 1.8, as it says, 1.8 trillion parameters, 120 431 00:32:31,840 --> 00:32:32,280 layers. 432 00:32:32,440 --> 00:32:36,280 So a million lines of code to get it to, and what alignment means is thinks and 433 00:32:36,280 --> 00:32:37,600 reasons like a human. 434 00:32:37,920 --> 00:32:39,520 That's what they're going for. 435 00:32:39,520 --> 00:32:43,880 Just to say that her kind of stand for prompt engineering was just basically 436 00:32:43,880 --> 00:32:45,960 that's not something on enterprise level. 437 00:32:46,520 --> 00:32:49,800 That is a part of the whole bigger problem. 438 00:32:49,840 --> 00:32:54,480 It's maybe a feature that a whole team of people have put into a single button. 439 00:32:54,720 --> 00:32:55,760 This is very similar. 440 00:32:55,760 --> 00:33:00,160 What I was looking at Prom Perfect, their new feature prompt as service. 441 00:33:00,200 --> 00:33:00,840 Yeah, I love that. 442 00:33:00,880 --> 00:33:01,360 Exactly. 443 00:33:01,400 --> 00:33:03,680 This is what you place behind the button. 444 00:33:03,800 --> 00:33:07,280 You click it, everything is optimized for your specific use case. 445 00:33:07,280 --> 00:33:11,240 When we were drinking wine, she was celebrating that there is a reorg. 446 00:33:11,400 --> 00:33:16,760 She's moving up, can't disclose that actual title now, but it's insanity. 447 00:33:17,240 --> 00:33:19,240 She was kind of making us an analogy of it. 448 00:33:19,360 --> 00:33:25,800 Apple is going to make pretty shiny thing as their VR headset and everyone is 449 00:33:25,800 --> 00:33:26,760 going to want that. 450 00:33:26,800 --> 00:33:33,280 But again, if you compare money to be made from enterprise AI, where Microsoft 451 00:33:33,280 --> 00:33:39,200 is going 100% chips versus doing hardware VR headset and trying to compete with 452 00:33:39,200 --> 00:33:43,240 Apple on all the aspects, it's just something completely not focused. 453 00:33:43,280 --> 00:33:49,880 The whole VR experience or like augmentation of how we experience or 454 00:33:49,880 --> 00:33:55,720 interface with hardware, the final view, the final how we experience it is not 455 00:33:55,720 --> 00:34:00,760 going to be a headset sitting on your head, but to get there, that's one of the 456 00:34:00,760 --> 00:34:04,720 tools that we are going to use till we have something better technologies there. 457 00:34:04,920 --> 00:34:09,120 We talked about prompt engineering similarly, and we've been talking that 458 00:34:09,120 --> 00:34:12,800 this is going to be relevant for like two, one to two years. 459 00:34:12,880 --> 00:34:18,840 Immediately relevant for anybody to casually get the most out of AI. 460 00:34:19,080 --> 00:34:23,360 If you want to use AI at a high level now, you have to learn some prompt 461 00:34:23,360 --> 00:34:27,040 engineering to do that plunge different ways, pop over to learn prompting. 462 00:34:27,280 --> 00:34:28,680 Hey, we happen to have a course. 463 00:34:28,680 --> 00:34:33,240 On AI and chat GPT that we're launching here and with co-rise in August. 464 00:34:33,440 --> 00:34:34,640 That's another great way. 465 00:34:34,640 --> 00:34:37,200 There's tons of different resources, but you were right. 466 00:34:37,440 --> 00:34:41,840 I think the days of that being a requirement to interface with 467 00:34:41,840 --> 00:34:43,720 AI at a high level are numbered. 468 00:34:43,960 --> 00:34:49,360 I also gave her like my, what I shared with you was two visions of easy and 469 00:34:49,360 --> 00:34:54,520 effective and that it's 1% of people who are going to move to what is 470 00:34:54,520 --> 00:34:57,840 actually behind that button, how it's getting done, how it's getting done. 471 00:34:57,840 --> 00:35:02,360 And how it's getting optimized, how every organization has their own AI and 472 00:35:02,360 --> 00:35:07,800 prompts designed for that organization, maybe versus everyone else using it. 473 00:35:08,000 --> 00:35:10,960 But it's the same as it always been, right? 474 00:35:10,960 --> 00:35:12,800 It's the same pattern going forward. 475 00:35:13,080 --> 00:35:17,680 That 1% is amazing opportunity because this is also the reason why we 476 00:35:17,680 --> 00:35:22,520 called this how to talk to AI and not prompt engineering podcast. 477 00:35:22,520 --> 00:35:24,880 Yeah, we had no discussion. 478 00:35:25,960 --> 00:35:26,320 Right. 479 00:35:26,320 --> 00:35:32,480 So because how you talk with AI will be evolving to that crazy sci-fi levels. 480 00:35:32,920 --> 00:35:37,360 Where one day I can just grunt at a Microsoft PowerPoint document and 481 00:35:37,360 --> 00:35:39,160 it makes me my whole presentation. 482 00:35:39,560 --> 00:35:44,880 It was just fascinating to see at what scale she's thinking. 483 00:35:45,040 --> 00:35:49,960 It's completely different from anything I've seen. 484 00:35:50,160 --> 00:35:52,920 We are not talking millions, hundreds of millions. 485 00:35:52,920 --> 00:35:56,120 And if it's not hundreds of millions, then it's not interesting. 486 00:35:56,360 --> 00:35:58,160 So that's definitely something. 487 00:35:58,160 --> 00:36:02,280 And it was quite reassuring that the biggest companies are jumping on AI. 488 00:36:02,560 --> 00:36:07,400 She said 90% of them have no clue what to do, how to do it. 489 00:36:07,800 --> 00:36:10,120 So there's so much opportunity. 490 00:36:10,160 --> 00:36:13,640 Maybe they need an AI expert, a certified AI expert. 491 00:36:13,880 --> 00:36:20,680 This story of you and your journey, plus, you know, going from small $32 prompting job. 492 00:36:20,680 --> 00:36:22,720 Do you still know the client? 493 00:36:23,000 --> 00:36:24,400 We should bring them on the podcast. 494 00:36:24,440 --> 00:36:29,760 He was a marketing guy and he actually, he had a tool for internal tool that he had 495 00:36:29,760 --> 00:36:34,040 already built that would just basically like help his people respond to client 496 00:36:34,040 --> 00:36:39,960 inquiries way faster and just built three little prompts that, you know, did some 497 00:36:39,960 --> 00:36:44,840 specifics that he wanted from summarizing to email responses to different viewpoints. 498 00:36:45,160 --> 00:36:49,880 As you can see, I still remember it, but yeah, it was $96 and then it upworked 499 00:36:49,880 --> 00:36:52,200 you know, when you're a newbie, they take their cut. 500 00:36:52,520 --> 00:36:57,200 So they take a 20% on made all like 70, 74 bucks or something like that. 501 00:36:57,600 --> 00:37:00,440 But that was enough to get this train rolling. 502 00:37:00,800 --> 00:37:02,120 This is how things start. 503 00:37:02,120 --> 00:37:06,960 And nevertheless, this is absolutely inspiring and should be inspiring for anyone 504 00:37:06,960 --> 00:37:10,400 that it's not by any means too late. 505 00:37:10,440 --> 00:37:15,960 There is of course, first mover advantage, which you experienced, but the space is 506 00:37:15,960 --> 00:37:19,840 still so early, like this is going to be just like, you know, like a 507 00:37:19,840 --> 00:37:26,400 place growing and how starting with something smaller can lead you to almost 508 00:37:26,400 --> 00:37:28,000 working at enterprise level. 509 00:37:28,200 --> 00:37:32,440 This is just fascinating because this is where real problems and real money is 510 00:37:32,440 --> 00:37:33,160 going to be made. 511 00:37:33,320 --> 00:37:37,360 Is this place good as any, finish this podcast. 512 00:37:37,360 --> 00:37:43,000 This place is as good as any go to go for G for yourself and West the Synth Mind. 513 00:37:43,040 --> 00:37:45,640 That's me saying happy prompting everybody. 514 00:37:46,120 --> 00:37:48,200 Happy prompting everybody. 515 00:37:48,200 --> 00:37:53,280 Thanks for listening to How to Talk to AI with your hosts, 516 00:37:53,280 --> 00:37:56,000 Go2Go and Wes the Synth Mind. 517 00:37:56,000 --> 00:38:02,840 As always, you can check out the show notes and links at howtotalk2.ai. 518 00:38:02,840 --> 00:38:05,280 That's all for this week's episode. 519 00:38:05,280 --> 00:38:06,920 Happy prompting everyone. 520 00:38:06,920 --> 00:38:19,560 MUSIC