WEBVTT 1 00:00:00.080 --> 00:00:02.560 Do you know what AI actually stands for? 2 00:00:03.840 --> 00:00:07.680 It's Artificial Intern. And that's how 3 00:00:07.680 --> 00:00:11.520 you have to work with. It's an intern who strangely knows a lot 4 00:00:11.520 --> 00:00:14.960 of stuff. So they're straight out of university and they have 5 00:00:14.960 --> 00:00:18.560 studied pretty much everything, but they, 6 00:00:18.640 --> 00:00:22.360 they don't have any, any, any real world experience. They don't 7 00:00:22.360 --> 00:00:26.080 have the intuition that, that you have as a business person or a 8 00:00:26.160 --> 00:00:29.880 developer or whatever you do. They can do a lot of 9 00:00:29.880 --> 00:00:32.920 the repetitive work, a lot of the menial work of 10 00:00:33.720 --> 00:00:37.240 analyzing stuff, analyzing text, analyzing information, 11 00:00:37.720 --> 00:00:41.520 or writing templates, classifying something and 12 00:00:41.520 --> 00:00:45.320 kicking off backend algorithms. This is what they're really good at. 13 00:00:45.560 --> 00:00:49.080 But in most cases, if they make any 14 00:00:49.080 --> 00:00:52.600 decisions or produce any facts, 15 00:00:52.760 --> 00:00:55.450 you should always double check and make sure. 16 00:01:01.370 --> 00:01:04.410 Welcome to Startuprad IO, 17 00:01:04.970 --> 00:01:08.530 your podcast and YouTube blog covering the German 18 00:01:08.530 --> 00:01:12.010 startup scene. With news, interviews and 19 00:01:12.250 --> 00:01:13.530 live events, 20 00:01:15.850 --> 00:01:19.690 AWS is proud to sponsor this week's episode of startup raid 21 00:01:19.690 --> 00:01:23.440 IO. The AWS team compromises former 22 00:01:23.600 --> 00:01:26.800 founders, CTOs, venture capitalists, 23 00:01:27.040 --> 00:01:30.560 angel investors and mentors ready to help you prove 24 00:01:30.640 --> 00:01:34.440 what's possible. Since 2013, AWS has 25 00:01:34.440 --> 00:01:37.280 supported over 280,000 26 00:01:37.520 --> 00:01:41.360 startups across the globe and provided US$7 billion 27 00:01:42.640 --> 00:01:46.480 in credits through the AWS Activate program. 28 00:01:46.920 --> 00:01:50.520 Big Ideas Feel at home at AWS and with access 29 00:01:50.520 --> 00:01:54.320 to cutting edge technologies like generative AI, you can quickly 30 00:01:54.320 --> 00:01:58.120 turn those ideas into marketable products. Want your own 31 00:01:58.120 --> 00:02:01.920 AI powered assistant? Try Amazon Q. Want your own 32 00:02:01.920 --> 00:02:05.480 AI products? Privately customize leading foundation 33 00:02:05.480 --> 00:02:09.200 models on Amazon Bedrock. Want to reduce the cost 34 00:02:09.200 --> 00:02:13.000 of AI workloads? AWS Trainium is the silicon 35 00:02:13.000 --> 00:02:16.630 you're looking for. Whatever your ambitions, you've already had 36 00:02:16.630 --> 00:02:20.270 the idea. Now prove it's possible on AWS. 37 00:02:20.670 --> 00:02:23.470 Visit aws.Amazon.com 38 00:02:23.950 --> 00:02:27.550 startups to get started. Dennis Straub is a developer 39 00:02:27.550 --> 00:02:31.070 advocate at aws, where he guides companies through the safe 40 00:02:31.070 --> 00:02:34.790 adoption of emerging tech. With a deep background in cloud 41 00:02:34.790 --> 00:02:38.510 security, developer enablement and generative AI 42 00:02:38.510 --> 00:02:42.180 integration, Dennis helps teams test, iterate and 43 00:02:42.340 --> 00:02:46.020 learn without putting the data or business at risk. 44 00:02:46.340 --> 00:02:50.060 Today we unpack the AWS playbook for starting 45 00:02:50.060 --> 00:02:53.460 with gen AI. Even if you're just getting curious. 46 00:02:53.700 --> 00:02:57.220 Denis welcome to StartupRate IO and for every 47 00:02:57.540 --> 00:03:00.940 podcast aficionado, we may add that you have been the 48 00:03:00.940 --> 00:03:04.020 original voice of the German AWS podcast. 49 00:03:05.060 --> 00:03:08.620 Oh, thank you Joe. Thanks everyone for listening. I 50 00:03:08.620 --> 00:03:12.170 can't. I don't. Is it. Is it even still true with the AWS podcast? 51 00:03:12.240 --> 00:03:16.000 Podcast? It has been. That was during. That was during COVID I 52 00:03:16.000 --> 00:03:19.760 started that during COVID I put out I think 50 episodes or so 53 00:03:19.760 --> 00:03:23.400 until Traveling started up again and unfortunately I wasn't able to 54 00:03:23.400 --> 00:03:27.000 continue, but a few of my friends and colleagues here in Germany actually 55 00:03:27.000 --> 00:03:30.000 picked it up and are still continuing it. 56 00:03:30.720 --> 00:03:33.360 Anyway, thanks for having me on. On the show and 57 00:03:34.480 --> 00:03:38.240 right in the, in the introduction, you mentioned something I think that's really, 58 00:03:38.930 --> 00:03:42.370 really dear to my own heart and probably to most of your listeners. 59 00:03:43.090 --> 00:03:46.290 What's the ROI in 60 00:03:46.690 --> 00:03:50.530 AI? I think that's a question that many people have, including myself, 61 00:03:51.010 --> 00:03:54.210 quite often. So I'm happy to talk about this today. 62 00:03:56.130 --> 00:03:59.650 When people talk about AI, what comes to mind is 63 00:03:59.650 --> 00:04:03.370 ChatGPT doing everything with it, but 64 00:04:03.370 --> 00:04:07.060 it's a chat window. Plus what, 65 00:04:07.220 --> 00:04:10.740 what has been in the news on and off is Elon 66 00:04:10.740 --> 00:04:14.420 Musk's croc for either very great or very 67 00:04:14.420 --> 00:04:18.260 bad answers. So everybody who's only heard about 68 00:04:18.260 --> 00:04:22.100 that, how could you get started 69 00:04:22.660 --> 00:04:24.900 safely with Gen AI? 70 00:04:26.660 --> 00:04:30.020 Well, I think, most importantly, first of all, it's 71 00:04:30.260 --> 00:04:33.780 important to understand what generative 72 00:04:33.780 --> 00:04:37.580 AI actually is, how it works. Not in detail. 73 00:04:37.660 --> 00:04:41.460 You don't have to. I don't have a PhD in math. I don't really understand 74 00:04:41.460 --> 00:04:45.100 math. But you don't, you don't need to have that. But it's important 75 00:04:45.180 --> 00:04:49.020 to have a foundational understanding of how these models work 76 00:04:49.179 --> 00:04:52.940 and specifically what they are not. They are not people. 77 00:04:53.260 --> 00:04:56.940 They are not human beings, even though 78 00:04:57.340 --> 00:05:00.300 they talk like human beings. And 79 00:05:02.080 --> 00:05:05.880 Andre Karpathy, one of the, one of the people who do a lot of foundational 80 00:05:05.880 --> 00:05:07.600 work in AI, he actually said, 81 00:05:09.280 --> 00:05:12.640 LLMs are like stochastic 82 00:05:12.720 --> 00:05:16.400 simulations of people. So they behave like 83 00:05:16.400 --> 00:05:19.840 people in a certain way in terms of putting out text, saying something, 84 00:05:20.080 --> 00:05:23.800 but they behave like this friend that some of you may have 85 00:05:23.800 --> 00:05:27.560 had in the past. I certainly did. That person who knew 86 00:05:27.560 --> 00:05:31.280 every, everything. And when you, when you ask them anything, 87 00:05:31.600 --> 00:05:35.320 they, they would have an answer. And they were so convincing with what they 88 00:05:35.320 --> 00:05:38.960 said. But once you started questioning, you might have 89 00:05:38.960 --> 00:05:42.800 realized, well, maybe, maybe it's not 90 00:05:42.800 --> 00:05:46.360 really what they're saying, or maybe they are not as sure as 91 00:05:46.360 --> 00:05:50.160 they think they are. So I try to, I try 92 00:05:50.160 --> 00:05:54.000 to compare AI models, generative AI 93 00:05:54.000 --> 00:05:57.830 language models, compare them with this kind of friend who 94 00:05:58.310 --> 00:06:01.910 would like to know a lot and probably knows a lot as well, but sometimes 95 00:06:01.910 --> 00:06:05.430 confuses things and isn't really aware or 96 00:06:05.830 --> 00:06:09.670 doesn't want to, doesn't want to show any weakness and tries 97 00:06:09.670 --> 00:06:13.350 to bring across whatever they come up with as convincing 98 00:06:13.350 --> 00:06:16.470 as possible. And that's what's really important. AI 99 00:06:18.470 --> 00:06:22.230 does not know anything. AI has been trained 100 00:06:22.390 --> 00:06:26.190 on, on the entire Internet, basically on a lot of text 101 00:06:26.190 --> 00:06:29.750 Material and what they do internally is just whenever 102 00:06:29.750 --> 00:06:33.510 you type something, whenever you send something into the 103 00:06:33.510 --> 00:06:36.550 language model, it looks at what you wrote 104 00:06:37.350 --> 00:06:41.110 and then it compares it to what it has read in the past. 105 00:06:41.190 --> 00:06:44.710 And then it comes up with, well, when I had this sentence, 106 00:06:45.270 --> 00:06:48.710 most of the time the next word was this. 107 00:06:49.490 --> 00:06:53.050 So it learns during training, it learns to 108 00:06:53.050 --> 00:06:56.730 relate concepts with each other without actually 109 00:06:56.730 --> 00:07:00.450 understanding the concept. Take a cat and the word 110 00:07:00.450 --> 00:07:04.010 flurry. And an LLM sees these 111 00:07:04.010 --> 00:07:07.490 words together very often when being trained on the Internet. 112 00:07:08.130 --> 00:07:11.730 And then it knows with I'm doing air quotes here, it 113 00:07:11.730 --> 00:07:15.170 knows that cats and flurry somehow 114 00:07:15.760 --> 00:07:18.560 relate to each other and 115 00:07:20.080 --> 00:07:23.720 may create text that puts these two 116 00:07:23.720 --> 00:07:27.520 words together. This is very simplified, but that's effectively 117 00:07:27.520 --> 00:07:30.880 how it works. It does not understand anything. 118 00:07:31.200 --> 00:07:34.800 It is extremely well trained in terms of 119 00:07:34.800 --> 00:07:38.480 pattern recognition. And it repeats patterns 120 00:07:38.720 --> 00:07:42.080 that it originally saw. And many of these patterns 121 00:07:42.240 --> 00:07:44.770 have been scientific papers, 122 00:07:45.810 --> 00:07:49.610 lexical articles and all kinds of information where 123 00:07:49.610 --> 00:07:53.290 people convincingly describe what they are 124 00:07:53.290 --> 00:07:56.890 talking about because they are convinced. Because most of the time it's 125 00:07:56.890 --> 00:07:59.170 actually true. And the model just 126 00:07:59.970 --> 00:08:03.450 adapted this way of communicating. That's why it sounds 127 00:08:03.450 --> 00:08:07.130 convinced. It is hard to 128 00:08:07.130 --> 00:08:10.970 find any text on the Internet where somebody says, I don't know. This 129 00:08:10.970 --> 00:08:14.640 is why most models also do not respond with I don't 130 00:08:14.640 --> 00:08:18.360 know. They just come up with stuff. 131 00:08:19.000 --> 00:08:21.800 And that's what's really important. 132 00:08:22.360 --> 00:08:26.120 AI. AI models know 133 00:08:26.120 --> 00:08:29.640 a lot, but have often have a hard time to 134 00:08:29.640 --> 00:08:33.320 really put the things together in 135 00:08:33.320 --> 00:08:37.120 a way that's that it's really, that it's really factual. And that 136 00:08:37.120 --> 00:08:40.890 is something that you should be basically aware of. If you 137 00:08:40.890 --> 00:08:44.410 know that, then you can deal with it in a certain way, then you know, 138 00:08:44.410 --> 00:08:48.170 I shouldn't rely on it. It's not that they're not good enough 139 00:08:48.170 --> 00:08:51.810 yet. The way these models work, they 140 00:08:51.810 --> 00:08:54.850 will never have actual 141 00:08:54.850 --> 00:08:58.610 understanding of what they talk about. They will always 142 00:08:58.930 --> 00:09:02.450 be pattern recognition recognition 143 00:09:02.450 --> 00:09:06.250 algorithms. And if you understand that, you can work 144 00:09:06.250 --> 00:09:09.490 with them. Like I like to think about them as 145 00:09:11.170 --> 00:09:14.530 another thing is like, do you know what 146 00:09:14.610 --> 00:09:18.210 AI actually stands for? It's 147 00:09:18.290 --> 00:09:22.050 artificial intern. And that's how you have to 148 00:09:22.050 --> 00:09:25.690 work with. It's an intern who strangely knows a lot of 149 00:09:25.690 --> 00:09:29.410 stuff. So they're straight out of university and they have studied 150 00:09:29.570 --> 00:09:33.330 pretty much everything, but they don't have 151 00:09:33.810 --> 00:09:37.250 any real world experience. They don't have the intuition 152 00:09:37.490 --> 00:09:41.070 that you have as a business person or a developer or 153 00:09:41.070 --> 00:09:44.510 whatever you do. I have 154 00:09:44.670 --> 00:09:48.270 the exact example for this. For example, I'm using 155 00:09:48.510 --> 00:09:52.190 a lot of chatbots for 156 00:09:52.190 --> 00:09:56.030 many, many different functions and I've come to use them 157 00:09:56.510 --> 00:10:00.350 for evaluating pitches, guest pitches, for the very 158 00:10:00.350 --> 00:10:03.550 simple reason. I get up to 30 a week 159 00:10:03.870 --> 00:10:07.590 during summer, and during winter it can be 60 to 160 00:10:07.590 --> 00:10:11.230 100. That only makes sense for me to reply if it's 161 00:10:11.610 --> 00:10:14.810 template, if it's not AI generated. And 162 00:10:15.530 --> 00:10:19.290 so you mean people, people approaching you because they want to be on 163 00:10:19.290 --> 00:10:22.890 your podcast. Okay, exactly, exactly. 164 00:10:22.890 --> 00:10:25.770 And so basically, at first I was copying 165 00:10:26.410 --> 00:10:28.890 simply in the email and 166 00:10:30.330 --> 00:10:34.050 the AI of choice gave me back a potential reply, but then 167 00:10:34.050 --> 00:10:37.290 I told it, okay, I want you to first evaluate 168 00:10:37.880 --> 00:10:41.640 what this actually is, how likely is 169 00:10:41.640 --> 00:10:45.320 it that it's written by an AI, what percentage is 170 00:10:45.480 --> 00:10:48.440 by human, and then give me a pretty fair 171 00:10:48.840 --> 00:10:52.600 assessment of X, Y and Z and this and this and that. 172 00:10:53.479 --> 00:10:57.160 And then only when I know that there is less 173 00:10:57.160 --> 00:11:00.760 than 50% AI involved, I look into the 174 00:11:00.760 --> 00:11:04.440 email, tell the AI what to do, what kind of reply 175 00:11:04.440 --> 00:11:08.280 I want to send, and then it goes out. And that's the 176 00:11:08.280 --> 00:11:12.000 thing. One of the things that these language models are really good 177 00:11:12.000 --> 00:11:15.520 at is classifying text, because they have 178 00:11:15.520 --> 00:11:19.080 seen so much text during their training that it's very 179 00:11:19.080 --> 00:11:22.720 easy for them to classify text if you help them understand 180 00:11:24.560 --> 00:11:27.840 what the premises are, what the conditions are, what the 181 00:11:28.240 --> 00:11:32.000 requirements are that you're looking at so they can classify. What they're 182 00:11:32.000 --> 00:11:35.720 really bad with is coming up with facts because 183 00:11:35.960 --> 00:11:39.240 they do not understand the concept of facts. 184 00:11:39.480 --> 00:11:42.920 They just, they're internally, they're just numbers. 185 00:11:43.080 --> 00:11:46.600 And texts can be represented as numbers, which is why 186 00:11:46.600 --> 00:11:50.280 summarization, classification and similar tasks are very 187 00:11:50.280 --> 00:11:53.640 easy, but there is no actual understanding. 188 00:11:54.200 --> 00:11:57.080 And the use case you just described is, 189 00:11:59.320 --> 00:12:03.160 I would be interested in understanding how many, how many false positives you 190 00:12:03.160 --> 00:12:06.770 have in terms of how many supposedly AI generated 191 00:12:06.770 --> 00:12:10.050 pitches you unfortunately sort out. Because 192 00:12:10.930 --> 00:12:14.610 maybe the model or maybe the person actually writes like an 193 00:12:14.610 --> 00:12:18.250 AI and it's, it's getting, it's getting harder to 194 00:12:18.250 --> 00:12:21.490 actually distinguish between well trained 195 00:12:21.570 --> 00:12:24.530 AI language models and actual human beings. 196 00:12:26.130 --> 00:12:29.810 There are, there are some indicators right now 197 00:12:30.130 --> 00:12:33.890 they may be out of date by tomorrow because things evolve, and 198 00:12:33.890 --> 00:12:37.710 especially people who build systems based on AI, AI to 199 00:12:37.710 --> 00:12:41.510 actually obfuscate the fact, to 200 00:12:41.510 --> 00:12:45.270 hide the fact that they're using AI, they're working on this 201 00:12:45.270 --> 00:12:49.070 as well. But let's get back to the intern, the artificial intern, 202 00:12:50.430 --> 00:12:54.070 like every intern, especially if they know a lot, 203 00:12:54.070 --> 00:12:57.750 they're fresh from university, they are really excited about 204 00:12:57.750 --> 00:13:01.430 the job, and they are really excited about learning about what they 205 00:13:01.430 --> 00:13:05.230 can do about your industry, about your use case, your products, your customers. 206 00:13:05.310 --> 00:13:09.060 At the same time, they have no Intuition and they don't have any real 207 00:13:09.060 --> 00:13:12.460 world experience that they can reflect on 208 00:13:12.460 --> 00:13:16.300 whenever they approach a new task or have a new challenge to 209 00:13:16.380 --> 00:13:20.180 reflect on and say, well, I saw something similar sometime in the past 210 00:13:20.180 --> 00:13:23.580 and it worked this way. And they can start abstracting 211 00:13:23.820 --> 00:13:27.180 and mapping and matching and coming up with solutions. 212 00:13:27.500 --> 00:13:31.260 They just know their textbooks, they know what they studied, 213 00:13:31.340 --> 00:13:35.130 they don't have the intuition. So it's your job. If you work with, with 214 00:13:35.130 --> 00:13:38.850 an AI, it's your job. You can give them a lot 215 00:13:38.850 --> 00:13:42.610 of tasks, but you always have to double check. They can do 216 00:13:42.610 --> 00:13:46.170 a lot of the repetitive work, a lot of the menial work of 217 00:13:46.890 --> 00:13:50.410 analyzing stuff, analyzing text, analyzing information 218 00:13:50.970 --> 00:13:54.570 or writing templates, or kicking off, 219 00:13:54.650 --> 00:13:57.930 kicking off classifying something and kicking off backend 220 00:13:57.930 --> 00:14:01.270 algorithms. This is what they're really good at. But 221 00:14:02.390 --> 00:14:06.230 in most cases you should really have a look at what 222 00:14:06.230 --> 00:14:09.590 if, if they make any decisions or produce any 223 00:14:10.150 --> 00:14:13.750 facts, you should always double check and make sure. 224 00:14:15.510 --> 00:14:19.230 What I found is they're really, really good in for example, you 225 00:14:19.230 --> 00:14:22.990 give them like 10 bullet points and a lot 226 00:14:22.990 --> 00:14:26.790 of keywords and say, okay, write me a text from that. They're excellent. 227 00:14:26.790 --> 00:14:30.550 They, they spare. They saving me a hell lot of time by doing 228 00:14:30.550 --> 00:14:34.270 that. I see an email, I say, okay, I want this, I want this, 229 00:14:34.270 --> 00:14:37.870 I want that. Because when you get tired and English is not your 230 00:14:37.870 --> 00:14:41.350 native language at 10pm it gets really, really 231 00:14:41.430 --> 00:14:44.710 difficult to formulate a straight, easy to read 232 00:14:45.430 --> 00:14:48.630 email. And that's when it comes in quite handy. But 233 00:14:49.830 --> 00:14:52.550 I wouldn't necessarily hand over my 234 00:14:53.190 --> 00:14:57.030 mailbox to Gemini or ChatGPT or Claude. And that's 235 00:14:57.030 --> 00:15:00.430 the point. I just want to know. We cannot hand over 236 00:15:00.830 --> 00:15:04.670 the mailbox. So how do we get to really start 237 00:15:04.910 --> 00:15:08.590 safely as a company with AI? 238 00:15:10.270 --> 00:15:14.110 Most importantly, don't give any AI system the 239 00:15:14.110 --> 00:15:17.830 keys to the kingdom. And your mailbox, especially if 240 00:15:17.830 --> 00:15:21.510 you're a founder, most likely is your key to 241 00:15:21.510 --> 00:15:25.070 the Kingdom. There are many, there are 242 00:15:25.070 --> 00:15:28.750 many agentic orchestration systems out there 243 00:15:29.070 --> 00:15:32.110 that allow to just connect to your Gmail 244 00:15:32.590 --> 00:15:35.950 account or to Outlook, whatever, 245 00:15:36.590 --> 00:15:39.790 or your calendar to connect to your data sources. 246 00:15:42.830 --> 00:15:46.670 It's useful as long as 247 00:15:46.670 --> 00:15:49.550 you don't let it act on this information 248 00:15:51.460 --> 00:15:54.980 or at least not let it act without you double checking before it actually 249 00:15:54.980 --> 00:15:58.660 acts on this information because in fact it would 250 00:15:58.660 --> 00:16:02.460 be able to delete all your emails, 251 00:16:02.460 --> 00:16:05.060 including important information. 252 00:16:06.340 --> 00:16:09.860 It may be able to actually send an email to 253 00:16:09.860 --> 00:16:13.300 somebody and it may not be the mail that you want 254 00:16:14.020 --> 00:16:17.780 this person to get. So there's certain risks whenever you 255 00:16:17.780 --> 00:16:21.620 give an AI or anyone like the Intern. Would you give 256 00:16:21.620 --> 00:16:25.280 your intern access to your, to your mail account? 257 00:16:26.160 --> 00:16:29.880 I just had in mind, as when he said, 258 00:16:29.880 --> 00:16:33.600 for example, my wife has a pretty common first name, and 259 00:16:33.680 --> 00:16:37.520 instead of sending her an invitation to date, I send it to 260 00:16:37.520 --> 00:16:41.200 a potential client. The AI could 261 00:16:41.200 --> 00:16:44.960 do that because it's the same first name. Right, right, right, 262 00:16:44.960 --> 00:16:47.840 the AI could do that. But on the other hand, there's another risk, 263 00:16:49.050 --> 00:16:52.650 and that's actually data, data privacy and security. 264 00:16:53.530 --> 00:16:57.170 Because if 265 00:16:57.170 --> 00:17:00.930 you give a language model access to your 266 00:17:00.930 --> 00:17:04.650 email account, it has access to 267 00:17:05.050 --> 00:17:08.330 all private information that's inside this email account. This 268 00:17:08.330 --> 00:17:11.930 includes private information about you, but it may also 269 00:17:12.010 --> 00:17:15.330 include pii, personally identifiable information of 270 00:17:15.330 --> 00:17:18.739 customers of other people. It may even include 271 00:17:21.859 --> 00:17:25.539 specific sensitive information like health information or, 272 00:17:26.499 --> 00:17:30.019 or like, like relationships between a lawyer and 273 00:17:30.019 --> 00:17:33.739 client. And there are certain pieces of information that, that, at least in 274 00:17:33.739 --> 00:17:37.499 Germany, there's even more protection 275 00:17:37.499 --> 00:17:41.139 around it. And you can actually go to jail if you, if you 276 00:17:41.139 --> 00:17:44.860 expose your client's health information 277 00:17:44.940 --> 00:17:48.780 or your customer's health information, you can go to jail for that. And it's 278 00:17:48.780 --> 00:17:52.180 you, it's not your intent, intern who goes, it's you who goes to jail for 279 00:17:52.180 --> 00:17:56.020 it. So what's, what's really important to acknowledge 280 00:17:56.020 --> 00:17:59.580 is the fact that in your emails, there's a lot of personal 281 00:17:59.740 --> 00:18:03.460 information, whether it's sensitive or not, 282 00:18:03.460 --> 00:18:07.260 it is personal information. And as soon as 283 00:18:07.260 --> 00:18:10.940 you send it to a model, you need 284 00:18:11.100 --> 00:18:14.860 to be aware of how the provider of this 285 00:18:14.860 --> 00:18:18.620 model interacts with that data, 286 00:18:19.020 --> 00:18:22.700 whether they store it anywhere, whether they send 287 00:18:22.700 --> 00:18:25.899 it somewhere else, whether they keep it secure, 288 00:18:27.260 --> 00:18:29.980 whether they just throw it away and not store it at all. 289 00:18:30.940 --> 00:18:34.780 This information is really important. So as soon as you give a language model 290 00:18:34.780 --> 00:18:38.060 access or an agent access to your mailbox, 291 00:18:38.960 --> 00:18:42.800 you give the model provider access 292 00:18:43.360 --> 00:18:46.800 to personal information. You give your model provider 293 00:18:49.440 --> 00:18:52.800 technically access to all the private information 294 00:18:53.040 --> 00:18:56.840 that your customers, that your family, that you 295 00:18:56.840 --> 00:19:00.160 yourself have entrusted your mailbox with. And 296 00:19:01.120 --> 00:19:04.680 I mean, even if your customers wouldn't care, it would 297 00:19:04.680 --> 00:19:08.510 be a GDPR headache because as soon as you sent that to 298 00:19:08.510 --> 00:19:12.230 OpenAI or any other public provider 299 00:19:12.230 --> 00:19:15.190 that provides like a public interface, 300 00:19:16.230 --> 00:19:19.670 maybe even without any price tag, 301 00:19:20.470 --> 00:19:24.230 you send that information somewhere else. And you as a company 302 00:19:25.990 --> 00:19:29.790 have no information about what happened to this 303 00:19:29.790 --> 00:19:33.620 data. You lose control over this data. And, and 304 00:19:33.620 --> 00:19:36.020 by losing control over the data, you 305 00:19:37.140 --> 00:19:40.660 effectively violate GDPR and maybe 306 00:19:40.740 --> 00:19:44.380 other laws too. So it's really important to understand that 307 00:19:44.380 --> 00:19:47.460 if you build an AI system 308 00:19:48.180 --> 00:19:51.140 that has access to your MA box or any other 309 00:19:51.300 --> 00:19:54.180 proprietary or private information, 310 00:19:54.980 --> 00:19:58.830 you need to make sure that you understand the terms and conditions of, 311 00:19:58.900 --> 00:20:02.180 of the model provider that you're working with. You need to understand 312 00:20:02.660 --> 00:20:06.380 how, how, what do they do with your data? Do they, do they use 313 00:20:06.380 --> 00:20:10.140 it for model training? Do they store it somewhere to 314 00:20:10.140 --> 00:20:13.780 make it available to authorities when they 315 00:20:13.780 --> 00:20:15.940 ask for it, for instance, or 316 00:20:17.220 --> 00:20:20.980 maybe even outside the European Union? This is a very important piece of information, 317 00:20:21.220 --> 00:20:24.740 especially if you work with publicly available APIs 318 00:20:25.860 --> 00:20:29.700 and even more so if you work with APIs or 319 00:20:30.180 --> 00:20:33.140 providers that don't charge you. 320 00:20:34.020 --> 00:20:37.220 I mean, it should be a fairly well known 321 00:20:38.100 --> 00:20:41.940 fact nowadays that if you don't pay with money, you usually 322 00:20:41.940 --> 00:20:45.540 pay with data, especially with services on the Internet. And 323 00:20:45.540 --> 00:20:48.980 that's most likely what happens with 324 00:20:49.700 --> 00:20:53.290 many providers. I'm not going to name any 325 00:20:53.290 --> 00:20:57.130 names. It is up to you to have a look at the 326 00:20:57.130 --> 00:21:00.770 actual conditions and there are ways to work around that. 327 00:21:01.890 --> 00:21:05.010 Most major model providers provide 328 00:21:05.730 --> 00:21:09.490 ways to use their models that are GDPR compliant 329 00:21:09.890 --> 00:21:13.730 or that allow you to use these models. GDPR 330 00:21:13.730 --> 00:21:17.570 compliant? The models themselves are never. Or a tool that you use is 331 00:21:18.010 --> 00:21:21.730 never GDPR compliant by itself. It's always about the way 332 00:21:21.730 --> 00:21:25.290 that you use it. But most commercial model providers 333 00:21:25.690 --> 00:21:29.370 actually have options that allow you to build 334 00:21:29.370 --> 00:21:32.810 GDPR compliance systems with their models. But 335 00:21:33.690 --> 00:21:37.530 it's usually not their chat interface. That is 336 00:21:37.530 --> 00:21:40.250 exactly because what I was going for, because 337 00:21:41.370 --> 00:21:44.810 why wouldn't you start with a chatbot? And how, 338 00:21:45.860 --> 00:21:49.660 how would you look in a company, like on a meta level for 339 00:21:49.660 --> 00:21:53.380 the first real project they can 340 00:21:53.380 --> 00:21:57.060 use, they can do with AI. And I 341 00:21:57.060 --> 00:22:00.740 have to admit that made me a little bit nervous because 342 00:22:00.900 --> 00:22:04.620 currently I have somebody coding an AI based chatbot from a 343 00:22:04.620 --> 00:22:08.180 website here. So there's two 344 00:22:08.180 --> 00:22:11.460 questions. The first question why I wouldn't use a chat, 345 00:22:11.780 --> 00:22:15.490 why I wouldn't build a chatbot. And the 346 00:22:15.490 --> 00:22:19.330 second question is what I would look at when I would go 347 00:22:19.330 --> 00:22:23.090 into a company or I would think about a use case that would actually make 348 00:22:23.090 --> 00:22:26.730 sense. And that's an interesting question. First of all, I would not 349 00:22:26.730 --> 00:22:30.410 necessarily say I wouldn't build a chatbot. A chatbot can be a 350 00:22:30.410 --> 00:22:33.690 good use case, but most of the time it isn't. 351 00:22:35.770 --> 00:22:39.510 The thing is, let me step back just one a little 352 00:22:39.510 --> 00:22:43.190 bit. We, as 353 00:22:43.190 --> 00:22:46.830 in everybody who's trying to use AI or trying to figure 354 00:22:46.830 --> 00:22:49.990 out how to use AI in a useful way, are making, 355 00:22:51.190 --> 00:22:54.910 we're falling in a certain trap and it's 356 00:22:54.910 --> 00:22:58.110 completely normal to do that. We are trying to solve 357 00:22:58.110 --> 00:23:01.750 problems that we have. 358 00:23:03.190 --> 00:23:05.510 Most of them we have already solved 359 00:23:07.430 --> 00:23:11.190 or the problems are inside of what we can 360 00:23:11.190 --> 00:23:14.990 imagine fairly easily. And that reminds Me of back in the 361 00:23:14.990 --> 00:23:18.790 day, back in the 90s when the world Wide Web became a thing. 362 00:23:20.710 --> 00:23:24.110 I don't know if you had been around, I mean, Joe, you probably have been 363 00:23:24.110 --> 00:23:27.550 around. I have been around. I don't know about the listeners, but if you have 364 00:23:27.550 --> 00:23:31.110 been around. Back in the day, when we started building 365 00:23:31.350 --> 00:23:35.070 websites or web pages or home pages as we called 366 00:23:35.070 --> 00:23:38.800 them back then, back then we were mostly trying to 367 00:23:38.800 --> 00:23:41.280 replicate what we already knew. So 368 00:23:42.560 --> 00:23:46.240 we had yellow pages on the Internet where you could 369 00:23:46.240 --> 00:23:49.480 find web pages like yellow pages with 370 00:23:49.480 --> 00:23:52.560 classification systems that were based on 371 00:23:52.879 --> 00:23:56.720 traditional libraries. Everybody had a 372 00:23:56.720 --> 00:24:00.240 homepage which was more or less a business card. 373 00:24:00.720 --> 00:24:04.070 And we tried to replicate print material 374 00:24:04.470 --> 00:24:07.750 onto the screen, which was really hard because most MySpace, 375 00:24:08.790 --> 00:24:12.310 even before that, even before that, most screens only had 800 by 376 00:24:12.310 --> 00:24:16.110 600 pixels. Early HTML didn't really allow you 377 00:24:16.110 --> 00:24:19.870 to position stuff. It's still hard nowadays, but back then it just didn't 378 00:24:19.870 --> 00:24:23.070 work. The lineup lines, so the 379 00:24:23.070 --> 00:24:26.470 connections to the Internet were so slow that you couldn't really 380 00:24:26.870 --> 00:24:30.680 use images. And it was really hard. So everybody said, 381 00:24:30.680 --> 00:24:34.360 every company said, well, we know we need to be on the 382 00:24:34.360 --> 00:24:38.200 Internet now, just like they say nowadays, we need to use AI, 383 00:24:38.360 --> 00:24:42.080 but it doesn't really work. And where's the 384 00:24:42.080 --> 00:24:45.640 ROI in this? Where's the ROI in 385 00:24:45.640 --> 00:24:49.280 putting my brochure, my print brochure onto the 386 00:24:49.280 --> 00:24:52.840 screen? And it's terribly slow for, for customers to 387 00:24:52.840 --> 00:24:56.600 load and display, but. Or what sense does 388 00:24:56.600 --> 00:25:00.400 it make? Do you know what came to mind when you 389 00:25:00.400 --> 00:25:04.040 talked about terribly slow? The sound of a dial up connection. 390 00:25:04.200 --> 00:25:07.560 Right, exactly, 391 00:25:07.640 --> 00:25:11.200 exactly. And that's the thing. We tried to 392 00:25:11.200 --> 00:25:14.960 solve traditional problems with this new tool with this 393 00:25:14.960 --> 00:25:17.960 new technology. We tried to use 394 00:25:19.160 --> 00:25:22.640 traditional means and just map them onto the 395 00:25:22.640 --> 00:25:26.180 screen. And that didn't work because the screen is not 396 00:25:26.180 --> 00:25:29.700 made for printed stuff. The screen is 397 00:25:29.700 --> 00:25:33.220 not a thick yellow pages, a tome 398 00:25:33.220 --> 00:25:36.860 of addresses, of phone numbers for businesses. 399 00:25:37.500 --> 00:25:41.260 That's not what it is. And over time, more and more people, and took 400 00:25:41.260 --> 00:25:45.060 a few years, more and more people started realizing there's a 401 00:25:45.060 --> 00:25:48.620 completely new way of thinking about things. And that's where 402 00:25:48.700 --> 00:25:52.540 Google started. And Google started replacing traditional yellow 403 00:25:52.540 --> 00:25:55.730 pages. And that's when Amazon started, 404 00:25:55.810 --> 00:25:59.610 Amazon started replacing traditional brochures 405 00:25:59.610 --> 00:26:03.170 on the Internet where you could read what you could buy and then 406 00:26:03.650 --> 00:26:07.090 pick up the phone or send an email to the 407 00:26:07.090 --> 00:26:10.689 retailer to tell them, well, I want this as a mail order. 408 00:26:11.490 --> 00:26:15.250 That's when ebay came around, when Wikipedia or Wiki 409 00:26:15.250 --> 00:26:19.050 as a principle started to come around, when things like Facebook and 410 00:26:19.050 --> 00:26:22.900 Twitter came around, when we started to embrace 411 00:26:22.980 --> 00:26:26.780 the new medium and actually use it for 412 00:26:26.780 --> 00:26:30.500 things that we couldn't even imagine before. Traditional 413 00:26:30.500 --> 00:26:34.260 classification systems, like in libraries, they are 414 00:26:34.260 --> 00:26:37.940 important in libraries because they have to deal with shelf space. 415 00:26:38.260 --> 00:26:41.980 They have to put the book somewhere and they cannot 416 00:26:41.980 --> 00:26:45.300 put the book everywhere. But with the Internet, with 417 00:26:45.300 --> 00:26:48.580 hyper hyper data, with hypermedia, 418 00:26:49.240 --> 00:26:52.760 you can put the book literally everywhere. 419 00:26:52.920 --> 00:26:56.720 You don't need classification systems, or you can have adaptive 420 00:26:56.720 --> 00:27:00.440 classification systems, you can have dynamic classification systems. 421 00:27:00.520 --> 00:27:04.360 You can even create something that looks at how many 422 00:27:04.360 --> 00:27:08.120 people actually read your book and cited from 423 00:27:08.120 --> 00:27:11.840 it, which effectively is Google. How many people actually 424 00:27:11.840 --> 00:27:15.320 visit your homepage, your website, your application, 425 00:27:15.560 --> 00:27:19.110 and actually link to it from their page. That is what Google looks at. 426 00:27:20.780 --> 00:27:24.380 And it's the same situation all over 427 00:27:24.380 --> 00:27:27.820 again, in my opinion. With AI, we're still trying to solve 428 00:27:28.140 --> 00:27:31.860 old problems with the new tool, and we 429 00:27:31.860 --> 00:27:33.260 haven't really figured out 430 00:27:37.100 --> 00:27:40.740 what's the exciting new thing that we 431 00:27:40.740 --> 00:27:42.940 can build with this tool that was 432 00:27:44.220 --> 00:27:48.000 prohibitively costly in terms of money or in terms 433 00:27:48.000 --> 00:27:51.440 of time, so that we didn't even think about doing it. 434 00:27:51.440 --> 00:27:54.720 Imagine back in the day before we had the Internet and mobile phones. 435 00:27:55.360 --> 00:27:59.160 Imagine back in the day. You live in Germany. I 436 00:27:59.160 --> 00:28:02.960 do. When I called my relatives in the US or when my family 437 00:28:02.960 --> 00:28:06.480 called relatives in the US that happened once a year 438 00:28:06.720 --> 00:28:10.280 on Christmas. And every family members had about 439 00:28:10.280 --> 00:28:14.000 five seconds to talk to them because it was an 440 00:28:15.360 --> 00:28:19.160 intercontinental call, which was so expensive. So we never 441 00:28:19.160 --> 00:28:22.920 talked to our relatives except on Christmas when I 442 00:28:22.920 --> 00:28:26.640 went on vacation and I sent a postcard back 443 00:28:26.640 --> 00:28:30.400 home. Most of the time the postcard arrived two weeks after 444 00:28:30.400 --> 00:28:34.160 I arrived back home. And right 445 00:28:34.160 --> 00:28:37.760 now, with mobile devices, with the Internet and everything, 446 00:28:39.060 --> 00:28:41.700 we talk to people all over the planet all the time. 447 00:28:42.500 --> 00:28:45.940 Yes. I vividly remember what a revelation it was 448 00:28:46.260 --> 00:28:49.860 when I was studying in the US or working in China, when I could 449 00:28:49.860 --> 00:28:53.620 use Skype to call people for, for 450 00:28:53.620 --> 00:28:57.060 local, for local rates. So 451 00:28:57.620 --> 00:29:01.220 how could a company really identify 452 00:29:02.180 --> 00:29:06.020 what should be the first project? Because I wouldn't 453 00:29:06.020 --> 00:29:09.500 necessarily recommend to have like this really big 454 00:29:09.500 --> 00:29:13.340 hairy goal for the first AI 455 00:29:13.340 --> 00:29:16.060 project, but rather something small, something 456 00:29:17.020 --> 00:29:20.780 that really makes sense, takes a lot of maybe repetitive 457 00:29:20.780 --> 00:29:24.620 work out of the job of the employees. Right? And 458 00:29:24.780 --> 00:29:26.620 that's the most important question. 459 00:29:30.380 --> 00:29:34.180 What use case? What workflow? What 460 00:29:34.180 --> 00:29:37.900 item on your to do list gets 461 00:29:37.900 --> 00:29:38.700 never done? 462 00:29:41.580 --> 00:29:45.020 What are the painful things in your business that 463 00:29:45.580 --> 00:29:49.100 nobody ever took care about because it 464 00:29:49.100 --> 00:29:52.860 would take too much time, or because it would take too many 465 00:29:52.860 --> 00:29:56.660 people to work on it, or because it would be just too 466 00:29:56.660 --> 00:30:00.340 costly to do it? What are 467 00:30:00.340 --> 00:30:03.190 the things that 468 00:30:04.710 --> 00:30:08.550 if you had a magic wand and you could make them go away. 469 00:30:08.710 --> 00:30:12.470 So the daily tasks, the menial things, the things 470 00:30:12.470 --> 00:30:16.150 that bother you all the time, but 471 00:30:16.390 --> 00:30:19.990 they need to be done or they should be done, 472 00:30:20.790 --> 00:30:24.310 but I don't get around to doing them because I have so much to do. 473 00:30:24.470 --> 00:30:28.260 What is that thing that you would like to be to get 474 00:30:28.340 --> 00:30:32.140 done and it never gets done because there's no time for 475 00:30:32.140 --> 00:30:35.460 it. Make a list of these things, write down your 476 00:30:35.620 --> 00:30:39.300 most painful things that you have to deal with every day 477 00:30:39.300 --> 00:30:43.140 because they don't get done and they don't get done. And then have a 478 00:30:43.140 --> 00:30:46.580 look at it at them and think about, is this something 479 00:30:47.700 --> 00:30:50.500 I could hand over to an AI 480 00:30:51.780 --> 00:30:55.460 safely? Hand over to an AI, right, safely. Maybe not 481 00:30:55.460 --> 00:30:58.970 in its completeness, maybe just a small part of it. 482 00:30:59.210 --> 00:31:02.810 And if you want to build a startup for a certain industry, 483 00:31:03.690 --> 00:31:07.330 talk to the people in this industry, talk to them, ask them 484 00:31:07.330 --> 00:31:11.010 this, this question. What is the one 485 00:31:11.010 --> 00:31:14.730 thing that has been bothering you for the last 486 00:31:14.890 --> 00:31:18.650 30 years since you started in this industry? What is the 487 00:31:18.650 --> 00:31:22.330 one thing that's bothering you 488 00:31:22.490 --> 00:31:24.780 but nobody ever took care of it? 489 00:31:26.450 --> 00:31:29.170 And then think about 490 00:31:30.770 --> 00:31:34.530 whether this could be something that you could hand over 491 00:31:34.850 --> 00:31:38.690 either in part or maybe even completely to an AI 492 00:31:38.930 --> 00:31:42.730 and well, of course, don't start with the big hairy 493 00:31:42.730 --> 00:31:46.210 goal, don't start with the big thing. Try to find a small 494 00:31:46.290 --> 00:31:49.810 painful thing, create a solution for it 495 00:31:51.420 --> 00:31:55.140 and then go back to the customer, go back to the market, go back to 496 00:31:55.140 --> 00:31:58.900 the person you talk to in the industry or go back to yourself if it's 497 00:31:58.900 --> 00:32:01.980 for yourself and see does it actually. 498 00:32:02.620 --> 00:32:06.299 Do you know, Dennis, what my consultant mind was making of what 499 00:32:06.299 --> 00:32:10.140 you say? Basically you put a lot of your employees into 500 00:32:10.140 --> 00:32:13.820 brainstorming session, they come up with 20 problems, 501 00:32:13.820 --> 00:32:16.620 you cut it down to 10 problems that are really 502 00:32:17.610 --> 00:32:21.410 efficient if you could automate or partially automate them. And 503 00:32:21.410 --> 00:32:24.970 then you start with the easiest. Yes. 504 00:32:25.050 --> 00:32:28.730 And it's not only about they would be most 505 00:32:28.730 --> 00:32:32.490 efficient, but we could finally address 506 00:32:32.570 --> 00:32:36.170 them through automation. It wasn't possible before. 507 00:32:36.570 --> 00:32:40.410 That's the thing. If we try to address problems that we already 508 00:32:40.410 --> 00:32:43.790 automate and we could make them more efficient and that's not 509 00:32:43.790 --> 00:32:47.590 innovation, that's optimization. That's a good thing. I'm 510 00:32:47.590 --> 00:32:51.110 not saying we shouldn't optimize, we should optimize, but that's not 511 00:32:51.110 --> 00:32:54.750 innovation, that's not the breakthrough, that's not the next Google, 512 00:32:55.390 --> 00:32:59.230 that's not the next unicorn startup. The next unicorn startup 513 00:32:59.310 --> 00:33:02.950 will solve a problem that everybody has, but 514 00:33:02.950 --> 00:33:06.590 nobody even knew that they had it or nobody 515 00:33:06.590 --> 00:33:10.080 even thought of solving it because actually solving them, it 516 00:33:10.320 --> 00:33:13.880 wasn't even possible before. And this can be a small 517 00:33:13.880 --> 00:33:16.640 thing. This can be a tiny, small thing. 518 00:33:17.920 --> 00:33:21.640 It doesn't need to be a big thing. It can really be a small, tiny 519 00:33:21.640 --> 00:33:25.360 thing. And if you have something like this, you effectively 520 00:33:25.360 --> 00:33:29.080 have a money printing machine because everybody's going to tell, whoa, I didn't 521 00:33:29.080 --> 00:33:32.920 know that's possible. Right. I don't have the 522 00:33:32.920 --> 00:33:36.600 solution for you. So I cannot tell you it's this or that 523 00:33:36.600 --> 00:33:40.360 thing. That is something that you need to look into with your specific 524 00:33:40.360 --> 00:33:43.900 expertise, with your intuition, with your background, 525 00:33:44.140 --> 00:33:47.780 with your creativity. Creativity. But what we've been 526 00:33:47.780 --> 00:33:51.620 doing most of the time with AI in the last two 527 00:33:51.620 --> 00:33:55.380 or three years really was just trying to, trying to, trying 528 00:33:55.380 --> 00:33:59.100 to use AI to solve problem that we are already solving 529 00:33:59.340 --> 00:34:03.020 and making them more efficient, making them least 530 00:34:03.340 --> 00:34:07.140 less costly, reducing cost, unfortunately, firing people 531 00:34:07.140 --> 00:34:10.729 and replacing them with AI to then figure out, 532 00:34:11.929 --> 00:34:15.609 well, maybe. That was needed those people. Right. 533 00:34:15.609 --> 00:34:19.409 Maybe it was the best idea. Maybe we shouldn't have listened to the 534 00:34:19.409 --> 00:34:22.489 promises of AI now replacing everyone. 535 00:34:23.449 --> 00:34:27.129 AI is something that can help you solve new problems 536 00:34:27.849 --> 00:34:31.489 and AI shouldn't be used to solve a problem that's 537 00:34:31.489 --> 00:34:34.649 already been solved unless you are in the 538 00:34:34.649 --> 00:34:38.369 optimization stage. Big enterprises may be in that stage and 539 00:34:38.369 --> 00:34:41.790 big enterprises may be doing the right thing when they're looking at their 540 00:34:41.790 --> 00:34:45.630 processes and workflows and everything and think about well, where are the 541 00:34:45.630 --> 00:34:49.230 bottlenecks? Can we apply this to individual bottlenecks in 542 00:34:49.230 --> 00:34:52.710 here to make the overall process more efficient or more 543 00:34:52.710 --> 00:34:56.550 scalable or whatever. But especially in the startup space, 544 00:34:59.110 --> 00:35:02.670 you want to innovate and innovation. Innovation is creating something 545 00:35:02.670 --> 00:35:06.430 new or solving a problem that everybody 546 00:35:06.430 --> 00:35:10.110 thought was not solvable or didn't even think 547 00:35:10.110 --> 00:35:13.510 about solving because it didn't. Yeah, we didn't. 548 00:35:13.750 --> 00:35:17.430 We didn't think about calling our relatives in the US every 549 00:35:17.430 --> 00:35:21.190 day because it just wasn't possible. It was 550 00:35:21.190 --> 00:35:25.030 too expensive. Right, I see. 551 00:35:25.830 --> 00:35:29.310 I would be wondering what for our 552 00:35:29.310 --> 00:35:32.750 audience would be their first gen AI use case 553 00:35:32.750 --> 00:35:36.310 idea and what's holding them back from trying it. To top 554 00:35:36.310 --> 00:35:40.110 you, drop your comment or DM us on LinkedIn. 555 00:35:40.270 --> 00:35:42.510 We'll be back after a very short ad break. 556 00:35:48.590 --> 00:35:51.390 So let's talk a little bit more 557 00:35:52.270 --> 00:35:55.630 specific about the problems here. 558 00:35:57.310 --> 00:36:00.710 What do you think is the biggest misconception of non 559 00:36:00.710 --> 00:36:04.250 technical founders they have about AI, especially around 560 00:36:04.250 --> 00:36:05.570 model choice and privacy? 561 00:36:07.730 --> 00:36:11.490 There's two misconceptions, one I mentioned earlier that 562 00:36:11.490 --> 00:36:14.770 is that we mistake these things for 563 00:36:15.010 --> 00:36:18.850 humans because they use human language and that's the way our brains work. 564 00:36:19.090 --> 00:36:22.930 If somebody talks to it or to us or uses human language, 565 00:36:23.570 --> 00:36:27.330 our brain automatically thinks it's a human being. And by 566 00:36:27.410 --> 00:36:30.830 thinking this starts to make 567 00:36:30.830 --> 00:36:34.110 assumptions. And many of these assumptions just aren't 568 00:36:34.430 --> 00:36:38.070 true. And these assumptions lead us down a 569 00:36:38.070 --> 00:36:41.630 path where we get disappointed, where we get 570 00:36:41.630 --> 00:36:45.470 frustrated, where we feel like, well, it just 571 00:36:45.470 --> 00:36:49.230 doesn't work for me. AI just doesn't work for me. It isn't there yet. 572 00:36:50.350 --> 00:36:53.870 It will never be a human being. It will never be able to actually 573 00:36:53.950 --> 00:36:56.710 replace a human being in that sense. 574 00:36:57.750 --> 00:36:59.910 However, its capabilities 575 00:37:01.590 --> 00:37:05.310 are incredible, but they are slightly different. So that's 576 00:37:05.310 --> 00:37:08.630 the first misconception. And one of the things that we tend to do is give 577 00:37:08.630 --> 00:37:12.430 it names, which makes it even harder 578 00:37:12.430 --> 00:37:16.270 for us. Like I remember we had. I can say 579 00:37:16.270 --> 00:37:20.030 it because I don't have a device in here. There's Alexa from 580 00:37:20.030 --> 00:37:23.390 Amazon. Alexa, the device which uses human 581 00:37:23.390 --> 00:37:26.710 language, it talks to us. And 582 00:37:27.190 --> 00:37:30.990 when I talk to Alexa, at least the original version, 583 00:37:30.990 --> 00:37:34.670 not Alexa, when I talked to the original person and I asked it something 584 00:37:34.670 --> 00:37:38.349 and it didn't know or it didn't understand me because it was just rule based, 585 00:37:38.349 --> 00:37:42.150 more or less, it would say, I can't answer 586 00:37:42.150 --> 00:37:45.870 that question. And I would get annoyed. I 587 00:37:45.870 --> 00:37:49.680 would get, I would feel frustrated because due to the fact that 588 00:37:49.680 --> 00:37:53.440 it was speaking to me with a human voice, something inside of 589 00:37:53.440 --> 00:37:57.240 my brain thought, it's a human being. And 590 00:37:59.080 --> 00:38:02.800 the next thought was not even consciously, but probably 591 00:38:02.800 --> 00:38:06.560 unconsciously. How stupid are you? Why don't 592 00:38:06.560 --> 00:38:08.920 you understand? And that is, that is 593 00:38:09.960 --> 00:38:13.520 there's a break in communication happening 594 00:38:13.520 --> 00:38:17.070 because my brain makes some assumptions that the 595 00:38:17.070 --> 00:38:20.750 technology doesn't fulfill. So I was frustrated. 596 00:38:20.750 --> 00:38:24.470 The technology doesn't care, but I was frustrated. I felt 597 00:38:24.470 --> 00:38:28.230 like that doesn't really work until I really understood. Well, it works in a different 598 00:38:28.230 --> 00:38:31.870 way. I cannot, I cannot project 599 00:38:32.590 --> 00:38:36.190 human consciousness into it. And 600 00:38:36.270 --> 00:38:39.870 that's a misconception. Projecting human consciousness into the 601 00:38:39.870 --> 00:38:43.590 thing is a misconception. This is something to be really aware 602 00:38:43.590 --> 00:38:47.310 of. The other misconception, and that's an entirely different, different thing, 603 00:38:47.310 --> 00:38:50.990 is that you need the most capable model. 604 00:38:51.310 --> 00:38:54.990 That you need to really make sure that you get 605 00:38:54.990 --> 00:38:58.270 the most capable model to get started. 606 00:38:58.990 --> 00:39:01.709 That is a way, and that's a kind of procrastination 607 00:39:02.750 --> 00:39:06.470 because all the models are really capable nowadays. Sure, if you 608 00:39:06.470 --> 00:39:09.870 look at the benchmarks, the models are different and every day there's a new one 609 00:39:09.870 --> 00:39:13.630 which beats some specific capability over all the others. 610 00:39:13.790 --> 00:39:17.210 There's a lot of progress going on. But if you wait for the perfect 611 00:39:17.210 --> 00:39:20.650 model you'll never get started. You can use 612 00:39:20.890 --> 00:39:24.730 literally any of the frontier models nowadays. It could be one of 613 00:39:24.730 --> 00:39:27.770 the open weights models like Llama or Mistral. 614 00:39:28.890 --> 00:39:31.770 It could be one of the commercial models like 615 00:39:32.570 --> 00:39:35.930 GPT or Claude or Nova. 616 00:39:36.810 --> 00:39:40.250 It doesn't really matter. These models are 617 00:39:40.250 --> 00:39:44.090 capable enough to experiment with the first 618 00:39:44.090 --> 00:39:47.730 use cases. And once you've experimented and once you've 619 00:39:47.730 --> 00:39:51.410 found a product market match, once you found 620 00:39:51.410 --> 00:39:55.090 a use case that really works, then it makes sense to think 621 00:39:55.090 --> 00:39:58.850 about, well, does it make sense to maybe use a different 622 00:39:58.930 --> 00:40:02.690 model that's a bit more capable 623 00:40:02.770 --> 00:40:06.450 in this specific use case? Or maybe it makes sense to 624 00:40:06.450 --> 00:40:10.090 introduce a second model which is less expensive for 625 00:40:10.090 --> 00:40:13.420 part of the use case? Because for instance, for 626 00:40:13.420 --> 00:40:16.780 summarization, I don't need any reasoning capabilities, 627 00:40:17.340 --> 00:40:20.860 I just need a good summarizer model. And for 628 00:40:21.740 --> 00:40:25.340 the actual workflow orchestration, for the agentic 629 00:40:25.340 --> 00:40:28.980 workflow maybe that I'm going to build, I need a model that's actually 630 00:40:28.980 --> 00:40:32.300 able to do planning and reasoning. These are two very different 631 00:40:32.300 --> 00:40:36.100 capabilities and they have very different costs. So 632 00:40:36.100 --> 00:40:39.740 I might need, at a point in the future, I might want to look at 633 00:40:39.740 --> 00:40:43.540 different models and at their price, structure, at their capabilities. But to 634 00:40:43.540 --> 00:40:46.520 get started, just pick one. Just pick one. 635 00:40:47.400 --> 00:40:51.240 If you have GDPR or other privacy issues that you 636 00:40:51.240 --> 00:40:55.000 need to take into account, pick a model service, 637 00:40:55.160 --> 00:40:58.440 a model hosting provider that provides you 638 00:41:00.440 --> 00:41:04.280 this functionality that guarantees you, that tells you we don't 639 00:41:04.280 --> 00:41:08.080 store your data and all the data is being encrypted and we don't use 640 00:41:08.080 --> 00:41:11.850 the data to train our models and we don't send the data to anybody else. 641 00:41:12.240 --> 00:41:15.120 If you come to AWS to use a model on Bedrock, 642 00:41:15.840 --> 00:41:19.600 and no matter whether you use our own Nova models or you use Claude or 643 00:41:19.600 --> 00:41:22.400 you use Llama or any of the other models, 644 00:41:23.920 --> 00:41:27.440 we host these models ourselves. We're not a 645 00:41:27.440 --> 00:41:31.200 gateway to the actual model provider. We're not a gateway to LLAMA 646 00:41:31.200 --> 00:41:34.640 or to the Llama API or to the Anthropic API or anything. 647 00:41:34.800 --> 00:41:38.320 We host versions of these models in air 648 00:41:38.320 --> 00:41:40.520 gapped accounts. 649 00:41:42.440 --> 00:41:45.960 Nobody gets into this. These models and these models 650 00:41:46.120 --> 00:41:49.600 don't send anything anywhere. Everything that 651 00:41:49.600 --> 00:41:53.360 happens is just your request gets 652 00:41:53.360 --> 00:41:57.200 sent into the air gapped account, gets handed over to the 653 00:41:57.200 --> 00:42:00.880 model. The model itself is stateless, it doesn't store anything. It 654 00:42:00.880 --> 00:42:04.680 just takes your data, loads it into its GPU along 655 00:42:04.840 --> 00:42:08.690 with the model algorithm, with the model weights and processes 656 00:42:08.690 --> 00:42:11.730 that, and then it sends the response back 657 00:42:12.370 --> 00:42:16.210 and everything else just goes back to sleep. There's nothing that we 658 00:42:16.210 --> 00:42:20.050 store and well, we 659 00:42:20.050 --> 00:42:23.729 do store, obviously we store telemetry, we store that 660 00:42:23.729 --> 00:42:26.930 you actually called the model and how many tokens you use because 661 00:42:27.410 --> 00:42:31.010 that's how you ultimately pay for that. But 662 00:42:31.010 --> 00:42:33.410 we don't do anything with your data. 663 00:42:35.770 --> 00:42:39.490 We even have models running in Frankfurt that you can use so that 664 00:42:39.490 --> 00:42:42.730 you don't even have to send your data to the us. 665 00:42:43.050 --> 00:42:46.490 We even provide access to these models through our own 666 00:42:46.490 --> 00:42:50.170 backbone, so you don't even have to use the public Internet if you 667 00:42:50.170 --> 00:42:53.770 want. That's what model provider or non 668 00:42:53.770 --> 00:42:56.810 model providers, model hosting providers 669 00:42:57.450 --> 00:43:01.220 like AWS provide. It's a little 670 00:43:01.220 --> 00:43:04.660 more costly than just going to ChatGPT or to Claude 671 00:43:04.660 --> 00:43:08.100 AI or to the Llama API. It's more costly. 672 00:43:08.340 --> 00:43:11.900 But on the other hand, we have different terms and 673 00:43:11.900 --> 00:43:15.460 conditions and we make sure that you will be able to build 674 00:43:15.460 --> 00:43:18.740 GDPR or HIPAA or whatever compliant workloads 675 00:43:19.060 --> 00:43:22.780 using these models. And that's an important point. If you have 676 00:43:22.780 --> 00:43:26.460 pii, if you need to take GDPR into account from 677 00:43:26.460 --> 00:43:30.270 day one, make sure you work with one 678 00:43:30.270 --> 00:43:33.910 of the providers that actually give you these capabilities, 679 00:43:34.310 --> 00:43:38.110 give you access, give you encryption, make sure that they don't use 680 00:43:38.110 --> 00:43:41.790 your data in any other way so that you can safely say in 681 00:43:41.790 --> 00:43:45.190 your own audit and to your own customers, I know 682 00:43:45.670 --> 00:43:48.070 where your data is going and I guarantee 683 00:43:49.270 --> 00:43:52.870 that it's not being handed over to somebody without my knowledge 684 00:43:53.680 --> 00:43:57.200 or without your knowledge as a customer. That's the important thing. 685 00:43:57.360 --> 00:44:00.800 But to get started, I might even really start 686 00:44:01.120 --> 00:44:04.600 with a use case that 687 00:44:04.600 --> 00:44:08.240 doesn't even need these complexities because it introduces 688 00:44:08.240 --> 00:44:11.920 complexities. And that's the thing. As soon as you need to 689 00:44:12.480 --> 00:44:16.040 work with sensitive data, you have 690 00:44:16.040 --> 00:44:19.700 to think about these things. You may have to think 691 00:44:19.700 --> 00:44:21.540 about. Even if you 692 00:44:23.540 --> 00:44:27.340 have a workflow that uses data from 693 00:44:27.340 --> 00:44:31.140 your database which goes through a model in a GDPR 694 00:44:31.220 --> 00:44:34.660 compliant way and gets displayed somewhere 695 00:44:34.980 --> 00:44:38.580 to a client, you still need to make sure that 696 00:44:39.860 --> 00:44:43.540 the data isn't being displayed by accident 697 00:44:43.620 --> 00:44:47.060 to somebody who shouldn't have access to them. So you need to be able to 698 00:44:47.540 --> 00:44:51.320 ensure authentication, authorization. You need all the 699 00:44:51.320 --> 00:44:54.840 security and compliance mechanisms that make sure that not a 700 00:44:54.840 --> 00:44:58.680 random person on the Internet or just a random person inside 701 00:44:58.680 --> 00:45:02.360 of your company is able to just use the agent and access 702 00:45:02.360 --> 00:45:03.360 your customer data. 703 00:45:06.959 --> 00:45:10.600 I see. I was wondering for our audience, if you 704 00:45:10.600 --> 00:45:14.400 could safely test any AI idea without 705 00:45:14.400 --> 00:45:18.250 risk, what would you build, tag us or reply to us 706 00:45:18.250 --> 00:45:21.850 on substack or with your moochart? I have 707 00:45:22.010 --> 00:45:25.610 two final questions for this interview, Dennis, because we are already 708 00:45:25.610 --> 00:45:29.250 recording for more than 45 minutes. But. 709 00:45:29.250 --> 00:45:32.490 But I do believe they're very important thoughts 710 00:45:32.810 --> 00:45:36.010 before you even start thinking about 711 00:45:36.250 --> 00:45:40.010 applying AI. And we already know you guys support different 712 00:45:40.090 --> 00:45:43.510 models. You are more the infrastructure provider for 713 00:45:43.670 --> 00:45:47.510 something like this. But I was wondering, have you seen any 714 00:45:47.510 --> 00:45:51.190 clever AI adoption stories where companies 715 00:45:51.270 --> 00:45:53.670 started small and then scaled rapidly? 716 00:45:57.190 --> 00:46:01.030 The first thing really that I would look at is 717 00:46:01.030 --> 00:46:04.430 do I even need AI for that? Many of the things that we're trying to 718 00:46:04.430 --> 00:46:07.590 solve with AI nowadays, they have already been solved 719 00:46:08.390 --> 00:46:11.990 and probably in a good and much less expensive and much less 720 00:46:13.130 --> 00:46:14.890 ecologically impactful way. 721 00:46:16.730 --> 00:46:20.530 If you have a calculator that can add up to numbers, 722 00:46:20.530 --> 00:46:24.290 use a calculator. Don't ask an AI to do it for you. First of all, 723 00:46:24.290 --> 00:46:27.770 it isn't very good at it. Well, they're getting better at math. But 724 00:46:28.890 --> 00:46:32.170 why would you start up an entire cluster of big 725 00:46:32.170 --> 00:46:35.130 Nvidia GPUs to get 726 00:46:35.770 --> 00:46:39.380 the sum of two numbers? You shouldn't be doing 727 00:46:39.380 --> 00:46:42.900 that. So first of all, don't try to solve already 728 00:46:42.900 --> 00:46:46.660 solved problems. And the second thing really is 729 00:46:46.660 --> 00:46:49.940 again, look at 730 00:46:50.980 --> 00:46:54.260 the painful things that nobody ever tackled. 731 00:46:54.580 --> 00:46:58.180 Look at something that has been bothering 732 00:46:58.180 --> 00:47:01.940 you or your customer for a long time 733 00:47:02.500 --> 00:47:06.030 and it hadn't been addressed because 734 00:47:06.030 --> 00:47:08.870 everybody said, well, it doesn't just doesn't work and we don't have the time to 735 00:47:08.870 --> 00:47:09.630 do it ourselves. 736 00:47:12.990 --> 00:47:16.190 It's actually pretty good closing words. 737 00:47:16.830 --> 00:47:20.510 We will be back for one, the Founders 738 00:47:20.510 --> 00:47:24.150 Vault for our premium subscribers on substack and YouTube. And second, 739 00:47:24.150 --> 00:47:27.630 you'll be back for a second interview where you get more 740 00:47:27.630 --> 00:47:31.070 hands on when you go through all the thoughts you had 741 00:47:32.190 --> 00:47:35.830 that you need to think through before you can even get started on 742 00:47:35.830 --> 00:47:39.390 AI. Great. I'm looking forward to it. Me too. 743 00:47:39.470 --> 00:47:40.750 Have a good day. Bye Bye. 744 00:47:45.790 --> 00:47:49.310 That's all folks. Find more news, streams, 745 00:47:49.550 --> 00:47:50.590 events and 746 00:47:50.590 --> 00:47:54.850 interviews@www.startuprad.IO. 747 00:47:55.560 --> 00:47:57.560 remember, Sherry is caring.