1 00:00:00,240 --> 00:00:04,040 Imagine waking up tomorrow to discover a thousand AI agents are making 2 00:00:04,040 --> 00:00:07,720 decisions across your company and nobody knows exactly what 3 00:00:07,720 --> 00:00:11,120 they're accessing. Sound dramatic? Perhaps 4 00:00:11,680 --> 00:00:13,840 impossible? Not anymore. 5 00:00:23,520 --> 00:00:27,160 Hello and welcome back to Data Driven, the podcast. We explore the 6 00:00:27,160 --> 00:00:30,770 emergent and industry of AI 7 00:00:30,850 --> 00:00:34,250 data science. And of course, without the underpinnings of data 8 00:00:34,250 --> 00:00:37,970 engineering, none of this would be possible. And with me today, I'm very 9 00:00:37,970 --> 00:00:41,810 happy to once again have my favoritest data engineer in the world, Andy 10 00:00:41,810 --> 00:00:45,330 Leonard. How's it going, Andy? It's going well, Frank. How are you doing? 11 00:00:45,570 --> 00:00:49,330 I'm doing all right. I realized that this is no longer an emerging industry 12 00:00:49,330 --> 00:00:53,010 after, what, 10 seasons? Now we're in season 10, and 13 00:00:53,010 --> 00:00:56,630 pretty wild journey this has been. And I'm sure 14 00:00:56,630 --> 00:01:00,350 by the time you're hearing this, we already launched our 10 season 15 00:01:00,350 --> 00:01:03,950 10 retrospective and looking to the future episodes. I won't bore you 16 00:01:03,950 --> 00:01:07,710 again with that. Today we have a special treat. We 17 00:01:07,710 --> 00:01:11,270 have Ron Longo, who is the CEO of Trustlogix, 18 00:01:11,750 --> 00:01:15,590 and I love the marketing material. 19 00:01:15,830 --> 00:01:19,190 This is your AI control plane with a kill switch. 20 00:01:19,750 --> 00:01:22,900 So welcome to the show, Ron. Great to be here. Thanks so much for having 21 00:01:22,900 --> 00:01:26,620 me. This is what a control plane is, particularly around AI a little differently. So 22 00:01:26,620 --> 00:01:30,380 what in your mind is an AI control plane? So from 23 00:01:30,380 --> 00:01:33,820 our perspective, there's security, right? You want to keep AI 24 00:01:33,820 --> 00:01:37,220 secure, and there's a variety of ways of doing it. And if you look at 25 00:01:37,220 --> 00:01:40,940 the typical guardrails, they're making sure that there's not prompt injections. 26 00:01:41,020 --> 00:01:44,740 There's no problems with the actual LLM, but what we focus 27 00:01:44,740 --> 00:01:48,540 on is the data, because, I mean, that's the fuel 28 00:01:49,190 --> 00:01:52,950 of AI. And so we basically put in place 29 00:01:52,950 --> 00:01:56,310 a control, an access control plane. 30 00:01:56,550 --> 00:02:00,390 So we ensure that no matter if it's an agent, a human, 31 00:02:00,390 --> 00:02:04,230 a non human, if it's accessing multiple data sources, we just 32 00:02:04,230 --> 00:02:07,590 ensure least privilege. And. And we can do that 33 00:02:07,750 --> 00:02:10,950 in runtime. We can provide full auditability. 34 00:02:11,750 --> 00:02:15,550 And as I think you read in that same description, we also 35 00:02:15,550 --> 00:02:19,090 apply a kill switch because if you think about it, what happens if you. 36 00:02:19,320 --> 00:02:22,840 Your agents run amok? You have to be able to control 37 00:02:22,840 --> 00:02:26,680 them. And so we've introduced this concept of a kill switch just in 38 00:02:26,680 --> 00:02:30,440 case something happens. So kill switch 39 00:02:30,440 --> 00:02:33,920 tends to be a loaded word. This isn't like a 40 00:02:33,920 --> 00:02:37,520 Skynet kill switch, right? Like, how dramatic is the kill switch? Is it 41 00:02:37,520 --> 00:02:41,280 just a process kill request, or is it. What happens. Do you 42 00:02:41,280 --> 00:02:44,600 reset the entire state? Run what. What exactly 43 00:02:44,760 --> 00:02:48,450 happens when you press the emergency button? That's a great question. 44 00:02:48,450 --> 00:02:52,250 So it's proportional to the risk. Right. So if it's a nominal risk 45 00:02:52,330 --> 00:02:55,970 and a particular agent, and you can isolate the agent and 46 00:02:55,970 --> 00:02:59,730 it's accessing data, it shouldn't be accessing. The kill switch can be something 47 00:02:59,730 --> 00:03:03,330 as simple as, hey, we're going to force data masking or we're going to revoke 48 00:03:03,330 --> 00:03:06,930 privileges for that particular agent. But if you've got 49 00:03:06,930 --> 00:03:10,210 a massive threat, something that's actually going to take down your 50 00:03:10,210 --> 00:03:14,030 systems, your SaaS service, or it's going to have access 51 00:03:14,270 --> 00:03:17,830 across a broad swath of data sources, we have the 52 00:03:17,830 --> 00:03:21,550 ability to basically block all of those, all of the agents. 53 00:03:21,790 --> 00:03:25,590 And then once you at least eliminate the major threat, then you 54 00:03:25,590 --> 00:03:29,069 have time to basically step back and say, okay, is it this 55 00:03:29,069 --> 00:03:32,510 particular data source or is it this particular group? 56 00:03:32,990 --> 00:03:36,550 And then you can basically allow the rest of the agents to operate 57 00:03:36,550 --> 00:03:40,310 normally. But how you would look at it is it could either be 58 00:03:40,310 --> 00:03:43,870 a pause, it could be a revocation, it could be data masking. 59 00:03:44,600 --> 00:03:48,400 But the point is you gotta keep the data safe. And if you think about 60 00:03:48,400 --> 00:03:52,120 there was this recent breach, I think it was anthropic 61 00:03:52,200 --> 00:03:55,920 and it took them hours to isolate it. Enterprises don't 62 00:03:55,920 --> 00:03:59,240 have hours. And so we introduced something that you can do at runtime. 63 00:03:59,959 --> 00:04:03,640 Nice. Nice. Yeah. With the speed of bandwidth being 64 00:04:03,640 --> 00:04:07,360 what it is, hours is. You could. How many gigs can be offloaded in 65 00:04:07,360 --> 00:04:11,010 that period of time? Exactly, exactly. An 66 00:04:11,010 --> 00:04:11,810 alarming amount. 67 00:04:14,690 --> 00:04:18,410 Most enterprises won. Sounds like a 68 00:04:18,410 --> 00:04:22,130 great idea though. Ron, I was going to ask, is there some 69 00:04:22,130 --> 00:04:25,849 particular story, client, friend, some 70 00:04:25,849 --> 00:04:29,570 news story you read that inspired this? It was a 71 00:04:29,570 --> 00:04:33,370 couple of things actually. It was our founder, Ganesh Kirti actually had 72 00:04:33,370 --> 00:04:37,100 read about that anthropic breach and 73 00:04:37,100 --> 00:04:40,860 he then wrote a blog that followed after that happened. And 74 00:04:40,860 --> 00:04:44,540 he said, look, what needs to be put in place is this concept of effectively 75 00:04:44,540 --> 00:04:47,900 of a kill switch. Well, we happen to have that in our product. 76 00:04:48,380 --> 00:04:52,020 As marketing got together with the founder, we said, hey, let's make that front and 77 00:04:52,020 --> 00:04:55,820 center. We here's the reality in the world of 78 00:04:55,820 --> 00:04:59,660 AI, it is so noisy, right? Just walk through rsa. 79 00:04:59,660 --> 00:05:03,420 Or we've got Databricks Summit coming up, Snowflake Summit coming up. 80 00:05:03,420 --> 00:05:06,700 What is everyone gonna do going to talk about? It's all about AI 81 00:05:07,100 --> 00:05:10,820 and it's so confusing. And when we talk to our customers, which are 82 00:05:10,820 --> 00:05:14,340 the some of the largest enterprise customers on the planet, they look, they come to 83 00:05:14,340 --> 00:05:17,940 us and they say, look, we don't know what everybody's doing 84 00:05:17,940 --> 00:05:21,780 because it all starts to sound the same. And what I wanted to 85 00:05:21,780 --> 00:05:24,740 do and what I've asked the marketing team to do is look, we've got to 86 00:05:24,740 --> 00:05:28,220 punch above the noise threshold and we have to make sure that we have a 87 00:05:28,220 --> 00:05:32,030 message that resonates that people can look at. And, and it's very simple to 88 00:05:32,030 --> 00:05:35,790 understand. And that concept of kill switch and that came from 89 00:05:35,790 --> 00:05:39,470 our founders and from our marketing team was hey guys, that makes a lot of 90 00:05:39,470 --> 00:05:43,230 sense because. And it also came from 91 00:05:43,230 --> 00:05:46,910 an actual customer that came to us and said our CEO 92 00:05:46,910 --> 00:05:50,150 wanted to develop an agent and run it and we had to tell him no. 93 00:05:51,270 --> 00:05:55,030 And by the way, we've got a thousand agents, so we don't really know what 94 00:05:55,030 --> 00:05:58,750 they're accessing. Can you help us? And when that came up, 95 00:05:58,750 --> 00:06:02,030 we knew that we had to jump in and we had to offer things like 96 00:06:02,030 --> 00:06:05,830 that access control plane. But also the concept of a kill switch 97 00:06:05,830 --> 00:06:08,870 for protection, that makes sense. 98 00:06:09,430 --> 00:06:13,030 So you talk about in that same thing, right underneath Kill Switch, 99 00:06:13,189 --> 00:06:16,830 it says govern every agent, every mcp, every 100 00:06:16,830 --> 00:06:20,310 tool and every data request. Right? Now speaking, 101 00:06:20,390 --> 00:06:23,790 that really doubles down on the noise, Right. The noise of AI has gotten 102 00:06:23,790 --> 00:06:27,370 ridiculous. Yeah. So how do you, would you 103 00:06:27,370 --> 00:06:30,730 govern an agent? The same way as you would govern 104 00:06:30,890 --> 00:06:34,730 a MCP server data request, things like that. 105 00:06:35,690 --> 00:06:39,410 How does your platform, does your platform treat them the same or 106 00:06:39,410 --> 00:06:43,130 does it treat them differently based on what they are? Yeah, so 107 00:06:43,130 --> 00:06:46,610 that, that's a great question. And you have to treat them based on their 108 00:06:46,610 --> 00:06:50,330 purpose, based on their intent. You can't treat 109 00:06:50,330 --> 00:06:53,930 everybody the same. And so that's one, one of our capabilities, 110 00:06:53,930 --> 00:06:57,630 right. So we're at the data layer, so we' we basically 111 00:06:57,630 --> 00:07:01,350 ingest all the policies and we understand what the 112 00:07:01,350 --> 00:07:05,190 agent's purpose is. We, we understand what the least 113 00:07:05,190 --> 00:07:08,550 privileges are that the user that's accessing that particular 114 00:07:08,630 --> 00:07:12,310 agent because the agents have broader access, right. 115 00:07:12,710 --> 00:07:16,470 Service level access into multiple data sources. And 116 00:07:16,630 --> 00:07:19,790 just by the very nature of the fact that they're there to basically help and 117 00:07:19,790 --> 00:07:23,540 optimize. But the users may have privileges just to 118 00:07:23,540 --> 00:07:27,220 get to one particular data source or one type of data. Maybe they're the financial 119 00:07:27,700 --> 00:07:31,460 analyst and they can't get access to HR data. So we 120 00:07:31,460 --> 00:07:35,300 treat that differently, right. We treat the intent and the 121 00:07:35,540 --> 00:07:39,220 capability and the access privileges that particular 122 00:07:39,380 --> 00:07:43,100 user has while, whether it's human or non human, while 123 00:07:43,100 --> 00:07:46,500 it's accessing that agent or those MCP tools. 124 00:07:46,820 --> 00:07:50,560 And all of that is effectively follows the concept of 125 00:07:50,560 --> 00:07:53,920 least privilege all the way through the stack. Whether it's 126 00:07:53,920 --> 00:07:57,560 agent mcp, whether it's agent orchestration 127 00:07:57,560 --> 00:08:00,960 layer or data source. We have to make sure. That least 128 00:08:00,960 --> 00:08:04,600 privilege is basically enforced all the way down to the data source. 129 00:08:05,480 --> 00:08:09,080 Is that what the protocols was? A Spiff Inspire or something like that? 130 00:08:09,800 --> 00:08:13,600 Oh, now you're getting into an area that. Okay, I'm just curious because 131 00:08:13,600 --> 00:08:17,260 I. It was a non technical. Oh, okay. Don't 132 00:08:17,260 --> 00:08:20,980 worry. I only heard about Spiff and Fire two weeks ago and I 133 00:08:20,980 --> 00:08:24,300 know I'm mispronouncing them, but it's basically protocols that basically 134 00:08:24,300 --> 00:08:27,740 immediately based on certain flags will immediately revoke an 135 00:08:27,740 --> 00:08:31,459 agent's privileges or an agent has to request to do something and then 136 00:08:31,459 --> 00:08:35,020 that window opens and then as soon as the task is done, that window 137 00:08:35,020 --> 00:08:38,820 closes. Right. So you have really fine grain control. Andy's 138 00:08:38,820 --> 00:08:42,100 nodding his head like he's heard of these protocols, but I don't know. 139 00:08:42,949 --> 00:08:46,549 I've heard the terms, but I like you, I haven't dug into them. I've got 140 00:08:46,549 --> 00:08:50,309 about what you got out of it, Frank. That. Okay. The idea and 141 00:08:50,629 --> 00:08:54,109 what you're saying, Ron too. Governing the principle of least 142 00:08:54,109 --> 00:08:57,589 privilege. And that sounds like a fantastic idea. 143 00:08:58,789 --> 00:09:02,629 I'm not experienced with any of this type of automation in 144 00:09:02,629 --> 00:09:06,069 practice, although I do interact with AI 145 00:09:06,309 --> 00:09:09,920 genic AIs often. I'm not familiar with how this 146 00:09:09,920 --> 00:09:13,680 works in practice. I was following your earlier explanations with 147 00:09:13,680 --> 00:09:17,400 Frank there. It sounds good, I'll say that part. 148 00:09:17,400 --> 00:09:20,720 It also sounds like it might present a bit of an obstacle to 149 00:09:21,280 --> 00:09:25,120 the YOLO lifestyles that some AI 150 00:09:25,120 --> 00:09:28,720 assistant developers are used to. Just tripping that 151 00:09:28,720 --> 00:09:32,480 bit and having it just go. I imagine some of the 152 00:09:32,480 --> 00:09:36,220 horror stories are related, but I don't know for sure. And you only 153 00:09:36,220 --> 00:09:39,420 get what you get, right? Yeah, I think this is 154 00:09:39,500 --> 00:09:43,020 definitely addresses that one issue because that is 155 00:09:43,180 --> 00:09:46,220 prevalent and that goes back to the story I was telling earlier about the company 156 00:09:46,220 --> 00:09:49,740 that came to us and said, we've got a ousand agents, we're AI first. However, 157 00:09:50,060 --> 00:09:53,380 we've got to put some sort of control in place. And so that's why we 158 00:09:53,380 --> 00:09:56,980 introduced this concept of registering the agents and making 159 00:09:56,980 --> 00:10:00,700 sure that those agents behave. The Spiff 160 00:10:00,700 --> 00:10:04,410 Inspire. While I don't. I'm not feeling familiar with those terms, I'm definitely 161 00:10:04,410 --> 00:10:08,210 familiar with Just in time. We are. We do support that because you want 162 00:10:08,210 --> 00:10:11,890 to be able to support time restricted privileges and then be able to 163 00:10:11,890 --> 00:10:15,210 revoke that privilege over. And we're going to see that more and more 164 00:10:15,530 --> 00:10:19,290 because agents are developed for very specific purposes. You may 165 00:10:19,290 --> 00:10:23,050 launch this for a week, a day, an hour, and then you want to revoke 166 00:10:23,050 --> 00:10:26,690 it and so that we fully support that. Yeah. And 167 00:10:26,690 --> 00:10:30,530 it's funny because that's an old. That that whole it's an old thing, right? Open 168 00:10:30,530 --> 00:10:34,330 the window, close the window. All these security best practices, this isn't new. 169 00:10:34,330 --> 00:10:38,030 And I think speaking to your point about AI noise, right. Everything 170 00:10:38,030 --> 00:10:41,750 old is new again, Right. The fundamentals are there. Right. I have a lot 171 00:10:41,750 --> 00:10:45,590 of debates in the same meeting I was at where they were talking about. 172 00:10:45,750 --> 00:10:49,430 And it's actually spiffy. S, P, I, F, F, E. 173 00:10:49,430 --> 00:10:53,230 And the other one is spire. That's like spire, like a cathedral 174 00:10:53,230 --> 00:10:56,790 spire or something like that. But it also. People, I think, are 175 00:10:56,790 --> 00:11:00,470 realizing now that the AI hype wave, it's not crashed, but 176 00:11:00,470 --> 00:11:03,740 we're definitely towards the crest of it. People are realizing as they 177 00:11:03,740 --> 00:11:07,500 operationalize these workloads that fundamentals matter. 178 00:11:07,660 --> 00:11:11,380 Fundamentals are boring, but they do matter. And like this whole 179 00:11:11,380 --> 00:11:15,060 least privilege and things like that, rotating keys and all of that is, 180 00:11:15,060 --> 00:11:18,780 is. It's everything old is new again, right? 181 00:11:18,780 --> 00:11:22,380 Yeah. The more things change, the more they 182 00:11:22,540 --> 00:11:25,660 stay the same. That's right. And it's the reality. 183 00:11:26,220 --> 00:11:29,580 Yeah. And it's funny because I find myself using 184 00:11:29,970 --> 00:11:33,450 more, more often in my career. I'm using the phrase garbage in, 185 00:11:33,450 --> 00:11:36,290 garbage out than I have in 20 years. 186 00:11:37,250 --> 00:11:40,930 Because if you put garbage into an AI, you're going to get 187 00:11:40,930 --> 00:11:44,690 garbage out. And all that goes from the data it was trained on. Right. 188 00:11:45,570 --> 00:11:49,330 All the way through the prompt that you give it. I was talking with someone 189 00:11:49,330 --> 00:11:52,970 who, he recently lost his job and he's not an AI guy and he's 190 00:11:52,970 --> 00:11:56,770 very, very stuck in a previous era. That's what I'll say. And 191 00:11:56,770 --> 00:12:00,530 he was saying like, oh, when I asked the AI questions or to help me 192 00:12:00,530 --> 00:12:04,090 with my resum. It gives me really a bunch of text that doesn't make 193 00:12:04,090 --> 00:12:07,370 sense. And he said, well, what do you expect from AI? And I turned to 194 00:12:07,370 --> 00:12:11,170 him, I'm like, Dude, this isn't 2023 anymore. What did you 195 00:12:11,170 --> 00:12:14,010 prompt it? What did you prompt it? He goes, I just said, give me your 196 00:12:14,010 --> 00:12:17,210 resume. No, no, no, no. You have to really spike the prompt 197 00:12:17,689 --> 00:12:21,490 or not. That sounds bad, like prompt injection. But the 198 00:12:21,490 --> 00:12:25,050 more information you give the AI going in, the 199 00:12:25,050 --> 00:12:28,370 better quality you're going to go out. Right. It's kind of like talking to my 200 00:12:28,370 --> 00:12:31,740 kids. Right. If I ask them how their day was, they'll say, fine, but if 201 00:12:31,740 --> 00:12:34,300 I ask them what was your favorite part of your day? Right. 202 00:12:35,660 --> 00:12:39,180 Find is not a good answer to that. Right. They'll have to answer me with 203 00:12:39,180 --> 00:12:42,300 something more. Right. So it's like being a parent almost prepares you to be A 204 00:12:42,300 --> 00:12:46,060 prompt engineer. Yeah, you have to architect and prompt engineering is a great way to 205 00:12:46,060 --> 00:12:49,900 put it. You have to architect the prompt. And the problem with AI, I think, 206 00:12:50,140 --> 00:12:53,740 and totally to your point, garbage in, garbage out. But they speak 207 00:12:53,980 --> 00:12:57,500 with such authority, so much confidence 208 00:12:58,090 --> 00:13:01,810 that you tend to believe what it says even though it's just delivering 209 00:13:01,810 --> 00:13:05,570 back to you garbage. And that is one of the, like 210 00:13:05,570 --> 00:13:09,050 you said, old principles apply, best practices apply. 211 00:13:09,210 --> 00:13:12,930 Here's the difference, here's the foundational difference. And not 212 00:13:12,930 --> 00:13:16,730 to date myself, but I've been through the broadband era, 213 00:13:17,050 --> 00:13:20,890 the Internet era, the cloud era, the SaaS era. So I've been 214 00:13:20,890 --> 00:13:24,250 part of, I've been fortunate to be part of companies that have driven these 215 00:13:24,830 --> 00:13:28,670 sea changes. This sea change, I think despite 216 00:13:28,910 --> 00:13:32,590 some of the hype, is significant and one of 217 00:13:32,590 --> 00:13:36,310 the biggest issues that makes a difference is the fact that 218 00:13:36,310 --> 00:13:40,110 it's running at machine speed. And you can run with a huge 219 00:13:40,270 --> 00:13:43,870 number of these things. Just picture 10,000 agents 220 00:13:43,950 --> 00:13:47,750 or even a thousand agents supporting at least that many 221 00:13:47,750 --> 00:13:51,560 users or non human users running machine 222 00:13:51,560 --> 00:13:55,200 speed. So not only do you have to apply best practices, 223 00:13:55,200 --> 00:13:59,000 but you have to apply best practices very quickly. And 224 00:13:59,000 --> 00:14:01,960 so that's one of the key differences in terms of 225 00:14:03,320 --> 00:14:06,519 how do you have to evolve best practices from yesterday 226 00:14:06,920 --> 00:14:10,720 to meet the demands of today. Yeah, and that 227 00:14:10,720 --> 00:14:14,400 actually calls back to maybe two or three episodes ago depending on what 228 00:14:14,400 --> 00:14:18,180 order releases in where we talked about the neuro resilient leader 229 00:14:18,500 --> 00:14:22,140 where basically AI has basically made it. So the hustle culture has to 230 00:14:22,140 --> 00:14:25,860 change because you said it machine speed, Machine speed 231 00:14:25,940 --> 00:14:29,700 changes everything. Right. I don't care how fast you could run, you're 232 00:14:29,700 --> 00:14:33,100 not going to outrun a Camaro. Right. Like it's really that 233 00:14:33,100 --> 00:14:36,740 mentality, it's just not going to happen. Or I think Andy's a Dodge 234 00:14:36,740 --> 00:14:40,580 fan, so whatever that Dodge Challengers, Dodge Challenger, 235 00:14:40,580 --> 00:14:44,100 you're never going to outrun it. Right, but, but to your point, the 236 00:14:44,100 --> 00:14:47,840 governance I think also like you said, has to happen at that, that, that 237 00:14:47,840 --> 00:14:51,480 speed. So I don't know what the exact Latin 238 00:14:51,480 --> 00:14:55,320 phrase is, but it's something like who watches the watchers? What's your gov like? 239 00:14:55,320 --> 00:14:59,120 Who watches the governance on that you provide here? Is that something the customer 240 00:14:59,120 --> 00:15:02,880 decides or is that built into your product? Yeah, I 241 00:15:02,880 --> 00:15:06,080 want to make sure I understand the question. Obviously we apply 242 00:15:06,080 --> 00:15:09,680 enforcement and we have to apply enforcement at machine 243 00:15:09,680 --> 00:15:13,440 speed, but there's always going to be a human enforcement, the 244 00:15:13,440 --> 00:15:17,080 middle. If you think about just the fact that AI 245 00:15:17,160 --> 00:15:20,760 can hallucinate, things can happen. Going back to the kill 246 00:15:20,760 --> 00:15:24,600 switch concept, Obviously we're not going to automate that. We have 247 00:15:24,600 --> 00:15:28,360 the capability of automating that. But I don't think any enterprise in their right 248 00:15:28,360 --> 00:15:32,120 mind is going to say if something happens, automatically kill 249 00:15:32,120 --> 00:15:35,320 everything. So there has to be someone that's actually overseeing this, 250 00:15:35,720 --> 00:15:38,920 but you can't turn it into a human speed 251 00:15:39,560 --> 00:15:43,240 action. Right? Say we're going to display all this information. You got to read 252 00:15:43,240 --> 00:15:46,960 through this information and make a decision who you're going to, what agents 253 00:15:46,960 --> 00:15:50,320 are you going to revoke. It has to basically be a single 254 00:15:50,320 --> 00:15:54,080 button. And that's what you have to simplify. You have to give them the 255 00:15:54,080 --> 00:15:57,840 facts and then you have to say, look, we've discovered a 256 00:15:57,840 --> 00:16:01,120 breach or someone's trying to get access to data as we narrowed it down as 257 00:16:01,120 --> 00:16:04,680 an example to your S3 buckets in this particular, for this 258 00:16:04,680 --> 00:16:07,840 particular group, then you can make a decision. I'm going to, I'm going to kill 259 00:16:07,840 --> 00:16:11,440 that group of agents. But back to your question of 260 00:16:11,760 --> 00:16:15,040 who's governing the governance. How our solution 261 00:16:15,200 --> 00:16:18,720 works is we're running the enforcement at machine speed, 262 00:16:18,720 --> 00:16:22,440 but everything is auditable. So we track all of this, provide the 263 00:16:22,440 --> 00:16:26,280 logs, and then we provide the audit trail. And you have to do 264 00:16:26,280 --> 00:16:29,440 that because what a customer is requiring 265 00:16:30,000 --> 00:16:33,720 is it needs that audit trail down to the agent, down 266 00:16:33,720 --> 00:16:37,520 to the user that's accessing the agent. Right. If you're just looking at the 267 00:16:37,520 --> 00:16:41,280 agent that's got broad access to everything, you're missing all of the 268 00:16:41,280 --> 00:16:44,840 access to that agent. So we have to provide that level of 269 00:16:45,160 --> 00:16:47,960 auditability so that the enterprises can govern. 270 00:16:48,760 --> 00:16:52,480 Okay, that makes sense. So you're not really. The human is in the loop, but 271 00:16:52,480 --> 00:16:56,280 the human has a much more powerful hammer to, to break the glass 272 00:16:56,280 --> 00:17:00,040 and stop the process line. Cool. If you think about how enterprises 273 00:17:00,040 --> 00:17:03,590 have historically worked. So the problem that we solved, making it really 274 00:17:03,590 --> 00:17:06,390 simple, enterprises have lots of data sources. 275 00:17:07,190 --> 00:17:10,550 Anyone that says, oh no, I've got everything in one data 276 00:17:10,550 --> 00:17:14,150 platform, so they're lying to you. That, that just doesn't exist. 277 00:17:14,470 --> 00:17:18,270 Right, because they've got on prem Oracle SQL servers, they've 278 00:17:18,270 --> 00:17:22,030 got S3 buckets. And they may, their aspiration 279 00:17:22,030 --> 00:17:25,270 may be to put everything into a single data lake or a single 280 00:17:25,510 --> 00:17:29,230 data source, but that's not how it is today. And so what we do is 281 00:17:29,230 --> 00:17:32,870 we apply this unified control plane 282 00:17:32,870 --> 00:17:36,590 across all those data sources so that you can manage all of 283 00:17:36,590 --> 00:17:39,950 that simply from a single pane of glass. How it's done today, 284 00:17:40,430 --> 00:17:44,030 customers will say, we solved that problem. We have JIRA tickets, 285 00:17:44,030 --> 00:17:47,470 we've got folks that are running, doing Python scripting and SQL 286 00:17:47,950 --> 00:17:51,430 and we've solved that problem, but they solved it at human 287 00:17:51,430 --> 00:17:54,990 speed. Now all of a sudden you start adding in agents that are accessing those 288 00:17:54,990 --> 00:17:58,670 data sources. That process is broken. And 289 00:17:58,670 --> 00:18:02,510 so that's what we do. We basically take that manual human speed process 290 00:18:02,750 --> 00:18:06,350 and we apply this unified control plane to basically 291 00:18:06,350 --> 00:18:09,870 match the human speed requirement. Match the machine speed requirement. 292 00:18:10,590 --> 00:18:14,430 Interesting. Yes. My question, and I 293 00:18:14,430 --> 00:18:17,950 think Frank and I have been dancing around this, but I don't know if we 294 00:18:18,030 --> 00:18:21,870 understand. And the answer may very well be it depends. One of 295 00:18:21,870 --> 00:18:25,300 my favorite answers as a consultant, but the actual 296 00:18:25,300 --> 00:18:29,020 trigger of what sets it off. So I've heard part of the answer 297 00:18:29,020 --> 00:18:32,740 and I believe part of the answer is definitely human in the loop, 298 00:18:32,820 --> 00:18:36,500 perhaps human in control. And then I've heard, then I've 299 00:18:36,500 --> 00:18:40,140 heard machine speed. And I feel like there's, there's some 300 00:18:40,140 --> 00:18:43,740 piece in the middle there. And again the answer may very well be it 301 00:18:43,740 --> 00:18:47,220 depends. Is it? When Frank asked about who's watching the watcher 302 00:18:47,380 --> 00:18:50,260 is the watcher. If the watcher is an AI, 303 00:18:51,340 --> 00:18:55,180 then certainly that becomes the actuator, if 304 00:18:55,180 --> 00:18:58,980 you will, of the kill switch, the actor that works on that. And 305 00:18:58,980 --> 00:19:02,780 I could see that being a very important piece. If there's 306 00:19:02,780 --> 00:19:06,220 some signal received or some pattern detected 307 00:19:06,780 --> 00:19:10,300 that looks like a hack. Let's say you're, 308 00:19:10,460 --> 00:19:14,060 it's two weeks ago and you're sitting in the dock at Canvas. 309 00:19:14,620 --> 00:19:18,200 Okay. I have a daughter at the computer science program 310 00:19:18,280 --> 00:19:22,000 at Virginia Tech and it's two days before exams and 311 00:19:22,000 --> 00:19:25,320 she can't get to her material to study. That happened. 312 00:19:25,960 --> 00:19:29,680 So if you're there and you're watching these 313 00:19:29,680 --> 00:19:33,520 various signals pop up I monitoring 314 00:19:33,520 --> 00:19:37,360 that that sort of system that has access to a kill switch 315 00:19:37,360 --> 00:19:41,000 would have come in handy if data had been exfilt or if 316 00:19:41,000 --> 00:19:44,320 heuristics detection was occurring and see a network 317 00:19:44,880 --> 00:19:48,400 bombardment, denial of service or however 318 00:19:48,720 --> 00:19:52,560 Trojans were making their way into the space behind 319 00:19:52,560 --> 00:19:56,000 a firewall. If you got to rely on a human to do that, 320 00:19:56,240 --> 00:19:59,840 then that could be a problem. And again, the answer may 321 00:19:59,840 --> 00:20:03,480 be for Canvas or a scenario like that in the 322 00:20:03,480 --> 00:20:06,880 future. Your product is sitting there and it is automated. 323 00:20:07,360 --> 00:20:11,200 There is an AI with its virtual hand over hovering over 324 00:20:11,200 --> 00:20:14,920 the virtual switch that does the killing. It may be for 325 00:20:14,920 --> 00:20:18,680 another, another use case and I want venture to guess which 326 00:20:18,760 --> 00:20:22,600 where that may still be there. But a human 327 00:20:22,680 --> 00:20:26,040 has to make that final call. That's kind of what it is. 328 00:20:26,280 --> 00:20:29,000 But if I can interject. Yeah, but if I can. I'm sorry, I Cut you 329 00:20:29,000 --> 00:20:32,080 off. Andy, if I can interject and Ron can tell me if I'm off base 330 00:20:32,080 --> 00:20:35,840 or on base, is that because enterprise systems 331 00:20:35,840 --> 00:20:39,680 are so spread out and disparate across multiple systems? And that's not really 332 00:20:39,680 --> 00:20:43,320 a new thing, it's just gotten worse. So I may have a dependency 333 00:20:43,320 --> 00:20:47,020 in Aw bucket, I may have something in Azure SQL, I may have 334 00:20:47,020 --> 00:20:50,540 something on Prem, I may have something God knows where else. 335 00:20:50,540 --> 00:20:54,340 Right this way. I have a kill switch to kill agents X, Y 336 00:20:54,340 --> 00:20:57,940 and Z and their system, correct me if I'm 337 00:20:57,940 --> 00:21:01,660 wrong, would know. Oh, if you press that button, I know where all the 338 00:21:01,660 --> 00:21:04,900 bodies are buried, so to speak. That's a terrible analogy, sorry about that. 339 00:21:05,860 --> 00:21:09,460 But it knows where all the bad. Makes sense. But it knows 340 00:21:09,540 --> 00:21:13,060 all the things it needs to shut down as opposed to if you're a sysadmin 341 00:21:13,140 --> 00:21:16,520 and you get paged at 2 in the morning, you're like oh, I gotta do 342 00:21:16,520 --> 00:21:19,960 this or I have to pull up a document and things like this. Whereas this 343 00:21:19,960 --> 00:21:23,720 way I just say kill this process. It's kind of like task 344 00:21:23,720 --> 00:21:27,400 manager. Right. But it kills all the related pattern. Yeah. Is that kind 345 00:21:27,400 --> 00:21:31,200 of what. Let's back up a little bit because we were talking, 346 00:21:31,360 --> 00:21:35,040 we were talking best practices. So when we talking about machine speed or 347 00:21:35,040 --> 00:21:38,720 run, keep in mind that this would be 348 00:21:38,720 --> 00:21:42,560 for access controls and access control governance of the 349 00:21:42,560 --> 00:21:46,350 agents, right? So if you've got 100 agents or a 350 00:21:46,350 --> 00:21:49,590 thousand agents and they're all accessing different pieces of data 351 00:21:49,910 --> 00:21:53,710 based on the user, that's what we're providing basically at 352 00:21:53,710 --> 00:21:57,190 runtime. We're doing all of that at runtime and ensuring 353 00:21:57,670 --> 00:22:01,390 that you've got least privilege, which is a task. Right. If you think 354 00:22:01,390 --> 00:22:04,070 about thousand agents accessing multiple data sources. 355 00:22:04,950 --> 00:22:08,750 But because we have that automatically in 356 00:22:08,750 --> 00:22:12,380 the system and we have all those agents are registered, 357 00:22:12,780 --> 00:22:16,420 the kill switch portion of it is exactly what you said, Frank, is that 358 00:22:16,420 --> 00:22:20,140 we have that visibility and we can then say, okay, wait a minute, 359 00:22:20,140 --> 00:22:23,620 it's all the agents that are accessing the 360 00:22:23,620 --> 00:22:27,420 databricks data sources. So let's, let's basically 361 00:22:27,900 --> 00:22:31,620 isolate those particular agents, let's revoke them, let's 362 00:22:31,620 --> 00:22:35,460 block them. And we can do that very quickly because 363 00:22:35,460 --> 00:22:39,240 of this concept of this unified control plane. Let's look 364 00:22:39,240 --> 00:22:43,000 at what happens if you didn't have a unified control plane. Right. 365 00:22:43,000 --> 00:22:46,800 You have to have experts on data bricks, right? Access controls for databricks. 366 00:22:46,800 --> 00:22:50,080 You'd have experts that understand snowflake experts that understand 367 00:22:50,160 --> 00:22:53,799 S3, the databases. And if this 368 00:22:53,799 --> 00:22:57,640 happened to be an agent that was accessing multiple data sources. 369 00:22:57,640 --> 00:23:01,320 You'd have to coordinate between multiple experts to do 370 00:23:01,320 --> 00:23:04,920 this. And that's what causes the hour long times, the 371 00:23:04,920 --> 00:23:08,320 JIRA tickets going back and forth, the humans talking on the phone 372 00:23:09,090 --> 00:23:12,810 versus I have a unified control plane that has the ability 373 00:23:12,810 --> 00:23:16,050 to control all of this support across those data sources. 374 00:23:16,370 --> 00:23:19,970 Press this. And I don't want to oversimplify, but press this 375 00:23:19,970 --> 00:23:23,650 button. Yeah, yeah, that makes sense. Right? Because things, when things break or they go 376 00:23:23,650 --> 00:23:27,370 sideways, they never happen during convenient business hours. Right. At 377 00:23:27,370 --> 00:23:30,170 three in the morning, I'm like, oh God, do I have to call the databricks 378 00:23:30,170 --> 00:23:34,010 admin or. Oh, I've been in jobs where I 379 00:23:34,010 --> 00:23:37,330 did have a sheet. We would have a binder. It'd be like if something breaks, 380 00:23:37,940 --> 00:23:41,700 you had to do your initial first cut, obviously years before AI, and 381 00:23:41,700 --> 00:23:44,300 you'd be like, all right. And then you go down the list. The further down 382 00:23:44,300 --> 00:23:48,100 you go to list, the less you really want to call those people. 383 00:23:49,220 --> 00:23:52,820 Yeah. So if you can automate away a lot of that, and 384 00:23:52,980 --> 00:23:56,780 I think that that's really the power I see is that these dependencies 385 00:23:56,780 --> 00:24:00,500 can be embedded in your product and then the humans make the decision to 386 00:24:00,500 --> 00:24:04,100 kill. No kill. But the triaging and the troubleshooting can, 387 00:24:04,100 --> 00:24:07,820 like, you know, wait till the dawn breaks and then the next day people 388 00:24:07,820 --> 00:24:11,620 can do this as opposed to after the data goes out. Yeah. 389 00:24:12,420 --> 00:24:15,980 So because the kill switch is actually an artifact of the 390 00:24:15,980 --> 00:24:19,500 unified control plane, it wasn't that we said, hey, let's build a kill switch and 391 00:24:19,500 --> 00:24:23,180 let's build all the scaffolding behind it to be able to do that. No, we 392 00:24:23,180 --> 00:24:26,820 started off with the basically best practices as a, how do you 393 00:24:26,820 --> 00:24:29,940 unify access control, whether human or non human, 394 00:24:30,560 --> 00:24:33,920 across this? And then the artifact of this was, hey, we have this 395 00:24:33,920 --> 00:24:37,040 centralized control plane that we can isolate 396 00:24:37,760 --> 00:24:41,360 and revoke. And so it's just an artifact of 397 00:24:41,360 --> 00:24:44,560 the, I think of the benefit of having a unified control plane. 398 00:24:45,200 --> 00:24:48,560 Oh, okay. That makes sense that. Because when this 399 00:24:49,440 --> 00:24:52,800 doing some OSINT, as the coolest kids would say on 400 00:24:52,800 --> 00:24:56,000 Trustlogic. Trustlogic has been around since 2020ish. 401 00:24:56,850 --> 00:25:00,450 Right. So clearly it really was not a twinkle in Sam Altman's eye, so to 402 00:25:00,450 --> 00:25:03,970 speak. But so it sounds like this originally was 403 00:25:04,050 --> 00:25:07,650 a control plane setup and then you've pit. I think it's brilliant because I think 404 00:25:07,650 --> 00:25:11,450 it's exactly what the market needs. Because I've been in a lot of 405 00:25:11,450 --> 00:25:15,090 conversations of late where if they have agents, 406 00:25:15,090 --> 00:25:18,890 they haven't. They say they have governance, but you really only know 407 00:25:18,890 --> 00:25:22,700 if you have governance when something hits the fan, right? Yeah, 408 00:25:22,940 --> 00:25:26,500 everyone's looking at the fan. And you did make a great 409 00:25:26,500 --> 00:25:30,100 comment in terms of. We were around in 2020, we were providing 410 00:25:30,100 --> 00:25:33,900 services to these large, large banks, large pharmaceutical companies 411 00:25:34,540 --> 00:25:38,260 that have foundationally this messy underlying 412 00:25:38,260 --> 00:25:41,940 data ecosystem. Right. And our founders are all former Oracle 413 00:25:41,940 --> 00:25:45,740 folks. They understand these messy eco data environments. And so 414 00:25:45,740 --> 00:25:49,560 we solved that problem with this unified control plane. And 415 00:25:49,720 --> 00:25:53,440 naturally, as you. What is an agent? An agent is just a 416 00:25:53,440 --> 00:25:57,240 much faster, algorithmic, controlled thing that 417 00:25:57,800 --> 00:26:01,400 can act like a human. Now you got thousands of them and they act 418 00:26:02,040 --> 00:26:05,800 much faster. So it's a natural extension of 419 00:26:05,960 --> 00:26:09,600 hey, let's apply this to agents. I will say 420 00:26:09,600 --> 00:26:13,440 this, that there are a number of AI companies and we've heard this, and I 421 00:26:13,440 --> 00:26:17,230 don't mean to disparage these AI companies that are just not being form, 422 00:26:17,230 --> 00:26:20,950 but we were talking to a large hospital and they came to us and said, 423 00:26:20,950 --> 00:26:24,630 hey, look, we just met with dozens of these companies, 424 00:26:24,710 --> 00:26:28,550 these Y Combinator companies that, that are solving AI 425 00:26:28,550 --> 00:26:32,350 problems, but they don't understand our environment. They want 426 00:26:32,350 --> 00:26:36,110 everything in a single cloud and they can solve that problem really well 427 00:26:36,110 --> 00:26:39,870 and very quickly and they understand the models, et cetera, but 428 00:26:39,870 --> 00:26:43,590 they don't understand this messy environment that we have. 429 00:26:44,070 --> 00:26:47,750 And that's one of our advantages. And why I think we've been 430 00:26:47,750 --> 00:26:51,470 successful in introducing our trust AI into these larger enterprises 431 00:26:51,470 --> 00:26:54,710 is that we have that foundational, messy data 432 00:26:54,710 --> 00:26:58,470 ecosystem background. We understand that that's the life 433 00:26:58,470 --> 00:27:02,150 that we live. Interesting. That does 434 00:27:02,150 --> 00:27:05,830 fit. And that explains how the kill switch concept came about. 435 00:27:06,710 --> 00:27:10,350 And I agree with Frank. I think it's pretty cool that 436 00:27:10,350 --> 00:27:13,630 you had a solution that was solving some of these other problems 437 00:27:13,950 --> 00:27:17,510 already. And then all of a sudden AI starts going off the 438 00:27:17,510 --> 00:27:20,670 rails. You're already in the business of detecting, 439 00:27:20,910 --> 00:27:24,750 governing best practices and you found a new use case 440 00:27:24,750 --> 00:27:28,390 for it. That's pretty cool. Yeah, yeah. And it's such a 441 00:27:28,390 --> 00:27:32,030 catalyst right now. Look everywhere you look, these enterprises 442 00:27:32,510 --> 00:27:36,360 know they have to be AI first, but in 443 00:27:36,360 --> 00:27:40,160 doing it, they're discovering things. And the reality is this 444 00:27:40,160 --> 00:27:43,800 is very nascent. And I'd love to say that we have 445 00:27:43,800 --> 00:27:47,480 every answer. The trustlogix knows what's going to happen. 446 00:27:47,880 --> 00:27:51,639 But even in this conversation you brought up this concept of why not 447 00:27:51,639 --> 00:27:55,480 have an AI bot that's doing some level of control and 448 00:27:55,480 --> 00:27:59,120 triaging. These are things I think that will evolve with solutions 449 00:27:59,120 --> 00:28:02,820 like ours. One of the things that I 450 00:28:02,900 --> 00:28:06,620 really enjoy about this job is that 451 00:28:06,620 --> 00:28:10,380 every day is a new day. Right. I used to be 452 00:28:10,380 --> 00:28:14,060 in networking. I worked for the phone company and it was pretty rough. You go 453 00:28:14,060 --> 00:28:17,220 off, put a telephone pole there, hang this cable 454 00:28:17,860 --> 00:28:21,540 and every day you knew what was going to happen in 455 00:28:21,620 --> 00:28:24,900 this world. Every day changes. 456 00:28:25,300 --> 00:28:28,880 When I was at Fortanix was right when LLMs 457 00:28:28,880 --> 00:28:32,680 came out and then what was the next thing rag. Oh shoot, 458 00:28:32,680 --> 00:28:35,960 we gotta figure out how to do rag. Next thing you know is agentic. 459 00:28:36,520 --> 00:28:40,040 Things have changed so rapidly that 460 00:28:40,360 --> 00:28:43,880 it makes every day exciting. And that's what I enjoy about 461 00:28:43,960 --> 00:28:47,080 being in this particular. At this particular time, 462 00:28:47,560 --> 00:28:51,320 in this particular startup, addressing these particular problems. 463 00:28:52,200 --> 00:28:55,970 Because you don't know what's around the corner. You really don't. And 464 00:28:55,970 --> 00:28:59,810 anyone that says they do is lying to you. Right. And you can make 465 00:28:59,810 --> 00:29:03,050 good guesses but. And kind of educated guesses and be 466 00:29:03,210 --> 00:29:06,650 directionally. Right. But no. And another thing you brought up, a lot of this really 467 00:29:06,650 --> 00:29:10,330 goes back to the fundamentals, right. Tried and tested enterprise tech 468 00:29:10,330 --> 00:29:13,930 and best practices that we people really 469 00:29:14,010 --> 00:29:17,850 don't know how important they are until they try to make something 470 00:29:17,850 --> 00:29:21,130 an agent. Right. People don't realize how bad their data is until they try to 471 00:29:21,130 --> 00:29:24,520 wire up an AI to it. Right. Oh, because I think maybe it's the machine 472 00:29:24,520 --> 00:29:28,320 speed makes you see the inadequacies pretty 473 00:29:28,320 --> 00:29:32,040 quickly. Right. Using the car analogy, the 474 00:29:32,040 --> 00:29:35,800 Dodge Challenger, right. You don't know the Dodge Challenger has a problem until you're 475 00:29:35,800 --> 00:29:39,440 going maybe a little faster than the speed limit than you really know. 476 00:29:39,440 --> 00:29:43,000 Maybe they don't make carburetors anymore, but you really don't know 477 00:29:43,320 --> 00:29:47,040 what's wrong with your system until you. You're going until it's up and 478 00:29:47,040 --> 00:29:50,830 going. I think that's the same with enterprise tech and data particularly. This 479 00:29:50,830 --> 00:29:54,630 is why I always make a big deal to talk about data engineering as part 480 00:29:54,630 --> 00:29:57,750 of this podcast is it's foundational to this. 481 00:29:58,230 --> 00:30:01,990 I have a slide deck and I've talked about Maslow's hierarchy of needs. But for 482 00:30:01,990 --> 00:30:05,790 AI, right? AI is the very top. And I would say agentic AI 483 00:30:05,790 --> 00:30:08,790 is probably now the top layer, but at the bottom of it 484 00:30:09,430 --> 00:30:13,270 we have power, networking and infrastructure. The former networking guy, right. 485 00:30:13,350 --> 00:30:17,090 That's important. But also kind of like a big middle part is 486 00:30:17,170 --> 00:30:20,970 the data engineering and kind of just your best practices 487 00:30:20,970 --> 00:30:24,690 there. And I've seen organizations, both government and private, 488 00:30:25,170 --> 00:30:28,810 you really get to see how well the. How well the dish is put 489 00:30:28,810 --> 00:30:32,530 together when we start AI ing stuff because then it really 490 00:30:32,530 --> 00:30:36,290 exposes, I think the inaccurate pretty quickly. Yeah. 491 00:30:36,370 --> 00:30:40,210 Never heard anyone explain that using Maslov's 492 00:30:40,210 --> 00:30:43,530 hierarchy of needs. And if it's okay, I'm going to borrow that. Go ahead, use 493 00:30:43,530 --> 00:30:47,250 Ali. I'll even use the slide. Just give a shout out to the podcast. 494 00:30:47,250 --> 00:30:51,090 That's all I ask. I absolutely will. Great example. And 495 00:30:51,090 --> 00:30:54,850 I know it's cliche now and they'll say if you don't have a data strategy, 496 00:30:54,850 --> 00:30:58,370 you don't have an AI strategy, but it's absolutely the truth. You've got 497 00:30:58,930 --> 00:31:02,130 this pyramid of which you've got these foundational 498 00:31:02,210 --> 00:31:05,850 principles that all lead up to now I can deploy my 499 00:31:05,850 --> 00:31:09,410 AI. A, that it's going to work and B, 500 00:31:09,410 --> 00:31:13,260 it's going to be secure. The thing that you just brought up in terms of 501 00:31:13,260 --> 00:31:16,820 pressure testing the system, part of it is pressure testing the 502 00:31:16,820 --> 00:31:20,420 technology and the process. Right. Do you have a 503 00:31:20,420 --> 00:31:24,100 run a run speed process that can handle these? 504 00:31:24,500 --> 00:31:28,220 But the other thing is the knowledge. Right. It comes back to your, the 505 00:31:28,220 --> 00:31:31,580 point that you made of the guy that wrote the resume. He just didn't know 506 00:31:31,580 --> 00:31:34,740 how to prompt the LLM correctly. 507 00:31:35,140 --> 00:31:38,790 When people say AI is going to do away with all these jobs, the 508 00:31:38,790 --> 00:31:42,510 reality is you have this level of tribal knowledge and this level 509 00:31:42,510 --> 00:31:46,270 of human experience that AI just doesn't have. It 510 00:31:46,270 --> 00:31:49,470 doesn't have that context. And so it's going to require 511 00:31:49,870 --> 00:31:53,470 somebody that actually knows what the heck they're doing to be able to do that 512 00:31:53,470 --> 00:31:57,270 prompt engineering. You just can't get somebody that was a journalist 513 00:31:57,270 --> 00:32:00,350 major out of college to come in and say, okay, you're really good with words. 514 00:32:00,830 --> 00:32:04,480 Go make some genetic testing or some clinical trial. I'll 515 00:32:04,480 --> 00:32:07,880 LLM, no, no way. It's going to require an actual 516 00:32:08,280 --> 00:32:11,880 a doctor or a researcher to, to foundationally develop 517 00:32:11,960 --> 00:32:15,280 that LLM or that AI model to where it can do what it needs to 518 00:32:15,280 --> 00:32:18,280 do. Absolutely. I also think that 519 00:32:18,920 --> 00:32:22,400 you mentioned the journalist. Right. I think people who study 520 00:32:22,400 --> 00:32:25,880 language, whether they're literature majors, journalism 521 00:32:25,880 --> 00:32:28,840 majors, I think they actually have a unique position 522 00:32:30,050 --> 00:32:33,890 in terms of writing, their writing ability. Assuming they don't entirely rely 523 00:32:33,890 --> 00:32:37,210 on AI, they understand the nuance of language in ways that very few people can. 524 00:32:37,210 --> 00:32:40,930 So I think lawyers in particular also might have an advantage in being. 525 00:32:41,090 --> 00:32:43,930 I don't want to say that they would make great prompt engineers because I think 526 00:32:43,930 --> 00:32:47,570 that really undersells the, the ability. But you know where I'm going with that, 527 00:32:47,570 --> 00:32:51,090 right? Anyone who has understanding about linguistic 528 00:32:51,090 --> 00:32:54,610 nuances, whether that's for artistic purposes or legal purposes, 529 00:32:54,690 --> 00:32:58,360 I think has a unique advantage 530 00:32:58,360 --> 00:33:01,600 over normies who don't think about language in that way. 531 00:33:02,720 --> 00:33:06,320 Percent agree with that. And again, 532 00:33:06,400 --> 00:33:10,200 back to the point I was making, I think someone that's great at 533 00:33:10,200 --> 00:33:13,800 language link, linguistics, etc. They'll be great 534 00:33:13,800 --> 00:33:17,400 at figuring out how to make this prompt and shape the 535 00:33:17,400 --> 00:33:21,120 prompt such that it gets you the desired outcome. They 536 00:33:21,120 --> 00:33:24,960 just need to get that experience with what exactly 537 00:33:24,960 --> 00:33:28,800 are they writing that prompt for. But you could imagine that someone 538 00:33:28,800 --> 00:33:32,600 that's good with words, but also has a kind of a technical brain 539 00:33:32,600 --> 00:33:36,280 or a creative brain, they'll be able to. Things are going 540 00:33:36,280 --> 00:33:39,800 to be developed that are way beyond what we're 541 00:33:39,800 --> 00:33:43,560 envisioning today. And that's what I think is so exciting 542 00:33:43,560 --> 00:33:47,360 about AI. For the first time, anybody can actually 543 00:33:47,360 --> 00:33:50,330 be an application developer now. It's not that 544 00:33:50,570 --> 00:33:53,850 everybody will be, but anybody can be. 545 00:33:54,410 --> 00:33:57,970 And if you're exceptionally creative and you're good with words and 546 00:33:57,970 --> 00:34:01,570 articulating things and formulating ideas, think 547 00:34:01,570 --> 00:34:05,249 about the types of applications that can be built. You don't have to 548 00:34:05,249 --> 00:34:08,570 know C and Python and this and that. You're not 549 00:34:08,570 --> 00:34:12,330 trapped with the scaffolding that you have to know now 550 00:34:12,330 --> 00:34:15,450 you can just let your brain run free and you can start to 551 00:34:16,250 --> 00:34:18,880 ideate. Is that the right word? You can 552 00:34:19,280 --> 00:34:22,880 conceptualize these things. And this 553 00:34:22,880 --> 00:34:26,640 happened to me, right? I was telling Kim, I was trying to figure out how 554 00:34:26,640 --> 00:34:30,320 can I create someone to look at my blogs 555 00:34:30,400 --> 00:34:34,119 and see if they're reasonable blogs for could be published 556 00:34:34,119 --> 00:34:37,960 in EE Times or whatever. And so I went into Claude and I 557 00:34:37,960 --> 00:34:41,680 said, hey, can you help me create folks that can critique 558 00:34:41,680 --> 00:34:44,870 my blog? And it literally came back with an app. 559 00:34:45,420 --> 00:34:49,020 Is that nice? Here's your app. Upload your document 560 00:34:49,020 --> 00:34:52,860 here, and here's seven Personas, one from EE Times, one from Harvard 561 00:34:52,860 --> 00:34:56,700 Business Review, one from LinkedIn, and it will critique this. 562 00:34:57,420 --> 00:35:01,220 And so I thought, wow, this is cool. So I refined that app to 563 00:35:01,220 --> 00:35:04,740 where now I just upload my blog and I get 10 different reviews and it 564 00:35:04,740 --> 00:35:07,820 says, here's the red line of where they think your blog 565 00:35:08,220 --> 00:35:11,500 doesn't pass muster. And I was like, I didn't even go in 566 00:35:11,990 --> 00:35:15,270 intending to build an app, but I did. I said, look what I did. I 567 00:35:15,270 --> 00:35:18,990 built an app. No, it's very powerful, right? It's very powerful 568 00:35:18,990 --> 00:35:22,830 in terms of really leveraging AI, particularly if you have them do 569 00:35:22,830 --> 00:35:26,590 the actor critic kind of approach and you can iterate really quickly. 570 00:35:26,590 --> 00:35:30,390 I have a Claude skill that will I give it a presentation, 571 00:35:30,390 --> 00:35:34,190 Here's a presentation, make it more, or I get 572 00:35:34,190 --> 00:35:37,940 a template for an event, whether that's Red Hat Summit, Databricks Summit, and I'll 573 00:35:37,940 --> 00:35:41,740 be like, here's a template. Convert this and it does it now. Could 574 00:35:41,740 --> 00:35:44,900 I have done that? Yeah, I could have done that. I can do that. I 575 00:35:44,900 --> 00:35:47,820 can go get a cup of coffee, I can walk the dogs while it does 576 00:35:47,820 --> 00:35:51,140 that. It's that type of thing. We also have techniques where 577 00:35:51,540 --> 00:35:54,900 I actually wrote a command line tool a couple of years ago 578 00:35:55,540 --> 00:35:59,380 called Dingo and been wanting to get more 579 00:35:59,380 --> 00:36:02,580 people to use it because I find it very useful. It's a command line tool 580 00:36:02,580 --> 00:36:04,600 written in originally.net 581 00:36:06,440 --> 00:36:10,040 and then I migrated it to Python and then not everyone's into 582 00:36:10,040 --> 00:36:13,640 command line tooling. So I actually have now a 583 00:36:13,640 --> 00:36:17,480 process where we are going to make it like a SaaS, right? Where it basically 584 00:36:17,640 --> 00:36:20,440 you could train it to write like you, similar to what you did like with 585 00:36:20,440 --> 00:36:24,280 the critics. And that's basically a similar thing. If folks want to check it out, 586 00:36:24,280 --> 00:36:27,720 you go to thedingo AI. The Dingo 587 00:36:27,800 --> 00:36:31,380 AI. The Dingo AI. And we're going to launch 588 00:36:31,380 --> 00:36:35,140 out public beta soon, hopefully by the time this goes live. And 589 00:36:35,140 --> 00:36:38,620 it's actually named after one of my dogs who looks like a Dingo. There you 590 00:36:38,620 --> 00:36:42,380 go, There you go. But you think you're onto something, right? You have, you 591 00:36:42,380 --> 00:36:46,180 have the ability now to take an idea and then get it to 592 00:36:46,180 --> 00:36:49,500 where you need to go. Now a lot of the critics of 593 00:36:50,140 --> 00:36:53,700 vibe coding, which is, you know, is that. Will that scale to millions of 594 00:36:53,700 --> 00:36:57,420 users? Do you need it to scale to millions of users? Right? 595 00:36:57,870 --> 00:37:01,150 You built it for yourself, right? You have a tool, 596 00:37:01,790 --> 00:37:04,550 you're obviously a CEO of a tech company, you can probably find a way to 597 00:37:04,550 --> 00:37:07,390 make it scale a million. But do you 598 00:37:08,830 --> 00:37:12,270 need to? You don't. But I think to your point earlier 599 00:37:12,430 --> 00:37:16,070 about you still need the human experience, the 600 00:37:16,070 --> 00:37:19,550 human capability to basically pressure test it and to 601 00:37:19,870 --> 00:37:22,990 make it scale and make it enterprise grade and. 602 00:37:23,660 --> 00:37:27,220 But I do think it speeds up the process if you look at the 603 00:37:27,220 --> 00:37:31,020 amount of time savings, grab a cup of coffee while it's doing a lot 604 00:37:31,020 --> 00:37:34,340 of the mundane work, but you come back, you look at it, you're 605 00:37:34,340 --> 00:37:37,980 validating it, but you've saved so much time. But 606 00:37:37,980 --> 00:37:41,660 back to the foundational aspect of AI and why I 607 00:37:41,660 --> 00:37:45,340 currently exist at Trustlogix is I think what's 608 00:37:45,340 --> 00:37:48,860 happening is you've got, within enterprises, you have folks that are 609 00:37:48,860 --> 00:37:52,610 discovering this capabilities, ability and I, I can't remember 610 00:37:52,610 --> 00:37:56,170 if Andy, if you had mentioned it or. But you've got folks that are used 611 00:37:56,170 --> 00:37:59,210 to now doing anything they want 612 00:37:59,770 --> 00:38:03,450 and bypassing the governance, bypassing the security. Because look, 613 00:38:03,450 --> 00:38:06,930 I'm about to develop something and this is the most important thing for me to 614 00:38:06,930 --> 00:38:10,770 do. That's a problem when you get when you're in an enterprise 615 00:38:10,770 --> 00:38:13,850 environment, that's where we're trying to apply that 616 00:38:14,090 --> 00:38:17,630 protection level that says, hey, we don't want to cramp your style. 617 00:38:17,710 --> 00:38:21,430 We want you to be able to develop as fast as you want. Just register 618 00:38:21,430 --> 00:38:25,030 your agent and that will take care of everything for 619 00:38:25,030 --> 00:38:28,590 you. It'll make sure that the data is secure. It'll make sure you don't go 620 00:38:28,590 --> 00:38:32,430 to access to data that you shouldn't be accessing. So we're trying to do 621 00:38:32,430 --> 00:38:36,110 it in a very non intrusive way so that 622 00:38:36,110 --> 00:38:39,750 people can be creative and develop those applications and do it 623 00:38:39,750 --> 00:38:43,480 quickly without having to manually go in 624 00:38:43,480 --> 00:38:47,200 and override things. We're trying to, we're trying to make it very non 625 00:38:47,200 --> 00:38:50,080 intrusive. An AI sandbox, if you will. 626 00:38:52,720 --> 00:38:55,120 Yeah, these are great ideas, 627 00:38:56,400 --> 00:39:00,000 this concept, the sandbox. But as let's just 628 00:39:00,000 --> 00:39:03,720 apply the governance layer. Make it very easy for whether you're an 629 00:39:03,720 --> 00:39:07,560 engineer or whether you're a data analytics person. Just make 630 00:39:07,560 --> 00:39:10,000 it easy so that you don't have to think about it, you don't have to 631 00:39:10,400 --> 00:39:14,080 create a JIRA ticket and ask permission. We just make sure that least 632 00:39:14,080 --> 00:39:17,720 privileges is applied. But this concept of a sandbox, 633 00:39:17,720 --> 00:39:21,560 that's interesting. I'm getting lots of ideas on it anytime, come back whenever 634 00:39:21,560 --> 00:39:25,320 you need more ideas. But, but also too, I think what you're providing is 635 00:39:25,320 --> 00:39:28,840 the really the sweet spot of innovation and safety. Right? 636 00:39:28,840 --> 00:39:32,640 Yeah. And I think that is sorely needed in this world. Right. 637 00:39:32,720 --> 00:39:36,360 You look at the yolo. Is that what it's called, yolo or is that what 638 00:39:36,360 --> 00:39:39,800 they. You look at openclaw. Right. And I have an openclaw instance and things like 639 00:39:39,800 --> 00:39:43,650 that. And it was intentionally made without security in mind. So we can do 640 00:39:43,650 --> 00:39:47,170 anything it wanted to. Now there are times when, 641 00:39:47,250 --> 00:39:50,610 hey, that's fun. You have a little home lab, you're messing around, it tells you 642 00:39:50,610 --> 00:39:53,410 for me, I have it tell me the pollen report and the weather every day 643 00:39:53,730 --> 00:39:56,090 and then at the end of the day it gives me all the AI news 644 00:39:56,090 --> 00:39:59,290 stories, right. I don't have it hooked up to my email. I don't have it 645 00:39:59,290 --> 00:40:02,530 hooked up to my home, my bank account. 646 00:40:02,930 --> 00:40:06,580 I'm okay in that sense. But I would love to have it read my email. 647 00:40:06,580 --> 00:40:10,220 Right. You see this little thing there? That's the data. That little lamp there, for 648 00:40:10,220 --> 00:40:13,860 those who are listening is a, a lamp I just picked up on Amazon 649 00:40:13,860 --> 00:40:17,540 and I actually am going to vibe code it. Whatever stocks I'm tracking that day, 650 00:40:17,620 --> 00:40:20,540 if it goes up, it'll turn green when it goes up and then it'll turn 651 00:40:20,540 --> 00:40:24,220 red when it goes down. Just a little fun thing to do. Right. And I 652 00:40:24,220 --> 00:40:27,780 get, I Vibe coded that, right? There's no published SDK for that. There is some 653 00:40:27,780 --> 00:40:31,620 guy on GitHub who has a similar product that he's kind of reverse 654 00:40:31,620 --> 00:40:35,410 engineered it. But I basically have the AI while we're talking 655 00:40:35,410 --> 00:40:39,130 is going and hitting the endpoint because it gets an 656 00:40:39,130 --> 00:40:42,250 IP address and then messing around with it and seeing what it does. And it 657 00:40:42,250 --> 00:40:45,330 basically will ask me, hey, did it blink this color? Did it do that? These 658 00:40:45,330 --> 00:40:48,890 are all things I'm doing while I'm doing something of higher value, which is talking 659 00:40:48,890 --> 00:40:52,530 to you. Right. And this is something that I think is really the power of 660 00:40:52,850 --> 00:40:56,610 this AI coding. Yeah, absolutely. One thing I 661 00:40:56,610 --> 00:41:00,420 do want to mention. Yeah. We were talking about non intrusive and how do you 662 00:41:00,420 --> 00:41:04,180 do that kind of vibe coding? One of the things that we found and 663 00:41:04,180 --> 00:41:07,580 we were talking to development engineers and when you have to, 664 00:41:07,900 --> 00:41:11,620 we have to stop and you have to put in security. Right? 665 00:41:11,620 --> 00:41:15,260 You have to make specific MCP calls. You have to put certain 666 00:41:15,340 --> 00:41:18,620 data access restrictions. You're doing that at the code level. Right. 667 00:41:19,020 --> 00:41:22,860 And so our concept was how can we shift left and 668 00:41:23,020 --> 00:41:26,810 give engineers a simple MCP call, One 669 00:41:26,810 --> 00:41:30,530 MCP call that basically then can secure that 670 00:41:30,530 --> 00:41:34,170 entire line of code. So the engineer can sit there and either vibe 671 00:41:34,170 --> 00:41:37,770 code or just develop the agent without having to worry about 672 00:41:38,570 --> 00:41:42,410 making very specific access control calls, MCP 673 00:41:42,410 --> 00:41:46,210 calls and effectively shift left. And so that's how 674 00:41:46,210 --> 00:41:49,890 we're, how we're trying to drastically simplify and make it 675 00:41:49,890 --> 00:41:53,510 non intrusive for a development engineer that frankly doesn't give a crap about 676 00:41:53,510 --> 00:41:56,990 security. They're just trying to get their job. They just want their lamp to turn 677 00:41:56,990 --> 00:42:00,670 green or red, depending on the stock. And so that's our 678 00:42:00,670 --> 00:42:04,390 mindset, is making sure it's non intrusive. No. 679 00:42:04,390 --> 00:42:07,510 And that's honestly the best way for security because developers, 680 00:42:08,069 --> 00:42:11,670 despite everything that's happened and if they haven't learned by now, they're not going to 681 00:42:11,670 --> 00:42:15,190 learn. Right. Security is often an afterthought, 682 00:42:15,190 --> 00:42:18,310 Right. I hate to throw fellow developers under the bus, 683 00:42:18,810 --> 00:42:22,410 but it's true, right. My wife works in cybersecurity. 684 00:42:22,570 --> 00:42:26,330 Funny story, actually. Again, this I think ties to the whole Vibe coding thing. 685 00:42:26,650 --> 00:42:30,250 So I have dingo lives on Vercel. Vercel had their breach. 686 00:42:31,050 --> 00:42:33,690 But one of the things that I was thinking in the back of my head 687 00:42:33,690 --> 00:42:37,450 was I should probably have things encrypted at rest on 688 00:42:37,450 --> 00:42:41,050 the credentials database. And I Asked 689 00:42:41,130 --> 00:42:44,570 the, I asked Claude, how long will it take? I'll take an extra 30 minutes. 690 00:42:44,920 --> 00:42:47,920 Like just do it because God forbid there's ever a breach. My wife will never 691 00:42:47,920 --> 00:42:51,720 let me live it down. And fortunately I did. Somebody 692 00:42:51,720 --> 00:42:54,680 asked me hey, or were you affected? I was like no, we had everything encrypted 693 00:42:54,680 --> 00:42:58,160 at rest. Now in the past that would have been more than a 30 694 00:42:58,160 --> 00:43:01,840 minute detour for an agent. Right. It would have been actual work 695 00:43:01,840 --> 00:43:05,480 which as it when I was doing software engineering I 696 00:43:05,480 --> 00:43:08,400 probably would have said well you know what that'll be. I'll put that on the 697 00:43:08,400 --> 00:43:11,670 backlog. I'll deal with that later. Right now you can just kind of. Or if 698 00:43:11,670 --> 00:43:15,270 you really want to get proactive you can have the, you can tell with 699 00:43:15,270 --> 00:43:18,910 the Claude file, tell it to no always enforce security best principles. 700 00:43:18,910 --> 00:43:22,670 So it'll kind of work around that. So I think that despite all the fear, 701 00:43:22,670 --> 00:43:25,630 despite all the loathing, I think vibe coding or 702 00:43:26,349 --> 00:43:30,110 this sort of thing will make software engineering easier 703 00:43:30,110 --> 00:43:33,830 and safer and accessible to a lot more people. Will that 704 00:43:33,830 --> 00:43:36,910 eliminate jobs in the near term? Historically, 705 00:43:37,450 --> 00:43:40,890 automation has grown the economy and grown the job market. 706 00:43:41,450 --> 00:43:45,130 I have faith in the trend line. Yeah, I think there's. It is a sea 707 00:43:45,130 --> 00:43:48,890 change but I think it's just going to change job 708 00:43:48,890 --> 00:43:52,250 descriptions. I think that's what's going to happen. It's going to change job 709 00:43:52,250 --> 00:43:55,890 descriptions and you have to have the AI talent 710 00:43:55,890 --> 00:43:59,730 to stay alive in the new economy. But it 711 00:43:59,730 --> 00:44:03,050 doesn't preclude the fact that you also need to understand 712 00:44:03,370 --> 00:44:07,220 foundational principles. Back to your earlier point, it is 713 00:44:07,220 --> 00:44:10,900 interesting that you had talked about encryption. That was my life for about five years 714 00:44:10,900 --> 00:44:14,500 and the fact that you've made it much easier because it's a 715 00:44:14,500 --> 00:44:18,180 scary subject and a lot of folks are intimidated by encryption 716 00:44:18,180 --> 00:44:21,780 and the concept of post quantum and all that. And you're right, 717 00:44:22,100 --> 00:44:25,780 this can drastically turning it into a 30 minute problem. 718 00:44:26,340 --> 00:44:29,940 That is unbelievable. And the AI and again 719 00:44:29,940 --> 00:44:33,640 encryption is something that does scare a lot of people. Many moons ago I'm way 720 00:44:33,640 --> 00:44:36,360 older than I care to admit but even before I knew Andy and I known 721 00:44:36,360 --> 00:44:40,160 Andy 20 years or more, I was applied for a 722 00:44:40,240 --> 00:44:44,080 small business incentivis sbir like it was sba thing 723 00:44:44,080 --> 00:44:47,360 and it was about detecting steganographic basically 724 00:44:47,600 --> 00:44:50,320 steganography's idea. You hide data and other packets of data. 725 00:44:51,200 --> 00:44:54,880 Oh wow. And it was an RFP for I believe it was the Air force. 726 00:44:54,960 --> 00:44:58,800 And I didn't win the contract but I learned so much about 727 00:44:58,800 --> 00:45:02,480 cryptography and mathematics and how to apply that Actually, in Net, 728 00:45:02,820 --> 00:45:05,780 that opened up so many other doors for me, including 729 00:45:06,420 --> 00:45:09,540 my wife. Actually, I met my wife when we were dating. 730 00:45:10,180 --> 00:45:13,940 We were talking and she. I mentioned steganography and she knew 731 00:45:13,940 --> 00:45:17,340 exactly what steganography was. And I'm like thinking in the back of my head, that's 732 00:45:17,340 --> 00:45:20,940 the woman I want to marry, is the lady who knows who 733 00:45:20,940 --> 00:45:24,780 steganography is. Because nobody knew. Not even techies know what steganography is. Most. Most 734 00:45:24,780 --> 00:45:28,460 normal people don't know that. But you're right, encryption does scare off a lot of 735 00:45:28,460 --> 00:45:32,250 folks, but it's so vital and. And we're getting close to time here. I 736 00:45:32,250 --> 00:45:35,170 want to be respectful. I actually have another podcast called Impact Quantum where we do 737 00:45:35,170 --> 00:45:38,690 talk about quantum computing and PQC post. Quantum computing comes up a 738 00:45:38,690 --> 00:45:42,490 lot because it definitely. For those that really want 739 00:45:42,490 --> 00:45:46,169 the two, an easy introduction to what that 740 00:45:46,169 --> 00:45:49,530 means. The Y files. And we'll put this in the show note. The Y files, 741 00:45:49,530 --> 00:45:53,370 which is an excellent YouTube channel, has a whole show based on. It's 742 00:45:53,370 --> 00:45:57,090 a little dramatically punched up, but it has a whole thing of what would happen 743 00:45:57,090 --> 00:46:00,530 when Q day happens. The idea that a quantum computer could break all conventional 744 00:46:00,530 --> 00:46:04,230 encryption and the proverbial, you don't know what hits the fan. 745 00:46:04,230 --> 00:46:07,990 It's an interesting thing. It's a little overly dramatic, but it's not 746 00:46:07,990 --> 00:46:11,750 that far off from the truth. Just add AI plus 747 00:46:11,910 --> 00:46:15,550 quantum computing together. Kaboom, man. There's your 748 00:46:15,550 --> 00:46:19,310 next. Yeah, there's your next thing. So I put. I'll put the graphic 749 00:46:19,310 --> 00:46:22,950 in the show notes, but in the comments, I did paste the slide of 750 00:46:22,950 --> 00:46:26,710 Maslow's hierarchy of needs where the AI hierarchy of AI needs. 751 00:46:27,150 --> 00:46:30,550 So feel free to use that. And where can folks find out more about you 752 00:46:30,550 --> 00:46:33,870 and what you're up to? Yeah, so just go to TrustLogix 753 00:46:34,030 --> 00:46:37,830 AI and we are introducing new products at 754 00:46:37,830 --> 00:46:41,070 the speed of AI. But definitely. And then join us at 755 00:46:41,230 --> 00:46:44,350 Snowflake Summit. Join us at Databrick Summit. We'll be there. 756 00:46:44,750 --> 00:46:48,310 I'm hosting, folks. I'm actually sitting on my boat on the 757 00:46:48,310 --> 00:46:51,990 bay. I'll be hosting customers on Monday and 758 00:46:51,990 --> 00:46:55,620 Tuesday. And then also at Databricks, we take folks out. 759 00:46:56,020 --> 00:46:59,660 We have fine wine and hors d' oeuvres, and then we talk about the latest 760 00:46:59,660 --> 00:47:03,060 trends in AI and in data governance. Nice. 761 00:47:03,380 --> 00:47:06,940 That sounds cool. Yeah. And with that, I'll let the outro music 762 00:47:06,940 --> 00:47:10,580 play. That was great. That was great, 763 00:47:10,580 --> 00:47:11,620 man. That was awesome.