1 00:00:14,560 --> 00:00:18,480 Hello, this is Eva and you're listening to the In Between Tech 2 00:00:18,480 --> 00:00:21,400 and Trust podcast. And in this week's episode, I 3 00:00:21,400 --> 00:00:24,640 get quite exciting because we talk about one of the most 4 00:00:24,640 --> 00:00:27,640 pressing topics in the tech and the ice era. 5 00:00:27,720 --> 00:00:33,160 And these are AI agents. And if I had to give it a title, 6 00:00:33,160 --> 00:00:37,240 then it's like the one O 1 of agentic systems for 7 00:00:37,240 --> 00:00:40,600 organizations. And for that, I have Anthony 8 00:00:40,600 --> 00:00:45,240 Alcatraz, author of the O'Reilly book on agentic graph rag, a 9 00:00:45,240 --> 00:00:49,760 senior figure of AWS pushing agentic AI adoption at 10 00:00:49,760 --> 00:00:53,360 enterprise scale. And he's also a business Angel 11 00:00:53,440 --> 00:00:56,880 working with early stage startups that are actually 12 00:00:56,880 --> 00:01:01,400 building, scaling and incorporating layers of the AI 13 00:01:01,400 --> 00:01:04,440 economy. And so with his background and 14 00:01:04,440 --> 00:01:09,120 his daily operations, Anthony sits Anon quite unusual in 15 00:01:09,120 --> 00:01:13,080 between an intersection because he did deep enough in the 16 00:01:13,320 --> 00:01:17,600 technical architecture to write a book and to co-author it, but 17 00:01:17,600 --> 00:01:22,880 he's also close enough to the business reality to know where 18 00:01:22,880 --> 00:01:27,120 organizations, corporations and teams currently are struggling. 19 00:01:27,280 --> 00:01:32,720 And so in our conversation, we get like so knee deep into what 20 00:01:32,720 --> 00:01:38,120 actually changes when it becomes more of a systems perspective 21 00:01:38,160 --> 00:01:42,640 and when it also starts making decisions and taking actions on 22 00:01:42,640 --> 00:01:47,240 your behalf when agents take over and do the research and 23 00:01:47,280 --> 00:01:49,960 feed it back into an organizational output. 24 00:01:50,080 --> 00:01:53,840 And so the tension that we looked into is that a genetic 25 00:01:53,840 --> 00:01:58,160 systems, if they are already, you know, incorporated into the 26 00:01:58,160 --> 00:02:02,280 organizational structures in a way that users cannot see them. 27 00:02:02,360 --> 00:02:06,080 The biggest question that we need to discuss is how we build 28 00:02:06,080 --> 00:02:11,039 trust into something that is invisible and that is yet to be 29 00:02:11,039 --> 00:02:17,040 defined and to be educated. Welcome Anthony to the In 30 00:02:17,040 --> 00:02:19,520 Between Tech and Trust podcast. Thank you for the invite. 31 00:02:19,600 --> 00:02:23,480 I want to start with you at the intersection of technology and 32 00:02:23,480 --> 00:02:27,720 trust, and particularly when you think about it and from a 33 00:02:27,720 --> 00:02:31,520 context of the Gentek AI, but also from your collaboration 34 00:02:31,520 --> 00:02:34,200 with start-ups, what's the first thing that comes to your mind? 35 00:02:34,480 --> 00:02:37,080 In term of the technology and trust that is to say especially 36 00:02:37,080 --> 00:02:41,440 for for I think currently it's a very clear change, OK. 37 00:02:41,800 --> 00:02:45,920 Before we had system that were mostly reactive, I would say. 38 00:02:46,080 --> 00:02:48,560 So you, you would ask a question, you would get an 39 00:02:48,560 --> 00:02:51,040 answer. Now, which change considerably 40 00:02:51,040 --> 00:02:54,600 with agent CKI, we can't define agent CKI, but but system 41 00:02:54,600 --> 00:02:57,360 basically that's observe, reason and act. 42 00:02:57,400 --> 00:02:59,920 So when will you have such system? 43 00:02:59,960 --> 00:03:03,080 I would say the trust, you need much more trust than before 44 00:03:03,120 --> 00:03:07,000 because basically you, you are delegating what we used to do as 45 00:03:07,360 --> 00:03:12,000 I mean, people human to a system that is automated that will make 46 00:03:12,000 --> 00:03:15,600 that, that which reason for us and that will take decision. 47 00:03:16,240 --> 00:03:20,200 So if you think about implementing that in regulated 48 00:03:20,280 --> 00:03:24,400 industries like, I don't know, healthcare or even legal, legal 49 00:03:24,400 --> 00:03:28,440 system or insurance and that it says you, you need to, to have 50 00:03:28,720 --> 00:03:32,160 a, a clear trust in what's the system is deciding. 51 00:03:32,600 --> 00:03:35,600 And I saw to put, I would say the human expertise that is 52 00:03:35,600 --> 00:03:39,600 still needed at the right place. So it changed considerably. 53 00:03:39,600 --> 00:03:43,360 I would say the data governance I wanted and even the governance 54 00:03:43,360 --> 00:03:44,840 I wanted. I think there is a problem 55 00:03:44,840 --> 00:03:49,320 currently because we don't have, I would say currently there is 56 00:03:49,320 --> 00:03:54,080 no clear framework or governance that has emerged to manage at 57 00:03:54,080 --> 00:03:58,640 large scale such system. I mean currently we can 58 00:03:58,640 --> 00:04:01,760 generate, we can like build, build the system, scale it. 59 00:04:01,920 --> 00:04:05,120 This is what we are doing at Entropic, open air etcetera. 60 00:04:05,240 --> 00:04:09,040 But I would say the governance infrastructure of it is not 61 00:04:09,240 --> 00:04:12,080 quite there. We, we, we struggle for example, 62 00:04:12,080 --> 00:04:14,400 to secure the security risk etcetera, so. 63 00:04:14,800 --> 00:04:18,240 Particularly when it comes to building a genetic systems, you 64 00:04:18,240 --> 00:04:22,200 work with startups, but you also help them scale into like 65 00:04:22,200 --> 00:04:25,280 environments that then also enable them to grow and to 66 00:04:25,280 --> 00:04:27,840 transition into this enterprise context. 67 00:04:28,400 --> 00:04:33,200 What fundamentally changes when AI shifts from a single use case 68 00:04:33,200 --> 00:04:37,680 and maybe also in an early stage company into a system setup that 69 00:04:37,680 --> 00:04:41,080 then also need to be robust, need to be trusted and need to 70 00:04:41,120 --> 00:04:45,840 be able to also be, you know, as flexible as the market evolves, 71 00:04:45,880 --> 00:04:49,400 but as stable as the company or organization requires? 72 00:04:49,520 --> 00:04:52,240 Yes, yes, yes. So, so it's complicated, OK, 73 00:04:53,240 --> 00:04:56,200 frankly to to put into production agents. 74 00:04:56,360 --> 00:04:59,360 So you need to understand that before we, we had system that 75 00:04:59,360 --> 00:05:02,160 were stateless, OK. So basically between 76 00:05:02,160 --> 00:05:05,840 interaction, it didn't have to remember, didn't have to 77 00:05:05,840 --> 00:05:08,440 forcibly connect with internal data, etcetera. 78 00:05:08,520 --> 00:05:10,600 Agent 6 needs to be stateful to work. 79 00:05:11,040 --> 00:05:14,440 So you need to build a a system that's basically maintain 80 00:05:14,440 --> 00:05:16,760 context. So to maintain context, you need 81 00:05:16,760 --> 00:05:20,000 to access a context, you need to store it, you need to maybe 82 00:05:20,000 --> 00:05:23,680 classify it to understand it. On top of that, most of this 83 00:05:23,680 --> 00:05:26,880 system are building up as some memory capabilities. 84 00:05:26,880 --> 00:05:29,960 So you need a memory system, you need to manage this memory. 85 00:05:30,280 --> 00:05:33,880 OK, currently I will cite an example from Munich Cogni, for 86 00:05:33,880 --> 00:05:37,840 example, this it's a graph databases building memory system 87 00:05:37,840 --> 00:05:40,360 for for agents. And this agent need also to 88 00:05:40,360 --> 00:05:42,600 learn from VZ. So basically you need to build 89 00:05:42,600 --> 00:05:44,960 an evaluation situation system. You need to make sure that the 90 00:05:45,280 --> 00:05:48,080 agent that you are putting into production are are performing 91 00:05:48,080 --> 00:05:50,560 well. And to build such evaluation 92 00:05:50,560 --> 00:05:52,640 system is still is still difficult. 93 00:05:52,680 --> 00:05:56,200 I will set you one problem with the complexity of evaluating 94 00:05:56,200 --> 00:05:58,440 agent is that agent are multi step. 95 00:05:58,440 --> 00:06:00,480 OK? It's not it's not only I give 96 00:06:00,480 --> 00:06:02,520 you an answer. Is it right or wrong? 97 00:06:02,880 --> 00:06:04,600 You have multiple dimension in this. 98 00:06:04,600 --> 00:06:08,120 You need to to to measure as it come with the right step. 99 00:06:08,200 --> 00:06:11,120 Are these steps secure? Are these steps efficient? 100 00:06:11,160 --> 00:06:14,800 Did it consider enough source when he when he looked into your 101 00:06:14,800 --> 00:06:17,160 data, internal databases, etcetera. 102 00:06:17,160 --> 00:06:20,720 So it's it's much more complex and This is why I think so. 103 00:06:20,720 --> 00:06:24,000 This is my thesis and this is what I'm writing about with my 104 00:06:24,000 --> 00:06:26,480 book in my O'Reilly agency graph rag. 105 00:06:26,680 --> 00:06:30,680 I think graph architecture are pretty useful to manage this 106 00:06:30,680 --> 00:06:33,760 complexity. But before we like hint there 107 00:06:33,760 --> 00:06:38,440 and also dive a bit into the context of it, I want to linger 108 00:06:38,480 --> 00:06:43,760 with the agentic systems because you've now explored on how the 109 00:06:43,760 --> 00:06:46,800 impact is in applying it. But when we come back to an 110 00:06:46,800 --> 00:06:50,280 architectural point of view, what is the big shift for 111 00:06:50,280 --> 00:06:53,640 organizations that they need to do, that they need to apply and 112 00:06:53,640 --> 00:06:57,840 that they need to evolve into when they wanna translate their 113 00:06:57,840 --> 00:07:00,560 structures into energetic flow or Yeah. 114 00:07:02,000 --> 00:07:05,560 This is big OK, because basically what is happening is 115 00:07:05,560 --> 00:07:10,160 that everything that can be accessed, OK is automobile. 116 00:07:10,160 --> 00:07:12,360 You can automate everything that can be accessed. 117 00:07:12,440 --> 00:07:16,880 So enterprises need to think about this accessibility of our 118 00:07:16,880 --> 00:07:21,880 current databases, tools, etcetera, because everything 119 00:07:21,880 --> 00:07:25,840 that can be accessed, consumed by agent is like food for it. 120 00:07:26,360 --> 00:07:30,560 So basically you need to have like a clear strategy about what 121 00:07:30,560 --> 00:07:34,160 does my agenting system need to access. 122 00:07:34,200 --> 00:07:37,120 So you need to build like some sort of gateway. 123 00:07:37,200 --> 00:07:41,960 For example, I know publicly companies like like big banks, 124 00:07:41,960 --> 00:07:46,080 Morgan Stanley built a huge gateway of all of their tool, 125 00:07:46,200 --> 00:07:48,960 all of their software are are accessible by their agents. 126 00:07:48,960 --> 00:07:52,520 There is a a public like an MCP gateway, OK. 127 00:07:52,760 --> 00:07:54,800 And there is this concept of course of MCP. 128 00:07:55,000 --> 00:07:58,560 I think we don't need to go in into detail technically, but 129 00:07:58,560 --> 00:08:02,080 basically MCP it's like a USB key for for agent. 130 00:08:02,080 --> 00:08:06,320 That is to say it's a protocol that allow agents to access a 131 00:08:06,320 --> 00:08:09,760 certain tools or a certain piece of software. 132 00:08:09,760 --> 00:08:11,920 So you can transform your existing software like 133 00:08:11,920 --> 00:08:14,840 Salesforce, for example, MCP, Salesforce allow an agent to 134 00:08:14,840 --> 00:08:18,280 access basically Salesforce. So the current opportunities, 135 00:08:18,280 --> 00:08:20,360 the, the customer notes, etcetera, extra. 136 00:08:20,440 --> 00:08:23,480 So if you think about it at large scale, you need to map out 137 00:08:23,480 --> 00:08:26,200 what are your current capabilities, what's your, your, 138 00:08:26,320 --> 00:08:29,400 your, your, your user, your, your, your, your employee 139 00:08:29,520 --> 00:08:33,120 accessing their work. To rationalize this into I would 140 00:08:33,120 --> 00:08:36,039 say a main gateway. And when you think about this, 141 00:08:36,039 --> 00:08:39,159 there is complexity because you need to manage your identity. 142 00:08:39,280 --> 00:08:44,240 Maybe when I work in a share, I got access to tools related to a 143 00:08:44,240 --> 00:08:48,520 share, but I don't want my sales agent to access this, a share 144 00:08:48,520 --> 00:08:50,960 tool. So you need to build this 145 00:08:51,040 --> 00:08:55,200 identity layer that is to say preventing agent to accessing 146 00:08:55,280 --> 00:08:57,440 tools that are not authorized for them. 147 00:08:57,520 --> 00:08:59,240 You have many complexity around it. 148 00:08:59,240 --> 00:09:04,080 And for example, so I don't speak as an AWS employee, but 149 00:09:04,280 --> 00:09:08,520 AWS for example, has deployed agent core OK, that help builder 150 00:09:08,520 --> 00:09:11,920 build agentic system. And here in agent core, we have 151 00:09:12,000 --> 00:09:15,680 agent core identity that's manage this credential identity 152 00:09:15,680 --> 00:09:19,720 layer for for agents. So, yeah, to sum up is access is 153 00:09:19,800 --> 00:09:24,200 agentic and it creates problem because you have many system in 154 00:09:24,200 --> 00:09:28,280 enterprises currently that are not accessible by IPI or MCP. 155 00:09:28,320 --> 00:09:32,920 So you have this integration layer that get that can get 156 00:09:32,920 --> 00:09:35,760 complexified by by this old system. 157 00:09:35,760 --> 00:09:39,360 And you will need to have and I think it's already the case and 158 00:09:39,360 --> 00:09:42,840 the engineering of this old system to be accessible by 159 00:09:42,840 --> 00:09:46,240 agents if you want to identify such processes. 160 00:09:46,760 --> 00:09:49,800 And interestingly, you now talk about this unifying or 161 00:09:49,800 --> 00:09:53,200 interoperability of those layers of integrating it with a tech 162 00:09:53,200 --> 00:09:56,680 infrastructure that combines the scattered use case that you 163 00:09:56,680 --> 00:09:59,440 currently find in heritage organizations, right? 164 00:09:59,600 --> 00:10:03,160 Yeah, I assume that this is a challenge, but what are the 165 00:10:03,160 --> 00:10:06,040 three biggest challenges that you currently see for 166 00:10:06,040 --> 00:10:10,360 organizations when they want to deploy a genetic systems or also 167 00:10:10,480 --> 00:10:15,040 transition from experimenting with AI to fully operationalized 168 00:10:15,040 --> 00:10:17,480 systematic collaboration? Yeah, yeah, yeah. 169 00:10:17,760 --> 00:10:21,760 So first one that come into mind and it's it's always the same 170 00:10:21,760 --> 00:10:25,320 for AI. OK, Well, whatever AI you you 171 00:10:25,320 --> 00:10:29,360 consider like machine learning or or no modern LLM surgency 172 00:10:29,360 --> 00:10:33,080 system, it is that architecture of fragmentation and it's data 173 00:10:33,080 --> 00:10:35,600 quality basically. One problem is that the system 174 00:10:35,600 --> 00:10:39,360 of record that we built for example, it doesn't contain and 175 00:10:39,360 --> 00:10:43,760 it was a recent thesis from Foundation Capital in the US 176 00:10:43,760 --> 00:10:46,640 around context graph. What we are saying is that in 177 00:10:46,640 --> 00:10:49,320 the current system of record, you don't have the why of the 178 00:10:49,320 --> 00:10:52,520 decision you have you have the what happened, but you don't 179 00:10:52,520 --> 00:10:55,880 know all such decision out such data was created. 180 00:10:55,880 --> 00:10:59,320 If you want to have autonomy agents that make decision, we 181 00:10:59,320 --> 00:11:02,600 need to access the why, the criteria, the factors that's 182 00:11:02,600 --> 00:11:06,560 that's that's allow such data to be produced etcetera. 183 00:11:06,760 --> 00:11:10,040 So this is the first problem. Second problem is that we 184 00:11:10,200 --> 00:11:16,000 studies demonstrate that agents like particularly for example 185 00:11:16,000 --> 00:11:18,880 graph architecture, because graph architecture allow to 186 00:11:18,880 --> 00:11:21,400 connect. So once you have an anthology, 187 00:11:21,400 --> 00:11:25,480 you you archive rationalize your processes to to the facts. 188 00:11:25,480 --> 00:11:29,200 To have it within a graph architecture diminish diminish 189 00:11:29,240 --> 00:11:33,920 complexity to access this data. It's reduced for example the SQL 190 00:11:33,920 --> 00:11:39,520 query that you would make if you were only to use tabular a beta. 191 00:11:39,520 --> 00:11:42,480 If you take the example of vector search only, the problem 192 00:11:42,480 --> 00:11:46,520 with vector search is that not everything can be retrieved with 193 00:11:46,560 --> 00:11:49,880 only semantic similarity. So basically the first problem 194 00:11:49,880 --> 00:11:52,880 is that you need to think of that architecture to be agency 195 00:11:52,880 --> 00:11:54,760 friendly. And this would be my my first 196 00:11:54,760 --> 00:11:57,840 problem. The second problem is basically 197 00:11:57,840 --> 00:12:00,720 the governance that I mentioned earlier here, here there is 198 00:12:00,720 --> 00:12:04,240 many, many thing I, I, I would demonstrate the the governance 199 00:12:04,240 --> 00:12:07,520 problem with the security risk recently, so there has been an 200 00:12:07,520 --> 00:12:11,280 explosion of this thing called clobot. 201 00:12:11,640 --> 00:12:14,640 We know for a fact that there are risk reversions. 202 00:12:14,640 --> 00:12:17,160 For example, there is this risk called tool poisoning. 203 00:12:17,520 --> 00:12:22,360 Tool poisoning is is that if see an agent access a tool that is 204 00:12:22,360 --> 00:12:26,680 online OK and that this tool contain for example malicious 205 00:12:26,840 --> 00:12:29,040 instruction that are hidden to human OK. 206 00:12:29,040 --> 00:12:31,880 This could be like instruction that can't be read by human. 207 00:12:32,080 --> 00:12:36,440 You could basically contaminate the agent, maybe modify his main 208 00:12:36,440 --> 00:12:40,360 behavior and have like a Trojan horses, that is to say an agent 209 00:12:40,360 --> 00:12:45,320 that's send your data or modify your system without you notice 210 00:12:45,320 --> 00:12:48,760 it or access databases etcetera. So basically agents are so 211 00:12:48,760 --> 00:12:52,520 powerful is that if you contaminate it, if you low 212 00:12:52,560 --> 00:12:57,560 external threat to act it, the consequences can be really big 213 00:12:57,560 --> 00:13:02,280 in terms of data exfiltration, data modification, ETC. 214 00:13:02,320 --> 00:13:05,560 So all of these governance gap that is to say is still 215 00:13:05,560 --> 00:13:08,120 emerging. So you you have a lot of startup 216 00:13:08,200 --> 00:13:11,720 in Israel, for example, the Israel Israelian are the best 217 00:13:11,720 --> 00:13:15,240 around this security risk and in governance I did not mention, 218 00:13:15,240 --> 00:13:19,520 but you have the evaluation so many of our of our problem. 219 00:13:19,640 --> 00:13:24,080 And the third because you ask for free is basically So coming 220 00:13:24,080 --> 00:13:27,840 up to your, your first point is a trust that is to say that is 221 00:13:27,840 --> 00:13:32,800 to say you need, I think to you need to demonstrate because the 222 00:13:32,800 --> 00:13:37,640 problem is that this system of course for sure will impact 223 00:13:37,640 --> 00:13:40,200 current occupation. I estimate myself that maybe 224 00:13:40,200 --> 00:13:44,160 1995% of knowledge worker job will be impacted. 225 00:13:44,560 --> 00:13:48,280 What will stay I think is I impact customer meetings this 226 00:13:48,280 --> 00:13:52,800 kind of stuff, but to say what is intrinsically human for the 227 00:13:52,800 --> 00:13:56,920 rest all of this, all of that is administrative task that 228 00:13:56,920 --> 00:14:00,400 analysis etcetera. All, all of the knowledge deep, 229 00:14:00,720 --> 00:14:04,040 I mean the routine knowledge work or even the deep one, OK 230 00:14:04,120 --> 00:14:07,400 can either be supported or automated by agentic system. 231 00:14:07,480 --> 00:14:12,320 So you will need, I saw in term of change management, I think 232 00:14:12,320 --> 00:14:15,800 rework around the current processes, what people are 233 00:14:15,800 --> 00:14:19,040 doing, what are their roles and will amplify. 234 00:14:19,040 --> 00:14:21,160 So I am not pessimistic about this I think. 235 00:14:21,200 --> 00:14:25,120 I think this system allowed to intensify the work. 236 00:14:25,200 --> 00:14:28,560 It's allow everyone to do much more, not less. 237 00:14:28,560 --> 00:14:31,960 But interestingly, I want to just briefly iterate on the 238 00:14:31,960 --> 00:14:35,440 trust topic that you touched upon because connected to your 239 00:14:35,440 --> 00:14:37,680 second challenge that you pointed out. 240 00:14:37,920 --> 00:14:42,960 Yeah, when it comes to the design of them, you explain to 241 00:14:42,960 --> 00:14:46,800 us that they can also be designed for misbehavior. 242 00:14:46,880 --> 00:14:53,280 So how would we, as appliers of an agent that has not been 243 00:14:53,280 --> 00:14:57,600 deployed or built by ourselves, how can we trust them that they 244 00:14:57,600 --> 00:15:00,520 act with our best intention in mind? 245 00:15:00,520 --> 00:15:05,200 Is there anything that you could advise us to either look for or 246 00:15:05,200 --> 00:15:09,200 make sure that once we apply it that we are prepared for the 247 00:15:09,200 --> 00:15:10,800 next or for the output? Yes. 248 00:15:10,880 --> 00:15:14,480 So basically there is like basic, I would say security 249 00:15:14,480 --> 00:15:17,360 tools that can be used to prevent, I would say most of the 250 00:15:17,360 --> 00:15:18,960 risks. For example, if you take the 251 00:15:18,960 --> 00:15:23,200 identity one, the fact that your agent is only allowed to access 252 00:15:23,200 --> 00:15:27,040 your own tools or exercise, it's it's reduce the, I would say the 253 00:15:27,040 --> 00:15:29,840 blast radius of of adacking basically. 254 00:15:30,040 --> 00:15:34,400 But you have like tools like for example sandbanks sandboxing, 255 00:15:34,560 --> 00:15:39,640 because before your agent is maybe doing a query or doing a 256 00:15:39,640 --> 00:15:44,040 coding activity, what you can do is to 1st sandbox it, observe 257 00:15:44,040 --> 00:15:47,640 what he's doing and honest is validated by an external system 258 00:15:48,160 --> 00:15:51,120 can do the actual coding. So there are many, I would say 259 00:15:51,120 --> 00:15:54,200 technologically, I think there will be more and more solution 260 00:15:54,200 --> 00:15:58,480 to prevent either the plus reduce of a of a problem or or 261 00:15:58,480 --> 00:16:02,600 even completely prevented. OK, I will set you an example 262 00:16:02,600 --> 00:16:05,520 that is this company called the TST, They work. 263 00:16:05,520 --> 00:16:09,400 They they work currently. So they are was a creator of of 264 00:16:09,400 --> 00:16:12,640 a library called outlines. It's it's structure output for 265 00:16:12,640 --> 00:16:15,280 LLMS basically. So it's very simple for everyone 266 00:16:15,360 --> 00:16:19,400 because I know our listener won't be forcibly technical. 267 00:16:19,400 --> 00:16:22,840 But what it does is that it forces the LLM to respect 268 00:16:22,840 --> 00:16:25,520 certain rules, basically when it's watching, when it produce 269 00:16:26,360 --> 00:16:30,280 prediction and and output. So this in itself, when it is 270 00:16:30,280 --> 00:16:35,400 placed at sensible points of the agentic system, prevents most of 271 00:16:35,400 --> 00:16:38,640 the risk because you make sure that the agent is respecting 272 00:16:38,880 --> 00:16:41,480 your schema, your rules, your procedure. 273 00:16:41,720 --> 00:16:45,000 And this company, by the way, is working on, on something very 274 00:16:45,000 --> 00:16:47,000 interesting. They are working basically on 275 00:16:47,000 --> 00:16:50,160 the control layer for agent. So basically with your question, 276 00:16:50,160 --> 00:16:55,320 I think trust will come from the emergence of this, this, this 277 00:16:55,520 --> 00:16:59,000 building up of this control layer for agent that I know many 278 00:16:59,000 --> 00:17:00,960 startup are are working, working on. 279 00:17:00,960 --> 00:17:05,480 But for enterprises, I mean, of course you need red teaming. 280 00:17:05,480 --> 00:17:08,520 So you need a team. So red teaming is 1/2 people 281 00:17:08,520 --> 00:17:10,560 internally that are testing for bad behavior. 282 00:17:10,560 --> 00:17:15,640 Basically you need to test for the bad usage of the agent. 283 00:17:15,640 --> 00:17:18,720 So to work against it basically, of course should you need to 284 00:17:18,720 --> 00:17:22,119 have that in mind, you start up are are also proposing that to 285 00:17:22,280 --> 00:17:24,440 to enterprises. But I'm sure many enterprises 286 00:17:24,440 --> 00:17:27,800 that have that have a strong data science team or data 287 00:17:27,800 --> 00:17:30,840 engineering team or data engineer team are doing doing 288 00:17:30,840 --> 00:17:32,960 that by by themselves. Yeah. 289 00:17:32,960 --> 00:17:36,760 So, yeah. Sorry, because you just explain 290 00:17:36,760 --> 00:17:41,040 to us the difference between the startups and also the enterprise 291 00:17:41,040 --> 00:17:44,000 setup. And I would be interested to 292 00:17:44,000 --> 00:17:47,320 dive a bit into the business Angel aspect of yourself and 293 00:17:47,320 --> 00:17:50,320 also how you collaborate with startups at early stage. 294 00:17:50,880 --> 00:17:53,480 Yes. If we look at them from an early 295 00:17:53,480 --> 00:17:57,120 stage perspective, what would be a differentiator for startups 296 00:17:57,120 --> 00:18:01,680 that understand how and trust architecture is being set up 297 00:18:01,680 --> 00:18:06,960 versus those who chase like the hype or also the progress of AI 298 00:18:06,960 --> 00:18:09,360 and itself? I mean, the best doctor that I 299 00:18:09,360 --> 00:18:14,880 met basically first, they they understand very, very precisely 300 00:18:14,880 --> 00:18:18,320 all the current processes are are are working within 301 00:18:18,320 --> 00:18:21,520 enterprises that are they're trying to improve or or 302 00:18:21,520 --> 00:18:22,800 disruptive. OK. 303 00:18:22,800 --> 00:18:27,200 So they they passed most of their first months or time to 304 00:18:27,200 --> 00:18:30,360 meet to meet basically the business owner of these 305 00:18:30,360 --> 00:18:32,080 processes. They speak to them. 306 00:18:32,080 --> 00:18:35,400 They have a clear understanding of all all it is going on 307 00:18:35,400 --> 00:18:37,200 currently within the auto prices. 308 00:18:37,280 --> 00:18:40,840 Most of the time one of the Co founder is coming also from 309 00:18:40,840 --> 00:18:42,880 these processes. So he has he has already a 310 00:18:42,880 --> 00:18:47,200 network within the auto prices to support his first go to 311 00:18:47,200 --> 00:18:49,480 market movement, this kind of stuff. 312 00:18:49,520 --> 00:18:52,160 So this is very important because you need to clearly 313 00:18:52,160 --> 00:18:55,480 understand, but what is your it's regulation landscape, all 314 00:18:55,480 --> 00:18:58,880 the processes are what will be the point of integration I saw 315 00:18:59,440 --> 00:19:02,320 about as mentioned earlier, maybe there will be such 316 00:19:02,320 --> 00:19:05,360 challenges to integrate with the system currently put in place. 317 00:19:05,360 --> 00:19:08,280 So you need to all of this inter standing at this background to 318 00:19:08,280 --> 00:19:12,080 make sure that after all you can have a clear go to market to 319 00:19:12,080 --> 00:19:14,400 disrupt or improve the current enterprises. 320 00:19:14,400 --> 00:19:18,840 So I, I have example recently, I mean, I must see this kind of 321 00:19:18,840 --> 00:19:22,080 start up every week. And when I evaluate them, I have 322 00:19:22,080 --> 00:19:26,040 this this clearly in mind. It's a social angle, I would 323 00:19:26,040 --> 00:19:28,400 say. Does this start up as design 324 00:19:28,400 --> 00:19:31,840 partner or as access to the enterprises they are trying to 325 00:19:31,840 --> 00:19:33,640 improve all the processes they are trying to improve. 326 00:19:33,680 --> 00:19:37,240 You can't work in isolation. Basically, what will be valuable 327 00:19:37,240 --> 00:19:40,320 is access to I I have this example of entropics Entropic 328 00:19:40,320 --> 00:19:44,280 recently did Nakatan and what is funny about the Sakaton is that 329 00:19:44,280 --> 00:19:48,880 the third so the three guys that were wanted were not developers, 330 00:19:48,960 --> 00:19:51,640 they were people. I think it was like a doctor. 331 00:19:51,680 --> 00:19:55,400 So they were basically the the business expert and and this 332 00:19:55,400 --> 00:19:58,880 tells something like basically I think the most of the creation 333 00:19:59,000 --> 00:20:02,080 as a value will come from business experts that were 334 00:20:02,080 --> 00:20:04,960 backward from virtual and processes virtual and problem 335 00:20:05,480 --> 00:20:08,400 that have no the mean with agent decoding to code it. 336 00:20:08,920 --> 00:20:12,480 We'll cut it and we'll come up with a start up ID and try to 337 00:20:12,480 --> 00:20:16,040 describe the current way things are done. 338 00:20:16,200 --> 00:20:17,520 So basically. Yeah. 339 00:20:17,760 --> 00:20:22,720 And now I guess when we look and also come back to your book that 340 00:20:22,720 --> 00:20:27,280 you've written and you just highlighted about the entrance 341 00:20:27,280 --> 00:20:30,840 of startups into, you know, markets, but also how they 342 00:20:30,840 --> 00:20:34,520 approach the orchestration from the infrastructure layer and so 343 00:20:34,520 --> 00:20:36,560 on. In your book like Agenda Graph 344 00:20:36,560 --> 00:20:40,680 Rack, you explore also how it advanced retrieval or general 345 00:20:40,760 --> 00:20:42,760 orchestration architecture is being set up. 346 00:20:43,400 --> 00:20:47,080 Yeah, writing to you and also of what you've written in the book. 347 00:20:47,080 --> 00:20:51,440 Why is retrieval architecture so central when you wanna build 348 00:20:51,440 --> 00:20:55,040 trust in AI systems? So basically if you have like a 349 00:20:55,400 --> 00:20:59,160 metaphor, LLM are just they are not databases. 350 00:20:59,160 --> 00:21:03,000 OK, LLM are used when an energetic framework only to 351 00:21:03,000 --> 00:21:07,080 reason open some data. So because and in most 352 00:21:07,080 --> 00:21:10,600 industries your data will evolve over time, OK, if you take the 353 00:21:10,600 --> 00:21:14,520 legal system every month, every week you have like a new 354 00:21:14,520 --> 00:21:18,640 regulation, you have courts, courts that are that are doing 355 00:21:18,640 --> 00:21:21,400 their business. So you need a way for agent to 356 00:21:21,400 --> 00:21:25,520 access freshness in the data. You need a way for agents to 357 00:21:25,840 --> 00:21:28,640 have an infrastructure that allow them to reason correctly 358 00:21:28,640 --> 00:21:32,360 on what is happening and I found myself by building such system. 359 00:21:32,360 --> 00:21:34,800 That's the best way to manage it long term. 360 00:21:35,040 --> 00:21:39,640 OK is to to build it with graph technology basically why because 361 00:21:39,640 --> 00:21:42,880 graph technology I lost something that is very powerful. 362 00:21:43,000 --> 00:21:45,000 It's called the multi op reasoning. 363 00:21:45,240 --> 00:21:48,120 That is to say you can start with a first inquiry about I 364 00:21:48,120 --> 00:21:52,640 don't know. So we you have maybe so I need 365 00:21:52,640 --> 00:21:56,400 to find a customer of my customer, OK. 366 00:21:56,480 --> 00:22:00,200 And I want to understand the customer of the customer of my 367 00:22:00,200 --> 00:22:02,080 customer. This is multi op reasoning. 368 00:22:02,080 --> 00:22:04,560 OK. And you can only do that with 369 00:22:04,960 --> 00:22:08,880 with rough basically. And agents are really good at 370 00:22:09,160 --> 00:22:13,120 navigating and all topic has demonstrated with cloud cuts for 371 00:22:13,120 --> 00:22:17,920 example, are really good at navigating this kind of multi op 372 00:22:17,920 --> 00:22:22,120 reasoning and trace back basically the information. 373 00:22:22,240 --> 00:22:27,640 So increase the access accessness, the assetsness of 374 00:22:27,640 --> 00:22:29,320 the data. I mean, it has been measured 375 00:22:29,320 --> 00:22:33,760 currently most of the industry is has been has been doing a 376 00:22:33,760 --> 00:22:37,440 vector search, OK, and only vector search because basically 377 00:22:37,440 --> 00:22:39,080 you can use a vector such as over graph. 378 00:22:39,080 --> 00:22:43,880 OK, if you measure so you you put a both by comparison, there 379 00:22:43,880 --> 00:22:48,600 is a 30% accuracy gap between only using vector search by 380 00:22:48,600 --> 00:22:51,200 itself or using graph rag solution. 381 00:22:51,480 --> 00:22:55,440 OK, and 30% is very big. And then so in the book what 382 00:22:55,440 --> 00:22:59,600 what we explained with my core for what we explained with my 383 00:22:59,600 --> 00:23:03,240 core for is that there is also this self evolution self 384 00:23:03,240 --> 00:23:06,280 evolution layer. So cognitive strays on the the 385 00:23:06,280 --> 00:23:10,040 memory, but memory system are very complex, OK, because when 386 00:23:10,040 --> 00:23:13,160 you have memory, you need to have a system, but forget on 387 00:23:13,160 --> 00:23:17,120 Casey information, OK, you need to choose what to remember. 388 00:23:17,240 --> 00:23:22,160 You need to and make sure that when you remember something it 389 00:23:22,160 --> 00:23:25,520 is attached to an existing memory etcetera, extra and what 390 00:23:25,520 --> 00:23:28,880 most of the of the memory builder are doing, so there is 391 00:23:28,880 --> 00:23:30,480 cognitive, there is memo etcetera. 392 00:23:30,480 --> 00:23:33,040 They are using graph to do that. OK, It's the only way. 393 00:23:33,040 --> 00:23:36,400 It is the only data structure that allowed to have such a 394 00:23:36,640 --> 00:23:40,000 complex external memory to achieve this. 395 00:23:40,040 --> 00:23:43,800 This basically. And in the industry we talk a 396 00:23:43,800 --> 00:23:49,840 lot about reducing hallucination or so you know, being precise on 397 00:23:49,840 --> 00:23:55,040 the output and very accurate on like the explainability of the 398 00:23:55,040 --> 00:23:58,040 output itself. What role does graph based 399 00:23:58,040 --> 00:24:00,160 reasoning play in? Yes, yes, yes. 400 00:24:00,400 --> 00:24:03,640 So it's really good because yes, so you, you increase the 401 00:24:03,640 --> 00:24:05,240 accuracy of the of the retrieval. 402 00:24:05,360 --> 00:24:09,560 But if you think logically, even if you retrieve correctly, maybe 403 00:24:09,560 --> 00:24:15,360 the LLM will interpret it wrongly or do approximation or 404 00:24:15,360 --> 00:24:18,520 over interpretation, etcetera. But what is great is that 405 00:24:18,520 --> 00:24:22,240 currently technically you can make sure that the LLM of the 406 00:24:22,240 --> 00:24:26,760 agents will respect the ontology of the, I would say the semantic 407 00:24:26,760 --> 00:24:28,480 layer of the graph that has been built. 408 00:24:28,480 --> 00:24:31,320 So what I am seeing emerging, and This is why I'm writing this 409 00:24:31,320 --> 00:24:35,720 book, is that the control layer for agents we come mostly and 410 00:24:35,720 --> 00:24:40,640 from a concrete and organized ontology and graph and to be 411 00:24:40,640 --> 00:24:44,840 sure that the agent that is interpreting it or using it is 412 00:24:44,840 --> 00:24:47,360 respecting it. OK at the token level. 413 00:24:47,600 --> 00:24:50,520 That's to say, currently the problem is that LLM are black 414 00:24:50,520 --> 00:24:53,400 box. But if you retrieve correctly 415 00:24:53,400 --> 00:24:57,920 from the graph, you no, no thanks to the possibility of the 416 00:24:57,920 --> 00:25:00,560 graph. And I mean, it is symbolic. 417 00:25:00,560 --> 00:25:03,840 So you know which, which nodes has been queried. 418 00:25:03,840 --> 00:25:06,360 So it is part of the graph that is related to the concept X 419 00:25:06,360 --> 00:25:08,840 star. So you can have a layer of 420 00:25:08,840 --> 00:25:11,040 explainability thanks to this graph. 421 00:25:11,040 --> 00:25:13,520 So you, you get it from SQL research. 422 00:25:13,520 --> 00:25:16,600 OK, but it's less less efficient and, and, and let's, let's, 423 00:25:16,600 --> 00:25:19,760 let's powerful relevant graph, but you don't get it with vector 424 00:25:19,760 --> 00:25:21,280 set. So basically it's increase. 425 00:25:21,280 --> 00:25:25,760 I saw the explain ability of it increase the traceability of the 426 00:25:25,760 --> 00:25:28,240 reasoning chain and the explainability of the of the 427 00:25:28,240 --> 00:25:30,880 graph. You know, while you're speaking 428 00:25:30,880 --> 00:25:34,920 and while I appreciate if we have this hands down very deep 429 00:25:34,920 --> 00:25:38,760 technical conversation right now, I'm just like, I, I'm so 430 00:25:38,760 --> 00:25:41,680 humbled of all the experience. And at the same time, I'm 431 00:25:41,680 --> 00:25:45,360 currently like, I'm asking myself how much of the knowledge 432 00:25:45,360 --> 00:25:48,840 that we just discussed and also, you know, translate into how 433 00:25:48,840 --> 00:25:52,560 it's applied and where it's also brought into the organization. 434 00:25:52,640 --> 00:25:57,320 How do we ensure that people within organizations feel 435 00:25:57,440 --> 00:26:00,680 empowered to understand of what we're actually talking about and 436 00:26:00,680 --> 00:26:05,920 how they also can shape it? Because in my concern that I 437 00:26:05,920 --> 00:26:10,200 have is that within organization says currently this friction 438 00:26:10,200 --> 00:26:14,560 between people using infrastructure technology to do 439 00:26:14,560 --> 00:26:18,920 their job and IT departments are like experts that are dedicated 440 00:26:18,920 --> 00:26:21,840 to provided. However, with AI and the genetic 441 00:26:21,840 --> 00:26:26,680 systems, we all come empowered to use do apply and develop the 442 00:26:26,680 --> 00:26:29,320 infrastructure of how we want to work with it ourselves. 443 00:26:29,520 --> 00:26:33,400 But a lot of employees do not yet have the knowledge of how to 444 00:26:33,400 --> 00:26:35,800 do it. So they also ask, you know, the 445 00:26:35,920 --> 00:26:39,920 experts in the organization to provide them with the respective 446 00:26:39,920 --> 00:26:43,360 infrastructure, which is like a huge gap that's currently not 447 00:26:43,360 --> 00:26:46,080 being moderated or also no solved. 448 00:26:46,120 --> 00:26:49,840 And so I will be interested in first, what's your take on how 449 00:26:49,840 --> 00:26:52,880 deep of a knowledge to employees going forward need to have an 450 00:26:52,880 --> 00:26:56,560 order for them to apply it and how to second solve this? 451 00:26:56,680 --> 00:26:59,720 I don't want to call it limbo, but at least I want to call it 452 00:26:59,760 --> 00:27:02,800 responsibility loop that is not solved yet. 453 00:27:02,880 --> 00:27:06,440 I think it's key, OK, what you are saying, I won't pretend to 454 00:27:06,440 --> 00:27:09,120 have a clear solution. It's a clear problem, OK? 455 00:27:09,440 --> 00:27:12,800 And it is even worse when you are considering that the fact 456 00:27:12,800 --> 00:27:16,280 that people that are not technical but have the business, 457 00:27:16,280 --> 00:27:20,200 business knowledge and the business insight will be going 458 00:27:20,200 --> 00:27:23,160 forward, the more valuable. Because basically you need this 459 00:27:23,160 --> 00:27:28,240 business knowledge to create powerful agentic system and and 460 00:27:28,240 --> 00:27:31,280 the tech guys won't be able to create it without them. 461 00:27:31,360 --> 00:27:34,480 So basically what you are saying this limbo as you call it, you 462 00:27:34,480 --> 00:27:38,160 need to solve this to create powerful with the sprayed 463 00:27:38,160 --> 00:27:41,640 agentic system all over the across the organization. 464 00:27:41,640 --> 00:27:45,480 So one advice I would have, so there's been this theme with the 465 00:27:45,480 --> 00:27:47,680 industry around forward deploy engineer. 466 00:27:47,760 --> 00:27:49,840 What are forward deploy engineer? 467 00:27:49,840 --> 00:27:53,160 They are engineer that are embedded within the business 468 00:27:53,160 --> 00:27:55,680 teams to understand what they are doing and to create a 469 00:27:55,680 --> 00:27:59,480 genetic system on top of it. I think this I would say 470 00:27:59,480 --> 00:28:03,760 intellectual framework to put tech guys within business units, 471 00:28:03,760 --> 00:28:06,960 so sales guys, marketing guys, whatever to understand what they 472 00:28:06,960 --> 00:28:11,240 are doing, maybe educate them, unpower them, OK, when why they 473 00:28:11,240 --> 00:28:13,400 are doing so is a very powerful idea. 474 00:28:13,720 --> 00:28:17,160 And I think it's This is why open AI, Entropic and all of 475 00:28:17,160 --> 00:28:20,960 these guys are basically are basically building up and 476 00:28:21,000 --> 00:28:25,520 scaling such, such such system. It has been the first deployed 477 00:28:25,520 --> 00:28:30,160 by Patanti, but it's a right now it's it is, it is more and more 478 00:28:30,160 --> 00:28:32,680 we desperate. But what is changing now there 479 00:28:32,680 --> 00:28:35,840 has been the release of cloud cohort by Entropic. 480 00:28:36,000 --> 00:28:39,520 What people need to understand is that agentic coding, so 481 00:28:39,520 --> 00:28:42,520 agents that allow you to code are accessible by anyone 482 00:28:42,760 --> 00:28:46,720 basically because because they are coding, you don't need to 483 00:28:46,720 --> 00:28:50,040 know I would say. I would say the art of coding 484 00:28:50,040 --> 00:28:53,960 will disappear, but what will stay is system design. 485 00:28:54,000 --> 00:28:57,000 Basically, you need to have a clear sense of what you are 486 00:28:57,000 --> 00:29:00,640 doing basically to, to make make the the best out of it. 487 00:29:00,640 --> 00:29:03,360 Myself, for example, on my personal usage, I must have, I 488 00:29:03,360 --> 00:29:06,720 don't know, daily, I must have a 50 agent that are working for me 489 00:29:06,760 --> 00:29:09,800 to code what I want. OK, I don't code anymore, but I 490 00:29:09,800 --> 00:29:13,040 understand what I'm doing. I, I can catch errors, I can 491 00:29:13,240 --> 00:29:17,040 Orient, reorient, I can come up with better ideas. 492 00:29:17,200 --> 00:29:19,480 So it is a mixture between me and the agents. 493 00:29:19,600 --> 00:29:22,560 And I think we need to educate, we need to arm power people to 494 00:29:22,560 --> 00:29:25,320 leverage this technology. Because what I am observing here 495 00:29:25,320 --> 00:29:29,280 on the field is that this combination of human plus agent 496 00:29:29,280 --> 00:29:33,680 or coding agent, it creates 5050 time your productivity 497 00:29:33,720 --> 00:29:38,480 basically, if not more so. So basically we need a clearer 498 00:29:38,520 --> 00:29:40,680 organization. We need to have a clear program 499 00:29:40,680 --> 00:29:43,360 of empowerment. Once we have the basis, that is 500 00:29:43,360 --> 00:29:46,600 to say the basis is, is what we mentioned earlier, you need to 501 00:29:46,640 --> 00:29:48,880 access. So you need to create gateway 502 00:29:48,880 --> 00:29:52,200 for the agents to work on your data to work on your. 503 00:29:52,360 --> 00:29:55,680 Once you have that, we will need to empower people to use it at 504 00:29:55,680 --> 00:29:57,640 large scale. And this is by the way. 505 00:29:57,720 --> 00:30:00,560 So what I'm saying is this is what I'm doing, OK? 506 00:30:00,560 --> 00:30:04,680 I'm pushing for agentic solution, a really structural 1 507 00:30:04,800 --> 00:30:08,080 and I try, I'm trying to empower people because what people need 508 00:30:08,080 --> 00:30:10,840 to understand is that this modernization is impacting 509 00:30:10,840 --> 00:30:13,640 everyone, OK? There is no, not in the world 510 00:30:13,720 --> 00:30:16,960 maybe except at entropic organization that I'm using it 511 00:30:16,960 --> 00:30:19,000 at large scale. So yeah, so it will transform. 512 00:30:19,040 --> 00:30:22,360 I don't want to be a prophet, but the way we know our 513 00:30:22,360 --> 00:30:25,800 workplace, so how we are doing business will change 514 00:30:25,920 --> 00:30:27,360 considerably in the coming months. 515 00:30:27,640 --> 00:30:30,480 OK. And not leaning into the profit 516 00:30:30,480 --> 00:30:33,400 side of it too much, but still looking in the future, yeah. 517 00:30:34,160 --> 00:30:37,280 What does enterprise AI look like for you in 2020, let's say 518 00:30:37,280 --> 00:30:40,240 8 or 9? I mean, whatever long you want 519 00:30:40,240 --> 00:30:44,480 to look into the future and how do we transition into this multi 520 00:30:44,480 --> 00:30:48,120 agent ecosystem, organization, collaboration, what do you want 521 00:30:48,120 --> 00:30:52,040 to call it? And also how specific, like how 522 00:30:52,040 --> 00:30:56,800 would you envision of how being productive would look like in 523 00:30:56,800 --> 00:30:58,520 that? Yeah, yeah, yeah. 524 00:30:58,960 --> 00:31:02,200 So I will give you an image. I think most people will do what 525 00:31:02,200 --> 00:31:06,120 I'm doing OK every morning you will have like a set of outcome 526 00:31:06,120 --> 00:31:09,040 coming from your agent because my agent are working by night. 527 00:31:09,040 --> 00:31:10,880 That's all OK. It's, it's practical and 528 00:31:10,880 --> 00:31:15,360 basically all of day long you will switch between high impact 529 00:31:15,360 --> 00:31:19,280 customer meetings because we, we, we will, it will stay OK as 530 00:31:19,280 --> 00:31:23,040 a fact of going to town, having a nice lunch with your high 531 00:31:23,040 --> 00:31:26,080 profile customer, etcetera. But what would be cool is that 532 00:31:26,240 --> 00:31:29,960 everything that is around such high impact meetings will be 533 00:31:29,960 --> 00:31:32,560 managed by agents. So before this meeting, I will 534 00:31:32,560 --> 00:31:35,920 have prepped my meeting, your customer meeting with all the 535 00:31:35,920 --> 00:31:38,680 data that I need. After the meeting, everything 536 00:31:38,680 --> 00:31:42,400 that is around follow-ups, connecting with design partner, 537 00:31:42,400 --> 00:31:45,240 connecting internally or externally will be managed by 538 00:31:45,240 --> 00:31:49,000 agents, Gatner says. So I, I, I fetched up of course 539 00:31:49,000 --> 00:31:52,720 for this podcast, they said that 50% of work decisions will be 540 00:31:52,720 --> 00:31:57,560 made autonomously by agency Ki by 2028. 541 00:31:57,800 --> 00:32:02,240 I think there is one condition is a governance layer is secure 542 00:32:02,600 --> 00:32:05,800 and we have security etcetera that this number is will be 543 00:32:05,800 --> 00:32:10,880 considerably more and maybe 50%. But would that end up in agents 544 00:32:10,880 --> 00:32:13,200 talking to agents? Because if we would now have 545 00:32:13,200 --> 00:32:17,080 lunch, let's say in Munich or your hometown Paris. 546 00:32:17,200 --> 00:32:20,760 And then we met up, we had this great conversation and then each 547 00:32:20,760 --> 00:32:28,200 of us, you know, returns to or the Skype translate it anyways. 548 00:32:28,360 --> 00:32:31,880 And then my agent would send your agent a follow up of the 549 00:32:31,880 --> 00:32:34,240 conversation and your agent would answer or what? 550 00:32:34,240 --> 00:32:37,720 How would that look like? Of course, I mean, yeah, if I 551 00:32:37,720 --> 00:32:41,840 it's already happening, OK, I can't I can't speak about the 552 00:32:41,840 --> 00:32:46,520 detail, but myself I already built system where yes, of 553 00:32:46,520 --> 00:32:49,960 course everything that is follow up understanding you like like 554 00:32:49,960 --> 00:32:53,680 following the state, OK and updating the state, it's managed 555 00:32:53,680 --> 00:32:56,080 by agents. So of course, I think everything 556 00:32:56,080 --> 00:32:59,480 that is related to schedule, everything that is related to 557 00:32:59,480 --> 00:33:03,920 updating the schedule maybe with new data or coordinating an 558 00:33:03,920 --> 00:33:06,200 event. For example, recently I work on 559 00:33:06,200 --> 00:33:10,400 such system, that is to say, maybe matching people with with 560 00:33:10,400 --> 00:33:13,320 an event, event during an event you have multiple meetings, 561 00:33:13,320 --> 00:33:16,320 multiple schedule and you want to perfectly personalized the 562 00:33:16,320 --> 00:33:19,240 invite to the people you want to invite to what is happening 563 00:33:19,240 --> 00:33:21,120 during the event. All of this can be managed by 564 00:33:21,120 --> 00:33:23,640 agent. So yeah, and more books we have, 565 00:33:23,640 --> 00:33:27,040 we had recently an experiment which is called the MUD. 566 00:33:27,240 --> 00:33:29,320 So there has been a lot of fuse around it. 567 00:33:29,320 --> 00:33:31,400 That is to say it's a social network for agent etcetera. 568 00:33:31,800 --> 00:33:35,120 A lot of so a false promise. That is to say most of his 569 00:33:35,120 --> 00:33:38,880 agents were prompted to do this. OK about the end of the world or 570 00:33:38,880 --> 00:33:41,520 we will take over extra. So it was funny, OK to watch. 571 00:33:41,720 --> 00:33:45,400 But what's demonstrate is that you can of course have like it's 572 00:33:45,400 --> 00:33:49,360 an event based system. That is to say an agent is is 573 00:33:49,360 --> 00:33:54,040 like subscribed to the to a flow of event and react to the change 574 00:33:54,040 --> 00:33:56,680 of this event. So you can think all of the 575 00:33:56,680 --> 00:34:00,040 enterprises system around this. That is to say myself, what I I 576 00:34:00,040 --> 00:34:03,560 do normally is what I would follow my e-mail and react to 577 00:34:03,560 --> 00:34:07,400 the e-mail I received by checking up the databases. 578 00:34:07,400 --> 00:34:11,360 The news can do that perfectly. So basically takes the example 579 00:34:11,360 --> 00:34:14,600 of emails and following up e-mail. 580 00:34:14,800 --> 00:34:18,159 Currently, that's the future. Currently all of this stuff can 581 00:34:18,159 --> 00:34:20,600 be automated. Once we stay, I think we will 582 00:34:20,600 --> 00:34:24,040 have an explosion of software, OK, basically an explosion of ID 583 00:34:24,040 --> 00:34:26,199 and an explosion of execution of those IDs. 584 00:34:26,239 --> 00:34:29,360 Because what's what's this technology empower people is to 585 00:34:29,360 --> 00:34:32,159 directly go from IDs to execution. 586 00:34:32,239 --> 00:34:37,000 So what will happen I think is an explosion of of personalized 587 00:34:37,000 --> 00:34:42,000 software, OK, that's a fit to your use cases, fit to your 588 00:34:42,000 --> 00:34:44,960 problems, adapt to your situation, etcetera, etcetera. 589 00:34:45,040 --> 00:34:49,000 If that's the case, doesn't then also the human side of the 590 00:34:49,000 --> 00:34:53,080 connection and interaction become even more valuable and 591 00:34:53,080 --> 00:34:56,320 also more appreciated? Yeah, yeah, yeah, yeah. 592 00:34:56,360 --> 00:35:01,040 I think, I think what will stay is this is I impact, I impact 593 00:35:01,040 --> 00:35:04,520 customer meeting or I impact partner meetings, but these 594 00:35:04,520 --> 00:35:08,000 meetings will be so empowered. This meeting will will be like 595 00:35:08,000 --> 00:35:11,400 the the last step of of a process where we will have 596 00:35:11,440 --> 00:35:15,000 everything to decide or or do business or exactly. 597 00:35:15,280 --> 00:35:19,680 So I think this technology will basically arm power to have much 598 00:35:19,680 --> 00:35:23,040 more quality time gathering with people like Star star. 599 00:35:23,160 --> 00:35:25,040 I don't navigate anymore on Internet. 600 00:35:25,040 --> 00:35:27,600 OK, So people will will find it funny. 601 00:35:27,960 --> 00:35:30,200 It's the only thing that I navigate are social media. 602 00:35:30,200 --> 00:35:33,480 OK, because you can scrap it. You can't scrap it at large 603 00:35:33,480 --> 00:35:37,480 scale, but everything that is related to website YouTube news, 604 00:35:37,600 --> 00:35:41,480 everything can be analysed by agents and and everyday I 605 00:35:41,480 --> 00:35:46,040 received like a newsletter of what I need to understand what I 606 00:35:46,040 --> 00:35:48,840 need baby to to look up. I have synthesis across 607 00:35:48,840 --> 00:35:51,760 analysis, etcetera. All of this time it is gain is 608 00:35:51,760 --> 00:35:56,560 gain on on on on on being being more outdoor meeting people, 609 00:35:56,560 --> 00:35:59,440 meeting customer extra. OK, so I think this 610 00:36:00,000 --> 00:36:03,840 paradoxically and this is my optimistic vision, this 611 00:36:03,840 --> 00:36:07,440 technology will allow to get more human less, less time, less 612 00:36:07,440 --> 00:36:12,320 time spent on front of a screen and most time spent with real 613 00:36:12,320 --> 00:36:15,640 human interaction where you create a trust where where you 614 00:36:15,640 --> 00:36:18,440 haven't saw all of this. I would say I'm that sick 615 00:36:18,440 --> 00:36:24,440 understanding everything that AI can do currently and I hope I 616 00:36:24,440 --> 00:36:28,960 hope won't be as powerful as us, because I think it will be 617 00:36:28,960 --> 00:36:32,360 really, really complicate to automate what we get. 618 00:36:32,360 --> 00:36:36,280 The intuition, the personal feeling, the energy feeling 619 00:36:36,280 --> 00:36:38,000 everything. All of this stuff is very would 620 00:36:38,000 --> 00:36:39,480 be really really located for the. 621 00:36:39,480 --> 00:36:41,800 Neuroscience side of. It exactly. 622 00:36:41,800 --> 00:36:46,280 Since we're now come at the end of our of our episode, I do want 623 00:36:46,280 --> 00:36:49,160 to like jump into some in between moments that I want to 624 00:36:49,160 --> 00:36:50,760 have. And the first one that I put 625 00:36:50,760 --> 00:36:54,920 towards you is what does Gentek AI feel like to you? 626 00:36:55,200 --> 00:36:59,440 It's funny because so when I was a kid, I used to play a game 627 00:36:59,440 --> 00:37:02,000 called StarCraft. So it's, it's really geeky. 628 00:37:02,000 --> 00:37:05,360 But basically it's a game where you are commander and you 629 00:37:05,800 --> 00:37:10,400 basically you lead units to gather resources to attacks the 630 00:37:10,400 --> 00:37:12,800 opponent, etcetera. An agent TKI is like this, OK? 631 00:37:12,800 --> 00:37:15,280 You have like units are managing, you are managing your 632 00:37:15,280 --> 00:37:18,520 workforce basically, and you are managing your workforce maybe of 633 00:37:18,520 --> 00:37:24,240 5040 agents. So it's it's it's required from 634 00:37:24,240 --> 00:37:27,400 you strategic thinking, some sort of management, also 635 00:37:27,400 --> 00:37:29,720 organization, etcetera. It's feel like this, but it's 636 00:37:29,720 --> 00:37:31,520 feel like a game. But it's very fun, OK? 637 00:37:31,880 --> 00:37:35,040 And what does trust in enterprise system require? 638 00:37:35,480 --> 00:37:39,920 I think transparent failure mods, yes, I think basically you 639 00:37:39,920 --> 00:37:43,200 need to have a transparency of what is happening and what it 640 00:37:43,200 --> 00:37:45,440 fail. You need to precisely understand 641 00:37:45,440 --> 00:37:47,680 why. And what is the biggest risk for 642 00:37:47,680 --> 00:37:50,120 enterprises? I will be provocative but 643 00:37:50,120 --> 00:37:54,480 staying at the current stage where you only have human in the 644 00:37:54,480 --> 00:37:56,840 loop and calling it transformation. 645 00:37:57,360 --> 00:37:58,960 And what? What should start-ups stop 646 00:37:58,960 --> 00:38:01,680 doing? Building technology in search of 647 00:38:01,680 --> 00:38:03,720 a problem. Start by with a problem. 648 00:38:03,800 --> 00:38:08,680 Well, thank you for pushing ideas and boundaries and also 649 00:38:08,680 --> 00:38:11,200 moving us into the future. Thank you. 650 00:38:11,720 --> 00:38:14,400 I appreciate the conversation and all the insights that you 651 00:38:14,400 --> 00:38:14,960 share. Thank you. 652 00:38:15,000 --> 00:38:15,440 Thank you, Eva.