WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=SUqQlvyI3N8 NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/SUqQlvyI3N8.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:00.960 --> 00:00:04.920 You work with your AI agent for the last 2 years. s2 00:00:04.920 --> 00:00:07.160 You gained 2 years of experience. s3 00:00:07.160 --> 00:00:14.400 Your agent No. 10 people on your team work with the same AI agent. s4 00:00:14.400 --> 00:00:18.720 They cumulatively gained two decades of engineering experience. s5 00:00:18.720 --> 00:00:21.000 Your agent still there. s6 00:00:21.000 --> 00:00:25.760 It keeps on repeating the same mistakes again and again. s7 00:00:27.840 --> 00:00:38.560 I was talking to a senior executive at the world's biggest hedge fund and he said, "Sparsh, the problem is not that the AI of the agents are not smart. s8 00:00:38.560 --> 00:00:43.440 The problem is that it keeps on repeating the same mistakes again and again. s9 00:00:43.440 --> 00:00:50.040 If I course corrected today, then in the next session or with another team, it just makes that same mistake again. s10 00:00:50.040 --> 00:00:59.280 And the issue is that there are so many mistakes that we cannot improve those things at all times. s11 00:00:59.280 --> 00:01:03.160 We need a mechanism or we need a system to resolve this issue." s12 00:01:03.160 --> 00:01:07.560 And at that point of time, we realized that everyone is talking about context. s13 00:01:07.560 --> 00:01:09.960 Everyone is talking about memory. s14 00:01:09.960 --> 00:01:21.600 And the answer is we just need to take all the context, all the memory and give it to agents so that it will figure out. s15 00:01:21.600 --> 00:01:31.920 It was good in the start, but soon we realized at enterprise scale the issue is the context can be too big and giving all the facts to the agents just not increase the performance. s16 00:01:31.920 --> 00:01:42.720 It increases the cost, it increases the effort, it increases the overall time taken and on top of that it decreases it decreases the overall accuracy that is there, it decreases the performance. s17 00:01:42.720 --> 00:01:47.400 And there we realized that more context doesn't equal to better intelligence after a point. s18 00:01:47.400 --> 00:01:53.240 When you have When we are starting, when we have a smaller context base, yes, it means that, but after a point of time, it doesn't mean that. s19 00:01:53.240 --> 00:01:57.880 And we need a better mechanism, a better way to solve that problem. s20 00:01:57.960 --> 00:02:07.092 Here, we went to the back the drawing board and think like how why exactly we humans we get better with time. s21 00:02:07.092 --> 00:02:13.772 And at that point of time, we realized we humans have the skill of experience, of a judgment. s22 00:02:13.772 --> 00:02:22.372 And this judgment clears gets better with time when we learn something, when we make mistakes, and when we course correct. s23 00:02:22.372 --> 00:02:28.012 And that is a missing piece from a lot of these agents. s24 00:02:29.012 --> 00:02:36.812 And there we realized that at this point of time earlier, the memory was not learning for us. s25 00:02:39.132 --> 00:02:46.052 When we came back, when we built the system, the performance gain for us was 10x. s26 00:02:46.052 --> 00:02:54.972 And what exactly we came upon as the understanding of the memory is it's divided into two parts. s27 00:02:54.972 --> 00:02:58.012 The short-term memory and long-term memory. s28 00:02:58.012 --> 00:03:01.692 Short-term memory is the working memory or active session memory. s29 00:03:01.692 --> 00:03:06.012 The long-term memory is divided into two parts. s30 00:03:06.012 --> 00:03:08.772 Fact-based memory and experiential memory. s31 00:03:08.772 --> 00:03:14.372 This layer of experiential-based memory was something that has been missing now. s32 00:03:14.372 --> 00:03:18.452 And these are the abstracted out learnings from whatever we do with agents. s33 00:03:18.452 --> 00:03:23.732 And this is what makes we humans different from the AI that was here earlier. s34 00:03:23.732 --> 00:03:30.092 And further, this experiential layer can be doubled down into three different categories. s35 00:03:30.092 --> 00:03:33.812 The case-based, strategy-based, and the skill-based. s36 00:03:33.812 --> 00:03:38.572 Case-based is the detailed case and the detailed learning. s37 00:03:38.572 --> 00:03:44.692 The strategy is if you were in that particular situation, then what kind of strategy worked for us and what we should use. s38 00:03:44.692 --> 00:03:48.652 And then finally, is something that can be abstracted out, generalized very well. s39 00:03:48.652 --> 00:03:54.852 We abstract that thing out into a skill which an agent can use again and again. s40 00:03:54.852 --> 00:03:59.452 And this is how we built this experiential layer in the agents. s41 00:03:59.452 --> 00:04:03.614 This experiential memory is something that helps. s42 00:04:05.734 --> 00:04:11.614 At this point, we thought we have solved the problem of memory infrastructure. s43 00:04:12.334 --> 00:04:26.214 But, solving this problem made us realize that this is just a start of the problem and there are so many other bigger problems out there that form just because we solved this problem and we are able to see it because we reached this stage. s44 00:04:26.454 --> 00:04:32.214 And that is in the current times, the systems are changing very fast. s45 00:04:32.214 --> 00:04:37.854 The data drift, the system drift, the overall the whole environment drift. s46 00:04:37.854 --> 00:04:40.014 The world is changing every week, every month. s47 00:04:40.014 --> 00:04:42.374 An year feels like a decade in the AI years. s48 00:04:42.374 --> 00:04:51.534 So, the agent that works today doesn't work in a month or a year's time frame and we need to think through how we update that thing. s49 00:04:51.614 --> 00:05:06.174 And once we ran through a lots and lots of experiments, we came up on this understanding of the memory is that it's not just you make the system or the memory once, but it's just you have to evolve or make it better with time. s50 00:05:06.174 --> 00:05:10.334 And then in the final case, we came upon this architecture. s51 00:05:10.334 --> 00:05:19.574 That memory infrastructure pipeline is divided The lifecycle is divided into three parts. s52 00:05:19.574 --> 00:05:23.934 The formation, the updation, and the retrieval. s53 00:05:23.934 --> 00:05:26.254 Retrieval is very important. s54 00:05:26.254 --> 00:05:29.694 Now, formation we already understood right now. s55 00:05:29.694 --> 00:05:37.614 The updation is the part which evolves the system in which you add new notes when there are new experiences or facts. s56 00:05:37.614 --> 00:05:44.694 You update the older notes or like deactivate the notes which are no longer useful. s57 00:05:44.694 --> 00:05:47.374 And then finally, you have this forgetting. s58 00:05:47.374 --> 00:05:53.454 In updation, only storing the final layer of judgment is not enough in all cases. s59 00:05:53.454 --> 00:06:02.174 Because if you just store that, the agent can make a mistake and go back to the older point which had which was earlier and we just removed it. s60 00:06:02.174 --> 00:06:06.335 So, a better way is if it's a surface level research, it just looks at active nodes. s61 00:06:06.335 --> 00:06:09.975 And if it's a deeper research, it looks at the active nodes and the inactive nodes. s62 00:06:09.975 --> 00:06:16.695 And inactive nodes are the temporal journey that the system took to get the final layer of the judgment that is there. s63 00:06:16.695 --> 00:06:26.215 And then finally is the forgetting node, which is a must of very important here, and we cannot forget it. s64 00:06:26.215 --> 00:06:28.295 Let me give you an example. s65 00:06:28.295 --> 00:06:33.575 In the late 20th century, there was a person named Sheriv Sherosky. s66 00:06:33.575 --> 00:06:38.735 This is a human with near perfect memory with 100% recall. s67 00:06:38.735 --> 00:06:46.255 Now, you might be thinking, if we have got that particular memory, our life would have been amazing. s68 00:06:46.255 --> 00:06:57.895 But when the scientists and researchers studied his life, they started to realize a lot of times, when he even has to make a smaller decisions, his life was miserable s69 00:06:57.895 --> 00:07:09.735 because every point used to compete for an equal prioritization and an equal winning in things for something which we humans can take in a decision in a quick second. s70 00:07:09.735 --> 00:07:12.695 It was 100 thoughts fighting in his mind. s71 00:07:12.695 --> 00:07:14.895 And he gets overwhelmed. s72 00:07:14.895 --> 00:07:17.055 And that is the issue with our agents. s73 00:07:17.055 --> 00:07:24.015 If you keep on adding everything, if you keep on accumulating everything, and if you do not forget things which are no longer true or which are no longer required, s74 00:07:24.015 --> 00:07:37.055 then you might end up in a state where the system gets overwhelmed, and it the results just doesn't get better with time. s75 00:07:37.055 --> 00:07:41.295 So, you need to have a mechanism to forget things. s76 00:07:41.975 --> 00:07:47.735 Now, once you build this pipeline, we started with providing this memory to agents. s77 00:07:47.735 --> 00:07:49.815 We build this experiential agents. s78 00:07:49.815 --> 00:07:53.495 Now, we build this evolving experiential agents. s79 00:07:53.495 --> 00:07:55.935 But still, everything is manual. s80 00:07:55.935 --> 00:08:02.975 We thought we have solved the problem, but what's possible today, it was just half of the solution. s81 00:08:03.015 --> 00:08:05.375 Let me ask you a question. s82 00:08:05.375 --> 00:08:11.537 In the past 6 months, what has changed in the AI world? s83 00:08:13.497 --> 00:08:16.617 The models have become smarter. s84 00:08:17.137 --> 00:08:19.297 But that's only a smaller part. s85 00:08:19.297 --> 00:08:25.097 The bigger breakthrough is that the model is so smart enough that it can be the verifier. s86 00:08:25.097 --> 00:08:31.257 Until now, the situation was we always have to have the humans in the loop to review everything. s87 00:08:31.257 --> 00:08:41.777 But now for the first time, the model is smart enough such that we can bake that thing into a layer in itself as a verifier into the system so that the system gets better with time by itself. s88 00:08:43.337 --> 00:08:47.697 And then finally, we have this retrieval part. s89 00:08:48.617 --> 00:08:59.057 Now, in the retrieval aspect of things, your system's maximum performance can be just equivalent to the maximum performance of the retrieval. s90 00:08:59.697 --> 00:09:01.577 You might be asking, "Why?" s91 00:09:01.577 --> 00:09:03.177 Let me give you an example. s92 00:09:03.177 --> 00:09:06.777 Let's say I work in San Francisco and I have a home here. s93 00:09:06.777 --> 00:09:09.097 I also have a home in New York. s94 00:09:09.097 --> 00:09:16.417 I got off out of my office and I asked my AI agent, "Pull the maps for the home." s95 00:09:16.417 --> 00:09:24.057 Midway through the route, I realize, "This is not a place where I should be going." s96 00:09:24.057 --> 00:09:27.737 And then I realize I wanted to go to my San Francisco home. s97 00:09:27.737 --> 00:09:33.217 For the agent, the home is home, whether it's New York or San Francisco, and it put the map for the New York home. s98 00:09:33.217 --> 00:09:46.817 And there I realize, no matter how good my system is, the formation and the updation and the knowledge I have and the experiences I have, if my retriever doesn't give me the right information, s99 00:09:46.897 --> 00:09:48.777 it's just of no use. s100 00:09:48.777 --> 00:09:50.457 The whole system is of no use. s101 00:09:50.457 --> 00:09:56.937 And that's where the retrieval part of our system is very, very important. s102 00:09:57.777 --> 00:10:07.057 And then finally, once we have this thing, we have the formation, the updation, and the retrieval, this is how we complete the life cycle of the memory infrastructure we have. s103 00:10:07.057 --> 00:10:13.338 And let me tell you, earlier for gathering context for our agent for a lot of cases, it used to take 20 minutes. s104 00:10:13.338 --> 00:10:21.538 But now if we have gathered that context earlier and we know what all together, that same agent just takes 2 minutes of time and that's not only the improvement in the timing. s105 00:10:21.538 --> 00:10:35.418 It It is improving in the efforts it has to make It can make it It improves in the tokens it costs and it overall improves the performance of our agents in the long horizon task because it doesn't have to think through too much in the things we have already thought s106 00:10:35.418 --> 00:10:37.178 earlier. s107 00:10:37.178 --> 00:10:40.538 And this is just the start of the system. s108 00:10:40.538 --> 00:10:48.338 Now the next level comes, that is how we made the system self-evolving and experiential. s109 00:10:48.338 --> 00:10:54.578 And in that particularly, we built this concept of cycles, our dreaming cycles. s110 00:10:54.578 --> 00:11:04.618 Now we humans have a tendency when we take rest or when we sleep, our mind, it cleans the things which are no longer important. s111 00:11:04.618 --> 00:11:11.498 It takes a prioritized items to the front and whatever is left, it rearranges and cleans everything. s112 00:11:11.498 --> 00:11:15.218 Such that next day when I wake up, I feel very refreshed. s113 00:11:15.218 --> 00:11:17.098 I can think clearly. s114 00:11:17.098 --> 00:11:19.858 And that is something we built within the agents. s115 00:11:19.858 --> 00:11:33.138 We built this looping and the eval infrastructure such that we clean the things which are no longer important, we update the things which should be prioritized and we rearrange the things such that the next day and whenever the context is getting accumulated, s116 00:11:33.138 --> 00:11:37.778 the agent is able to perform better and better. s117 00:11:38.578 --> 00:11:45.818 And then finally, once we build this infrastructure, earlier the issue for us was the agents cannot remember anything. s118 00:11:45.818 --> 00:11:56.018 But now with the 2 years of experience, we are able to give more experience to the agents than any one of us humans. s119 00:11:56.018 --> 00:12:01.338 And we learned that there are three golden rules when we are building this experiential memory. s120 00:12:01.338 --> 00:12:04.658 First, do not dump everything. s121 00:12:04.658 --> 00:12:07.698 You have to remember you have to disclose progressively. s122 00:12:07.698 --> 00:12:13.059 Second, you cannot make a system once and assume it will always perform good. s123 00:12:13.059 --> 00:12:15.539 You have to improve the system continuously. s124 00:12:15.539 --> 00:12:26.779 And then finally, instead of writing prompts, you have to move towards writing these loops because that is how you make this whole infrastructure for you autonomous. s125 00:12:26.779 --> 00:12:34.779 An example for us is once you build this memory layer formation, updation, and retrieval, you build an observability layer around it. s126 00:12:34.779 --> 00:12:49.379 Once we have observability, we build this looping layer combined with the eval layer such that which after every session, after every day, extracts the learnings, improves the systems, test it independently, s127 00:12:49.379 --> 00:12:53.619 and then finally make the changes in the real system. s128 00:12:54.419 --> 00:13:03.499 Now, this looping, this evals, and the security gardens around the whole memory infrastructure is a very interesting topic. s129 00:13:03.499 --> 00:13:07.139 But that's a topic for my next talk. s130 00:13:07.139 --> 00:13:11.619 And coming back to the problem we had that we were not getting experiences. s131 00:13:11.619 --> 00:13:17.059 Now, with this thing, we whenever the agent work with me, it gains that experience. s132 00:13:17.059 --> 00:13:21.659 Whenever it it works with any one of on my team, it gains that experience, too. s133 00:13:21.659 --> 00:13:31.699 So, for me as a human, I'm get gaining getting only 1 year of experience in 1 year, but this agent, since it's working with all those 10 members, it's getting that 10x experience, and it's getting better s134 00:13:31.699 --> 00:13:33.499 than any one of us in the team. s135 00:13:33.499 --> 00:13:39.539 And now, because of that, we have started to see this autonomous teams of humans plus AI agents combined. s136 00:13:39.539 --> 00:13:49.939 And in some of the teams I know, like their team is like 10x or 100x because this agent that they have created, this agentic infrastructure and the memory behind it, it's just that is powering them, it's just s137 00:13:49.939 --> 00:13:59.579 empowering these agents such that some of these agents are far better than any of the best engineers or people or employees out there in that team. s138 00:13:59.579 --> 00:14:10.899 And soon, like in the next 6 to 9 months, we have we will start to see more and more of these autonomous AI companies in which agents are collaborating with each other, and they just remember everything. s139 00:14:10.899 --> 00:14:20.781 Such that now, if you implement these things, whatever performance you have, it just will compound exponentially with time because it's not learning only with you. s140 00:14:20.781 --> 00:14:29.861 It learns with every member of the team and it gets better and better with time such that in the near future there will be a stage within your enterprise s141 00:14:29.861 --> 00:14:37.581 where this agent has learned more than any one of our person in the different domains in the different systems. s142 00:14:38.141 --> 00:14:50.061 And I would like to tell this one thing is people who do not remember from the mistakes or the agents who do not remember from their mistakes are condemned to repeat it. s143 00:14:50.061 --> 00:14:59.581 Build this memory infrastructure and get back to growing exponentially better and build a 10x or 100x team.