1 00:00:00,000 --> 00:00:07,699 What are AI agents? 2 00:00:07,699 --> 00:00:11,079 And to start explaining that, we have to look at the various shifts that we're seeing in 3 00:00:11,079 --> 00:00:12,479 the field of generative AI. 4 00:00:12,479 --> 00:00:19,520 And the first shift I would like to talk to you about is this move from monolithic models 5 00:00:19,520 --> 00:00:28,680 to compound AI systems. 6 00:00:28,680 --> 00:00:33,299 So models on their own are limited by the data they've been trained on. 7 00:00:33,299 --> 00:00:42,360 So that impacts what they know about the world and what sort of tasks they can solve. 8 00:00:42,360 --> 00:00:47,040 They are also hard to adapt, so you could tune a model, but it would take an investment 9 00:00:47,040 --> 00:00:54,040 in data and in resources. 10 00:00:54,040 --> 00:00:57,380 So let's take a concrete example to illustrate this point. 11 00:00:57,380 --> 00:01:02,779 I want to plan a vacation for this summer, and I want to know how many vacation days 12 00:01:02,779 --> 00:01:08,300 are at my disposal. 13 00:01:08,300 --> 00:01:22,059 What I can do is take my query, feed that into a model that can generate a response. 14 00:01:22,059 --> 00:01:26,900 I think we can all expect that this answer will be incorrect, because the model doesn't 15 00:01:26,900 --> 00:01:32,860 know who I am and does not have access to this sensitive information about me. 16 00:01:32,860 --> 00:01:37,419 So models on their own could be useful for a number of tasks, as we've seen in other 17 00:01:37,419 --> 00:01:38,419 videos. 18 00:01:38,419 --> 00:01:40,220 So they can help with summarizing documents. 19 00:01:40,220 --> 00:01:44,319 They can help me with creating first drafts for emails and different reports I'm trying 20 00:01:44,319 --> 00:01:45,459 to do. 21 00:01:45,459 --> 00:01:50,459 But the magic gets unlocked when I start building systems around the model and actually take 22 00:01:50,459 --> 00:01:54,980 the model and integrate them into the existing processes I have. 23 00:01:55,059 --> 00:02:00,660 So if we were to design a system to solve this, I would have to give the model access 24 00:02:00,660 --> 00:02:05,099 to the database where my vacation data is stored. 25 00:02:05,099 --> 00:02:09,419 So that same query would get fed into the language model. 26 00:02:09,419 --> 00:02:15,380 The difference now is the model would be prompted to create a search query. 27 00:02:15,380 --> 00:02:20,300 And that would be a search query that can go into the database that I have. 28 00:02:20,300 --> 00:02:26,100 So that would go and fetch the information from the database, output an answer, and then 29 00:02:26,100 --> 00:02:31,179 that would go back into the model that can generate a sentence to answer. 30 00:02:31,179 --> 00:02:35,639 So Maya, you have 10 days left in your vacation database. 31 00:02:35,639 --> 00:02:44,339 So the answer that I would get here would be correct. 32 00:02:44,339 --> 00:02:47,460 This is an example of a compound AI system. 33 00:02:47,460 --> 00:02:51,699 And it recognizes that certain problems are better solved when you apply the principles 34 00:02:51,699 --> 00:02:57,779 of system design. 35 00:02:57,779 --> 00:03:00,139 So what does that mean? 36 00:03:00,139 --> 00:03:04,100 By the term system, you can understand there's multiple components. 37 00:03:04,100 --> 00:03:06,899 So systems are inherently modular. 38 00:03:06,899 --> 00:03:07,899 I can have a model. 39 00:03:07,899 --> 00:03:13,820 I can choose between tuned models, large language models, image generation models. 40 00:03:13,820 --> 00:03:17,020 But also I have programmatic components that can come around it. 41 00:03:17,580 --> 00:03:20,020 So I can have output verifiers. 42 00:03:20,020 --> 00:03:25,059 I can have programs that can take a query and then break it down to increase the chances 43 00:03:25,059 --> 00:03:26,860 of the answer being correct. 44 00:03:26,860 --> 00:03:29,580 I can combine that with searching databases. 45 00:03:29,580 --> 00:03:32,059 I can combine that with different tools. 46 00:03:32,059 --> 00:03:38,059 So when we're talking about a system approach, I can break down what I desire my program 47 00:03:38,059 --> 00:03:42,500 to do and pick the right components to be able to solve that. 48 00:03:42,500 --> 00:03:46,779 And this is inherently easier to solve for than tuning a model. 49 00:03:46,820 --> 00:03:50,179 So that makes this much faster and quicker to adapt. 50 00:03:56,580 --> 00:03:57,460 OK. 51 00:03:57,460 --> 00:04:03,139 So the example I used below is an example of a compound AI system. 52 00:04:03,139 --> 00:04:09,020 You also might be popular with retrieval augmented generation, which is one of the most popular 53 00:04:09,020 --> 00:04:13,539 and commonly used compound AI systems out there. 54 00:04:13,539 --> 00:04:19,859 Most RAG systems, and the example I used below, are defined in a certain way. 55 00:04:19,859 --> 00:04:24,380 So if I bring a very different query, let's say I ask about the weather in this example 56 00:04:24,380 --> 00:04:31,660 here, it's going to fail because the path that this program has to follow is to always 57 00:04:31,660 --> 00:04:36,179 search my vacation policy database, and that has nothing to do with the weather. 58 00:04:36,179 --> 00:04:40,899 So when we say the path to answer a query, we are talking about something called the 59 00:04:40,899 --> 00:04:45,140 control logic of a program. 60 00:04:45,140 --> 00:04:51,899 So compound AI systems, we said most of them have programmatic control logic. 61 00:04:51,899 --> 00:04:57,540 So that was something that I defined myself as the human. 62 00:04:57,540 --> 00:05:01,980 Now, let's talk about where do agents come in. 63 00:05:01,980 --> 00:05:07,940 One other way of controlling the logic of a compound AI system is to put a large language 64 00:05:07,940 --> 00:05:09,820 model in charge. 65 00:05:10,059 --> 00:05:14,579 And this is only possible because we're seeing tremendous improvements in the capabilities 66 00:05:14,579 --> 00:05:16,940 of reasoning of large language models. 67 00:05:16,940 --> 00:05:21,500 So large language models, you can feed them complex problems, and you can prompt them 68 00:05:21,500 --> 00:05:25,739 to break them down and come up with a plan on how to tackle it. 69 00:05:25,739 --> 00:05:30,459 Another way to think about it is on one end of the spectrum, I'm telling my system to 70 00:05:30,459 --> 00:05:36,339 think fast, act as programmed, don't deviate from the instructions I've given you. 71 00:05:36,339 --> 00:05:42,140 And on the other end of the spectrum, you're designing your system to think slow. 72 00:05:42,140 --> 00:05:47,540 So create a plan, attack each part of the plan, see where you get stuck, see if you 73 00:05:47,540 --> 00:05:49,420 need to readjust the plan. 74 00:05:49,420 --> 00:05:53,739 So I might give you a complex question, and if you would just give me the first answer 75 00:05:53,739 --> 00:05:57,600 that pops into your head, very likely that answer might be wrong. 76 00:05:57,600 --> 00:06:01,940 But you have higher chances of success if you break it down, understand where do you 77 00:06:01,940 --> 00:06:05,779 need external help to solve some parts of the problem, and maybe take an afternoon to 78 00:06:05,779 --> 00:06:07,100 solve it. 79 00:06:07,100 --> 00:06:13,140 And when we put LLMs in charge of the logic, this is when we're talking about an agentic 80 00:06:13,140 --> 00:06:15,420 approach. 81 00:06:15,420 --> 00:06:21,899 So let's break down the components of LLM agents. 82 00:06:21,899 --> 00:06:26,500 The first capability is the ability to reason, which we talked about. 83 00:06:26,500 --> 00:06:31,339 So this is putting the model at the core of how problems are being solved. 84 00:06:31,339 --> 00:06:35,579 The model will be prompted to come up with a plan and to reason about each step of the 85 00:06:35,579 --> 00:06:37,899 process along the way. 86 00:06:37,899 --> 00:06:41,779 Another capability of agents is the ability to act. 87 00:06:41,779 --> 00:06:47,739 And this is done by external programs that are known in the industry as tools. 88 00:06:47,739 --> 00:06:52,940 So tools are external pieces of the program, and the model can define when to call them 89 00:06:52,940 --> 00:06:57,540 and how to call them in order to best execute the solution to the question they've been 90 00:06:57,540 --> 00:06:58,839 asked. 91 00:06:58,839 --> 00:07:05,940 So an example of a tool could be search, searching the web, searching a database at their disposal. 92 00:07:05,940 --> 00:07:10,100 Another example could be a calculator to do some math. 93 00:07:10,100 --> 00:07:15,279 This could be a piece of program code that maybe might manipulate the database. 94 00:07:15,279 --> 00:07:19,880 This can also be another language model that maybe you're trying to do a translation task 95 00:07:19,880 --> 00:07:23,119 and you want a model that can be able to do that. 96 00:07:23,119 --> 00:07:25,519 And there's so many other possibilities of what can do there. 97 00:07:25,519 --> 00:07:30,160 So these could be APIs, basically any piece of external program you want to give your 98 00:07:30,160 --> 00:07:32,320 model access to. 99 00:07:32,320 --> 00:07:37,160 Third capability, that is the ability to access memory. 100 00:07:37,160 --> 00:07:39,480 And the term memory can mean a couple of things. 101 00:07:39,480 --> 00:07:45,200 So we talked about the models thinking through the program, kind of how you think out loud 102 00:07:45,200 --> 00:07:47,799 when you're trying to solve through a problem. 103 00:07:47,799 --> 00:07:52,760 So those inner logs can be stored and can be useful to retrieve at different points 104 00:07:52,760 --> 00:07:53,760 in time. 105 00:07:53,760 --> 00:07:58,760 But also this could be the history of conversations that you as a human had when interacting with 106 00:07:58,760 --> 00:07:59,760 the agent. 107 00:07:59,760 --> 00:08:03,480 And that would allow to make the experience much more personalized. 108 00:08:03,480 --> 00:08:07,640 So the way of configuring AI agents, there's many ways to approach it. 109 00:08:07,640 --> 00:08:13,760 One of the most popular ways of going about it is through something called React, which 110 00:08:13,760 --> 00:08:20,940 as you can tell by the name, combines the reasoning and act components of LLM agents. 111 00:08:20,940 --> 00:08:23,519 So let's make this very concrete. 112 00:08:24,279 --> 00:08:25,839 What happens when I configure a React agent? 113 00:08:25,839 --> 00:08:28,959 You have your user query. 114 00:08:28,959 --> 00:08:31,920 That gets fed into a model, so an LLM. 115 00:08:31,920 --> 00:08:33,599 The LLM is given a prompt. 116 00:08:33,599 --> 00:08:39,599 So the instructions it's given is, don't give me the first answer that pops to you. 117 00:08:39,599 --> 00:08:40,719 Think slow. 118 00:08:40,719 --> 00:08:44,559 Plan your work. 119 00:08:44,559 --> 00:08:46,760 And then try to execute something. 120 00:08:46,760 --> 00:08:48,299 Try to act. 121 00:08:48,299 --> 00:08:53,039 And when you want to act, you can define whether if you want to use external tools to help 122 00:08:53,039 --> 00:08:58,320 you come up with the solution, once you call a tool and you get an answer, maybe it gave 123 00:08:58,320 --> 00:09:03,479 you the wrong answer or it came up with an error, you can observe that. 124 00:09:03,479 --> 00:09:08,400 So the LLM would observe the answer, would determine if it does answer the question at 125 00:09:08,400 --> 00:09:13,640 hand or whether it needs to iterate on the plan and tackle it differently up until I 126 00:09:13,640 --> 00:09:18,859 get to a final answer. 127 00:09:18,859 --> 00:09:22,340 So let's go back and make this very concrete again. 128 00:09:22,340 --> 00:09:25,340 Let's talk about my vacation example. 129 00:09:25,340 --> 00:09:29,159 And as you can tell, I'm really excited to go on one. 130 00:09:29,159 --> 00:09:31,619 So I want to take the rest of my vacation days. 131 00:09:31,619 --> 00:09:33,859 I'm planning to go on to Florida next month. 132 00:09:33,859 --> 00:09:36,340 I'm planning on being outdoors a lot. 133 00:09:36,340 --> 00:09:37,739 And I'm prone to burning. 134 00:09:37,739 --> 00:09:43,979 So I want to know, what is the number of two-ounce sunscreen bottles that I should bring with 135 00:09:43,979 --> 00:09:44,979 me? 136 00:09:44,979 --> 00:09:46,260 And this is a complex problem. 137 00:09:46,260 --> 00:09:49,700 So there's a first thing, there's a number of things to plan. 138 00:09:49,700 --> 00:09:52,659 One is how many vacation days am I planning to take? 139 00:09:52,659 --> 00:09:57,099 And maybe that is information the system can retrieve from its memory because I asked that 140 00:09:57,099 --> 00:09:58,700 question before. 141 00:09:58,700 --> 00:10:01,059 Two is how many hours do I plan to be in the sun? 142 00:10:01,059 --> 00:10:02,880 I said I plan to be in there a lot. 143 00:10:02,880 --> 00:10:08,059 So maybe that would mean looking into the weather forecast for next month in Florida 144 00:10:08,059 --> 00:10:12,340 and seeing what is the average sun hours that are expected. 145 00:10:12,340 --> 00:10:17,640 Three is maybe going to a public health website to understand what is the recommended dosage 146 00:10:17,640 --> 00:10:19,760 of sunscreen per hour in the sun. 147 00:10:19,760 --> 00:10:24,299 And then four, doing some math to be able to determine how much of that sunscreen fits 148 00:10:24,299 --> 00:10:25,859 into two-ounce bottles. 149 00:10:25,859 --> 00:10:27,580 So that's quite complicated. 150 00:10:27,580 --> 00:10:31,799 But what's really powerful here is there's so many paths that can be explored in order 151 00:10:31,799 --> 00:10:32,880 to solve a problem. 152 00:10:32,880 --> 00:10:35,080 So this makes the system quite modular. 153 00:10:35,080 --> 00:10:39,020 And I can hit it with much more complex problems. 154 00:10:39,020 --> 00:10:45,280 So going back to the concept of compound AI systems, compound AI systems are here to stay. 155 00:10:45,280 --> 00:10:49,479 What we're going to observe this year is that they're going to become more agentic. 156 00:10:49,479 --> 00:11:00,880 The way I like to think about it is you have a sliding scale of LLM autonomy. 157 00:11:00,880 --> 00:11:08,460 And the person defining the system would examine what trade-offs they want in terms of autonomy 158 00:11:08,460 --> 00:11:10,119 in the system. 159 00:11:10,119 --> 00:11:15,200 For certain problems, especially problems that are narrow, well-defined, so you don't 160 00:11:15,200 --> 00:11:19,440 expect someone to ask about the weather when they mean to ask about vacations. 161 00:11:19,440 --> 00:11:24,159 So a narrow problem set, you can define a narrow system like this one. 162 00:11:24,159 --> 00:11:28,679 It's more efficient to go the programmatic route because every single query will be answered 163 00:11:28,679 --> 00:11:29,679 the same way. 164 00:11:29,679 --> 00:11:35,700 If I were to apply the agentic approach here, there might be unnecessarily looping and iteration. 165 00:11:35,700 --> 00:11:40,460 So for narrow problems, programmatic approach can be more efficient than going the agentic 166 00:11:40,460 --> 00:11:41,460 route. 167 00:11:41,460 --> 00:11:48,380 But if I expect to have a system accomplish very complex tasks, like, say, trying to solve 168 00:11:48,380 --> 00:11:55,020 GitHub issues independently and handle a variety of queries, a spectrum of queries, this is 169 00:11:55,020 --> 00:11:59,179 where an agentic route could be helpful because it would take you too much effort to configure 170 00:11:59,179 --> 00:12:01,659 every single path in the system. 171 00:12:01,659 --> 00:12:04,820 And we're still in the early days of agentic systems. 172 00:12:04,940 --> 00:12:09,299 So we're seeing rapid progress when you combine the effects of system design with agentic 173 00:12:09,299 --> 00:12:10,979 behavior. 174 00:12:10,979 --> 00:12:14,979 And of course, you will have a human in the loop in most cases as the accuracy is improving.