0:00 Music 0:06 the levels of kind of maturity of the AI system and also the level of complexity. 0:16 Just the basic kind of understanding of how things work is like, again, very crash understanding. 0:23 If something's unclear, let me know, I'll clarify. 0:27 So the first level is like basic prompting. 0:31 Hey, ChatGPT, how do I solve this problem? 0:36 And this is the level operation that kind of we all started like back when ChatGPT came out in the market. 0:45 It's basically... 0:47 Request response interface and this is a level when you interact with system most 0:55 the second is like system prompt so what is the system prompt system prompt is the are part of the instructions that are always provided before your first prompt even arrives to a lab 1:10 So LEM-like is just archive, but every chat GPT, Claude, every chatbot is driven by system prompt. 1:19 System prompt is injected at initial level and very often are, like as an example, giving it some additional instructions to 1:30 challenge user more is something that you wouldn't want to do in the system prompt. 1:35 Because if you do it in basic prompt, it will likely be overridden with the system prompt instructions that are by default tuned to be like gentle, polite, and produce some results that are more agreeable to the user. 1:49 So it's like... 1:50 Important thing to understand that you can give as many like additional rules instructions in your basic prompts but if your system prompt is intrinsically designed 2:02 to tune system to behave differently it will likely point to failure so 2:09 You have your system prompt design well, but your basic prompts are kind of keep producing inconsistent results. 2:16 And this is like goes into the domain of prompt engineering. 2:21 So you can, they're like separate area of knowledge where you can explore more. 2:27 But typical most successful way to make your prompts work better is give a few short examples of how you would like the outputs to look like. 2:40 Hey, if I ask this, this is an example how I would expect the output. 2:44 If I ask that, that's an example. 2:47 In that case, that's how you would perform. 2:50 And this is just a kind of additional technique to improve the results of your queries. 2:58 So the... 3:00 First kind of level is chain of thought. 3:03 And then when additionally to few shot, you give LLM a multi-step instructions, how it should process your request. 3:15 You kind of think how you would approach this task and you'd give it basically a playbook. 3:20 Hey, first, look up all the competitors in the market. 3:26 Second, for every competitor, retrieve key facts like their volumes, their presence in the market, their social media. 3:36 Fourth, build a comparison table with the most fifths. 3:40 Identify how each of the competitors are different to our brand presence in the market. 3:46 This is kind of example of chain of thought, how it can be kind of used to improve the quality of your prompts. 3:54 So the fifth kind of tool to be employed is tool use. 3:58 It's how you can equip this autonomous system with access to data, with access to execution, with access to additional anything basically that you want. 4:09 You probably heard about MCP a lot. 4:13 And MCP is basically a USB for AI to connect to other systems and connect to the world. 4:22 So it's like basically superseding API level when API was built for engineers to plug and integrate with other systems. 4:30 systems, MCP was built to connect together AI and give AI access to external tools and communications. 4:38 So it can talk to your, for example, meeting platform, to your CRM, to weather forecast. 4:47 Or it also can execute some actions on your behalf. 4:50 It can turn light on and off. 4:52 So we can technically build MCPs that will be responsible for managing electricity grid. 5:00 So anything that you can give and you want your AI to perform has to be given like as a tool. 5:07 So most common way to do that is through a tool use. 5:12 And that's how you kind of equip AI with powers, not just to reason within a chatbot interface, but actually to perform some actions. 5:20 So that's a tool use. 5:23 But until that, it's still limited within the same chat interface. 5:29 You interact, you give requests, and you kind of receive and process the response and judge whether it kind of makes sense for you or not. 5:37 And then kind of it goes to autonomy is when you define some kind of goal and let the AI reason on this goal, perform 5:47 a loop of sequence tools, calls, chain of thoughts, reasoning, few short examples until it achieves its mission. 5:56 Or continuous exploration. 5:58 So exploration can also be a mission that's 6:00 Kind of an example of how open clock can be set up, for instance. 6:06 Okay, that was a lot I tried to cover in the shot, very narrative.
0:00 Music 0:06 the levels of kind of maturity of the AI system and also the level of complexity. 0:16 Just the basic kind of understanding of how things work is like, again, very crash understanding. 0:23 If something's unclear, let me know, I'll clarify. 0:27 So the first level is like basic prompting. 0:31 Hey, ChatGPT, how do I solve this problem? 0:36 And this is the level operation that kind of we all started like back when ChatGPT came out in the market. 0:45 It's basically... 0:47 Request response interface and this is a level when you interact with system most 0:55 the second is like system prompt so what is the system prompt system prompt is the are part of the instructions that are always provided before your first prompt even arrives to a lab 1:10 So LEM-like is just archive, but every chat GPT, Claude, every chatbot is driven by system prompt. 1:19 System prompt is injected at initial level and very often are, like as an example, giving it some additional instructions to 1:30 challenge user more is something that you wouldn't want to do in the system prompt. 1:35 Because if you do it in basic prompt, it will likely be overridden with the system prompt instructions that are by default tuned to be like gentle, polite, and produce some results that are more agreeable to the user. 1:49 So it's like... 1:50 Important thing to understand that you can give as many like additional rules instructions in your basic prompts but if your system prompt is intrinsically designed 2:02 to tune system to behave differently it will likely point to failure so 2:09 You have your system prompt design well, but your basic prompts are kind of keep producing inconsistent results. 2:16 And this is like goes into the domain of prompt engineering. 2:21 So you can, they're like separate area of knowledge where you can explore more. 2:27 But typical most successful way to make your prompts work better is give a few short examples of how you would like the outputs to look like. 2:40 Hey, if I ask this, this is an example how I would expect the output. 2:44 If I ask that, that's an example. 2:47 In that case, that's how you would perform. 2:50 And this is just a kind of additional technique to improve the results of your queries. 2:58 So the... 3:00 First kind of level is chain of thought. 3:03 And then when additionally to few shot, you give LLM a multi-step instructions, how it should process your request. 3:15 You kind of think how you would approach this task and you'd give it basically a playbook. 3:20 Hey, first, look up all the competitors in the market. 3:26 Second, for every competitor, retrieve key facts like their volumes, their presence in the market, their social media. 3:36 Fourth, build a comparison table with the most fifths. 3:40 Identify how each of the competitors are different to our brand presence in the market. 3:46 This is kind of example of chain of thought, how it can be kind of used to improve the quality of your prompts. 3:54 So the fifth kind of tool to be employed is tool use. 3:58 It's how you can equip this autonomous system with access to data, with access to execution, with access to additional anything basically that you want. 4:09 You probably heard about MCP a lot. 4:13 And MCP is basically a USB for AI to connect to other systems and connect to the world. 4:22 So it's like basically superseding API level when API was built for engineers to plug and integrate with other systems. 4:30 systems, MCP was built to connect together AI and give AI access to external tools and communications. 4:38 So it can talk to your, for example, meeting platform, to your CRM, to weather forecast. 4:47 Or it also can execute some actions on your behalf. 4:50 It can turn light on and off. 4:52 So we can technically build MCPs that will be responsible for managing electricity grid. 5:00 So anything that you can give and you want your AI to perform has to be given like as a tool. 5:07 So most common way to do that is through a tool use. 5:12 And that's how you kind of equip AI with powers, not just to reason within a chatbot interface, but actually to perform some actions. 5:20 So that's a tool use. 5:23 But until that, it's still limited within the same chat interface. 5:29 You interact, you give requests, and you kind of receive and process the response and judge whether it kind of makes sense for you or not. 5:37 And then kind of it goes to autonomy is when you define some kind of goal and let the AI reason on this goal, perform 5:47 a loop of sequence tools, calls, chain of thoughts, reasoning, few short examples until it achieves its mission. 5:56 Or continuous exploration. 5:58 So exploration can also be a mission that's 6:00 Kind of an example of how open clock can be set up, for instance. 6:06 Okay, that was a lot I tried to cover in the shot, very narrative.