AI Interview Questions - LLMs vs. Agents with @aloo_explains

[00:00:01]Hello, sir. Hi. Hi, Gaurav.

[00:00:03]Please don't call me sir. You can call me.

[00:00:05]So, can we start with a short intro first?

[00:00:07]So I am applying for an engineer position at your company.

[00:00:12]I have some experience as a software engineer.

[00:00:14]I have built some AI applications and some workflows have automated in my company.

[00:00:20]So basically you are telling me that you have automation experience along with some software dev? Yes. Yes. Okay.

[00:00:26]So by automation, if I ask you this question, you know that you might have heard about agents and all of that because you have tried out AI agents and stuff.

[00:00:36]Sometimes. Yeah. Yeah.

[00:00:37]So can you tell me what is the basic difference between an LLM and an agent?

[00:00:42]An agent is something which can perform actions.

[00:00:47]And so the API calls that it makes is can be to external systems, can be to our own internal system.

[00:00:55]Also while an LLM cannot make API calls.

[00:00:59]Yeah.

[00:00:59]Can you take it to whiteboard.

[00:01:00]Then we can have an explanation there.

[00:01:03]Let's say you have a user who says, plan my trip itinerary.

[00:01:08]I'm going to someplace for a vacation.

[00:01:09]This query reaches your system.

[00:01:11]So we have a TripAdvisor system.

[00:01:12]What we can do is go and ask an LLM to make the plan for us.

[00:01:16]But there is a problem.

[00:01:17]The trip query can be something as simple as I am going to Austria, please plan a trip itinerary.

[00:01:21]But I do not know about the user.

[00:01:23]I don't know who they are, whether they're going with family, what are their flight preferences?

[00:01:27]What are the hotel preferences? Various things that are necessary.

[00:01:30]What would be ideal is if the LM could counter question the user, asking them for a form to be filled with their details, like what flights are you looking for? What are the dates?

[00:01:39]All of this stuff.

[00:01:40]After this form has been filled, the LM will go and make a search.

[00:01:44]Let's say a flight system which tells me the prices.

[00:01:47]So this would be my query to the flight system.

[00:01:49]And I would also connect with, let's say, a hotel system, which would tell me which rooms are available and at what price.

[00:01:57]Once I have the relevant context along with the prompt, that's when the LM can actually make a decision as to how should I plan the itinerary of this user, which would be forwarded to our system.

[00:02:11]And our system is going to give that as a response to the user.

[00:02:15]So, Gaurav, can you tell me, like you said, that LM is doing this query to the flight system, let's say any, any website or LM is doing this query to the hotel system or any website.

[00:02:23]So is LM making the call or is like LM deciding to make the call.

[00:02:29]The LM is deciding to make the call. That's a good point.

[00:02:31]This hotel system that we have, and the flight system is actually not being directly called by the the LM cannot make API calls.

[00:02:38]You have to have this interaction between your system, which has the right permissions to call the hotel management system, and your system which has the right permissions or context.

[00:02:50]This can be a system prompt on how to call.

[00:02:53]When should you call?

[00:02:54]These rules can be given and also you might need auth keys.

[00:02:57]You might need a way to pass this response.

[00:02:59]All of those things are tools.

[00:03:02]The best way to describe a tool is like a function call.

[00:03:05]Instead of thinking of this as an API call, it's going to be called a tool.

[00:03:08]And the LM decides that okay, I got the user query.

[00:03:12]I need more context.

[00:03:13]I need to know which flights the user can book so they can make a call to the flight system.

[00:03:20]Then they say, I also need some context around the hotel system.

[00:03:23]Is the LM making the call?

[00:03:24]Know they are telling the system in the response.

[00:03:27]That tool call search flights.

[00:03:32]Okay, but searchlights is like a function.

[00:03:34]It's actually a tool where we have described what the search flights do, which parameters it needs, the description of every parameter, and also what kind of values can be passed in.

[00:03:42]Once you get a response of those flights, like a function call, you take that and you pass it back to the LM.

[00:03:47]So all of the information of, you know, what are to what tools are available.

[00:03:51]And you just said that you need the description out of that.

[00:03:54]Where are you actually going to tell the LM about all the description or what all tools are available.

[00:04:00]That definition is there in the system prompt.

[00:04:02]And this system prompt can change for every single interaction that you have with the LM.

[00:04:06]Usually it is static though. Okay okay, cool.

[00:04:08]So can you also tell me that you said that it is going to call multiple tools, right.

[00:04:12]Is deciding to call multiple tools.

[00:04:14]So it should be in like what format is it sequential.

[00:04:18]Is it loop.

[00:04:20]Oh yeah it could be I mean it's a loop in the sense that the LM is deciding should I continue making tool calls or should I exit the loop.

[00:04:30]And here's the response with the exit.

[00:04:33]So the final way for LM to exit the loop is to either say, I can't get this answer sorry, or I got the answer.

[00:04:41]Great. I have enough context to answer you.

[00:04:43]That's the time when you actually forward this plan to the user.

[00:04:46]And this loop has multiple parts.

[00:04:48]As you can see, if there is a hotel call and a flight call, these are not related to each other.

[00:04:53]So they can be asynchronous.

[00:04:54]But if the flight call depends on the hotel call or the hotel call depends on the flight call, then you want it to be ordered and synchronous.

[00:05:01]Cool. Yeah. Thank you so much.

[00:05:03]It was a good answer.

[00:05:04]I cleared. Is it? Yeah.

[00:05:13]Yeah, boy.

[00:05:17]Hello, sir.

[00:05:19]So can we start with a short intro first? Yes.

[00:05:24]My name is Gaurav.

[00:05:25]I have more than 700 K subscribers on YouTube.

[00:05:28]I'm a very famous person, but I don't have any money.

[00:05:31]So I was looking to become an AI engineer now.

[00:05:40]Getting a response.

[00:05:41]And what's the difference between both of them?

[00:05:44]Sure.

[00:05:44]And LM is like made by OpenAI or let's say Hugging Face has some models.

[00:05:52]We can actually just call those models to get some text.

[00:05:55]While an agent is something that we usually billed as the engineers.

[00:05:59]And yeah, so an agent is what the end user interacts with and the LM is what we interact with.

[00:06:08]So yeah, you're clearly you don't have any.

[00:06:22]Clearly you are.

[00:06:26]So clearly you have worked more with your software.

[00:06:33]Be more explanatory like try to include some examples and all of that.

[00:06:41]Agent is like a person.

[00:06:49]You can have a water.