EP 010 - Truth About "Smart" Agents You NEED to Know (intelligent automation agents)

00:00:01: Hello and welcome to our next episode of the AI Boardroom.

00:00:06: Today we have a pretty interesting topic which got a lot of buzz this week.

00:00:12: Serana, tell our audience what this is all about.

00:00:16: Yeah, so today we'll be talking about automated agents.

00:00:20: We'll be covering what they actually are, the types.

00:00:25: So there are different types of agents and what they mean for you, how to actually

00:00:30: implement them, and then what does it mean for your business.

00:00:34: And I think what Edgar was kind of alluding to as far as like being the buzz

00:00:37: of the week is because of Google Next.

00:00:40: And that was a hot topic that was discussed, it happened to be...

00:00:44: be able to join kind of the conference and observe all the events, talk to lots of

00:00:50: vendors, go to sessions and kind of learn, but it was a very strong theme, I think,

00:00:57: at the conference for sure.

00:01:01: So agents, I think in general, people are looking to automate their workflows.

00:01:05: They're looking to implement AI, not just in sales marketing and...

00:01:12: I want to say customer service agents, but we are seeing kind of a lot of promise of

00:01:19: using agents across different workflows.

00:01:21: And I think before we even joined the podcast, you know, Edgar and I were

00:01:25: talking about even for coding or automating workflows as it relates to your

00:01:29: day -to -day work.

00:01:30: You don't have to, again, use it for business necessarily, but how do you even

00:01:35: implement these agents in your day -to -day?

00:01:37: I mean, that is possible.

00:01:38: So I think...

00:01:39: Again, automation and efficiencies are something that can impact your day -to

00:01:46: -day work, but then you can also introduce them into your personal life.

00:01:49: So we'll focus on the business side, but just wanted to mention that it's

00:01:53: definitely something that's being talked about more this week, for sure.

00:01:59: Definitely, yeah.

00:02:00: Yeah, I was kind of curious.

00:02:02: I was also writing to you how it was in Vegas.

00:02:07: I think you still haven't told me the evening stories.

00:02:12: We have to get to this after the podcast.

00:02:16: Of course you didn't.

00:02:17: No, no, no, that's the version for your employer.

00:02:21: Yeah.

00:02:21: Save it.

00:02:26: Okay, no, and I was kind of curious because I've read a lot of headlines on

00:02:32: YouTube and agents were like the Google is all in agents and foreign agents and stuff

00:02:38: like that.

00:02:39: And then after watching the actual keynote and also I watched also the developer

00:02:44: keynote, I was kind of, yeah, yeah, maybe.

00:02:51: What is it?

00:02:52: Yeah, it directly came to me like what is an agent and it was kind of also the idea

00:02:59: of today's episode to dive more into what are agents in the AI world and also what

00:03:08: did Google mean and how did we solve stuff earlier and before.

00:03:13: So yeah, we have a bit of a roundup and deep dive today on what this actually

00:03:18: means.

00:03:20: So I think Solana, you brought it up.

00:03:23: So like we have, you said we have like, we have had already and still have today rule

00:03:30: -based agents.

00:03:31: And we now get into a realm where it's kind of slowly moves into not only rule

00:03:38: -based but also intelligence and intelligence agents, which you love to

00:03:43: elaborate on that.

00:03:44: Yeah, and so I think a lot of people when we talk about, or a lot of buzz that I

00:03:50: would say you would hear outside of these conferences is maybe LinkedIn, this is

00:03:55: where I hang out quite a lot, is automating workflows with Zapier.

00:04:01: And so I just wanted to kind of differentiate that with Intelligent Agent

00:04:06: because Zapier is a rule -based engine.

00:04:10: So and the way that you kind of define and...

00:04:14: differentiate between intelligence and rule base if you have to explicitly spell

00:04:18: out like when you experience this then you do this you're basically are spelling out

00:04:24: spelling out tasks and rules for it to follow and it's limited to the

00:04:31: integrations that it has so Zapier is really not that intelligent but a lot of

00:04:37: there again sales automation agents or i'm sorry the automated agent

00:04:41: experts in AI are selling you this as kind of automations that are intelligent.

00:04:47: And I've seen that quite a lot.

00:04:49: So I just want you to kind of emphasize that the Zapier and those types of tools

00:04:54: that are just automating when you have to spell out explicit rules, it's not

00:04:59: intelligence.

00:05:00: So.

00:05:00: also, like you have to, every branch has to be described as its own way to go.

00:05:06: Whereas if you have true intelligence, what does it actually mean?

00:05:10: Like we have, and that's basically also like, if you look at today's chatbots,

00:05:14: they look not that different, but they are a lot more flexible.

00:05:18: And that's because they are not pre -programmed to specific answers, but they

00:05:24: are programmed to a specific cause.

00:05:27: And...

00:05:28: following that course, they do what they have to and what they are able to with the

00:05:33: toolbox they have been given.

00:05:35: So that's basically like, that's another paradigm.

00:05:38: And that's the same for automation.

00:05:40: So we had Zapier, you also have on the Microsoft side, Power Automate.

00:05:45: I think make .com is also something that came up last year.

00:05:50: So these are all...

00:05:52: like automation workflow tools.

00:05:54: Like we had workflow tools basically from my whole career.

00:05:57: I started in IT in 2010 and a bit earlier like in school and stuff like that.

00:06:03: Like through all that time we had workflows and automations.

00:06:07: And what you mentioned the AI consultants that are running out there and AI

00:06:13: automation agencies, I think they're called.

00:06:19: What they...

00:06:20: just kind of emphasize maybe before you go on and the reason why they're calling them

00:06:23: AIs because you can actually integrate Zapier with AI tools, but that doesn't

00:06:28: make them intelligent.

00:06:29: Yeah, I'm sorry.

00:06:31: I stole your thunder, I'm sorry.

00:06:34: yeah, like Varna said, we have now the option to take these workflows and

00:06:40: automate parts of it.

00:06:42: One example would be I made a chatbot for, I think it was a website which needed

00:06:50: marketing, which needed to onboard marketing customers.

00:06:57: And through that onboarding, I had unstructured data in the form of...

00:07:02: They just typed in their CI and their basic company stuff, uploaded their logo,

00:07:11: that stuff.

00:07:12: So we had still a path to go, but instead of doing every path all the time or every

00:07:19: branch, I was able to integrate AI and significantly, significantly,

00:07:24: significantly, that's the word today, that's like significantly reduce.

00:07:31: reduce the amount of workflow, like parts I needed, because I could do it partly in

00:07:39: a loop.

00:07:39: So I replaced all the branches that were pre -programmed with one AI loop.

00:07:44: So there is intelligence in there, but it's not like the AI itself deciding where

00:07:51: to go and what to do, and maybe come up with new ways, which would be actually

00:07:55: intelligent.

00:07:56: And I like that definition.

00:07:58: I think hopefully for the audience, it's easier to like also decipher when you kind

00:08:01: of break down.

00:08:03: I think AI in general is better understood when you break these distill it down to

00:08:08: the essence.

00:08:08: So if you explicitly have to program rules and define workflows, very likely that

00:08:15: you're dealing with rule -based automation.

00:08:17: If there are things that you can rely on these services or tools that.

00:08:23: You don't have to explicitly spell out every possible scenario.

00:08:27: I think as Edgar was mentioning, that is AI enabled automation basically, or an AI

00:08:33: truly artificial intelligence automation tools.

00:08:36: And I think the other beauty behind it is that they learn based on the paths and

00:08:41: based on the workflows that you actually execute, and then they become better over

00:08:44: time.

00:08:44: And I think that's the intelligent piece is that there's some machine learning,

00:08:48: some enhancements and some learning mechanisms that are embedded in that

00:08:51: process.

00:08:52: So,

00:08:52: you don't have to explicitly spell out every possible scenario and that's where

00:08:57: the automation piece comes in.

00:09:00: right.

00:09:00: For example, I built one solution where the AI gets like a complete document and

00:09:08: makes a step -by -step tutorial from the document.

00:09:12: And there are no pre -programmed steps whatsoever.

00:09:14: It just decides what to give out, if anything, and also decides, okay, I

00:09:19: haven't found anything useful for you.

00:09:22: We have to solve it another way.

00:09:24: So, and that's where, and even that...

00:09:26: could debate if it's really intelligent, but yeah, it doesn't need tutorials, it

00:09:32: just needs a document describing the process and it makes a tutorial from it

00:09:36: automatically.

00:09:37: So that's kind of, but we could maybe take it later on the episode where I can

00:09:43: explain how you get this process intelligent because I thought already

00:09:47: about that.

00:09:48: But yeah, let's first of all talk again about Google.

00:09:53: Google was in the...

00:09:55: on the next keynote.

00:09:57: And as far as I understood, they basically call everything an agent because they have

00:10:01: customer support agents, which are basic chatbots with a knowledge base.

00:10:05: Like it's usually not called agent, it's a rack pipeline.

00:10:09: And you have a chatbot with retrieval argumented generation, which is fine

00:10:16: because the whole grounding is also a word which more and more gets into all the AI

00:10:22: stuff.

00:10:24: grounding is pretty pretty useful because it just grounds everything the chatbot

00:10:28: says on actual documents but it's not an agent at least not how I understand it

00:10:36: Yeah, I'm grounding just for the audience, I think, as you were kind of explaining

00:10:40: it.

00:10:40: So it avoids hallucination.

00:10:42: So if you're basically saying that, I think it's pretty big for businesses,

00:10:48: actually.

00:10:48: So if hallucination is a big risk for your organization, we've seen that where you

00:10:53: have airlines and, I want to say, automotive agents being, not agents, but

00:10:59: chat bots being customer facing, and they just make up stuff.

00:11:03: You want to make sure you limit.

00:11:05: their responses to whatever is available in your documents.

00:11:09: And so if you're, if it doesn't exist in the documents, they basically say, hey, I

00:11:14: don't have that knowledge, I can't respond.

00:11:16: So I think that's, that's what I think Edgar means.

00:11:19: But I think it's a huge consideration.

00:11:21: I heard a lot of that, especially if you're in the banking industries, or any

00:11:25: of the highly regulated industries, that's really huge.

00:11:28: So, and there was a lot of emphasis at the conference on that too.

00:11:33: Yeah, definitely.

00:11:34: So, and that's where kind of like the agent wording fell apart a bit on the

00:11:41: whole event because they use basically everything that integrates Gemini was an

00:11:47: agent.

00:11:48: So, no.

00:11:50: But if that's not an agent, what is?

00:11:53: And they had another really interesting example, which I found, which was shopping

00:11:59: assistant.

00:12:00: Mm -hmm.

00:12:00: And that was more of an agent to me than like just normal chatbot because they even

00:12:06: showed it.

00:12:07: They made a call to the call center, which also was an AI agent.

00:12:13: And it updated her chat during the call, which was, I think it was in the developer

00:12:20: keynote.

00:12:21: I found it fascinating because she was like, she uploaded a video and said, I

00:12:26: want the same shirt as the lead singer has.

00:12:30: And then it searched the shirt, which was looking like this.

00:12:33: And then she called to order, and it updated the stuff in the actual chat,

00:12:39: which I found, like, this was not only multimodal, it was also multi -channel.

00:12:45: And that's more of an agent, from my point of view, because there you have stuff

00:12:50: happening in the background, stuff is creating, stuff is searched, and it's more

00:12:54: of a system of...

00:13:00: AI tasks working together basically.

00:13:05: And I think that's also something I have to give Google props for.

00:13:10: They had this shop, which was the overarching theme or example they went

00:13:16: through.

00:13:16: And they had pretty good examples, I find, to how a business would use them.

00:13:21: And something maybe for people to test out because I did talk to someone on that

00:13:26: particular topic too, although I didn't see the developer keynote.

00:13:30: If you wanted to experience what Edgar is talking about as far as like being able to

00:13:36: shop for specific items, go on Pinterest.

00:13:38: I recently explored that.

00:13:40: Pinterest has that option.

00:13:41: So if you, let's say, you probably can't upload the image, but if you're browsing

00:13:45: something, and I thought it was pretty cool.

00:13:48: in your browsing of an image and you know sometimes it's not really that image is

00:13:52: not linked to a shopping page sometimes it's just an image that's associated like

00:13:56: a stock image with an article but you're like that's a really cool shirt and if you

00:14:00: can actually you can actually double tap it and it'll locate an item that looks

00:14:05: like that for you to shop.

00:14:07: it's nothing new in general.

00:14:09: Google had the stuff also implemented.

00:14:14: But now having...

00:14:18: I don't know.

00:14:19: That was not even their example from the demo that they had.

00:14:24: That was an actual customer of Google.

00:14:26: And they had their pictures indexed in a vector database.

00:14:32: And that was what I found interesting because...

00:14:34: I never thought about it, but it makes total sense if we have multimodal models

00:14:39: to also vectorize pictures and then find the nearest neighbor, basically.

00:14:44: So that was kind of cool.

00:14:47: I like that a lot.

00:14:51: But yeah, still not an agent for me.

00:14:58: So for me, AI agents, this is...

00:15:02: basically where the whole industry is heading into like in that direction.

00:15:08: And it's a topic which I follow like for like the first thing I wanted to build was

00:15:14: actually an agent in AI, but I wasn't, I was limited by the technology of my time

00:15:20: basically to quote Iron Man.

00:15:26: Yeah, to yeah, what I...

00:15:29: think an agent has to be something that works at least partly autonomous.

00:15:38: And there is a bit of autonomy in the search itself when you talk about the

00:15:45: shopping example.

00:15:47: But it's still a passive system.

00:15:50: I still have to say, hey, I have this request.

00:15:55: Please make something that solves this request.

00:16:00: Yes.

00:16:01: What I understand from an agent is, I have this goal, please reach this goal and do

00:16:10: everything you need on the way.

00:16:12: So it basically doesn't start with a problem, but it starts with a goal and all

00:16:19: the issues along the way which arise to reach that goal, it creates tasks for

00:16:26: itself and solves them.

00:16:28: And that's actually an agent for me.

00:16:31: Yes, and I totally agree.

00:16:32: So it would be multi -step, you define the goal.

00:16:35: And then, so for example, in the shopping example, it's actually, I want to buy the

00:16:40: shirt.

00:16:41: So the agent would be able to find based on the picture of whatever the prompt is,

00:16:47: being able to figure out which API to use to which store that is likely, and then

00:16:53: put that shirt into your cart and potentially take you to the...

00:16:58: basically checkout page, so all you have to do is just agree to pay for it.

00:17:02: So it resolved the issue of finding or browsing websites, finding the store,

00:17:09: finding all of these things.

00:17:11: So you don't have to check with a human along the way.

00:17:14: It basically takes you from what you want, from a problem to a solution, and in

00:17:21: however many steps it requires, and it just kind of figures out the rest.

00:17:25: And you can even bring it one step further if you have some cues or trigger which

00:17:34: actually, like the agent is able to like every day look at your emails and do the

00:17:40: work.

00:17:42: You know, like the goal is to keep your inbox clean and then the agent goes ahead

00:17:47: and does everything he needs to.

00:17:49: And I'm personally a big fan of...

00:17:54: partly autonomous agents which have some boundaries where they need human

00:17:59: intervention to keep going or to fulfill a task.

00:18:02: For example, if you have a social media agent, which every day scours your

00:18:09: internal company documents and looks for interesting stuff that happened and

00:18:13: creates a post, for example.

00:18:16: Let's just imagine it takes into consideration contents of meetings.

00:18:21: You might...

00:18:22: want to look over the post before it's posted.

00:18:26: So the last task, which is basically post on LinkedIn or something, there would be a

00:18:33: task where the boundary says only with human intervention, and it creates a task

00:18:37: where the human has to confirm for it to keep going.

00:18:41: Which doesn't mean it cannot work on other posts, but this one specific post is in

00:18:48: the line for being confirmed to then go on.

00:18:51: I think it's...

00:18:52: You have the same thing in the European systems where people change customer data,

00:18:57: for example.

00:18:58: And if it's finance data, then finance has to approve the change.

00:19:03: So stuff like that, it's common also for natural intelligence.

00:19:06: So why shouldn't it be for AI?

00:19:10: So go back to the social, I think, example, because I see a lot that maybe

00:19:13: this is going to be a hot topic for people to, that people are curious about.

00:19:18: So we do see a lot of agents being used on commenting.

00:19:23: And I think, you know, a lot of times people are looking to grow on LinkedIn,

00:19:27: and they try to make themselves visible on people's posts.

00:19:30: And so, and clearly, some of these comments are automated, like there's no,

00:19:37: I'm sure that there's no human,

00:19:39: in the loop involved because there's one post that was pretty funny that happened.

00:19:45: You know, someone, you know, I collaborated with someone and I think the

00:19:49: whole post was, you know, my collaborator basically kind of sharing his experience

00:19:56: and how thankful he was for the collaboration.

00:19:58: And then I think whoever posted the comment was like, oh, I too, I'm so glad

00:20:04: that you, Svetlana, have been, had a chance to collaborate with me.

00:20:08: So.

00:20:09: clearly got the context wrong.

00:20:11: And it was like completely.

00:20:15: So, so he didn't understand the context of like what was basically said or whatever,

00:20:22: but then there's also again, like a lot of discussions that are happening that people

00:20:27: are just kind of again, like there's like clear tell signs that this is automated,

00:20:33: like generation.

00:20:34: Can you maybe say a little bit more as to like how.

00:20:37: some of these agents actually work?

00:20:39: Because again, we see them a lot across social media platforms.

00:20:43: So how does it actually work?

00:20:46: So for me, it's all about task planning and execution.

00:20:51: So like one thing when talking about LLMs, oftentimes arises is that the systems

00:20:57: themselves are not able to plan.

00:21:00: That does not mean generally planning is out of the window.

00:21:05: It means just in a request, it's not planning ahead.

00:21:08: So it just generates the next token.

00:21:11: But what it can do,

00:21:13: And that's how you basically would start implementing it is first is how to, yeah,

00:21:19: like to implement a system, to tell the AI how to fill the task list and where to get

00:21:26: the tasks from.

00:21:28: Because like, like I always say, LLMs are text in, text out, and everything has to

00:21:34: be solved by language.

00:21:36: And you want basically to go from unstructured, your goal, to structured,

00:21:41: your task list.

00:21:42: And that's how it would start.

00:21:46: So you will go ahead and the first thing you have to solve is make one AI persona,

00:21:53: usually there are several AI personas working on the request to plan out all the

00:22:01: steps to get to the goal.

00:22:03: And then it goes ahead and basically with classical programming, you take one step.

00:22:09: after another, or maybe even you could do it with function calling and the AI gets

00:22:13: the next task itself.

00:22:16: But one way or another, you get one task, you execute on the task, you finish the

00:22:23: task, and then you go to the next task.

00:22:24: And that's basically the most basic implementation of that.

00:22:30: So if the context is write me LinkedIn comments, then one task would be, okay,

00:22:36: get the current LinkedIn feed.

00:22:39: Second task would be identify posts which are interesting and worth commenting on.

00:22:47: Third, generate the comments.

00:22:49: Fourth, ask the human if the comments are fine.

00:22:54: And then fifth would be post the task, post the comments.

00:22:59: And that's basically you use the AI as a brain for

00:23:08: planning out and doing all the text work, language work basically.

00:23:14: And then you go ahead and do basic classical programming to execute on that.

00:23:18: And you also can do, can prepare tools and tell the AI, hey, this is the tool you

00:23:25: have, you can execute on it.

00:23:27: But this basically you have to combine the thinking, LLM, with the execution, which

00:23:32: is like normal programming.

00:23:34: That's like the basic concept of how you would do it.

00:23:39: And I think you have to also focus on the tasks.

00:23:41: Again, I'll say this because I think I just see a little buzz also or kind of

00:23:47: talk on this on social media platforms.

00:23:49: I think using chat bots on social platforms is cheating.

00:23:56: You're taking claim for or leaving comments as if it's your presence and

00:24:02: making people feel like you were there to actually.

00:24:05: look at those comments and really think of an output to relationship build.

00:24:11: But then there's really none of you there.

00:24:15: And people, again, follow you and subscribe to you.

00:24:17: And this is why you kind of see these big followings.

00:24:21: But then people don't really communicate with them or don't relationship build with

00:24:27: these individuals because clearly they're their child.

00:24:30: So I feel like...

00:24:32: Again, there's use cases where chatbots should be included, but I don't think

00:24:37: chatbots on social media is one.

00:24:40: Yeah, I think it's cheating.

00:24:41: I'm always a bit...

00:24:46: Yeah, for example, a lot of times in my career, I had people helping me with

00:24:53: social media.

00:24:55: And because I was doing a lot of other stuff and I was just not putting in the

00:25:00: work for, or wasn't able most of the time to put in the work, create the post, craft

00:25:06: it, do a going camera, create the visuals for that.

00:25:11: I never did outsource commenting.

00:25:16: On that I'm kind of with you.

00:25:19: You either want to say something or you don't, but let AI comment on your behalf.

00:25:27: That has also been done by employees of influencers and companies and stuff, so

00:25:33: it's nothing new.

00:25:35: Is there actually a difference for it to be done by AI?

00:25:40: I think so.

00:25:41: I mean, even as you mentioned, I think that there's two schools of thoughts,

00:25:45: maybe one is you leave a comment because you truly are kind of reflecting on

00:25:50: someone's posts, like someone invested the time.

00:25:53: But why would even your employee comment on something that, you know, someone else

00:25:59: has written, right?

00:26:00: So like, what's what's kind of the point?

00:26:01: Like the only purpose, again, between using these.

00:26:05: automated agents basically that are running through social media platforms is

00:26:10: to gain visibility.

00:26:12: And so because you're into grow following, which again, I think it's again, a

00:26:18: cheating way to do that because there's really no authenticity behind it.

00:26:24: And so, and this is why I think I actually value even if you have a huge following

00:26:28: and we know that some of the...

00:26:33: bigger influencers, I think, in the industry.

00:26:35: They don't have the time to comment.

00:26:36: So guess what?

00:26:37: They don't release chat bots or their employees to comment on people's stuff.

00:26:42: Sometimes they'll leave a few comments, but I do think that authenticity is really

00:26:47: huge if you want to really stick to it.

00:26:49: So again, there's use cases.

00:26:50: And the reason why I'm saying this is there's use cases where you should

00:26:54: consider, is this something that I should be using AI for?

00:26:57: And this goes back to ethics.

00:27:01: But there are use cases where, I mean, it's game changing, right?

00:27:05: So things that are truly helping you build efficiencies into your workflows.

00:27:10: But again, when there is those questions like, should I be doing it?

00:27:14: Like ask a friend for an opinion sometimes.

00:27:18: Get a second opinion is all I'm basically hoping to inspire.

00:27:23: And one thing to always keep in mind, you always are in danger of posting stuff you

00:27:33: didn't want to post.

00:27:34: Back in the day, to be honest, I used Lampod, which is basically you are in the

00:27:41: community, and Lampod automates the interaction between the community members.

00:27:48: And I...

00:27:51: Yeah, you get likes and if you want to even get comments and you can even pre

00:27:56: -write the comments for your posts.

00:28:01: And you can comment on other people's posts automatically.

00:28:05: The thing is, I deactivated it after one day, I guess, or so.

00:28:10: Because if you comment on other people's posts and it's a post about right -wing

00:28:16: politics or something, you might not want to have your name under that.

00:28:20: It's like awesome!

00:28:27: Yeah, so that's definitely something to keep in mind or to have an eye on.

00:28:33: What are you actually commenting on?

00:28:37: Because we know how social media works.

00:28:42: You don't look properly and you're so quickly switching into social media

00:28:48: bubbles, which you didn't want to get in.

00:28:50: So yeah, I think there is a lot of risk involved too.

00:28:55: The same, like, yeah, you wouldn't do like completely automated sales emails without

00:29:00: actually looking over it.

00:29:02: So, or at least you shouldn't.

00:29:04: I think it's like kind of the same.

00:29:06: it's done.

00:29:07: And this is what I'm saying again, like before you automate these pipelines or

00:29:12: these workflows, just kind of consider the impact that those automations will create

00:29:19: as human in the loop, I think, a big portion of it.

00:29:23: And I would say maybe marketing sales, depending on industry, but I would say,

00:29:26: you know, some of these automations like human in the loop is like a must,

00:29:30: especially in highly regulated industries, you would never want to release.

00:29:35: an automated comment, nothing, I would say, like in healthcare and banking.

00:29:40: There's a human always involved, so make sure you consider that.

00:29:43: So don't release anything that's truly automated end to end without having kind

00:29:50: of human in the loop approved for at least that workflow.

00:29:52: Yeah.

00:29:54: There is also, we will come shortly to the challenges.

00:29:57: There's also another thing you have to keep in mind for that.

00:30:01: But before we get to that, I would love to just mention some of the frameworks.

00:30:06: So Google now created, or released, showed Vertex AI agent builder on their cloud

00:30:13: platform.

00:30:14: As far as I understood, and I didn't try it out, so it's like with a grain of salt,

00:30:19: it's more in the realm of GPTs.

00:30:22: Mm -hmm.

00:30:23: OpenAI, which also has to be mentioned here.

00:30:29: Yeah, like this is not an active system.

00:30:32: It's still more of a prompt builder kind of thing where you are able to argument

00:30:39: your context with actions, with third party sources, or with documents in form

00:30:44: of a rag pipeline.

00:30:45: So it's not as sophisticated.

00:30:48: It's even not as sophisticated as some of the demos they showed.

00:30:52: Mm -hmm.

00:30:53: cannot build that with the Vertex AI agent builder, which is kind of sad because it

00:31:00: would have been possible to put a bit more capability into that.

00:31:04: Yeah, I agree.

00:31:07: and then there's open source as always.

00:31:09: And there are two projects which I really, really want to emphasis, put a bit of an

00:31:14: emphasis on.

00:31:15: One is Autogen, which is a Microsoft framework which lets you build AI agents.

00:31:22: And one of the main things to be considered there is like you have

00:31:27: different roles and different personas working together.

00:31:32: And that's...

00:31:34: actually what the framework enables you to to easily or like more easily manage it to

00:31:40: not be the position to program everything yourself.

00:31:44: And there is another completely open source, I think at least there is no big

00:31:50: company behind that.

00:31:52: Or at least not released it because Autogen is basically released by

00:31:55: Microsoft.

00:31:56: And it's called Crew AI.

00:31:59: And I think Crew AI even has a like Pokemon like UI where

00:32:04: you see the people working together and stuff like that.

00:32:08: So yeah, Pokemon, yeah.

00:32:12: you if you like that, that may be a motivating factor for you to try it out.

00:32:18: Yeah, I think, if I'm not completely mistaken, but I think Crew AI was the one

00:32:23: agent platform which has also, might also be something else.

00:32:31: I might be mistaken on that one, I'm sorry.

00:32:34: But Crew AI is basically the latest and most interesting open source project I've

00:32:45: seen and it's...

00:32:47: If I see modern agent videos a lot of the times they use WhoAI because yeah, it's

00:32:53: just working well.

00:32:55: And I think we talked about GPTs also as some sort of, I guess, belonging to the

00:33:03: intelligent agent space.

00:33:06: So how so?

00:33:07: Because I thought it was, you know, a lot of the capabilities of GPTs right now are

00:33:13: basically glorified prompt agents, like prompted agents.

00:33:19: if we mentioned vertex, yeah, agent builder, it's just kind of like, to be

00:33:25: complete, we have to know the GPTs.

00:33:27: From our definition, it's not an agent, but they sell it like it, and people will

00:33:34: read agent in the context of that, so yeah.

00:33:39: Take it as you wish, I will not say like,

00:33:44: force anyone to use agents like we use it, but if you talk about AI agents, from my

00:33:52: point of view and my own opinion, you have to have at least partly autonomous

00:33:57: systems.

00:33:59: Yeah, I guess the playing the devil's advocate, you could potentially, so if

00:34:06: your workflow involves just creating content, right?

00:34:09: So you could potentially automate a big portion of your workflow of creating

00:34:15: content, if that's literally all you're kind of doing, which will probably involve

00:34:19: research, crafting an outline, and then generating.

00:34:22: So like there are a few steps in between for you to get from a blank sheet.

00:34:27: to a polished article, right?

00:34:31: So you could, again, depending on what your workflow is, but it's not going to be

00:34:34: able to automate a conversation.

00:34:37: You know, the GPT having a conversation with a client.

00:34:42: So there's certain tasks that I think it could be, and it's not even trained, but

00:34:47: like prompted or system prompted in order to kind of fulfill as far as multiple

00:34:52: steps.

00:34:53: But still, I think you are kind of...

00:34:55: providing it context, you're explicitly telling it what to do and how to

00:34:59: potentially achieve that task.

00:35:01: So again, depending on how you look at it, you could argue that there is some

00:35:05: intelligence that it's automate.

00:35:08: Some of the steps are automated, but still, it's not truly hands off.

00:35:14: Yeah, yeah, yeah.

00:35:16: And you have to activate it.

00:35:18: And that's where I would love to have its last topic, like what does it mean for you

00:35:24: and your business.

00:35:26: So we've said it at some point already, but we have passive stuff, which is

00:35:32: reacting to requests and try to solve the specific requests.

00:35:37: And then you have active stuff, which kind of is.

00:35:41: meant as autonomous agents which actively create their own tasks and solve them.

00:35:48: And like solve them one by one to reach a certain goal.

00:35:52: And this means basically you can do more complex stuff.

00:35:59: If you do it right, you can basically create virtual employees for that, which

00:36:06: are doing work and solving stuff for you.

00:36:11: And yeah, I think I don't have to tell any business owner how huge such an

00:36:16: opportunity might be.

00:36:19: But there are also challenges involved in that.

00:36:23: And yeah, would you elaborate on that?

00:36:28: Yeah.

00:36:29: So I would say cost and the precision, right?

00:36:34: So you want to make sure that if you're truly automating kind of workflows, that

00:36:41: depending on the solutions or what AI tools you've actually implemented, you can

00:36:48: run into some big inference costs or just costs in general, because again, you've

00:36:53: automated these workflows and you're kind of releasing them into the wild.

00:36:56: especially if you don't have human in the loop.

00:37:00: Some of the costs of these models, especially if you have Generative AI

00:37:03: plugged into those workflows, I think that's the biggest eye -opening kind of

00:37:10: situation is Gen .AI is quite expensive.

00:37:13: So you think the costs go up significantly depending on how many tasks you're

00:37:19: actually running.

00:37:20: So especially if you're doing that as part of your automation workflows.

00:37:24: they can run a big.

00:37:25: And so when you see your first bills, you want to make sure that whatever workflow

00:37:31: you're automating, it's actually worth the investment into the AI.

00:37:36: So the goal is for you to build efficiencies.

00:37:39: And efficiencies, the way they translate to business leaders, I think is ROI.

00:37:44: So what process are you actually trying to improve?

00:37:47: And what's the cost of that process end to end?

00:37:50: And then...

00:37:51: what would be kind of the projected cost for you to actually do this with agents.

00:37:55: So you want to make sure that you've introduced, you didn't introduce a process

00:38:00: that actually costs you more than to do it in kind of the old way.

00:38:04: So there's an ROI ballpark calculation that you'll need to do to make sure that,

00:38:09: you know, it's worth investing, not just for the sake of technology and saying,

00:38:14: hey, I have AI automating my workflows.

00:38:16: It actually has business sense for you to really implement.

00:38:20: And I think you want to make sure that you also look at precision.

00:38:26: And so, and I think, and maybe you'll kind of, I'd love to get your angle on this

00:38:31: too, but humans are really good at adopting.

00:38:35: And so one of the things that AI in general is looking to do is to kind of

00:38:41: replace how we think.

00:38:42: But it's not the same level of competence as we are.

00:38:46: So sometimes some of the processes...

00:38:48: that we run on a daily basis, yes, they're manual, but we are able to be flexible

00:38:53: enough to accommodate new situations and things that we would call edge cases.

00:38:58: We're just intelligent enough to be like, hey, this looks like, you know, this

00:39:02: process, let's do it.

00:39:03: So the accuracy, as you mentioned, like precision, sometimes with humans is higher

00:39:08: than these machines because they're still learning.

00:39:10: So again, depending on the maturity of your model and your process, you may

00:39:14: experience some.

00:39:15: issues as kind of these automated workflows encounters, especially like

00:39:21: these edge cases that they're not, have not been exposed to in kind of their, in

00:39:27: previous.

00:39:28: So yeah, I'd love to get your thoughts on that as well.

00:39:31: Yeah, the precision part is actually the one which is really hurtful because LLMs

00:39:40: are not 100%.

00:39:41: That's what we learned with all the usage of it over the time.

00:39:46: If you think about a system which is not 100 % and at some points even a bit lower,

00:39:57: like 85%, 80%, 60%, who knows?

00:40:01: on one task, and if you just add up enough tasks, and they are all a bit like, it's

00:40:08: like building a tower, and every bit is a bit misplaced.

00:40:13: And at some point, if the tower gets high enough, it just breaks.

00:40:17: And that's basically where we're at.

00:40:22: We are, like, this is something that will be solved with time.

00:40:26: So if you now implement an agent,

00:40:29: and it's maybe not as capable as you would love to have it because of that lack of

00:40:35: precision, this will solve itself by time.

00:40:40: We get newer models which will be more precise, which will do less mistakes and

00:40:44: less hallucinations, and then we get to it.

00:40:48: Another thing also connected to the complexity is context window, which at

00:40:56: least was an issue.

00:40:58: until this week or last week.

00:41:02: Because now we get to the models with a million context, but I think context is

00:41:07: not an issue anymore.

00:41:08: But that's something to be at least aware of that you have context that's building

00:41:15: up because like you create a task, the task had output, you have to consider the

00:41:19: output for the next task and to solve it and stuff like that.

00:41:23: So...

00:41:24: like splitting it up in separate tasks like reduces context sizes significantly,

00:41:28: what is it today?

00:41:31: Significantly, yeah, I don't know.

00:41:34: By a huge amount.

00:41:36: And that's really weird.

00:41:40: Yeah, it reduces it by a huge amount and but you still, if the task is complex

00:41:45: enough, for example, if you have like a lot of code which you wanna execute,

00:41:50: context might still be an issue, especially if you use.

00:41:53: like local models.

00:41:55: So yeah, that's that.

00:41:59: I'm...

00:42:00: note that I'll emphasize that the error being compounded, but I think it's not a

00:42:05: unique issue, which I'd love to talk about maybe one of the future episodes too is if

00:42:10: you're orchestrating multiple systems together, you have to look at the error

00:42:15: rates of each, not just tasks with automated agents, but those precision

00:42:21: rates are like these error rates compound.

00:42:25: And so you have to track kind of where...

00:42:29: and enhance each one of those steps.

00:42:30: So there's monitoring and observability that you have to kind of implement to make

00:42:34: sure that you have a good grasp as to what that's happening.

00:42:37: And so that you have an understanding, okay, well, what part of the process can

00:42:41: you really improve?

00:42:42: So there are things, again, we're not still at the level of maturity that these

00:42:46: automated agents are plug and play.

00:42:49: So there are some as you kind of build them, and I think this is where, Edgar,

00:42:53: you have a lot of experience.

00:42:55: There's some fine tuning that you have to do or some...

00:42:57: Yeah.

00:42:59: kind of molding of the system that you have to do so they're not completely out

00:43:03: of the box.

00:43:05: use fine tuning, stuff like that, yeah.

00:43:08: So do anything you can to avoid hallucinations, but it's the same for

00:43:14: every AI implementation whatsoever.

00:43:17: You have to keep the error rate down.

00:43:20: Like I always say, usually in classic programming, you had input A and you had

00:43:25: output B.

00:43:26: Now you have input A and you have output A to C or A to D.

00:43:31: a good play, I like that.

00:43:32: B, but you have to work with A to B.

00:43:34: But it's fine.

00:43:36: Error correction was something that we had at IT forever, and now we have just a

00:43:43: different way of error correction.

00:43:47: And yeah, it's solvable.

00:43:50: I think I still believe, and I'm a strong believer, and not everyone would agree

00:43:54: with that, but I'm a strong believer that we are...

00:43:58: really close to real intelligence just by engineering it right.

00:44:04: So I think it's not a matter, well, not completely the matter of only getting

00:44:10: better system, but it's also a matter of engineering the right implementation.

00:44:15: So that's why the correct integration of AI systems and knowing what's possible and

00:44:20: knowing the ins and outs is really valuable.

00:44:24: So before you start anything,

00:44:26: try to get this right and try to anticipate what issues you could face and

00:44:34: how to solve it.

00:44:36: And most of them are solvable.

00:44:38: So yeah, that's just a hint that I can give.

00:44:43: Yeah, I think value packed again, hopefully you've found something out of

00:44:51: this episode that you kind of stuck to you.

00:44:54: And we'd love to if there's anything that you want us to dive in a little a level

00:44:59: deeper, or if you want to come as a guest with us, please join us.

00:45:04: We always like to have different guests talking in different topics and speaking

00:45:07: of guests, we will have Mike.

00:45:10: joining us and we'll talk more about his background but on AI for recruiting next

00:45:17: week.

00:45:18: So really excited about that episode.

00:45:20: Yeah.

00:45:21: Because recruiting is a hard task and having it automated with AI is a sweet,

00:45:27: sweet dream.

00:45:30: Yeah.

00:45:31: And I think, I mean, he, we did meet with him.

00:45:34: So I think, again, something to look forward to.

00:45:36: He's developed some really cool tools in kind of his space that he's been able to

00:45:43: implement to support his workflow.

00:45:45: So speaking of agents, maybe if anyone's interested to understand how some of the

00:45:51: tools that we've talked about could be implemented and talk about specific use

00:45:55: cases, tune in on to next week's episode.

00:45:58: But other than that...

00:45:59: Please don't forget to subscribe.

00:46:01: We'd love your support.

00:46:03: It's validating for us to hear that they were on the right track and continue doing

00:46:06: what we're doing and we enjoy talking about this stuff.

00:46:09: So thank you for taking the time and joining us this week.

00:46:13: Yeah, also thank you from my side.

00:46:15: Thank you Lana again for your wonderful co -hosting.

00:46:20: And yeah, it's been 10 episodes now.

00:46:24: Plus we had even an extra video.

00:46:29: So talking about that, keep an eye on our YouTube channel.

00:46:33: We started uploading news episodes for covering the news.

00:46:37: You might notice that we didn't do that in the episode that was on purpose.

00:46:41: So yeah, keep your eyes open, follow us on YouTube and then you might get the latest

00:46:48: AI news in your beloved setting.

00:46:54: And yeah, and we also do screen shares and show stuff there, so stuff we cannot do on

00:46:59: the podcast.

00:47:00: So yeah, please go to YouTube, subscribe and leave a comment.

00:47:04: It helps us.

00:47:07: Thank you.

00:47:07: Thank you for tuning in.

00:47:09: We'll see each other next week.

00:47:10: Bye bye.