Enhancing AI Agents for Business Use, OpenAI Updates - EP 013

00:00:00: Hello and welcome to this week's episode of the AI Boardroom.

00:00:06: And with me again, my wonderful co -host Svetlana.

00:00:10: Tell us about today's topic.

00:00:13: Yeah, today we want you to talk about how Chagiviti and some of the recent releases

00:00:17: that OpenAI had announced in the last couple of weeks and how that is redefining

00:00:24: the human to AI interaction.

00:00:26: And what does that mean to you business owners?

00:00:30: And just I'd say, you know, we'll even cover it.

00:00:33: maybe everyone out there, but mostly your business owners.

00:00:36: Like.

00:00:37: and I think that some important topics will go into the practical all the way to

00:00:41: conceptual.

00:00:42: So yeah, if you're interesting in hearing more, stick around.

00:00:47: Yeah.

00:00:47: Yeah, so I would just try to introduce this today's topic a bit by just trying to

00:00:56: talk about what ways of communication are there.

00:00:59: Also a bit what my experience of different interactions with AI are, because some of

00:01:04: them are cool on paper, are cool in demos, not that cool in practice, which yeah, a

00:01:10: lot of it comes down to what people are used to.

00:01:13: And for me, I have an AI solution, which is basically a chatbot.

00:01:19: It does a bit something with images, but it's still a chatbot.

00:01:24: I am totally capable for some time now to do voice and to do real -time voice.

00:01:33: And I didn't do it because people are not used to it.

00:01:38: And even me,

00:01:40: Like I'm getting used to it, but even me, like I'm a digital native.

00:01:43: I grew up with computers.

00:01:45: I built my first computer, I was 10.

00:01:47: Like I really dig the whole AI stuff.

00:01:51: And still, like the chat interaction is like feeling the most natural to me.

00:01:56: And I'm still hesitant talking to my AI in a conversational manner.

00:02:04: Is it because you've been burned by the technology that maybe wasn't capable

00:02:09: enough for you to carry on the conversation?

00:02:12: help, to be honest.

00:02:14: I think that's what's probably messing a lot of folks up and adopting some of this

00:02:21: voice technology because of Siri on these Alexa, I mean Google Home and everything

00:02:26: like that.

00:02:26: They're unable to carry on the conversation.

00:02:28: actually a good experience from day one.

00:02:32: It's not perfect, but it's really fast.

00:02:36: It's just activated.

00:02:41: Yeah.

00:02:42: Yeah, it's just seeing the blue light in the corner of my eye.

00:02:47: So yeah, that's the better experience.

00:02:52: Also Google Assistant is also a really usable experience for quite some time now.

00:02:56: So it's mostly Apple that didn't update Siri since 2015 or something.

00:03:04: But rumors have it this year is the year.

00:03:08: So yeah, looking forward to the goal.

00:03:11: I love to, and honestly, I'm really rooting for it because GPT -4 .0

00:03:17: integrated into iOS, deeply into iOS on my phone.

00:03:22: I would love that.

00:03:24: And everyone will have Scarlett Johansson talking back to them.

00:03:28: Or not, yeah, we'll see.

00:03:31: see how the judge decides.

00:03:35: But yeah, I don't know what it is.

00:03:38: I found that older people, like in their 50s, even 60s, they oftentimes they don't

00:03:47: use the keyboard as much because they are not as fast in typing.

00:03:50: And they use voice to type a lot.

00:03:54: Mm -hmm.

00:03:55: And wait a second.

00:04:00: Keep going.

00:04:00: Was that about a sneeze coming on?

00:04:04: But I think that's a good point with, yeah, depends on people's maturity, as you

00:04:10: mentioned, like I have been grown up with this technology and what their comfort

00:04:13: level is, but I would argue that we shouldn't be building technology for the

00:04:17: people today because it's the technology who's gonna be most used by is millennials

00:04:22: and probably the upcoming generations who were born into the technology.

00:04:26: So they're much more likely to.

00:04:28: Like my kids are growing up with Alexa.

00:04:31: We have both Alexa and Google Home and they go and mind you, they don't speak

00:04:35: English.

00:04:37: They still try to like, they learn English with Alexa and Google Home because they

00:04:42: are motivated to like converse, con conversate with her.

00:04:46: Like if they want something to get done, like they go and they learn whatever task

00:04:50: that needs to be completed.

00:04:50: So there is, I think an emotional, like not emotional, but there's a need and

00:04:55: people will go out of their way to.

00:04:57: like in my kid's case, actually learn the language to be able to take advantage of

00:05:00: the system.

00:05:01: So again, I think always designing for people that, you know, of today is

00:05:06: probably not the best mentality, but I do think that over time, voice will become a

00:05:11: much more native experience.

00:05:14: the people today are most of the people.

00:05:15: So you cannot ignore them.

00:05:20: And you have to, it's not that, it's the majority, it's not the minority.

00:05:27: So yeah, I'm with you, you have to keep in mind that there will be a skill set, a new

00:05:34: skill set that is needed for the future of technology.

00:05:38: Always has been like that.

00:05:40: but still we have to at least show people a path to easily get on the train and to

00:05:47: not jump on the running thing.

00:05:52: And so, yeah, one thing that I think also is an issue is the whole notion of it

00:06:02: being clunky and kind of like, yeah, not really responsive, not really natural.

00:06:08: Mm -hmm.

00:06:08: that's, I think, where the latest updates of Google and OpenAI and Microsoft and

00:06:14: everyone came in, like multimodality and also small models that can be executed on

00:06:21: your device.

00:06:22: All of that gives you a new frame, a new canvas basically to work on.

00:06:29: And this canvas is latency free.

00:06:32: It gets your emotions.

00:06:34: It is really...

00:06:38: really starting to become helpful by quite a margin.

00:06:43: And that's what's getting up to us.

00:06:46: And to be honest, I'm a bit disappointed and I'm still not able to use the new

00:06:51: voice experience of Chetchipity because I'm really looking forward to it.

00:06:55: Because yeah, it got my attention.

00:06:59: actually starting to exhibit emotions too, right?

00:07:02: Like, and personality and things like that.

00:07:04: So it's too cited.

00:07:05: So it not only interprets.

00:07:07: Yeah.

00:07:07: right?

00:07:08: Last week was Microsoft's Build conference, which was Microsoft Developer

00:07:15: Conference, and they showed a demo of GPT -4 .0 integration into Windows, or into Co

00:07:22: -Pilot for that matter, and they had a screen share running and...

00:07:29: the AI was watching the screen basically, and they were playing Minecraft and it

00:07:34: gave live updates on the Minecraft game.

00:07:38: And then in the game, he turned around and said, what's that?

00:07:42: And then the AI was like with a like urgency, a lot of urgency was, hey, it's a

00:07:48: zombie, go away, find something to hide, a shelter or something and get given like

00:07:52: really actual usable advice in an urgent tone, which was in...

00:07:58: pretty cool.

00:07:59: the latency, which was ridiculous.

00:08:02: Like really, it wasn't sped up.

00:08:03: It was kind of a live recording.

00:08:08: Interesting, really interesting.

00:08:11: So it also opens up so many new use cases.

00:08:16: And to be honest, like I'm pretty sure AI is better in customer service.

00:08:26: pretty quickly than most of the customer service people are, just in terms of

00:08:30: always being polite, never being in a bad mood, always having all the answers that

00:08:36: are available.

00:08:39: So, yeah.

00:08:40: ChadGPT 4 .0 model, then 20 agents being live, because it's available 24 -7.

00:08:47: Yeah.

00:08:48: it's a huge opportunity.

00:08:53: And yeah, I find myself like more and more rooting for companies to add AI to their

00:09:04: service, like customer communications.

00:09:08: Because honestly, like I know the information I want is really mundane, but

00:09:15: I need it.

00:09:17: why can't an AI explain it to me?

00:09:20: Just go ahead, launch the stuff, honestly.

00:09:24: And I think with even my experience calling agents, the one thing that I hate

00:09:29: doing, and if I can figure it out on my own, I would always go through the website

00:09:35: and try to find the answer, Control -F, if there's no agent, like, you know, chat bot

00:09:40: available.

00:09:41: I dread calling the 1 -800 number because I'm like, and it's going to ask me 500

00:09:47: times to press these buttons, and it's going to send me to a...

00:09:52: If you want the reception in English, press 2.

00:09:54: If you want the reception in Dutch, press 3.

00:10:00: Yeah.

00:10:01: I just have such a bad experience is that I think if you can, if you truly, if your

00:10:05: experience was called a 1 -800 number and you immediately talk to someone who

00:10:09: doesn't care what prompts you ask, like it's going to handle all of your numbers

00:10:13: and select.

00:10:14: I mean, that's huge, I think for folks.

00:10:18: But one of the other use cases that I think I'm really excited about, which is

00:10:23: like outside of customer service, but like,

00:10:26: Any products and services, whether it's like an Alexa type of speaker or like a

00:10:31: personal robot, whatever have you, being embedded with these AIs?

00:10:36: Because we're talking about like, you know, this human to interaction or elder

00:10:40: population, because a lot of them are aging in their homes and they are by

00:10:45: themselves oftentimes and they don't see a lot of people kind of in their day to day.

00:10:50: And they've started to adopt, you know, these like robotic pets.

00:10:54: that of course don't talk.

00:10:56: Sometimes they, again, I've worked on some products that actually, this is true, and

00:11:00: we've done some research on this, that they talk to Alexa as if it was a person

00:11:04: because they had very minimal interaction.

00:11:07: How cool would it be if they had access to this chat GPT -40 omni version like Sky,

00:11:15: they would carry on like a full conversation.

00:11:17: Like...

00:11:19: Yeah, like, oftentimes, like, I'm, lately, like last year, I'm not really good at

00:11:25: following through on my personal goals, like eating behavior, stuff like that.

00:11:30: And like, just like having the AI know that and just talk to it about that, just

00:11:37: to get some reflection on my own thoughts, I think that could be hugely valuable.

00:11:44: And now that it has video, it can be like, Edgar, put that cake down.

00:11:47: yeah, don't go to that restaurant.

00:11:49: You know, you don't want Burger King or put put in random fast food name here.

00:11:58: No, accountability, I think, is going to be also huge because it can have that

00:12:04: longer term memory to understand your goals and then have like track your

00:12:08: progress towards them.

00:12:09: So I think that the use cases and possibilities, like you can integrate that

00:12:13: into an app and have the LLM kind of run the tracking with the voice and kind of

00:12:20: like this personality that again, you rely on.

00:12:23: You don't need a coach.

00:12:25: You don't need like an accountability partner.

00:12:28: when you have access to a powerful tool like that.

00:12:32: and then he takes you ransom with your browser history.

00:12:41: But first, I would love to take a small step back.

00:12:45: GPT -4 .0 is the new model of OpenAI.

00:12:49: For you that didn't know, your chat GPT got a lot better lately.

00:12:53: That's because of the new model.

00:12:56: The GPT -4 .0 model is not...

00:12:59: released to its full capabilities as of now, but will be in the upcoming weeks as

00:13:04: far, like at least as they say.

00:13:09: And O stands for Omni, and Omni meaning multi -channel or multimodality.

00:13:16: And what is multimodality in the first place?

00:13:20: So right now, chat with AI was mostly like, I type in text and it gives me text

00:13:25: back.

00:13:26: It's a large language model in the end, so yeah.

00:13:29: It's trained on a lot of text data.

00:13:33: Multi -modal models are able to not only read text, but also stuff like images, or

00:13:40: in the case of GPT -4 .0, also audio, natively.

00:13:44: Because right now you could also, like in the old version, you could talk to

00:13:50: Chetjubete, and you could get a talking answer back.

00:13:57: This was...

00:13:58: basically taking your voice, running through another model, making text from

00:14:03: the voice, putting it into the inference, like the text to text, giving text back.

00:14:12: This text is then synthesized back with the voice to give you like the voice

00:14:17: answer, the audio answer.

00:14:19: Now you give audio in, it gives audio out.

00:14:23: And it gives video in, it gives audio out.

00:14:26: And that's also the beauty of it.

00:14:29: If it's multimodal, you can switch between modalities basically.

00:14:34: And yeah, that's a huge leap.

00:14:37: What it gives you is recognition of, speaker recognition, recognized that there

00:14:44: are different people speaking.

00:14:47: It's the ability to interrupt, which was shown a lot in the LGBT4 .0 demos.

00:14:55: Mm -hmm.

00:14:56: give the AI the option to get the information in between the content, which

00:15:02: is your tone.

00:15:05: Do you speak in a loud voice?

00:15:08: Are you anxious?

00:15:10: Whatever.

00:15:11: It gets the feeling, the non -spoken part of the audio.

00:15:17: And that's, yeah, I think that's hugely valuable to get even better applications

00:15:23: that.

00:15:25: also have a low latency because you don't have to do all the transformation steps

00:15:28: and the answer comes quickly and that's what we also seen in the demo of GPT -4

00:15:35: .0.

00:15:36: I think with one other thing that one other benefit that I'm surprised you

00:15:41: aren't speaking of, because I think out of the two where you're the more technical of

00:15:46: us, but if you think about historically, if this multimodal version did not exist,

00:15:51: you would be integrated in order to support that type of an experience.

00:15:54: You would have multiple AI capabilities that would be a spaghetti of an

00:16:00: architecture kind of behind the scenes in order to kind of again support that same

00:16:04: experience.

00:16:04: Now you have to...

00:16:06: points.

00:16:08: One system, you basically select, and now you're going to have choices.

00:16:12: I think now we have Astra from Google, we have OpenAI, but I'm sure there's going to

00:16:17: be competitors emerging in the market.

00:16:18: But the users of this are the ones benefit.

00:16:22: So you'll have choices.

00:16:24: So you don't have to think about, well, what model and how am I going to architect

00:16:29: this and stuff like that?

00:16:29: You kind of have to think about which model am I going to insert?

00:16:33: What are some use cases?

00:16:34: What is the experience that I want to support?

00:16:37: And some of these tools are already available to you and they're pretty easy

00:16:42: to plug and play from proof of concept.

00:16:44: I think Edgar can speak much more at length as to like operationalizing this

00:16:49: tool is probably like more 90 % of the effort of actually bringing it into

00:16:53: production.

00:16:54: But most of it is very streamlined because you're only needing to rely on one model

00:17:00: in order to support that customer kind of like end -to -end experience.

00:17:03: And I think that's really powerful.

00:17:04: And the more capable the model is and the more it understands just from the get -go,

00:17:11: the easier it is for you to just prepare the right context.

00:17:17: And also you have more context from the same data if you want to.

00:17:22: So because you had the audio signal before, but you could only take the text

00:17:26: from it.

00:17:27: You couldn't take.

00:17:28: emotions from it without like there was also there's I think, Hume AI, they are

00:17:33: pretty huge in recognizing emotions from any modality basically.

00:17:39: And you could also like take into consideration that but you would only

00:17:43: build text from that and also and then get the answer out as text which then has to

00:17:51: be translated.

00:17:52: So yeah.

00:17:54: Yeah, and of course, the less moving parts you have, the better, always, always was

00:17:59: like that, always will be.

00:18:02: So yeah, having one model which can do it all gives you the ability to do more with

00:18:12: what you have already.

00:18:14: And generally, I have used cases you haven't even dreamed of.

00:18:19: And there has been a lot of discussion, which I think you're really strong in, is

00:18:23: agentic AI.

00:18:25: And so, following Google Next, there was a huge talk, and I think the word agentic AI

00:18:33: has been trending ever since.

00:18:36: They even released use cases that these companies, their customers are using

00:18:42: Google's technology and some of the models they make available through their cloud

00:18:47: technology.

00:18:48: But I think...

00:18:49: you know, none of the use cases actually came close to what the potential with the

00:18:55: 4 .0 is going to enable you with agentic AI.

00:18:59: Yeah, what do you think is, what do you think the 4 .0 will enable us to do more

00:19:06: of from a perspective of productivity and how kind of you envision maybe even

00:19:13: businesses using it, maybe internally even, like, is that an option?

00:19:18: What are the...

00:19:19: What are your thoughts?

00:19:20: So generally speaking, that's a really good question.

00:19:27: Because one thing that GPT -4 is worse than the model before is reasoning.

00:19:35: Reasoning capabilities are hugely important for agentic behavior because the

00:19:41: model has to decide based on reason what to do with the given information.

00:19:49: and how to plan out and stuff like that.

00:19:51: So the worse your model is in reasoning, the worse it will perform on agentic tasks

00:19:57: and the performance, like if you go even a little bit beyond easiest task, will give

00:20:07: you poor results as of now.

00:20:10: That said, having a better zero -shot performance, which...

00:20:15: definitely is the case in this GPT -4O model because of course you can't hand

00:20:19: over the same data, like I said, the same audio file you had before, but it just

00:20:24: gets you a better zero -shot result.

00:20:27: That's hugely, hugely useful because then you have to do less calls to the AI, which

00:20:38: if you build any application, of course will help you.

00:20:41: The AI itself gets faster.

00:20:44: the result, you get a faster result because, yeah, like you have only to do it

00:20:51: once.

00:20:52: But then it goes back to the reasoning part.

00:20:59: It might fail on like executing the tasks and sticking with the plan because it

00:21:05: hallucinates somewhere along the way and then arrows compound and it breaks or

00:21:11: influences.

00:21:12: the outcome in a bad way.

00:21:15: So yeah, like from application building perspective, you definitely have more

00:21:20: options and honestly, I'm really, I close to every day look up if there's any news

00:21:27: on when actually the API will allow me to send audio and get audio back because

00:21:32: yeah, the applications are really huge.

00:21:34: If you can combine it with image generation.

00:21:38: which also will be a part of that multimodality because you can now, you

00:21:42: don't need Dali 3 for image generation because the model itself can do it.

00:21:48: So right now you would say, JJBT give me a prompt for that and then you would take

00:21:54: that prompt and put it into mid journey and get the image.

00:21:57: Now you just get out what you want and get the image back from the same model.

00:22:04: And yeah, and all that is really...

00:22:07: hugely, hugely beneficial if you want to build applications because I told it a lot

00:22:14: in the past episodes, I tell it a lot on LinkedIn, it's all about context.

00:22:20: And the better you can use your context, the better you understand your context,

00:22:23: the better the result is.

00:22:25: In terms of...

00:22:27: automate the tasks I'm assuming.

00:22:29: So is that kind of where you're leading?

00:22:31: Yep.

00:22:32: And more tasks, because I spoke to a company which is doing real -time voice on

00:22:41: a telephone already, like already as of now, before GPT -4 and everything.

00:22:47: And they have to add stuff like arms and keyboard typing to just bridge the time

00:22:55: the inference takes.

00:22:56: They got it pretty low, give them that.

00:22:59: but it's not instant.

00:23:01: So it's not like a real conversation.

00:23:05: And these applications might have been possible already, but they get really,

00:23:11: really humanized, basically, with this new release.

00:23:17: And I think that's really interesting because I think with, we talk about like

00:23:22: the lines blurring ultimately between the humans and AI because we're adding these

00:23:28: things that even, I mean, production of podcasts, videos that you watch on

00:23:33: YouTube, like that's ultimately what, you know, there are systems that are developed

00:23:37: to remove the ums and the ifs because like that's just our normal tendency to like

00:23:43: speak that way.

00:23:44: Or like my tendency is always to start like what's so.

00:23:48: So you have these additional keywords that you're, not keywords, but words that you

00:23:55: add in your normal conversations that we're now adding to AI for it to sound

00:23:59: more human -like, which I think interesting, or as you mentioned, like

00:24:04: typing behind the scenes.

00:24:05: Because again, these are the things that we hear.

00:24:07: So at some point, and we even started to experience this with, there's been a lot

00:24:11: of discussions with character AI now.

00:24:14: people starting to form bonds with these avatars that people are creating because

00:24:18: they've created and they chat with them.

00:24:21: Just think about the potential of having a voice, kind of AI agent, basically

00:24:28: speaking to you with all of these additional kind of more human -like

00:24:32: interactions.

00:24:33: I think, I mean, we're probably gonna hear more on this once the AI system is

00:24:38: released, but I do anticipate that people are gonna start forming bonds.

00:24:41: with this AI, as I mentioned, like people who are lonely, people who, you know,

00:24:46: maybe taking the system even to the extreme of developing a relationship with

00:24:51: Sky.

00:24:52: already thanking it, basically.

00:24:58: Yeah, but this is so natural language that I can't feel inclined to do what I'm

00:25:08: raised to and that's being polite.

00:25:13: Yeah, it's weird, but in the end, as it gets emotions now, maybe I should.

00:25:20: thank Skye.

00:25:22: better things, Kai.

00:25:26: But yeah, like you said, the line gets pretty blurry.

00:25:30: I'm already at a point where I would love to just speak to an AI, because oftentimes

00:25:36: humans don't do a good job with random situations if they're not highly skilled

00:25:42: to deal with them.

00:25:44: But if you think about also, there's research that actually wrote an article, I

00:25:49: have a blog post on humanizing AI.

00:25:52: I think it's interesting that people develop a form of bond and they actually

00:25:58: trust it because one of the things why kind of going to a therapist, there's not

00:26:03: as many people going to a therapist because they're like, what if I say

00:26:07: something and people are going to think badly of me?

00:26:09: Or I don't want to disclose this because it's going to end up, I don't know.

00:26:14: And they're going to form the wrong picture of why my motivations were such.

00:26:20: But now you have AI.

00:26:22: So what if you taught it to be not judgmental or I could truly be open to

00:26:28: this agent, with some caveats that my information will not be shared and it will

00:26:33: be anonymous.

00:26:34: But you could truly use this for therapies and being a psychologist, because it could

00:26:39: recommend some of the same principles that...

00:26:42: these therapists again, like people are going to hate me for saying this, but like

00:26:45: some of the things that maybe on like mild cases, like the day -to -day therapy, just

00:26:49: to keep your kind of mind flowing or inspired, there's certain things that I

00:26:53: think AI could be delegated to do.

00:26:55: And then there's certain things where, you know, you are going to need more hands -on

00:26:59: like therapy, deep dive sessions that AI cannot handle.

00:27:03: But I think that you're going to get, it's an opportunity, I think, for people to

00:27:08: have this bond to be more open.

00:27:11: which they otherwise would not either seek help or would not even have anyone else to

00:27:16: talk to and kind of voice out their concerns or get help.

00:27:19: So.

00:27:20: to be completely honest, I was with the current AI voice integration already at a

00:27:27: point where sometimes I drive home, like I have a thought, I wasn't able to do it on

00:27:32: my notebook.

00:27:33: And on the way home, I just used the speech, like the microphone system in my

00:27:39: car and I talked to it and I tried to conversate properly.

00:27:47: And now I can also interrupt it if I have the feeling it mis -enlisted me without

00:27:52: pressing any buttons whatsoever and being able to edit.

00:27:55: One thing also, Chachapiti related, is it uses, like they added memory across

00:28:03: different conversations, if you want to.

00:28:07: And this is also hugely helpful because it starts building up like a knowledge base

00:28:12: about you and what you talked about and stuff like that.

00:28:16: And that's where it gets me excited.

00:28:19: Also, Microsoft did this recall function Windows, which we will talk about at

00:28:24: another episode, I guess.

00:28:28: But AI needs context to really feel like a personal assistant.

00:28:33: It needs to know you, it needs to know what you do, and the more it knows, the

00:28:38: more you're exposed, of course, but the more it can help you.

00:28:42: Because otherwise, if it doesn't know what I'm talking about,

00:28:45: most of the time, how should it help me, right?

00:28:48: Yeah.

00:28:49: I think it's only fair, right?

00:28:50: So if I was thinking about this the other day, so if you go out of the blue without

00:28:56: signing in and without AI opening, I even knowing your contacts, you're just going

00:29:00: to prompt the system to be like, hey, what color should I choose between this dress

00:29:06: and this dress?

00:29:06: And you upload like a picture of the dress.

00:29:09: How should it know how to answer?

00:29:12: It's just going to be like an arbitrary choice, which is similarly like you

00:29:15: walking up to a stranger.

00:29:17: But that person, like understanding who you are, what you stand for, what your

00:29:22: preferences are, you're just going to be like, hey, what dress can I choose?

00:29:27: It's going to basically, that person is going to rely on their own experience and

00:29:30: they're going to provide their preferences.

00:29:31: In this case, large language models just give you the average tendency.

00:29:35: So it's like, they just rely on kind of its knowledge base to give you the

00:29:38: recommendations.

00:29:39: So if you want a much more personalized output, you kind of have to like,

00:29:44: similarly to the stranger, to like...

00:29:46: give them at least a little bit of your life story and like your preferences in

00:29:49: order for them to have a contextual feedback, something that's helpful to you.

00:29:53: So I do think that providing that context, whether it's with system prompting or

00:29:58: whatever have you, or through this kind of conversational history, you're truly

00:30:02: building this context about what your preferences are, what you stand behind,

00:30:06: and like what kind of questions you tend to like gravitate towards, or how you're

00:30:11: following up on like...

00:30:13: fine tuning the responses or maybe iterating on the responses.

00:30:17: So it's learning some of these things and that's kind of the power of AI to be able

00:30:22: to do that like infinitely and to be able to kind of handle it for different people.

00:30:28: But that's kind of the power of personalization.

00:30:30: But you have to provide that context in order if you want a much more accurate

00:30:34: response that will make you happy.

00:30:37: So in a way, I think that's like a fair ask.

00:30:40: And also to all the business leaders out there or project leaders for AI projects,

00:30:45: same for you.

00:30:46: The more information you gather, the more information you also curate or let it be

00:30:52: curated by the AI.

00:30:54: If you want to know how this works, feel free to ask.

00:30:59: Because if AI is building up memory, it can structure it in a way where it can be

00:31:06: recalled easily.

00:31:08: and also be already prepared to be recalled better than if it would be the

00:31:15: plain text.

00:31:17: And also stuff like context gives you the option to rephrase a question.

00:31:24: So you have a context which might be only your chat history from the actual chat, or

00:31:29: like the current chat.

00:31:30: And then you go ahead and then you get a new question.

00:31:34: Like you were talking about the dresses.

00:31:35: Like...

00:31:38: And then your next question would be like, and what about a scarf, for example?

00:31:44: And then AI goes ahead, looks at your history, and then it can rephrase the

00:31:52: question.

00:31:53: It's like, what scarf would be good for a dress which is in that or that color, for

00:31:58: example, right?

00:32:00: And then it rephrases the question for itself before it does the actual

00:32:06: inference.

00:32:07: And that's how you can technically solve or make context and more context usable by

00:32:18: looking at what's getting in, looking at what is around it and just rephrase what

00:32:23: is in there to get a more precise answer.

00:32:25: And that's how you get rid of hallucinations.

00:32:28: That's how you get rid of false information.

00:32:30: That's how you get proper grounding for your information to be really accurate.

00:32:37: It's not perfect by any means and you need a lot of tuning to get this dialed in and

00:32:41: right, but it's hugely helpful.

00:32:45: Yeah, totally agree.

00:32:48: Yeah, if you enjoy this episode, would love for us to dive into any particular

00:32:53: agents.

00:32:54: So I know that we covered about, you know, Chagy PTV focused on 4 .0 and kind of the

00:32:58: multimodality that it now offers.

00:33:00: But if you want us to kind of talk to you about maybe a specific use case around

00:33:05: sales, marketing, and kind of lay out the groundwork, maybe with some, some tangible

00:33:11: use cases.

00:33:12: Happy to do it.

00:33:13: As we always say, we want to grow with you and we want to deliver content that you

00:33:18: find useful.

00:33:19: So let us know in the comments if there's a particular topic you'd like for us to do

00:33:24: a deep dive on.

00:33:25: But don't forget to subscribe.

00:33:27: want us to do this live on stream in a session where you can ask live questions,

00:33:36: yeah, just let us know down in the comments.

00:33:39: And yeah, that's it for today, I would say.

00:33:43: Yeah, well, don't forget to subscribe again as I mentioned, like this video and

00:33:48: I guess we'll see you on the next one.

00:33:51: Thank you.

00:33:51: bye.