S1|Ep81 AI Evolution: Theory & Practice Behind Sustainable and Responsible Implementation

WEBVTT
00:00:00.320 --> 00:00:05.040
Hello and welcome to Digital Transformation and AI for Humans with your host, Annie.
00:00:05.280 --> 00:00:12.800
In this podcast, we delve into how technology intersects with leadership, innovation, and most importantly, the human spirit.
00:00:13.519 --> 00:00:26.160
Each episode features visionary leaders who understand that at the heart of success is the human touch, nurturing a winning mindset, fostering emotional intelligence, and building resilient teams.
00:00:26.719 --> 00:00:36.560
In today's episode, we are diving into the evolution of artificial intelligence with my brilliant guest from Dubai, UAE Prakash Singhani.
00:00:37.359 --> 00:00:51.439
Together, we'll explore the journey from LLMs to agentix solutions to the MCP protocol, and take a closer look at the theory and practice behind sustainable and responsible implementation.
00:00:51.600 --> 00:01:04.879
Prakash is a digital construction thought leader, startup founder, early stage investor, board advisor, co-founder and CEO at NOAATEC and general partner of Empedacapital.
00:01:05.519 --> 00:01:18.000
I'm honored to have Prakash as a part of the executive group of the AI Game Changers Club, an elite tribe of visionary leaders redefining the rules and shaping the future of human AI Synergy.
00:01:18.640 --> 00:01:19.760
Welcome, Prakash.
00:01:19.920 --> 00:01:22.560
It's a great pleasure to have you here in this studio.
00:01:23.359 --> 00:01:23.920
Thank you so much.
00:01:24.000 --> 00:01:25.280
And the pleasure is all mine.
00:01:26.400 --> 00:01:32.799
Let's start the conversation and transform not just our technologies, but our ways of thinking and leading.
00:01:32.959 --> 00:01:38.319
If you are interested in connecting or collaborating, you can find more information in the description.
00:01:38.480 --> 00:01:41.920
And don't forget to subscribe for more powerful episodes.
00:01:42.079 --> 00:02:00.959
And if you are a leader, business owner or investor ready to adapt, thrive, and lead with clarity, purpose, and wisdom in the era of AI, I would love to invite you to learn more about AI Game Changers, a global elite hub for visionary trailblazers and change makers shaping the future.
00:02:01.280 --> 00:02:05.599
Prakash, I've been waiting for this conversation for such a long time.
00:02:05.760 --> 00:02:09.520
And today we are going to dive into what really matters.
00:02:09.759 --> 00:02:19.599
And I'm sure that our listeners and viewers will not only learn something new, but also get many insights.
00:02:19.840 --> 00:02:21.840
Let's start with your story.
00:02:22.000 --> 00:02:29.280
I would love to hear more about yourself, about your journey, and everything you would like to share with us.
00:02:29.759 --> 00:02:32.800
Yeah, so my journey started as a civil engineer.
00:02:32.879 --> 00:02:34.879
So I have got a master's in civil engineering.
00:02:34.960 --> 00:02:36.800
I was born and brought up in the UK.
00:02:37.120 --> 00:02:46.080
Being part of the construction industry, I guess that's where I understood the role of construction professional, is a lot to do with managing people, right?
00:02:46.159 --> 00:02:54.960
It's understanding what drives people, how people react and do things is a lot to do with what's happening in their lives, what's going on.
00:02:55.120 --> 00:03:00.560
And in order to motivate them to succeed, you need to understand who they are as people.
00:03:00.719 --> 00:03:05.840
Most of my career, although I graduated as a civil engineer, was spent doing digital transformation.
00:03:06.000 --> 00:03:14.319
So looking at how the construction industry can take and adopt technology to help make it safer, more productive, improve the quality.
00:03:14.479 --> 00:03:31.120
And even that space, although it was digital transformation, was the title of my role, most of that, I would say 80% of that time, is looking at people and how those people can adapt and utilize the technologies to be able to get some of the improvements that we were looking at doing.
00:03:31.280 --> 00:03:35.360
I was really fortunate to also have been given the opportunity to travel around the world.
00:03:35.520 --> 00:03:43.280
So I worked in India, I've done a little bit in Indonesia, and then for the last 12 years or so, I've been here in the Middle East, based out of Dubai.
00:03:43.439 --> 00:03:50.080
And again, having worked in these different places, you get to understand how people's perception of things are different, right?
00:03:50.240 --> 00:03:52.240
People's perception of risk is different.
00:03:52.400 --> 00:03:57.680
People's perception of technology and how to use technology and how to adopt technology is different.
00:03:57.840 --> 00:04:10.800
And so I think these kind of unique elements to my career have really helped with what I'm doing now in running Narvotech, where we've developed AI tools that put the frontline workers first, right?
00:04:10.879 --> 00:04:16.720
We want to make sure that people who ordinarily technologies like AI might pass them by.
00:04:16.879 --> 00:04:27.600
We help them to get access to this technology, we help them to, you know, understand this technology, and we help them to interact with business systems using AI as the main interface.
00:04:28.399 --> 00:04:29.199
Amazing.
00:04:29.360 --> 00:04:31.920
Thank you so much for sharing your story.
00:04:32.000 --> 00:04:54.000
And I agree, working with different cultures, seeing different places, and working with different structures helps seeing the opportunities and finding those common connectors in order to enrich your product and create something truly amazing.
00:05:05.439 --> 00:05:14.879
In your view, what marks the true turning point in this evolution and what mindset and skills should leaders adopt to stay ahead of it?
00:05:15.600 --> 00:05:25.199
So I think you're right that is that the technology has rapidly moved from some of the basic things that LLMs do, where um you know they became really good at writing, right?
00:05:25.279 --> 00:05:30.959
So a basic LLM, I think most people know, it's a really good, essentially a really prediction engine.
00:05:31.120 --> 00:05:38.160
What it does is is that um you ask it a question, it predicts what the first word would be to the response to that question.
00:05:38.319 --> 00:05:43.839
Then based on the question and the first word, it'll predict what the second word is, and then what the third word is, and et cetera, et cetera, right?
00:05:44.000 --> 00:05:48.160
And so most people still, unfortunately, are still only using it to do that.
00:05:48.319 --> 00:05:54.480
Whereas I guess there's probably a lack of education in terms of what it can do now as we've moved to agentic systems.
00:05:54.639 --> 00:06:04.639
So I guess the true turning point was we went from you know being able to create better emails, write better responses, or you know, create content for our social media posts.
00:06:04.800 --> 00:06:10.879
We've now switched into what we call agents that are capable of taking autonomous actions over time.
00:06:11.040 --> 00:06:19.519
So they're able to have a memory, they can remember what was said previously, they can act autonomously to go and fetch information for the internet and other sources.
00:06:19.759 --> 00:06:25.120
And that evolution introduces these things called persistence, planning, and purpose.
00:06:25.360 --> 00:06:34.319
So this changes the way that we use LLMs, not just as a almost like a passive tool to give us things, but as a true collaborator.
00:06:34.399 --> 00:06:38.079
So we can start working with it and do meaningful things, right?
00:06:38.160 --> 00:06:42.160
For the agent, you're gonna do meaningful things for you and with you.
00:06:42.319 --> 00:06:53.600
And so the moment we let AI operate across these multiple steps and carry out action and you know, give it feedback on how we can improve, uh, the more autonomous we can get these agents to become.
00:06:53.920 --> 00:07:04.160
And so as we enter this kind of new phase where uh we have alignment, control, and obviously protocols, you have to apply leadership over it to make sure that you know you're understanding what it's going to do.
00:07:04.240 --> 00:07:24.079
As soon as you start giving up some of this autonomy, giving it to the agents, or you start giving up agency to these agents, you need to really understand what the boundaries are and how these are going to act and react so that you can therefore predict what's gonna uh come out from and be sure of the responses and what the actions that the LLMs are gonna take.
00:07:24.399 --> 00:07:26.079
I couldn't agree more.
00:07:26.399 --> 00:07:43.519
And just to dive a little bit deeper, I want to ask you the biggest misconception about agentic AI systems and how can businesses practically prepare for their integration without losing control, clarity, or accountability.
00:07:44.480 --> 00:07:54.079
So I think and linked to what we were just talking about, I think the biggest myth is that agentic systems are just you know super duper LLMs or LLMs on steroids.
00:07:54.639 --> 00:07:55.600
They're absolutely not.
00:07:55.759 --> 00:07:58.560
Agents are usually goal-oriented.
00:07:58.720 --> 00:08:07.680
Agents are designed to do something specific and particular, and it's more than just give you a text-based response or you know, give you back a generated image.
00:08:07.920 --> 00:08:13.680
There are systems that are an interface, but sometimes interfaces to uh external tools.
00:08:13.839 --> 00:08:27.519
So you can connect them to things like APIs, um, and we're going to talk about this in a second, but you can also uh start connecting them to other agents, and even what we're starting to see now with agents being uh connected to robots, um, so even external physical environments, right?
00:08:27.600 --> 00:08:39.759
So you're now starting to see this crossover between AI and LLMs that used to be very computer-based, whether that's your phone or your PC, now moving more into the physical world for you to interact with.
00:08:40.000 --> 00:08:49.600
I think in terms of some practical steps and things, maybe you know, how to prepare, as organizations, you probably need to start looking at assessing whether organization is ready for gentic systems.
00:08:49.759 --> 00:08:57.279
You know, do you have the right culture in place, the right protocols in terms of safety and guidance for your employees to be able to use this in a safe way?
00:08:57.600 --> 00:09:05.200
One thing I'm starting to see is organizations create sandbox environments for safe testing and experimentation with some of these agents.
00:09:05.279 --> 00:09:14.480
So if you're asking your agent to place orders for something on your behalf, you want to make sure that it's not going to over-order things or it's not going to order, it's going to repetitively order things, right?
00:09:14.559 --> 00:09:22.960
I see people talking about, if I bring it back to what we're doing in construction, is talk about giving an agent autonomy to order a concrete pool, right?
00:09:23.039 --> 00:09:27.120
Uh and depending on the weather, you then cancel it and then set it up for another day.
00:09:27.279 --> 00:09:35.600
Now imagine like there was there was some glitch in the agent or there was some misunderstanding with the agent, and it ended up ordering that same delivery of concrete three times.
00:09:35.759 --> 00:09:44.000
Obviously, there's real-world implications, both in terms of cost, but environmental and all this, you that concrete would go to waste, and there's a whole bunch of things.
00:09:44.159 --> 00:09:45.600
So you have to be really careful.
00:09:45.679 --> 00:09:52.960
And so creating some of these safe spaces to do the testing and making sure you get trust and build trust in it is really important.
00:09:53.120 --> 00:09:57.840
I think the other part of it is being able to have cutoffs and like dead switches with agents, right?
00:09:57.919 --> 00:09:59.679
In case they do end up doing things wrong.
00:09:59.840 --> 00:10:06.480
So putting safeguards in place in the expectation that the agent might go awry and do something that you don't expect it to do.
00:10:06.639 --> 00:10:12.480
So I think these are some of the things that I'm starting to see, and some of the safeguards I think as organizations, we need to start looking at.
00:10:12.879 --> 00:10:20.399
I like that you highlighted the culture needs to be in place among all the other things and technologies.
00:10:20.879 --> 00:10:35.039
And of course, it is important to test it before you implement it in real life, because otherwise, the implications might be so impactful that it might be even dangerous.
00:10:35.200 --> 00:10:45.840
And too much concrete is, I believe, not the worst case scenario of what might happen when you trust your agents blindly.
00:10:46.240 --> 00:10:46.639
Yeah.
00:10:46.879 --> 00:10:51.759
But I think the culture, you're right, the culture is really important, but it also goes both ways, right?
00:10:51.919 --> 00:10:59.759
So most organizations, if they're going to be looking at deploying agents and agentic workflows, are probably going to be spending some amount of money to put this in place, right?
00:11:00.000 --> 00:11:00.960
And other resources.
00:11:01.120 --> 00:11:06.960
And so if you don't have a culture that's ready to adopt it in the first place, that's one element of it as well, right?
00:11:07.039 --> 00:11:14.720
So making sure that you've got the right culture where you know people want to experiment, want to innovate, want to look at different things.
00:11:14.879 --> 00:11:22.399
And then you've also got the culture where people are understanding the limitations of what the agent can do and some of the things that could go wrong.
00:11:22.639 --> 00:11:30.879
So I think that of all of the things, I think the culture is definitely one of the most important elements that you need to understand and work on getting right.
00:11:31.279 --> 00:11:31.919
Absolutely.
00:11:32.080 --> 00:11:33.519
It goes both ways.
00:11:33.679 --> 00:11:56.240
And I love that you just highlighted it and explained it a little bit better because sometimes it tends to become just a buzzword, and leaders try to bypass that topic because it is seen as something one-sided, but it is deeper than that, and it has a huge impact on the consequences.
00:11:56.399 --> 00:12:06.000
That's why it is important to understand that it's not that straightforward and it is more powerful than we would like to see it.
00:12:06.480 --> 00:12:09.440
But let's talk about the emerging MCP.
00:12:09.679 --> 00:12:15.759
How will it change how we build, scale, and collaborate with AI across industries?
00:12:16.320 --> 00:12:26.559
So I think this is probably going to be one of the most fundamental changes that we've seen in AI since its inception, since it started becoming into the public consciousness in what back end of 2022.
00:12:26.799 --> 00:12:34.960
So MCP is model context protocols, and sometimes it's called agent-to-agent, so A to A or multi-agent context protocols.
00:12:35.200 --> 00:12:44.000
It's essentially a set of standards that allows different agents to be able to talk to each other, share content, sometimes even share memory.
00:12:44.159 --> 00:12:47.840
So this is now, you can imagine, becoming really, really powerful.
00:12:48.000 --> 00:12:54.639
So if you think about agents that are being designed to do one task, and usually they're designed to do one task really, really well.
00:12:54.720 --> 00:12:56.879
Um, and that task might have multiple steps.
00:12:57.120 --> 00:13:02.720
Here now you can get uh with MCP, you can get agents that talk to uh other agents.
00:13:02.799 --> 00:13:08.080
And so they can do things in different systems outside of the confines of of where they've been set up.
00:13:08.320 --> 00:13:11.919
And so they can also pass tasks from one to another.
00:13:12.080 --> 00:13:17.039
So some people create an analogy like MCPs are like the APIs for LLMs.
00:13:17.200 --> 00:13:20.000
I think that's a little bit short-sighted, not quite like that.
00:13:20.159 --> 00:13:21.600
It's a lot more powerful than that.
00:13:21.759 --> 00:13:26.320
And so you can get basically multi-agent systems that can do uh multiple different things.
00:13:26.399 --> 00:13:41.279
So you can get agents that are really specialized in one thing and another one that's specialized in another, and then you can get orchestration agents then that that kind of sit over these multiple agents and are able to work with the inputs and the outcomes from these different agents.
00:13:41.440 --> 00:13:46.159
That's not currently possible with straightforward LRMs, and it's very difficult to do with a single agent.
00:13:46.320 --> 00:13:52.000
So I think this move into this kind of multi-age environment is extremely powerful.
00:13:52.159 --> 00:13:57.840
And I think this also allows us to do things like um use different models for different things.
00:13:58.000 --> 00:14:04.399
So, you know, we've seen recently all these uh latest models coming out, like Grok and the latest GPT models and things like that.
00:14:04.559 --> 00:14:08.559
And as these models evolve, they're gonna become better and better at certain things.
00:14:08.720 --> 00:14:20.399
So even you've got this menu of models that you can access ChatGPT, because some of them are good at certain things, some of them are good at reasoning, some of them are good at creating images, some of them are good at um giving you summaries.
00:14:20.559 --> 00:14:31.840
And so once you've got that, you can create agents, utilize these different models, and then bring them together to carry out tasks that are much more complex and nuanced and bring all of these things together.
00:14:32.000 --> 00:14:41.840
So the way I like to do to kind of see MCP or the way I like to describe MCP is that it's very akin to basically building a human team, right?
00:14:42.000 --> 00:14:53.679
So, you know, you have a manager, then you have next level down, and then you've got workers, and each of them have got certain skills and certain specialisms, and then the manager is kind of there to kind of coordinate and orchestrate between them to get an output.
00:14:53.840 --> 00:15:04.399
And so if I apply it to you know some of the things that we're doing, we're actually experimenting with MCP in construction and in safety, we have to create multiple documents for regulation, for communication, and things like that.
00:15:04.480 --> 00:15:09.919
So, what we're trying to do is look at how we can create MCPs that are really good at specific tasks.
00:15:10.080 --> 00:15:23.120
So you might have an MCP that's really, really good at understanding how to work with ladders, another one that's really, really good at understanding how to work with scaffolding, another one that's really good understanding how to do excavation, et cetera, et cetera.
00:15:23.279 --> 00:15:31.200
And then you'll get an orchestration agent that sits over the top that you know is really good at writing method statements or uh or doing a risk assessment.
00:15:31.279 --> 00:15:38.159
And so what you do there is you can talk to one agent and say, you know, write me a risk assessment for doing this piece of work.
00:15:38.320 --> 00:15:43.120
It will then go to the relevant expert agent, they will produce their content.
00:15:43.360 --> 00:15:58.240
The orchestration agent would look at each of them and ask questions of the others to make sure it's all alright, and then finally produce and give you an output that is going to be contextualized, it's going to give you the correct information, the validated information, and it'll be of a high quality.
00:15:58.399 --> 00:16:02.720
So that's one kind of very simple example of how MCBs can be applied.
00:16:02.799 --> 00:16:04.720
We're kind of experimenting with some of these.
00:16:04.799 --> 00:16:06.159
Um, there's lots of other things.
00:16:06.240 --> 00:16:11.039
Um, you could use it as like a planning agent, you could use it for you know doing vision interpretations.
00:16:11.120 --> 00:16:12.159
There's a whole bunch of things.
00:16:12.399 --> 00:16:20.799
Without MCPs, these agents would probably operate in silos, and sometimes it'll be difficult to achieve a greater goal to try and do something useful.
00:16:20.960 --> 00:16:27.679
So with a single agent, you're probably limited, or your prompts would get so complex that there'll be edge cases where it would fall down.
00:16:27.759 --> 00:16:31.759
So using MCPs, I think, is going to be a great expansion of what LLMs can do.
00:16:32.080 --> 00:16:33.360
Sounds fantastic.
00:16:33.519 --> 00:16:36.639
What an amazing world we're living in.
00:16:37.120 --> 00:16:41.120
Just a few years ago, we couldn't even dream about all this.
00:16:41.200 --> 00:16:47.440
And today we are discussing such topics, and it is also applied in reality.
00:16:47.679 --> 00:16:51.840
By the way, you mentioned that you are applying it in your business.
00:16:52.080 --> 00:17:04.720
So I wonder if you have a success story around your positive return on investment while you applied those technologies in real business, in real life.
00:17:05.039 --> 00:17:08.160
So we're still at the experimental stage with things like MCP.
00:17:08.240 --> 00:17:10.160
So I think there's a couple of risks with it as well.
00:17:10.240 --> 00:17:14.240
So because it's so new, some of the security protocols are not well established.
00:17:14.400 --> 00:17:31.200
So, you know, being able to ensure that the data flowing from one agent to another is controlled, what it can do back in your own systems, you can potentially um inadvertently expose your internal systems because of the way these MCPs interact, and especially because some of these agents have a little bit of autonomy.
00:17:31.279 --> 00:17:41.039
And so I guess at the moment, the ROI for us is look how we can um enhance our product lines and also look at doing internal processes more efficiently.
00:17:41.200 --> 00:17:44.240
But currently I think it's a bit too early to get any definitive ROIs.
00:17:44.319 --> 00:17:47.920
It's it's very much for us anyway, is in is in the experimental state.
00:17:48.160 --> 00:17:58.720
I love this, and I totally understand and support the approach where it's better to be on the safe side and test it a little bit longer.
00:17:58.880 --> 00:18:06.480
And of course, the topic is quite new in our reality, in the way we're running business.
00:18:06.799 --> 00:18:23.279
So obviously, it will take a little bit more time, but once it is applied to the real cases, it is going to be fantastic, and then we are mitigating our risks along the journey as well.
00:18:23.759 --> 00:18:36.079
And to all our listeners and viewers, if this conversation sparks something for you, hit like, follow, or subscribe, and share it with one person you know would be inspired by this episode.
00:18:36.640 --> 00:18:37.759
Share in scaling.
00:18:39.200 --> 00:18:52.400
Prakash, in your experience, what are the most common mistakes companies make when trying to implement next gen AI, especially when jumping from traditional automation to more autonomous agents?
00:18:52.640 --> 00:19:04.319
We just mentioned that sometimes it is smart to slow down and not rush into the unknown, but at the same time, I would like to hear more about it.
00:19:05.039 --> 00:19:09.039
So I think one of them is giving the agents too much trust too early.
00:19:09.200 --> 00:19:28.880
Organizations sometimes skip some of the things that we talked about right at the beginning, about you know, uh understanding what the governance models are, understanding the culture in the organization, setting yourself up with sandbox environments and things like that, and thoroughly testing what the agents can do and also what they shouldn't be able to do, uh, putting some of these guardrails in.
00:19:28.960 --> 00:19:39.279
So I think that's one thing that organizations should definitely make sure that they don't jump too quickly into trusting agents with a rush to try and just deploy them.
00:19:39.519 --> 00:19:52.799
So the way I like to describe it is you know, if you've got an intern in into the organization, someone who's got relatively low experience and you know, just have just come out of university or something, you wouldn't necessarily have them making the most key decisions in your business, right?
00:19:52.960 --> 00:19:59.680
Particularly ones that are going to potentially impact revenue or um, you know, safety or even relationships with your clients.
00:19:59.839 --> 00:20:05.920
You'll put them as shadowing somebody else, train them, make sure, and then you know, slowly start giving them responsibility over time.
00:20:06.160 --> 00:20:10.640
Once you've established that trust, then you let them do more and more things and give them more responsibility.
00:20:10.880 --> 00:20:36.960
I think we should be looking at agents in in very much the same way, where we we essentially treat them as if they were interns and drink feed them levels of responsibility over time so that you can build that trust yourself, but also you know, start uh understanding where the agents are going to be useful and are not, and where they could potentially create harm and introduce unwanted or unrecognizable risk into your organization.
00:20:37.279 --> 00:20:42.480
Um, some of the other things I think are like having no context management layer.
00:20:42.640 --> 00:20:51.680
I think agents to be able to understand and improve agents, you need to have a layer of context management to be able to make sure what the agents are producing.
00:20:51.759 --> 00:20:55.519
Um, if it is wrong, you've got a feedback mechanism to be able to go back into it.
00:20:55.599 --> 00:21:00.960
And that feedback then enhances and it learns from uh so the agents learn from these things.
00:21:01.279 --> 00:21:07.200
A lot of organizations almost treat agents as they do with LLMs like a one-shot, right?
00:21:07.279 --> 00:21:12.000
You expect it, you give it something, you get something back, and then you blindly take and trust that.
00:21:12.160 --> 00:21:19.039
More and more people are starting to understand that just as LLMs hallucinate, agents can also hallucinate and misinterpret things.
00:21:19.200 --> 00:21:32.640
And so making sure that you've got protocols in place to be able to identify those, but more importantly, reinforce the agents and uh the underlying LLMs to make sure that those things don't continuously keep happening.
00:21:32.880 --> 00:21:42.000
I think from recommendations similar to what we were talking about, you start with supervised, making sure you're keeping the human in the loop at every single stage.
00:21:42.160 --> 00:21:44.319
So you don't go to full autonomy straight away.
00:21:44.480 --> 00:21:47.920
Again, similar to what we're seeing with you know driverless cars and things like that.
00:21:48.079 --> 00:21:56.640
You're starting with a driver in the seat, then the driver moves into the passenger seat and make sure that they've got control over these autonomous vehicles if they need to.
00:21:56.799 --> 00:22:02.799
And eventually we'll get to, I think it's level five autonomy where that the cars will drive themselves.
00:22:02.960 --> 00:22:12.400
So, in exactly the same way, we should be looking at agents within our organizations where we start with the human in the driving seat and you kind of seeing what the agent can do.
00:22:12.559 --> 00:22:24.720
Then you're probably sitting next to it, uh almost like a co-pilot, not to be mistaken, with Microsoft's product, um, and then letting it work alongside you, and then you start giving it autonomy over time and letting it do things by itself.
00:22:24.880 --> 00:22:30.480
The other thing is what I've been reading about is creating agent identities and then role protocols.
00:22:30.640 --> 00:22:43.039
So going back to my previous analogy about treating it as if it was an employee and starting it as an intern, and then as it develops better skills or you develop more trust in it, you almost give it promotion, right?
00:22:43.119 --> 00:22:49.519
So understanding the role that it has to play and giving it protocols that it can work within.
00:22:49.680 --> 00:23:00.559
And then, and as soon as it goes outside of those uh guidelines or protocols, there's you know mechanisms to stop it or ask for permissions or do certain things to make sure it doesn't go beyond those guidelines.
00:23:00.880 --> 00:23:03.920
It reminded me of another episode with Dr.
00:23:04.079 --> 00:23:10.559
Victor Monga from California, where we were discussing cybersecurity and zero trust journey.
00:23:10.960 --> 00:23:31.680
So it is exactly the same logic behind where you really need to start with something simple, something that can't bring too much damage, and then move from there and develop together and then co-create along your roadmap.
00:23:32.480 --> 00:23:47.359
But I wonder what's one project or scenario you've seen where Agentic AI didn't work as expected, and what lessons can we take from that to design more sustainable, human compatible systems?
00:23:48.480 --> 00:23:59.920
So I think the previous about ordering concrete or giving an agent you know autonomy to make decisions about when to order concrete, that came from I've seen uh organizations start to experiment with this.
00:24:00.079 --> 00:24:05.440
And so one of the use cases they gave there was that the agent was designed to look at the weather.
00:24:05.599 --> 00:24:21.759
And if the weather meant that um it was beyond the minimum requirements, if the temperature was too high or you know, there was uh precipitation or rain uh expected on that day, it would cancel the order and then rebook the order for when the weather was likely to be uh better again.
00:24:22.000 --> 00:24:25.200
And we what we saw was that a level of discretion wasn't there.
00:24:25.279 --> 00:24:38.480
So if it saw that the you know that the right conditions were not met at the exact time that the concrete was supposed to arrive, it wasn't contextual, it didn't have enough context to say, okay, in an hour's time it's going to be okay.
00:24:38.559 --> 00:24:41.519
So let the let the delivery come and we can still do the pool, right?
00:24:41.599 --> 00:24:45.119
So some of these nuances and that were, and again, this is learning, right?
00:24:45.200 --> 00:24:50.400
This is these are um just and a very simple example about uh over time we start building it.
00:24:50.480 --> 00:24:52.799
So um, and luckily this was a sandbox environment.
00:24:52.880 --> 00:25:02.240
And so it was able to learn, or the humans were able to teach it some of these edge cases and some of these nuances so that it can start making decisions like a human would, right?
00:25:02.319 --> 00:25:13.680
The human's decision making isn't uh finite, it's it's based on a lot of nuances, uh, heuristics, and trying to teach agents all of those, which we as humans have picked up over life, is gonna take a long, long time.
00:25:14.000 --> 00:25:20.160
The downside to some of that is by doing that, we also start embedding our own biases into the agents as well, right?
00:25:20.240 --> 00:25:26.559
So they'll also start inheriting some of the um biases that exist because of our own experiences of that.
00:25:26.640 --> 00:25:39.119
So I think there's a there's a really fine balance to be had between teaching the AI agents on how to navigate the real world without enforcing our own biases onto it.
00:25:39.359 --> 00:25:40.960
We've seen it in LLMs as well, right?
00:25:41.039 --> 00:25:49.039
Where you where the responses come back, and because they're based off of things off the internet, they start getting you know misogynistic or racist or whatever these these things are.
00:25:49.200 --> 00:25:59.839
And I think with agents, we've got a risk that the behaviors that we exhibit or we've been used to do um acting out are now going to start becoming embedded into agents.
00:25:59.920 --> 00:26:01.440
And so there's a fine line, I think.
00:26:01.519 --> 00:26:08.000
And then and I don't know what the answer is on on how we do that, but we just have to be really conscious that we don't end up going down that route.
00:26:08.400 --> 00:26:09.519
This is so interesting.
00:26:09.680 --> 00:26:15.839
But what do you think is the most dangerous bias in the long-term perspective?
00:26:16.559 --> 00:26:17.200
Dangerous?
00:26:17.359 --> 00:26:18.000
That's a good question.
00:26:18.319 --> 00:26:33.599
I guess if we transfer some of our insecurities into the AI agents where we limit the possibilities are, but sometimes because of our own fears, our own personal insecurities, then some of those not necessarily dangerous, but they'll they'll end up limiting what's possible from the AI agents, right?
00:26:33.839 --> 00:26:50.319
So again, going back to talking about balance, I think there's a real balance to be struck in terms of making sure these AI agents are useful and are productive and are doing things that can really help us, but at the same time, uh, we're not limiting what they can do just because we can't think about it, right?
00:26:50.400 --> 00:26:52.559
We we can't imagine what the possibilities are.
00:26:52.960 --> 00:26:53.599
You're right.
00:26:53.759 --> 00:27:03.200
This aspect is definitely going to define whether we're going to make it or break it in the new stage of development.
00:27:03.599 --> 00:27:14.160
But coming back to the leadership topic, what kind of leadership is needed to ethically scale MCP, especially in regions moving fast like the UAE?
00:27:14.480 --> 00:27:16.640
And what should leaders stop doing?
00:27:17.279 --> 00:27:19.200
What should they lean into?
00:27:20.319 --> 00:27:24.319
So I think this is true for all technologies, not just for uh AI.
00:27:24.480 --> 00:27:30.720
I think it's maybe exacerbated with AI agents and MCPs, is just chasing the novelty factor.
00:27:30.960 --> 00:27:46.400
I think there's a real risk that leaders will say another company or our competitors are doing this or everyone's looking at this, like we need to do this, we need to do something to do with agentic AI, we need to do something with MCPs just because everyone else is talking about it.
00:27:46.720 --> 00:27:55.839
I think understanding your business, understanding where some of these tools could be applied and utilized is really, really important.
00:27:56.240 --> 00:28:03.839
And understanding, like we said, your organization's capacity and culture to be able to accept and absorb and utilize these tools.
00:28:04.079 --> 00:28:10.960
So AI could quickly turn into an expensive mistake if you're not careful and really purposeful about what and how you deploy.
00:28:11.200 --> 00:28:18.000
And again, I always, I know, compare AI as digital transformation that's been happening for the last few decades.
00:28:21.599 --> 00:28:34.079
Where there'll be this new shiny technology or this new shiny digital tool, and they would want to press ahead and deploy it just so that you know they can tell their leadership or they can tell their stakeholders that they're doing something in this space.
00:28:34.319 --> 00:28:46.640
The hype around AI is potentially starting to create this risk where leaders kind of jump in and almost have this fear of missing out and want to do something for the sake of doing something.
00:28:46.799 --> 00:28:53.279
And I think that that's really, really dangerous because therefore, usually in those scenarios, you'll skip some of the things that we talked about, right?
00:28:53.359 --> 00:29:04.160
You know, set setting up the framework, setting up procedures, um, creating sandboxes just purely from the desire to get something done and start showing results, right?
00:29:04.480 --> 00:29:11.920
I think the other thing is we're talking about what we were saying before, is start treating agents as if they're like junior members of your team, right?
00:29:12.079 --> 00:29:23.519
Literally like they're an intern or a graduate, and that they'll need hand holding, they'll need integrating, they'll need understanding of your business and how you work and what you do.
00:29:23.680 --> 00:29:30.960
Uh, I think all of those things with oversight and then uh continuing the analogy, like doing things like performance reviews, right?
00:29:31.039 --> 00:29:33.440
So reviewing the outputs, making sure that they're contextualized.
00:29:33.839 --> 00:29:43.119
I think those types of things are going to be required to make sure that we control some of the risk and also get the most benefit out of these types of MCP systems.
00:29:43.359 --> 00:29:46.720
We should probably be leaning into is experimentation.
00:29:46.880 --> 00:29:59.839
Is so we should be encouraging people to experiment, but in sandbox environments, doing it, taking it one step at a time and kind of reviewing what the outputs are, doing almost like you know, security and safety reviews in the same way that you.
00:30:00.079 --> 00:30:15.359
You know, we've got cybersecurity protocols and that we should be having protocols around agentic review, agenc safety, and like we have cybersecurity, and you kind of have a agentic security and making sure that that's constantly being reviewed and understood about what's happening.
00:30:15.519 --> 00:30:29.440
And then I guess in the same way that we ended up when we were deploying things onto servers, and we ended up creating like uh development operations, so DevOps, um, and now we've got MLOps, you've got machine learning operations.
00:30:29.759 --> 00:30:33.839
Most likely we're gonna see the equivalent of something like agentic ops, right?
00:30:33.920 --> 00:30:34.799
Or agent ops.
00:30:34.960 --> 00:30:43.599
So we we're looking at you know the whole that where the infrastructure is, what where the data is going, what it is that is being stored, how those agents are interacting with other agents.
00:30:43.759 --> 00:30:58.079
I think there's gonna be a whole, probably a whole category of skill set under this kind of bracket of agent ops that's gonna start coming up, which will be monitoring, creating the protocols, assessing and constantly adjusting all these different elements, right?
00:30:58.240 --> 00:31:06.079
The protocols, the infrastructure, the accessibility, and all of those things around uh where and how agents operate.
00:31:06.960 --> 00:31:24.400
I really appreciate that you took a look into the future because of course everybody who is working with these topics is wondering what's coming next, how is it going to develop, and what are the other opinions in the world around the same subjects?
00:31:24.640 --> 00:31:36.559
And of course, it is valuable when you are sharing your vision and also warning about something we should keep in mind before we step into that future.
00:31:36.960 --> 00:31:42.559
Prakash, do you believe we need a new kind of AI operating philosophy to guide this shift?
00:31:42.799 --> 00:31:45.359
Something beyond policies and protocols?
00:31:45.599 --> 00:31:49.200
And if so, what core principles should it include?
00:31:50.319 --> 00:31:51.839
So I think absolutely.
00:31:52.000 --> 00:31:54.559
So I think just governance isn't enough.
00:31:54.720 --> 00:32:12.799
Um I talked a little bit about Agentic Ops or AgentOps creating a new set of skill sets that understand what agents can and can't do and how they what the interaction between them and existing business systems and people is this whole philosophy and understanding that this is different.
00:32:12.880 --> 00:32:19.759
This is very different to you know the types of infrastructure, the types of software, the types of tools that we've that we've ever used before.
00:32:19.920 --> 00:32:28.640
And so I think automatically that's going to start start for it has to create a different set of skills and different approaches uh within organizations.
00:32:29.039 --> 00:32:33.759
So I guess one of the things that we should do is look at context over computation.
00:32:33.920 --> 00:32:57.359
So one of the kind of core tenants of it is to prioritize relevance over the biggest model or the fastest model and things like that, is looking at making sure that the outputs from both LLMs, Agentic, and when you put it together, the outputs of MCP are relevant and contextualized to what you do or what you need as an organization.
00:32:57.599 --> 00:33:03.440
It's really easy to just use these tools, like I said, and then and just get tons and tons and tons of output.
00:33:03.599 --> 00:33:18.000
But if it doesn't really help you do anything, and actually potentially might even hinder and make your processes slower because you're now having to evaluate or do something with a whole bunch of more data that you never never really needed or never really had access to.
00:33:18.160 --> 00:33:26.559
So there are risks that it can it can actually slow you down and make you less efficient, and which is the uh the opposite or antithesis of what we're supposed to be doing.
00:33:26.799 --> 00:33:33.680
I think the other element is making sure or we're looking at um a philosophy where you get explainability, right?
00:33:33.759 --> 00:33:38.000
So the models are able to explain what they're doing and how they got to the answer.
00:33:38.400 --> 00:33:48.799
This is a really important area within the evolution of LLMs and agents, is that they're continuously designed to explain how they got to the answer.
00:33:48.880 --> 00:33:50.240
So it doesn't become a black box.
00:33:50.400 --> 00:34:05.759
I think it could be really dangerous if we let the LLMs and agents kind of continue moving on, and especially as we move to more towards more and more autonomy, if we don't know how they're arriving at the answers or at least aren't able to explain when queried.
00:34:05.920 --> 00:34:10.320
And so I think that plus interruptibility, right?
00:34:10.480 --> 00:34:24.000
Being able to interrupt an agent and almost like have a kill switch to be able to stop it from doing something as soon as you realize it's starting to go in the wrong direction, so that you can prevent any potential ultimate catastrophic uh outcomes from happening.
00:34:24.159 --> 00:34:27.920
And then linked to that is this agent-to-agent transparency.
00:34:28.000 --> 00:34:36.480
So you've got explainability within one agent, but when you when you get one agent talking to another agent, making sure that you've got transparency of those interactions.
00:34:36.719 --> 00:34:54.000
I think in the early days, there was um I think Google, uh I think it was Google, they they came up with a um a method of of letting DeepMind, uh one of the early AI kind of technology companies, uh translate languages from one to another and then and then another language to another.
00:34:54.079 --> 00:35:02.639
And what the AI did was it created its own language to translate between language two and language three, uh which the scientists didn't understand.
00:35:02.800 --> 00:35:04.960
And so that's obviously very dangerous, right?
00:35:05.039 --> 00:35:16.239
Because you could then you had two systems communicating with each other that uh that was uh completely beyond the scientist understanding, and and therefore you had no idea of how it was doing it on what way it was doing.
00:35:16.400 --> 00:35:26.159
And so these are some early examples of how AI systems will will learn to start doing things without human intervention, and that could potentially lead to negative outcomes.
00:35:26.480 --> 00:35:33.199
And then one other kind of philosophy that I think we should be looking towards is um getting consistency out of the agents.
00:35:33.360 --> 00:35:43.840
I think one of the things we've learned with you know human interactions and interacting with each other is you know, we build relationships and bonds because we know what to expect from another human being.
00:35:44.000 --> 00:35:52.400
So, so you know, a friendship or or any other type of relationship is built on trust, and and that trust is usually because you know what you're gonna get.
00:35:52.480 --> 00:35:55.760
There's a dependable outcome and interaction with these people.
00:35:55.920 --> 00:36:00.159
Similarly, I think we're gonna have to look at how we can get behavioral consistency.
00:36:00.239 --> 00:36:11.920
Because imagine if you are gonna go to an agent and it gives you a quality response once a day, or it gives you a quality output one day, but on another day it's really poor, and you can't extend changed, nothing's different.
00:36:12.000 --> 00:36:17.039
But just because you know the way that the models work and they work on probabilities, you're getting inconsistent answers.
00:36:17.119 --> 00:36:20.000
You're less likely to rely on it, you're less likely to trust it.
00:36:20.159 --> 00:36:25.280
And so I think that's another area and another philosophy I think that we're gonna have to start looking at.
00:36:25.360 --> 00:36:40.880
And I think there's there's there's a bunch of work being done to try and ensure that you know that agents don't behave differently every time you interact with them, and they don't behave differently depending on you know who's interacting with them, as long as as it's within it within a certain frame framework.
00:36:41.199 --> 00:36:59.760
I just imagined how dangerous it might be if the agents start depending on their mood, like humans do sometimes, and they offer different quality and uh vary in their outputs depending on their hair day problems.
00:37:00.639 --> 00:37:05.039
I feel like on this on this uh this conversation, there's quite a lot of negativity, right?
00:37:05.280 --> 00:37:11.199
And I don't want to come across negative, but I I think that there's a lot of unknowns, right, in this in this space.
00:37:11.280 --> 00:37:25.440
And so I guess the message I'm trying to get out is just be aware of you know what this technology can do and it's great potential, but also be wary of um some of the potential downfalls and some of the pitfalls.
00:37:25.519 --> 00:37:28.400
I think that the three things that have got to go hand in hand.
00:37:29.199 --> 00:37:30.079
Exactly.
00:37:30.320 --> 00:37:43.039
And I totally share your opinion that it might sound a little bit negative, but at the same time, this is a serious game, and uh we need to be real.
00:37:43.199 --> 00:37:50.000
And in order to be real, we have to look into the positive outcomes and into the dangers.
00:37:50.320 --> 00:38:26.400
And of course, everybody prefers talking about the positive outcomes and the results we might reach and grow our business and enable humans as well, but still at the same time, in order to come there in a good, sustainable way, we have to highlight all the negative sides or potential limitations and dangers so that we don't need to go through it in reality, but can eliminate it and reduce that risk through discussing it, through talking about it.
00:38:26.639 --> 00:38:28.960
And uh that's exactly what we are doing.
00:38:29.119 --> 00:38:43.599
So this is a great help to everybody who is entering the space of agentic solutions so that they can avoid all those pitfalls and go straight into the winning game.
00:38:45.119 --> 00:38:59.039
Prakash, what is your one piece of advice for leaders and business owners preparing to implement next generation AI in a way that is meaningful, scalable, responsible, and future-proof?
00:39:00.320 --> 00:39:06.800
I guess if it's just one piece of advice, don't ask, you know, can this agent or agency system work?
00:39:06.880 --> 00:39:07.039
Right?
00:39:08.000 --> 00:39:09.440
Like, don't ask, does it work?
00:39:10.239 --> 00:39:15.119
Ask, can this agent work with my my teams and my people?
00:39:15.280 --> 00:39:16.880
Can it work with my system?
00:39:17.039 --> 00:39:18.559
Can it work at scale?
00:39:18.639 --> 00:39:20.639
And and can it work with transparency?
00:39:20.800 --> 00:39:22.320
Is there explainability in it?
00:39:22.639 --> 00:39:39.519
Make sure that when you're putting the agency system in place, you're doing it for a reason that we understood, and that you're not you're not just expecting it to do something, you're expecting it to do something with the people in your organization, with the systems and protocols that you you have in place.
00:39:39.599 --> 00:39:49.039
Um, it's able to be scaled across your organization, and that it's it's your understanding how it works and that that there's that level of transparency.
00:39:49.280 --> 00:39:54.800
So I think you've got to think sometimes hard to think beyond LLM.
00:39:55.039 --> 00:39:59.360
And I I hear this a lot from people is that they use LLMs as if it's Google, right?
00:39:59.440 --> 00:40:02.400
As if it's a search engine and they're typing answers in there.
00:40:02.559 --> 00:40:12.800
And then as soon as they start using it for more than that to help them to you know uh refine their thoughts and define what they what they're about to do, and that it opens up a whole new world.
00:40:13.039 --> 00:40:39.199
And so I think um all of these things are really important to remember when you're looking at moving to the next level, right, from LLMs into you know agents and then MCPs, is things like context sharing, um, role delegation, understanding the role of the different agents and what what the inputs and outputs are likely to be from each of those, some of the ethical boundaries that we talked about, um, and and then always, always making sure that there's a human fallback, right?
00:40:39.360 --> 00:40:51.679
So no matter how much uh or how powerful some of these agentic systems uh will get, is I think we should never be in a situation where it's 100% independent of human intervention.
00:40:51.840 --> 00:41:11.039
There should always be some level of human in the loop to uh, even if that is uh just to make sure that there's a there's a safeguard or a uh a way of shutting it off, there should be a human in the loop at all at all times to ensure that we don't end up with this potential you know negative or disastrous uh circumstances.
00:41:12.239 --> 00:41:13.039
100%.
00:41:13.599 --> 00:41:18.400
I couldn't agree more, and that's uh truly great advice.
00:41:19.280 --> 00:41:32.960
And my last question for today, even though I would love to continue this conversation and hear so much more about it from you, but I trust we can come back to this topic in the future.
00:41:33.199 --> 00:41:38.639
So, what is one thing we all need to unlearn in the AI era?
00:41:40.239 --> 00:41:42.239
Oh, I think there's so many things.
00:41:42.400 --> 00:41:47.519
Uh one of the we just talked about is you know not treating LLMs like like they're search engines.
00:41:47.599 --> 00:41:53.519
I think we're so used to um, you know, uh Googling uh things and and then expecting an answer back.
00:41:53.679 --> 00:41:59.440
So unlearning uh that is uh is probably one of the kind of lowest hanging fruit.
00:41:59.679 --> 00:42:25.920
But I guess if we're gonna go a little bit more deeper, I guess unlearning some of the command and control mentality, historically, our relationship with digital tools and and computers and things like that is assuming that we we input something and then we get an output and we know you know what we expect, what the output is expected to be in actually probably quite quite a narrow band of expectation.
00:42:26.159 --> 00:42:34.400
And what we're seeing now in the AI era is we get agency systems that are we're no longer commanding a piece of software, right?
00:42:34.480 --> 00:42:40.880
We're commanding something that we're co-creating with uh and and it's using probability and probabilistic outcomes.
00:42:41.039 --> 00:42:48.239
And so so getting out of that command and control mentality and and almost getting into a co-authoring mentality, right?
00:42:48.320 --> 00:42:54.559
So so this thing is going to be there, uh your AI agent is going to be there to do things with you, not for you.
00:42:54.960 --> 00:42:59.679
I guess and that's a really there's a really difficult mental mind shift to do.
00:42:59.840 --> 00:43:04.400
So even now, I think a lot of people are still using LLMs to do things for them.
00:43:04.559 --> 00:43:17.840
Uh and the ones that get the best outcomes from LLMs are when they when they work with the LLM, do multiple iterations, do refinements, and then you know create whatever the outcome, uh the uh the output that you're expecting is.
00:43:18.000 --> 00:43:25.440
I think that's that's really, really important and probably one of the most difficult things to do because we've been so used to and grown up with doing it this way.
00:43:25.599 --> 00:43:34.960
So I think you know the the the younger generation who are gonna be exposed to this technology right from the beginning, they're not gonna know any different, right?
00:43:35.039 --> 00:43:41.760
This is going to be the way that they interact with the technology, and so they're gonna grasp it much, much quicker than the the likes of you and I.
00:43:42.159 --> 00:43:43.039
So true.
00:43:43.280 --> 00:43:51.599
Thank you so much, Prakash, for being here in the studio today, for sharing your experience, your vision, your wisdom.
00:43:51.920 --> 00:43:53.840
I truly appreciate you.
00:43:54.639 --> 00:43:55.360
Thank you so much.
00:43:55.519 --> 00:43:57.280
Thank you for giving me the opportunity to do this.
00:43:57.360 --> 00:43:58.639
It's been my absolute pleasure.
00:44:00.239 --> 00:44:00.639
Thank you.
00:44:02.079 --> 00:44:06.000
Thank you for joining us on digital transformation and AI for humans.
00:44:06.159 --> 00:44:09.440
I am Amy, and it was enriching to share this time with you.
00:44:10.159 --> 00:44:17.119
Remember, the core of any transformation lies in our human nature, how we think, feel, and connect with others.
00:44:17.679 --> 00:44:25.360
It is about enhancing our emotional intelligence, embracing the winning mindset, and leading with empathy and insight.
00:44:26.000 --> 00:44:35.440
Subscribe and stay tuned for more episodes where we uncover the latest trends in digital business and explore the human side of technology and leadership.
00:44:36.000 --> 00:44:47.519
If this conversation resonated with you, and you are a visionary leader, business owner, or investor ready to shape what is next, consider joining the AI Game Changers Club.
00:44:48.000 --> 00:44:50.639
You will find more information in the description.
00:44:51.119 --> 00:44:57.679
Until next time, keep nurturing your mind, fostering your connections, and leading with heart.