The Age of Hybrid AI

Speaker A: Welcome to Digital Value Creation. This is the channel with my brother and me. My brother works at an AI focused hardware company and I work at an AI focused software company. And we created this channel to discuss AI trends and how to drive impact at our work and our personal lives and uh, more importantly how to stay ahead and future proof our careers. So what's top of mind for you brother?

Speaker B: Well, today's um, interesting day. It's been two years now since GPT4 launched. I mean it's hard to believe how much changed and for me it's, I'm um, super happy to see that Google who started it all with the Transformer white paper just released their Gemini 2.5 which supposedly it seems outperforming both OpenAI and anthropic models. So on one hand it's amazing how much these models get better, more capable. On the other hand I wonder how much changed. I mean do companies really drive tangible value? And what are the areas. I know that you just came back from a CSO summit, I'm just curious, what did you hear? What others are doing, what companies are really who are making strategic investments see as a return on those investments.

Speaker A: So I want to come back to the, to the uh, summits I went to. But this is, you know, this, you and I talk about this all the time. How much more AI do we need as individuals in our daily lives? I mean GPT4 did a lot and uh, I was pretty happy with what I could do with it two years ago and now you know, we went through deep seq, we went through various uh, anthropic models and now Gemini is out and they finally got beyond the blue dock which is great. Um, but ultimately the question is do we need more AI for our personal use? And I'm almost at the point where I think I could just for most of the tasks I'm doing I think I would be fine with most text based models in GPT4. 4.5 is cool, um, and obviously 1.0 is awesome and all that but at the end of the day it looks

Speaker B: co interesting though because uh, I think especially for personal use the trend we have shift it's not about the best and the latest models is about how can you integrate it in our life. So um, I'm looking at super exciting. I just saw a demo where the latest AI glasses can do real time synchronous translation. Now that's interesting. And you don't need the GPT 4.5 or ah, a reasoning model uh, to do that. So some of these lighter weight models that are optimized for smaller form factors I think will make a difference for personal use.

Speaker A: Yeah, it's interesting. A friend of mine is, um, uh, you know, who has wanted to be a film producer or filmmaker.

Speaker B: Right.

Speaker A: So for him the new AI models with better text to video performance actually make a difference. So it depends on the individual use cases and we'll talk about more the business ones. But for my use for um, tax to text basically kind of use case, it's limited. But then your point is well taken. So there could be edge cases. And it's interesting at our businesses, most companies now have their own licenses for OpenAI or similar models and all the companies are developing on top of them. Maybe they could just use the corporate license. I'm wondering if one day the same thing will happen. I'm already paying 20 bucks a month here and 20 bucks a month there. Why can I just use my subscription to a single OpenAI model or Gemini model, what have you, and let all the other tools that are coming into our lives just tap into my subscription. So I'm wondering if that consolidation will ever happen. But going back to the question of, question of events is very interesting. So I went to different um, summits. I went to a summit of chief strategy officers. It was really interesting because they're not technologies. So it was a combination of uh, strategies from across industries. And there were a lot of different discussions. Um, I'll touch base on some of them. But one of the things that kept coming up, which also came up at a customer summit I recently went to at my company, is the nature of hype that exists in the tech industry. So the one way and because hype is a loaded word, but what it practically means is, is technology comes out ahead of its practical use. So we are, and we're making a big deal out of it and we all do it, all of us are in technology. So we like to announce and bring to market solutions that really don't have a well defined problems. And because I always lived in technology, you have too. I never thought about that being a problem. But then I ran into uh, uh, others, uh, that are from the pharma industry or chemical industry and they're saying, listen, we would not come up with a, a possible drug saying listen, we really don't know what it's good for. Why don't you go out and try it and see what it works, find out use cases for this molecule we came up with or a chemical like here's a chemical compound we Concocted. Why don't you take it out in the open and see how we can use it. But in the tech industry we do this all the time. So I don't know if you think about this kind of stuff, but I do. It sort of bugs me that, that we do this uh, instead of us going out and solving problems, which is what a pharmaceutical industry would do or a chemical. And you're saying this is a compound that's needed in order fix this kind of a problem. We generate a lot of uh, technology seeking problems. I don't know how you think about it.

Speaker B: Yeah, absolutely. I think we are amazing at creating solutions without defining the use case for them. And it's like well it can solve anything and everything. But because of that I think this is why we have venture hype cycle, this is why we have booms and busts because we over invest in the potential and a lot of times that potential just doesn't translate into tangible use cases which as you know my pet wave of how does that translate to real value creation? And uh, especially right Now, I mean two years ago ChatGPT changed everything, especially with GPT4. That was mature. Now the next wave is agents. Like the notion that all of a sudden we will just start to trust enterprise processes with autonomous agents. I'm pretty sure that came up in your conference as well. What was the opinion there?

Speaker A: It's interesting, um, uh, there was a lot of discussion because agents are everywhere right now, at least at this moment. Um, and one of the interesting things, I was um, at the strategy forum and um, um one of the presenters came from um, a semiconductor company, um, not unlike yours. And uh, his point was we're layering uncertainty on top of uncertainty. So we already have a challenge with the unexplainability or explainability problem with AI. With generative AI, we don't quite know how the algorithm works or because it's not an algorithm. Um, um, we don't know how the models really work so we can't quite explain it but we sort of like the outcome of it. So there was even a joke in the room saying, you know, like Igor, collect my money but I don't need to know how you did it. It's almost how AI works, right? Like you know, find, you know, perform task but I don't need to know or obviously cannot know how, how you perform it. And his point was something in addition to that. So we're layering this kind of technology and over time the expectation we're going to put AI on top of quantum computing because we're going to need a lot more compute. So it's a race for better and better, uh, or more and more compute. One great source of compute is going to be quantum computing. So, but quantum computing, we don't even know where the electrons are like or the photons like. It, we, we, we can't explain it at all. So if you layer on certain compute and on top of that you layer uncertain AI, you can't explain the world. So we're going to, and ultimately all of this agentic force, or whatever we're going to call it a year from now is supposed to perform tasks. So it's going to be autonomous more and more. So we're building autonomous enterprises on top of unexplainable quantum computing and unexplainable, um, uh, um, generative models. So it's sort of, you know, we're building the unexplainable enterprise. So that was one really thought provoking discussion. So I thought was pretty cool. I don't know. How do you think about that? You're in the hardware space.

Speaker B: Yeah. So uh, build on that. Just attended a couple CIO roundtables and AI forums and the discussion is top of mind, especially for enterprise. Because content creation is one thing where if something is not explainable, but we love the outcome, hey, we wanted to create the outcome. That was the whole purpose of content creation. Now when we want AI to take action, uh, make decisions and execute with real life consequences, that's a whole different game. So we have a lot of discussion around how AI agents migrate the renaissance on automation technologies or process orchestration. Because all of a sudden you don't want to build uncertainty after uncertainty. You don't want one agent's um, unexplainable results used by another agent as an input and use whatever tool they do. And it's interesting because in the past, um, a lot of these models, a lot of models were necessarily creating just content and you still had let's say uh, llama index or some other orchestration layer that uh, chained it together. But right now most of the reasoning model uh, has inherent agent capabilities. So you have even less control. It's more of a black box. So what we discussed with my peers is how enterprise systems will need a balance, uh, we call it hybrid AI that need a balance between uh, process controls, orchestration capabilities and leverage AI in a way where uncertainty or unexplainability is within a certain risk tolerance. Um, what was the solution that emerged besides just identifying uh, a problem in your discussions,

Speaker A: uh, I know you and I sort of settle in this term hybrid AI, uh, uh, which all of a sudden anthropic uses that word for slightly different purpose. Uh, but hybrid AI is really good description for me in moving on. This, this range from highly deterministic processes and highly probabilistic processes or explainable and unexplainable range.

Speaker B: Right.

Speaker A: So, and, and I actually believe we need both. So uh, so it was interesting. I used to work for uh, an RPA company and rpa supposed to be dead now, but if we think about it, there are many, many use cases where you actually need the if then else predictable RPA process or some kind of a process flow that you exactly understand. It can follow standard operating procedures because that's what you do. So I was at a conference in Europe, um, at a customer conference and uh, I was with some banks and they said, you know, it's interesting because they think about this. So on one end of the spectrum they need deterministic processes that are processing loans. You want the loans to process them. You process loans in a certain way because it's the law, you have to do it in a certain way. If you randomly have an agent make up a way of processing a loan and you violate some rules, that's going to be a problem. So you want to. One end of the spectrum loan processing will have a very deterministic element where you follow the seven step that's prescribed by your. Either by your policy or by law or combination thereof. But then you have things like fraud detection that is a fuzzy area because you don't know all the like you can't predetermine where fraud is going to happen because it's a cat and mouse game. But AI can get really good identifying edge cases where saying that looks fraudulent to me and it could be anything and it's an image or it could be a behavior. And this is going to. The agents will get really good or uh, generally AI will get really good at learning and determining and identifying these kind of. So your future process for lawn processing for example will have a combination of the two and that's the hybrid AI nature. So you don't want to let loose an agent saying hey, just go crazy. But maybe there's going to be something that orchestrates it all and there's going to say for this process I'm going to use more deterministic way of doing certain elements of it and probably more unexplainable or probabilistic way using Gen AI or AI Agents that I don't even know what they're going to, they have a goal like identify if this document is fraudulent loan. That's your assignment as an agent. You find a way to execute that task. That would be a goal seeking agentic performance versus process the loan following these seven steps. So these are the kinds of things we talked about. I thought it was very interesting because I think we in tech industry have a tendency to just say hey, everything's going to be using the latest technology. And that's not the case. Yeah, so that's what I'm finding. Um, uh, and there was an interesting thing as it relates to all of this, uh, because I was in Europe. Um, so Europe has an AI directive that's very strict. So it requires explainability, data protection. It has a lot of attributes that's sort of counter to um, a self determined goal seeking agentic kind of AI. So unless the law changes, I think there's going to be inherent conflict where you can explain it, you can prove that data was protected. So you have one of two choices and this came up in discussions. Either you dial down, create these guardrails, it's a very popular term. So let's put guardrails on the AI. And when you put guardrails on AI, all you're doing is you're putting AI into more of a deterministic corner. You're going to say I know it's exactly how it's going to perform, which is fine. But then it's not the AI you're trying to unleash. It's not using the latest capability of a probabilistic goal seeking kind of, kind of agent. So I'm wondering how do you think about those two ends of the, of the spectrum and maybe you think about, I don't know, around your, your company if you can talk about it like where you seeing into, you know the, the role of, of different, more deterministic or, or probabilistic AI very similar.

Speaker B: Not just in my company, as I said, um, that that was top of my forum because everybody wants to take advantage of finally going beyond content creation. Because while content creation creates value and I know we want, we discussed before, before this call, uh, the latest McKinsey survey, which is super, super insightful, what domains are really uh, adopting AI and not surprisingly initially started with the content rich domains but as we go beyond that you need some of these agent capabilities. So it's more about the focus on this orchestration layer, focus on how can you have checkpoints in the process so the uncertainties don't multiply, but you will have a chance to drive deterministic process elements. So we are looking at some legal processes where we do a lot of content creation. When there is a check, we do um, especially co generation is fantastic for us because uh, you can have actually test included in the process. Which is why we see a lot of potential of co generation. Not just test, but complete uh, orchestration because the tool itself can have checkpoints. Uh, you don't need necessarily human in the middle. But I know that you were passionate about some of the other insights from the survey. What strike you that?

Speaker A: Yes. Yeah, we'll drop the survey link in here because I certainly want to avoid the cease and desist from mackenzie. So we'll provide the link in the show notes. But there were a couple of really interesting insights. One is um, so this was a survey. This is brand new when we shoot this video, brand new in March 2025. And this is their annual survey on AI. This is done by um, Quantum Black, the AI unit of MacKenzie. And what I like about this is just the size of the survey. So about 1500 companies involved and they look at something, I don't know who else does it. They have consistently done it around automation and AI now is trying to figure out where um, the technology can have realistic business impact. So that's, that's sort of the angle I always liked about these surveys. So they look into one thing. In this survey they said, they survey, the company said, what percentage of your companies are using AI in a certain segment? And they look at this by industry and by business function. And it was actually surprising to me, uh, because the, the number one use case for generative AI use across industries in sales and marketing. And I did not expect that actually um, by far. So in the technology industry itself is 55% of the companies saying we are using generative AI in sales and marketing. And maybe all of this is copywriting for all we know or uh, some uh, agency work. But it's still very high percentage. Some of the things that I expected, for example in software engineering or IT, um, cogeneration, uh, which is very common is about 30% of the companies said we're using that in the tech industry. Across the board it was varied. So I thought that was very interesting. It was very low in financial services, which is an industry that's very focused on rolling their own. There's massive engineering resources typically in a financial services industry in banking or insurance. So that was surprising to me. But across the board there was a lot in product development which again was surprising uh to me IT product development in professional services I assume this could be methodologies or um surveys or whatever consultants do. I Forgot it was 20 years ago so that was really interesting. A ton in knowledge management again in professional services 40% of the consultants said we use uh AI for uh um knowledge management. And I want to come back to because you have a, you have a really good observation on this based on what you and I talked about. Um but it's really interesting I suggest um that listeners uh and viewers look at this survey just to help them guide like what are the areas the top three is. Again it's sales and marketing, customer service and IT or software engineering. So when you're looking for your AI um use cases that's where I would start across industries because that's clearly where, where the survey shows companies are adopting and you're going to find in survey also they're pointing out uh is there are companies that are using 4, 5, 6 different business processes are basically gen AI enabled which is high. So it's no longer just experimentation in one unit it's it's spreading across companies. Uh so that's a pretty good thing. But I want to come back to knowledge manager because you had a really interesting view on, on uh where genai helps um um in careers or in personal development and um depending on where you are in your stage of your career. So what were you thinking on that?

Speaker B: Actually before I answer that just want to confirm what you observe on a survey as well because this is actually how our company started as well. We started in marketing and sales which is uh, is very content heavy and you need to be adaptive. Time is always at the essence and then uh for us one of the major breakthrough and very significant successful around customer engineering which is dealing with hundreds of thousands of use cases uh that we need to respond to. And then it went into software which led me to think about how comes IT and engineering who technically pioneers these technologies were not the first one adopting it. Here is an interesting observation. We have seen where um these gen AI solutions really create value on kind of the opposite end of the spectrum. And what I mean by that is I'm a big believer of strength based leadership. We were taught that if you really want to multiply your impact, focus on your strength and delegate your weaknesses which to some extent works and sometimes you need to compensate for your weaknesses and learn it. And what we have seen that some of the early coding models really did not resonate with our top Developers because they created code snippet, there was some improvement in performance, not a true breakthrough. So they felt that it did not really help them in their strength area. It was helpful, not a breakthrough. But we've seen that changing in the last three, four months. I mean, I don't know as viewers, uh, or listeners whether you played with um, either anthropic Sony 3.7, which was until today the Gemini release the best encoding. But more importantly, some of the open source tools that are built on the top of it, like Klein, where you can actually describe a problem in generic terms and client will find the right library, actually is a fully automatic agent take approach to AI where it coordinates all the work from research to development to some testing to really create a fully functioning code. And this is where the light bulb went on for some of our leading developers that oh, this is no longer just code snippets or helping me create something. Now I can move from developer to product manager. All of a sudden I cannot build on my strength that I know how to architect and I can use this agentic AI to create a solution. But since I'm the expert, I can very quickly identify issues, problems and iterate very rapidly. It's almost like I just inherited the theme. So that's where we saw that right now these tools reach the level of maturity where you can really build on them to extend your strength. Now the other part of the spectrum is still true. Some of our more beginner developers actually leveraged AI earlier than the experts because for them, some of those core snippets how I best code an algorithm if I need to migrate from one, um, programming language to let's say angular, how I do that, that was super helpful for them. So what we find is really if somebody is new to an area, needed to learn a new language or add up to a new uh, code base, that was very helpful. Now the technology is at the level where if you are a top expert, it can really help you to elevate your career from a developer to, to product manager. But you still have a knowledge to provide oversight to make sure that the AI generates. So for me that was super interesting that now we are at the point where you can really delegate to AI to accelerate your strength and also to help you compensate your weaknesses. So super excited.

Speaker A: Yeah, it's interesting. And maybe we can close on this note and how people take that away because, um, um, I have, you know, a daughter that's leaving college and she wants to get ready for a new job and I think I Wish I had tools like this where you could just be. I don't think you become an expert but you could be productive very quickly with AI tools, uh, which just wasn't available. I don't remember how we got up to speed when we had a new job. Probably like struggling, struggling reading books. I don't know what we did. Um, so what do you think it means for somebody's professional career? Our usual comment on that for me

Speaker B: you just get a building consultant. I mean it's funny because sometimes our team want to bring in a consultant to even just frame up a problem or analyze the story. And I'm just not approving those type of consulting. These phase zero engagements, I think they are that. But basically each of us now have an ability to look at uh, a domain that maybe we are not an expert on. We need to learn from or we need to be get up to speed and be credible and have basically a consultant available for us. Not like any consultant. Sometimes they make up stuff and sometimes they are not accurate. But from zero to a good understanding. I think it's a phenomenal career too.

Speaker A: I think so too. So maybe. I know we keep saying this, I think in this channel that we're going to come, come back to the toolkits we do use because both of us use get up to speed tools. Um, I know at our company now the entire company uses basically uh, one of the research tools which we like. It's u dot com. It's not a plug. But uh, just, just before you open your mouth, please research the topic you're about to talk about. Just it's helpful. I think it, it levels the playing field. So and oftentimes if you. It's really interesting observation. Maybe it's my closing one. I mean none of us are experts in everything. So when I talk to different department and I'm not an expert on their area, I use AI tools like make me knowledgeable enough for this conversation. So they. This is a topic I don't talk about. It's important to them. I get up to speed. So at least I'm not an idiot on the call. So there are really interesting average daily use cases where you. We can, I mean at the end of the day we can communicate better if we have a sidebar. So maybe we'll dig into that next time. But, but for now we can close out like we usually do.

Speaker B: Philadelphia is out and San Diego is out. See you next time.