[Automated Transcript]

Ashwin: We have a lot of folks going into those fields because there is a tech revolution happening. They are being displaced because they don't have tech skills or they don't have adjacent skills and

Ashwin: they don't really have coaching skills, but because there is no barrier to entry there, everybody can go and be a coach. So now we have a flood in the market of effectively people skills, which means we have a degradation of the support that people get over time because now you have a flood in the market of mediocrity.

Ashwin: It's both a tangible and foreseeable outcome and it's also an ethical dilemma, and it's something that I don't see talked about a lot. When you displace people, where are they going to go that is meaningful?

Deep: Hello, I'm Deep Dhillon, your host, and today on your AI injection, we're joined by Dr. Ashwin Mehta, founder and CEO of Mehtadology 

Deep: Ashwin brings a blend of deep academic training and enterprise experience with a PhD from the University of Leeds and MBA from the University of [00:01:00] Northampton and senior roles spanning global organizations, including Bayer and Deloitte Today, he helps organization streamline and modernize their operations through automation for strategies and AI powered decision systems.

Deep: Ashwin, thanks so much for coming on the show. 

Ashwin: Thank you for having me. I'm looking forward to this discussion. 

Deep: Awesome. Maybe start us off by telling us about your time at Bayer and Deloitte and tell us like, you know, a little bit of about the environment that you were operating in and what type of AI problems you were seeing.

Deep: How did AI make its way into the org? What type of organizational dynamics were play, that sort of thing. 

Ashwin: Uh, yeah, so I'll, I'll start with the previous organization, which, uh, that was, that was Bayer and when I was there I was, it was really the emergence of the, the generative AI revolution. This was when Chat GPT came out pretty much as I was starting there, and.

Ashwin: A lot of the work at that time was [00:02:00] getting to grips with this new disruption. So we have a new disruption, everybody is talking about it, everybody's trying to figure out what it means for the business and how we can start to use it effectively for the things that we wanted to do. And at that time, a lot of the, the revolution was effectively in two areas.

Ashwin: So one was the generative side, which is, uh, can we make. Content. We can make media, we can make videos, we can make images. The tech was emerging at that point, but it was still very interesting. And you had, in communications, marketing and learning, you had a lot of these avatar tools, , which you type something in and you get a talking head video out of it.

Ashwin: So everybody was really enamored with those on the other side. So two sides. So one was this, the SaaS kind of tools that were there. And on the other side was the, uh, the kind of more foundational stuff, which is, it's kind of laid the groundwork for some of the things we do now, which is how do you think about data, how do you think about [00:03:00] leveraging that data within a chat bot kind of structure?

Ashwin: So you ask a question, it looks into your data, it gives you an answer that's. Hopefully pretty accurate now at the time that, that stuff was also emerging because in the old days, we used to make, let's face it, we used to make chat bots with Skype and Excel. Right. However many years ago that was. So it was the next level of that.

Ashwin: And we were looking at that for things like procedural, validation for things like SOPs and policies. And I imagine many organizations have been doing that over the last couple of years as well. So this is the kind of thing we were getting into, into, and we were very much in the Azure kind of, AI foundry, machine learning, that kind of, 

Deep: yeah.

Ashwin: Area. 

Deep: And were you guys, would you say, you know, this is like a technology org like that, that whose charter was to sort of track state-of-the-art moves and try to apply them internally? Or would you say there's like a very specific set of [00:04:00] problems that Bayer, for example, had that you found yourself looking at the tooling to address.

Ashwin: So the specific set of problems that I was looking at at the time with that team was, uh, there were big companies. Don't necessarily move as quickly as small companies. Let's just put that on the table. And that's not a Bayer thing, that's every big company. But the kind of thing we were looking at at the time was lots of, lots of SOPs, highly regulated environment, and every SOP requires people to be trained on the SOP.

Ashwin: So we had whole departments that were basically churning out small bits of e-learning about procedures, very minor procedures. Can 

Deep: you maybe just contextualize that a little bit for our audience? Yeah. Like what are the SOPs exactly. And what's sort of like, what, what function do they serve in a, in a pharmaceutical company like Bayer?

Ashwin: So the, the function that they serve in pretty much any highly regulated business, but let's talk about pharmaceuticals, is every step of assuring that medicines are [00:05:00] safe to use, uh, and every step of responding to queries about that, whether it's data, whether it's things that you find in the market, what, whatever happens to be every step has to be scrutinized, has to be looked at, has to be communicated, has to be validated.

Ashwin: Uh, and you had lots and lots of departments of people doing these things to make sure that everything that the end consumer, uh, has access to or is prescribed, is ultimately safe for use and improves health and things like that. 

Deep: Right? So part of your challenge was like a training challenge internally for employees to be able to understand what the operating procedures were and then maybe a, a regulatory.

Deep: Step to document that they actually were trained and prove that they actually mm-hmm. You know, paid attention a little bit. Stuff like that. 

Ashwin: Absolutely. Making sure that everybody's trained and is conforming with the relevant procedures. That that was the challenge. 

Deep: Yeah. 

Ashwin: And the way that the company had been tackling that was to do lots and lots of e-learning.

Ashwin: And that has a cost. That has a cost not [00:06:00] only in terms of development and all of the departments that are, that are mobilized to make these things, but also it has a cost in terms of learner time, in terms of systems and all of these other things that you would expect. So that was a problem statement.

Ashwin: There were others, but that was one of the key one. And another thing we were trying to do at the time was can you get to a point where if I am doing a, an activity that is governed by a procedure, can I just ask if I was to ask, I'm doing this activity, what do I need to do? And ideally, your chat bot structure, your agent structure, your um, assistant structure, whatever we were calling it at the time.

Ashwin: Based on that query should say, these are the procedures, these are the things you need to do. We have documented that in some kind of way that you've been shown that information, we've tested or assessed that in some way that you've understood that and we are confident that you can go and do that. And that's a level of detail on the data that goes way beyond the tick of, yes, I did an e-learning [00:07:00] course, so this was the problem statement that we were trying to address because drops costs makes it scalable.

Ashwin: You always, you're always using the source data. You always have accuracy and you're getting the thing, uh, the information you're getting when you need it, rather than at some time during your onboarding. So lots of benefits around this. But that was the generative AI 

Deep: stage. Yeah. And that started around GPT 2 i I think from your timeline.

Ashwin: That was, no, that was GPT-3.5 

Deep: 3.5 Okay. So around three, five. And so then presumably, so you have a corpus of all of the regulatory processes, sort of all documented and documents, presumably you set up like kind of a standard rag like system, um, upload the content, find kind of a narrowed, reduced search space, and then, you know, hit the LLM to like align the responses.

Deep: Mm-hmm. But within the regulatory world, I imagine it's kind of a big deal to make, to at least audit that what you presented to [00:08:00] your internal people was indeed what was actually present in the documents. You know, with respect to the hallucination problems or just the distortion effect that you can get from LLMs, certainly in those days, you know, three, five was a much weaker model than what we're looking at today with, five plus and the, and the new Google models, et cetera.

Ashwin: So the, I think there are three aspects of this that at the time made it a little bit untenable. Now it makes it kind of easy. So one aspect is the quality of the models, which you've rightly described, are very different. The second is the ability to tweak model parameters. So back then it was very difficult within the interfaces that we had using Azure at the time, it was very difficult to make sure that we had.

Ashwin: Sampling temperature for, as an example, we could control whether an automation response was going to be very deterministic or very probabilistic. So we, we have that control at, at our fingertips. Now, we didn't really have it easily back then. It still required programming It required machine learning, it required Python to direct what [00:09:00] we were doing.

Ashwin: And, so those, those two bits are, very clear. The third bit, which is less clear and it's still less clear, is the ability to prove in the data that something was done in a fairly binary way. Because in a highly regulated environment, really you want a yes or a no, was this person appropriately trained on the thing they needed to do?

Ashwin: And that's way more difficult when you're looking at, sentiment or semantic analysis of transcripts, which is a conversation that you're having with a chat bot. So. It becomes a little bit more difficult, I think now we could easily solve that problem in a rag structure. But back then this, this stuff was pretty new.

Ashwin: So we were doing all that we could to make that work. 

Deep: But help me understand what that means a little bit. Lemme let me unpack that. So, is the challenge that you need to audit that the questions were asked by the employee or presented somehow to the employee and that they acknowledged the responses? Or is the challenge [00:10:00] that they asked something and the bot said something and you have to align what the bot said?

Deep: That question answer paired to the actually correct response from the kind of documented process in other, 

Ashwin: so 

Deep: in other words, like yeah, you know, dealing with the guardrails and the hallucination problem and then quantifying it. Like what, what part of it was hard. 

Ashwin: Uh, the path that was hard was actually nothing to do with the tech.

Ashwin: It's the organizational appetite to move away from something that is tried, tested, and accepted by regulators towards the new thing, which is, let's face it, this was the same problem moving from training to e-learning, right? Um, so there is a higher burden of proof on the data that you have. So then you have to scrutinize the data in different ways.

Ashwin: And then that's where you start to get into this, what you are talking about. Yeah, what was retrieve, when was it retrieved? How accurately was it retrieved? How does it pair? And all those are tend 

Deep: to be technical conversations that you have to have with multiple parties in a, kind of, in an [00:11:00] organization.

Deep: Well, let's, let's dig into that a little bit because I think like with your background and exposure to a lot of different AI projects in different environments, what do you see as some of the barriers to, AI adoption? What are some of the like, frictionless points and what are some of the heavier friction points and on, and which of them are like tech and origin and which of them are maybe sociological or process oriented in origin?

Ashwin: Uh, so I did my PhD in technology adoption, so I'll talk first of all about what exactly are we talking about now. I've recently been doing a lot of work in the space of agentic readiness rather than AI readiness.

Ashwin: And that's how you prepare your organizational structures for accepting whatever it's that the tech is capable of doing. When we talk about adoption in the IT space, we tend to talk about the human, capability or human preference to. Click on a particular button. So if I'm using a particular [00:12:00] tech, I'm basically clicking buttons and IT teams will usually say, how many people are using this on a daily basis?

Ashwin: How many people are clicking through? How many people are logging in? All of these kinds of things. That's valuable in the AI SaaS space. So we, in the old days, we used to say we have software as a service. We have a thing that we buy, we license that, and we wanna know how many people are logging in, how many people are using it in the AI SaaS space.

Ashwin: So the co-pilots of this world, as well as some of the point solutions that make media. We want to know again, how many people are using this tool when we move to agents. So we have things that can autonomously take action in different systems. Then the concept of adoption is a little bit fallacious because.

Ashwin: Really, we're not talking about people clicking on things, we're talking about, uh, workflows executing autonomously. But across all of these things, we need to be slightly clearer on the terminology of adoption. So, in the corporate sense, we might be [00:13:00] saying how many users have clicked on a thing? And we also have the diffusion of innovation.

Ashwin: So how these, uh, new ideas and technologies propagate across a psychological system, system, which we call an organization. And then we also have individual acceptance. So that's how do we form the intention at the human computer interface to go and use a thing? And these are all slightly different things.

Deep: So, I, I have a few questions here. So one is, if you're measuring clicks, then the presumption is you've already adopted the technology, at least in a prototypical stage. Like you've already decided to use the thing, deploy the thing, and get it out there. Um, so. Part of me is like curious, like what happens further upstream, like where you even get to the point.

Deep: Um, so maybe like the meta architecture of a new tech adoption. Like first there's like stuff in the ether, people start talking about it, then it sort of winds up, on the desk of a few maybe, [00:14:00] kind of early, , evangelizers or something within the company, and people start getting curious. Then, eventually you get some kind of prototype that gets built.

Deep: Um, and then now I think post prototype stage, we're probably at the starting point of what you're describing with click tracking to figure out how much value is being brought to the company. So yeah, maybe, maybe let's just tackle that a little bit. Like describe that kind of upstream environment that you see and.

Deep: Did I get that flow right or you know, are you seeing something different? Um, and then is there like a big gap from getting the team to do some kind of early kind of prototypical exploratory type of thing? Whether they're building it or they're bringing it in from outta house and in these larger orgs and then getting it sort of adopted and ramped and spread across teams.

Ashwin: So I've got two, two opinions on that. So one, one is the way you should do it and the other is what I'm seeing in the market, which are probably slightly different things. 

Deep: Yeah. 

Ashwin: Uh, so the way you should do it is there should be the identification of business problems, a [00:15:00] technology match. So problem, solution, fit, and then we know we're bringing in a particular tech that's gonna solve a bunch of problems.

Ashwin: That's the strategic view that mature companies are taking that view. We have another thing that I see quite often in the market and that's. Mainly in the non-technical disciplines, of course, but that is where we have the mandate in, in our company to do ai, whatever that means. We have to do ai. So some vendors maybe have shown us some shiny tools and we're gonna pick one, and then we're gonna say, we've done AI and companies are spending money in this, in this kind of manner.

Ashwin: That means, of course, that the adoption cycle is very different because buying a shiny thing and hoping that everybody uses it is not the same as buying a useful thing that everybody wants to use. So we are gonna have very different adoption cycles when, if you take one view or the other, and I see a lot of this in the market, but the strategic view really should [00:16:00] be, uh, this, this kind of problem, solution fit.

Ashwin: And if we think about some of the more broad based AI capabilities, so we might say a company is, is implementing something like copilot or GPT Enterprise or something like that, Google workspace and how they're doing that. That tends to be driven by IT strategy. So we say we need to do something in the AI space.

Ashwin: We don't really know what the use cases are, but we'll make it available to everybody and hopefully we will get everybody to use it. And what I'm generally seeing is that doesn't tend to work. So lots of companies are saying, we have, we have copilot, everybody's using, uh, we've, we've given everybody a license, but nobody's actually using it.

Ashwin: And that's, that's inevitable if you have something that didn't have a use case in the first place, and potentially you haven't done a lot of, uh, either sensitization of the population or you haven't done a lot of education or folks don't really trust it. So there wasn't that [00:17:00] communication of strategy.

Ashwin: So you've got all of these soft factors, these kind of sociological factors, which do prohibit or inhibit, uh, the use of these new technologies. 

Deep: Let, let's take something like copilot and kind of double click on it. What's, what's driving the desire for these enterprises to even bother putting something like copilot in the enterprise?

Deep: Is it efficiency? Is it fear of lawsuits because they recognize that their employees are regardless, gonna go out to open AI and Gemini and start using the free version that doesn't have indemnification clauses in it? Like is it just being inundated with sellers showing up in the orgs and and it's in the zeitgeist and everybody feels like they're supposed to do something?

Deep: Like what's really, pushing it down into, you know, IT directors basically just adopting it enterprise wide because it feels like it's missing. Like, I've always wondered about enterprises and how they, 'cause, you know, most of my background's in startups and smaller companies and, and a much more bottoms up kind of approach.

Deep: And I've always [00:18:00] wondered about this worldview of like a very top down driven approach where, you know. Something as simple, like if you went into a startup and you said like, Hey, we're gonna mandate like every laptop of every developer be set up like X. You've already got a rebellion on your hands like that.

Deep: You know, like people don't join startups because they want somebody telling them what's on their laptop. So, but in big companies that that can happen, you know, that that definitely can happen. So I'm curious, you know, what you're seeing there. 

Ashwin: So I think enterprise IT strategy, uh, big companies and startup IT strategy tends to be very different.

Ashwin: And with the big company it, um. Departments, a lot of their decision making is going to be based not on. Not necessarily on is this the best tech? Is it the most versatile tech? Do we need 4, 5, 6 different models? Do we need orchestration capability? It's gonna be based on can we click a button and buy 20,000 licenses in one go?

Ashwin: And [00:19:00] that's very, very easy for them to 

Deep: go and it's safe. I mean, like, you know. Yeah. Because they, they're usually like, sort of change averse. Generally large orgs are so, and there, and there's, you know, the bigger the company, the larger the fleet of lawyers is, that's like concerned about all kinds of things that small companies could care less about.

Ashwin: So I, I half agree with that because the whole, the, the, the consideration of is it safe? I, is it secure? Is there appropriate governance? Yeah. Uh, around this. I think with many of the tech, uh, many of the technologies that are out there, I think it can be safe to the same level, but the effort is different when you already have a hyperscaler environment and you already have all of that security in place.

Ashwin: Even if you were to go with something else, you could make it safe. You could go through the governance, you could maybe take six months of due diligence around data workflows and, and processes and all of this stuff. And the safety would be the same level. It would be the same level of secure. It would be the same, same amount of governance, same, same [00:20:00] risk, uh, assurance.

Ashwin: But the effort is huge in comparison to when it's already in your hyperscaler. So I, I think the, the, the safety angle is, what do you 

Deep: mean by that? The unpack that for me, like the hyperscale or like what are, what like can maybe, maybe, can you use a specific example or something? 

Ashwin: So, uh, you are in a Microsoft environment, you are in a Google environment.

Deep: Yep. 

Ashwin: You already have all of the infrastructure that is required to make it. Make it secure for your, uh, for your company. You decide, for example, to go with open ai. 

Deep: Yeah. 

Ashwin: Uh, now if you go with open ai, you've got a new entity, you don't know that much about it. So 

Deep: yeah, 

Ashwin: it could be the same level of security as Microsoft or Google.

Ashwin: Yeah. But you would need to do three, six months of work potentially in a big company to satisfy yourself that it has everything in in place. So even if the levels of security are the same, the effort required to satisfy yourself with that is different. 

Deep: Yeah, no, that makes a lot of sense. Right? Because if you look at the success of Microsoft's Azure Open AI [00:21:00] instances, you know, like any companies that sort of, we in our consulting business deal with over a certain size, they generally, you know, like they started off in Microsoft Trust land and uh, you know, those models are delayed in getting inside that environment, but they.

Deep: There's now a track record of them making it in and, 

Ashwin: and 

Deep: being usable. And then you get the benefit of that sort of long arc of trust within the Microsoft ecosystem. That's, I don't know, 40, 40 plus years in, in the, in the making. A little less so with Google, but, um, that makes sense. So, okay. So what we've talked about so far are the uses of AI for , maybe employee efficiency improvements, I guess is the large bucket that we're under.

Deep: It's like stuff targeted at large enterprises internal employee uses. There's a whole other bucket, which is, you know, these enterprises actually make stuff that goes out and gets used by the outside world. Um, do you focus mostly on the internal kind of enterprise adoption [00:22:00] for internal users or. 

Ashwin: So I'm, I'm focusing on a couple of different things.

Ashwin: So the internal adoption of AI is, is not really a focal point of mine because I think if, if companies are taking the, the approach or they've decided on strategy where what they actually want to do is buy software as a service and have their human workforce unchanged in structure, clicking through software as a service, that's, you know, of course I can get into that, but that's not really the focus of my business.

Ashwin: Yeah. The focus of my business is helping clients to understand how the organization might change as a result of the new capabilities. So we now have capabilities, the, in the agent space, in the workflow space, in the orchestration space, in the auto automation space where, which. You know, process automation is not a new thing.

Ashwin: It's been around for a long time. But with that respect, we now [00:23:00] have a lot of appetite to start re revisiting processes, revisiting how things are done with a lens of efficiency and having, uh, you know, for want of a better example, having a media designer who used to draw things on, uh, with, with Photoshop or, you know, the equivalent, uh, Adobe software, who now uses an AI tool to draw things on a laptop.

Ashwin: The efficiencies aren't that great. You've still got person working with a machine, making, making some images, or, you know, achieving a certain level of output. That's a very different level of output too. If you have automated processes driven by either ai, um, capabilities, reasoning capabilities, image generation, video generation, or you go fully agentic, you see, you move away from the orchestration stuff.

Ashwin: Those efficiencies are huge. The comparison is, is night and day. And that means you have to reconsider many things. So one is what is your human [00:24:00] workforce structure? What does it look like? What are the skills that you actually need? 'cause you probably don't need people to go and click buttons on a laptop anymore to make these things.

Ashwin: So what do you need? How do you, um, conceptualize that differently? What should the value chain of your business be? How can it be augmented by some of the tech that we have? And then what actions do you need to take to go and deploy these things in a safe, secure, and effective way? These are the kinds of things I would help clients with.

Deep: So let me, let me sort of unpack that a little bit. And, and it seems like there's two parts. So let me start with the first part, which I think we probably want to double click on the second part. But the first part is you're saying like, human assistance is, um, you get the efficiency of one human interacting with their laptop.

Deep: Doing whatever they do, but more efficiently because, you know, they can code a little more efficiently. Stuff's being suggested to them. Things are maybe automatic, but the general workflow of one [00:25:00] human, one entity, maybe you can increase their output by, I don't know, 20, 40, 50, a hundred, 200%. But that's like a different bucket for you than this other case where you're talking about replacing entire workflows in an automated scene.

Deep: Something like that is, did I get that part right? Yeah, let's, absolutely. So let's, okay, so let's talk about the, the ladder. Like what are these entire workflows like, can you give us like some examples of entire workflows that where, where this kind of agentic buzzwords being thrown around? Like, you know, I'm, I'm kind of an old school machine learning AI guy and to me, like a lot of this stuff is just people sort of.

Deep: Co-mingle fantasy with reality. And, um, there's a clear misunderstanding of where they're losing control once they start going fully, uh, you know, with all of these like agentic systems. So I'm, I'm curious, like what are sort of the maybe no-brainer use cases that you're seeing where like a [00:26:00] more sophisticated agentic like architecture makes sense and is being adopted?

Ashwin: Yeah, so I'll, I'll keep going with the, the media example. So for me, you have, in most companies, you've got a learning department, you've got a marketing department. They're basically the same thing. They produce some content, they post it onto platforms. Now both of those are, um, whether they're useful or not is not really the point of this discussion, but the mechanisms of what they do, they have processes in place, and those processes tend to be, and I'll break it down in a process view rather than, um.

Ashwin: Uh, as a, as a learning department and a process view, uh, what they tend to do is you have input data, whatever that is. It could be about a product, it could be about a process. That input data usually is summarized or explained in some kind of natural language. Then, uh, from that, it's structured in some way from that structure, there is media applied to it.

Ashwin: So it [00:27:00] could be imagery, it could be, uh, videos. And then those things are compiled, uh, or aggregated in some way into a finished product, which is another media file. It could be a PDF, it could be, uh, an HTML. Could be, could be anything. And then they are, those things are posted to either platforms or, um, either internal platforms or public platforms.

Ashwin: So that's, that's an example of a process that usually requires if you, if you take, uh, an e-learning. Or an internal learning department to make a, make a piece of e-learning used to be about 12 weeks from 

Deep: marketing department. Maybe take an example of the example, like give, give, give you something super, super crisp.

Deep: Like, you know, somebody's taken all of the internal documentation and content for a company on, I don't know, hiring practices or like pick, pick the example and then like walk us through like, what is this other thing, this other asset that's gonna be created. [00:28:00] 

Ashwin: Okay. So you make ev I think every company out there has got, uh, some kind of cybersecurity training.

Ashwin: Right. Okay. So e-learning on your 

Deep: Yeah. 

Ashwin: A you click, click, click and you're done. Uh, the process of creating that is creating a text copy, creating images, compiling that into a file structure that goes onto a platform. 

Deep: Yep. 

Ashwin: If, if we take, uh, I'll give you a slightly different example. If we take, uh, something that both of both of us do and we're both going to do it is we are recording a podcast right now.

Ashwin: That is going to be edited into reels, and we're going to somehow post those onto social media, Instagram, LinkedIn, whatever it happens to be. That's also a process that many marketing departments do. 

Deep: Yep. 

Ashwin: So we have two processes that are effectively the same that some variants of, of course. But those are two examples.

Ashwin: Now, both of those can be automated to some degree, whether that's a Gentech or not, I'm not gonna talk about that just yet. Okay. But they can be automated to some [00:29:00] degree. We can also add AI into those processes, so we can make them AI workflows. IE um, if you take this podcast as an example, you could just load a video into a workflow and you could apply, uh, Gemini or maybe GPT to this and say, well watch the video.

Ashwin: Produce a transcript, produce a caption, produce a LinkedIn post, produce an Instagram post or whatever. And then you link to your API on the other side. Then you and tie it all those 

Deep: APIs and, and basically like, bam. 

Ashwin: Yeah, so that's, that's a process that has some AI in it. It's not a gensec, but it has some AI in it, and it is achieving what a marketing agency would probably charge, um, I don't know, a couple of thousand dollars a month for your company to go and do.

Ashwin: So now we've saved money, we've saved effort, we've automated the process. It's standardized every time, and all we have to really do is drop our input videos into a folder. So that's an example of a process. The learning department piece is pretty much the same process. So we [00:30:00] drop, uh, documents into a folder.

Ashwin: We take it through the process where previously a, somebody would manually read through that and try and summarize it and try and write it in some kind of language that people would understand. Some other people will go and design some media, make some pictures, lay layout in, uh, on, on the screen. Maybe there'll be a video that was made, and then eventually they would package that up into a file that, oh, go onto a learning management system where you could click through your cybersecurity course.

Ashwin: Again, with that, we've saved time, effort, and money by making all of that into an automated process. We add some AI for summarization. We, uh, maybe need to generate images. Maybe we need to generate videos and then API or potentially yeah, a document upload. Yeah. 

Deep: That's a great example. But I'm gonna push back for a moment on your earlier.

Ashwin: Mm-hmm. 

Deep: Um, your earlier kind of characterization as of this as like sort of radically time saving and [00:31:00] the, and the one thing that I think you're sort of presuming here in this conversation, and correct me if I'm wrong, is that there's sort of an acceptance of. The quality of the output of these, it's like in this particular example, right?

Deep: Like, let's, let's take this video. Mm-hmm. Right? We take this video when we're all done, you know, I'm gonna wind up with like a, you know, a 70, 80 minute, you know, MP four. I can take that MP four, I can shove it into, you know, a tool like pictory. Um, it'll get chopped up and there'll be a bunch of candidates to like, you know, turn it into things, assets.

Deep: We could totally automate that pipeline. But what I'm seeing, and correct me if you're also seeing this or not, or maybe this is a nuance of the example that we happen to pick. Is that people want control still, and they actually, and that, and, and I think that's reflected in the fact that these tools are migrating towards human assist, where certainly they save ti, you know, they, [00:32:00] they have the potential to save a lot of time and they can like reduce it.

Deep: But there's still this desire to have that, you know, that ad expert or that designer to be in there saying if that 32nd video that got generated is good. And then backing up on the core assets, the title. Should be X or Y and there's optionality. And so I'm seeing tooling generated towards option generation and ease of editing.

Deep: So call it a migration away from direct creation to maybe advanced curatorial and editorial capabilities, but not, I mean, I, I agree it's the holy grail for a lot of this stuff to get it fully automated, but I'm not seeing us really anywhere close to that on, at least on those couple of examples. And now, would you agree with that or would you totally reject my assessment?

Ashwin: Uh, so for me, I'll talk about myself, right? Marketing. Yeah. Uh, my, my reels, uh, they are, I guess the, it's a two [00:33:00] step process. So I don't use the same platforms as you, but we do have, uh, so I use Riverside for mine. 

Deep: Okay. 

Ashwin: And in, in Riverside, uh, when we are done recording, the reels are generated automatically with a score.

Ashwin: Not, not not selling Riverside, just to be clear, right? 

Deep: Yeah. 

Ashwin: Yeah. Um, so that's a step. Now all I, all I need to do in that is just download 10 reels and I'm done and put them into a, a drive, a folder. Yep. Once they go into the folder, they automatically get captioned up, posted on a schedule. So my entire involvement in that is recording the thing and downloading the reels.

Ashwin: So it's not fully automated, but it's touch of a button. 

Deep: So yeah, 

Ashwin: the amount of time that would take me otherwise is it, it's night and day. It really doesn't compare. 

Deep: Yeah. I, [00:34:00] I, I totally get that. You would save that time. But let's contrast that to my process, which is less efficient for sure. You're correct.

Deep: I, uh, have an intern. Um, that takes this video and is tasked not only with generating a splice, but taking that gener and, and they're using tooling, um, to go off and like facilitate the construction of these, but they're actually tasked with creating clever and interesting ones that, um, that go beyond what we're getting directly from our AI tooling.

Deep: And then they're also tasked with optimizing the social media, optimizing the kind of playbooks. It's much more efficient than it was a couple years ago, but much less efficient than what you're describing, which is, and I, I would say the difference is the comfort level with the output of the auto it made it, or in our case, pseudo automated process.

Deep: And so there's [00:35:00] like, hmm. There's like, people have different comfort levels with what they're willing to like, let an AI system do, um, on its own, right? Like, you know, for me personally, I wanna make sure that every single thing that goes out that has my name anywhere near it is like, is like vetted. But that's also not the case for, you know, for other folks who maybe don't have that control issue or maybe see the cost of an error as like lower for whatever reason.

Deep: And I would've probably say that there, you know, there's an, even, like, I'm still on the real, pretty liberal side of that fence. I think if you go to a law firm, you know, you would see it like much higher and more aggressive in terms of like the, like the, the, the, the validation steps. So how, how does that translate into your world when you sort of like educate your clients on the trade-offs between control?

Deep: And automation and how, how, how does that [00:36:00] manifest itself, you know, in these larger orgs? 

Ashwin: So with every process we've, we talked about two processes, and we've touched on this idea of human in or on the loop, and it has to match with the risk appetites of the business. So every, as you rightly said, every, everybody's got their own view on what, how, what they want the output to look like and how much control they want.

Ashwin: If you really want to have somebody, let's say with, let's go back to the training example. We're making e-learning. We're gonna make a digital thing that someone's gonna click through. You can, in a big company, which is highly regulated, you can have your lawyers go through every single line. If that's what your company needs to do, you can.

Ashwin: Stop your workflow at any point and you can say, okay, we've looked at the input documents, we've made the summary, we've made the copy, and it's just text. And that's gonna go to the lawyers and it's gonna come back [00:37:00] into the workflow before anybody does anything in terms of media generation. So these are things that are going to be client specific and it's gonna depend very much on risk, appetite, governance structures, et cetera, et cetera.

Ashwin: So yeah, you're right. It's not the same for every single client, but it depends on the use case. And it may be that, uh, we're looking at saving a small amount of time. We're looking at saving a massive amount of time. Depends on the process, depends on the department. Uh, it depends on reliability and depends on risk appetite.

Ashwin: So these things are all in the mix. 

Deep: So maybe like, what are you seeing out there in terms of the differences and what are you seeing in terms of the evolutionary timeline? So from two years ago or three years ago to today and projecting into the future. So, for example, I have a hypothesis that, um, that, that A, the models are getting better, uh, b for any given use case.

Deep: Um, if you do [00:38:00] nothing other than like throw out your system and then like attach yourself to the evolution of one of the foundation models, it's probably gonna get better to some extent. I, I, you know, there's differences. I personally think that we're probably, you know, starting to go down on the, on the rate of, uh, improvement on LLMs, uh, at least on this LLM based only architecture.

Deep: But, um, but I would sort of guess that in a conservative org you start off conservative, you start off with like a relatively less. Um, concerning use case, uh, you probably start off with the human in the loop, um, and you sort of monitor things with the human in the loop and then as the human start to just like say, yep, yep, yep, yep.

Deep: You know, like if you start seeing like 99 plus percent adoption, then all of a sudden you start changing your, you know, your, your, your tolerance and you let more and more go out automatically without the human in the loop. Um, are you seeing that kind of evolution? Are you seeing an [00:39:00] acceleration, an acceptance of stuff that maybe a year ago was considered pretty, you know, risk and you wanted humans in the loop constantly and there was a lot of auditing pressure to make sure that we're statistically measuring the outputs and now people are just letting things fly.

Deep: Um, or are you seeing kind of a relative, sort of static level of conservatism towards that, you know, if the org was like, I don't know, a six out of a 10 on conservatism before they still are, and that if they were a maybe a six, seven, or a five for a range of use cases, they still are? Or are you seeing that's kind of like we're letting more stuff fly 'cause we're sort of gaining trust in this stuff overall, 

Ashwin: so we are letting more stuff fly, but that's, I don't know if it's necessarily trust.

Ashwin: I think it's, uh, also, um, we do have a slight change in. Perception of certain things. So I posted this on LinkedIn maybe in the middle of last year that I noticed people had stopped asking to record calls in the uk. [00:40:00] It's certainly you would ask consent before you 

Deep: Oh, yeah. Here too. We did until six months ago.

Deep: Yeah. 

Ashwin: Yeah. Everybody just stopped and 

Deep: Yeah. And we also asked a lot before some bots started like jumping into meetings and now it's like, whatever, there's like four bots and two people in the, in the meetings. 

Ashwin: Yeah. Yeah. So these things are a reflection on how our appetite as a, as a society is changing in these, the, these respects.

Ashwin: And so that's that one example. I think another, another example is, uh, nearly everybody I talk to is, is is doing something with agents at the moment. And it's basically not true, right? Most, most folks are not. So they might be looking at automation, they might be looking at co-pilot, but they're not really having autonomous action in things and.

Ashwin: I think the appetite isn't there yet, coupled with a lack of information or a lack of educational awareness around what Ag agentic systems can actually do. And we had a little [00:41:00] bit of this with things like Comet and Atlas, maybe middle of last year, where you had the first taste of being able to take autonomous action in a browser.

Ashwin: And I don't think that really got adopted in any way. No. So I think, I think many organizations are still in the space of, we were conservative, we're sticking to being conservative, but we have pockets of innovation. Who are, uh, doing things that allow us all to talk as if we're doing ag agentic work. And on the other side you have some companies that are in it depends on the vertical, I think.

Ashwin: And it depends on the competitive landscape. You have some companies that are very clearly saying, we must compete in this space. If we don't do something that's radically innovative, then we won't exist in a handful of years. And again, depends on the company, depends on, you know, their appetite, their competitive landscape, what they want to do, and whether or not there is competitive pressure in their, in their landscape.

Ashwin: Uh, but all of these things [00:42:00] together, I think there are just two folks in the road, companies that have always been conservative, have to, a company that's always been conservative is probably conservative for a reason. 

Deep: Yeah. And I think you could probably, you can probably apply the length of leash that companies give employees at any given level on.

Deep: That length of leash for employees probably translates into the length of leash for AI systems, you know, on some level, right? Like if you're willing to let a junior, you know, someone straight at a, you know, university pops into a company and you give them access to your entire company's coase, you know, pretty much something Google does.

Deep: Then you get that, you guys, that's a long leash, but that's not the same leash that you know, that you're gonna get at a, at a much more conservative company. You know, like, and, and there's certainly, so, so I think that sort of breadth is there. I wanna change, direction slightly and talk [00:43:00] about the ethical sort of reality of what you're seeing out there.

Deep: So what are some of the kind of key ethics questions that, that you're seeing being asked by. You know, execs and what are the key ethics questions that you're seeing that should be asked that aren't being asked? 

Ashwin: So I think the big question generally when I, when I go into, uh, any kind of presentation workshop, client engagement is.

Ashwin: The impact on people, the impact on jobs, the impact on employment, the impact on things like displacement, redeployment skills, uh, all all of these things are basically the impact on people, on employees. And that's, that's usually a big concern because for, if you're an employee, you wanna know what, you have a job.

Ashwin: And if you are a leader, you want to know that you are going to have either to think about change management or communication, or you're going to have to think about, uh, the [00:44:00] structure, the futures targets, operating model of your business. So there are these, 

Deep: AKA managers want to cut, employees wanna stick around, basically like in normal.

Deep: I mean, they want efficiency. 

Ashwin: Say that. Yeah. 

Deep: Yeah. But yeah, I mean, 

Ashwin: yeah. So, so this is the balance. So the impact on people is the balance and the thing that I. I'm rarely seeing, uh, is I have this perspective that we have. You, you probably know this, we have in statistics, we have normal distribution of capabilities in a population.

Deep: Sure. 

Ashwin: And you have top performers and you have low performers, and you have a mean, and you have everybody. You kind of coalescing around the mean and standard deviations from that. And all of the things that we used to say two years ago around, uh, you know, people will be great at critical thinking and problem solving and judgments and, you know, all of this stuff, the people frankly aren't that good at.

Ashwin: We have a distribution of capabilities and there comes a point where the mean, [00:45:00] whatever we're talking about, whether it's problem solving or critical thinking or maths or reasoning or something, the mean for human beings is going to be at some point less than the mean of most of the frontier models that are out there.

Ashwin: And I think we've already reached that in certain categories of capability, which means we have a dilemma. And that dilemma is that if we, uh, we're gonna have a lot of displaced people who potentially don't have the skills that they need to do something. So they're going to go and do something else, and they're gonna go do something else badly because they don't have the skills or the capabilities to do that.

Ashwin: So we get gonna get a flood in the market. Now, one of the, I, I'll give a tangible example. Uh, coaching, mentoring, these kinds of soft to high touch, uh, soft, uh, areas of, of, uh, people management. 

Deep: Mm-hmm. 

Ashwin: We have a lot of folks going into those fields because there is a [00:46:00] tech revolution happening. They are being displaced because they don't have tech skills or they don't have associated, you know, um, uh, adjacent skills and.

Ashwin: They don't really have coaching skills, but because there is no barrier to entry there, everybody can go and be a coach. So now we have a flood in the market of effectively people skills, which means we have a degradation of the support that people get over time because now you have a flood in the market of mediocrity.

Ashwin: That I think is one of the, it's, it's both a tangible and foreseeable outcome and it's also an ethical dilemma, and it's something that I don't see talked about a lot. When you displace people, where are they going to go that is meaningful? Not where are they going to go? 

Deep: You mean at a societal level?

Deep: Maybe not at a individual company level, or do you mean at a company level like that? Execs should be thinking about where their people are gonna land. 

Ashwin: Execs should be thinking about this. [00:47:00] It is a societal problem as well, but execs should be thinking about, redeployment of people should be meaningful.

Ashwin: Not only because it's easy, so we're gonna redeploy all these people into this, but why be no reason? Just because it's easy and it's there. That causes a degradation of quality. So I think that's an issue that, uh, organizations and society needs to think about. 

Deep: I mean, let's, let's frame this a slightly different way.

Deep: So very, there exists a set of skills that companies need, and if these machines are sucking up, uh, and the, these agentic systems, these advanced AI systems are, are like leading to increased layoffs. Which personally, I have a question mark there. I'm not as convinced this is happening yet for the reasons that we think it is.

Deep: I think, um, we, you know, talk, we'll find out when the recession's over, like whether the layoffs have to do with the recession, whether they have to do with the fact that they're just making room to go. Pump cash into data centers. [00:48:00] Like, I don't actually know, but it's pretty normal cycle every, you know, we've gone a long ways without a major recession in the tech industry.

Deep: So, and it's a convenient thing to blame it on. So I, you know, I don't know, like I, I'm, I'm totally open to hearing data and supporting, you know, one argument over the other. But regardless, usually in the economy, in a, in a normal scenario, there's a, like if you take a, a bar chart of skills, like a, you know, all the skills that all your employees have, and you maybe just count how many employees have them, you'll see certain areas where they're accelerating up and you need more and more of them.

Deep: So, like, you know, eight years ago we needed a lot more data scientists. You know, um, and then you'll see areas where there's deceleration. You know, I don't know, a hundred years ago we saw a lot of deceleration around carriage, horse, drivers. What, what are we seeing? Acceleration. Like what would an exec be seeing accelerations in, like what skills are actually, could they even have the option of rerouting people into, um, or, and what's the [00:49:00] net?

Deep: What does that net look like? Like the net drops relative to the net increases. Do you have any sense of that or, you know, because I, I'm not, I feel like a lot of the concern is in the abstract, but I don't, I'm not seeing the manifestation of the terminations directly related to AI in, but I could be wrong, you know?

Ashwin: So you're seeing them directly related to something else? 'cause they're happening right. 

Deep: I think they're happening, but I, I mean like how does one tease that out from just, you know, overall we can't sell as much stuff, so we're chopping versus we're selling more, you know, we're selling more stuff, so we're chopping because we can do it with less.

Ashwin: Yeah. I think doing it with less is probably a key thing. And I think, uh, there is some noise in, in the market around, [00:50:00] uh, there are layoffs, but the layoffs could be due to offshoring, they could be due to, uh, reduction in revenue. They could be due to, uh, re you know, pivoting, refocusing in organization.

Ashwin: They could potentially be due to ai. But I don't think, I think right now there hasn't been enough actionable production scale or production grade. Um. Uh, meaningful AI projects that have caused, 

Deep: yeah, I mean, we have that, 

Ashwin: all these 

Deep: m mi t paper that came out a few months ago. It's like 95% of these projects fail.

Deep: I mean, as somebody who tries to get these projects to succeed and sees a lot of 'em fail, like I would agree with that. I mean, it's really hard to get people to do the right stuff to make these projects succeed. It's not as easy as it seems. It's not like, just like you, you pointed out, you don't just plop in copilot and everybody's using it and like stuff goes up.

Deep: I mean, people get distracted. They rat hole, they don't know how to, it's not even really the same skill. Exactly. Like somebody who's really good at writing well encapsulated, you know, algorithmic optimization [00:51:00] is not necessarily the same person that's willing to like spin up 38 processes to sit around and munch stuff like that's so 

Ashwin: yeah, I mean on the MIT paper, I think there was a bit of a reach.

Ashwin: Um, I think from my perspective, many companies have implemented things I'm not, I'm not particularly picking on Microsoft, but many companies have done things like implement copilot for no reason and then been surprised that the thing that had no reason behind it and had no metrics behind it and had no success criteria behind it.

Ashwin: Somehow they couldn't prove that it, it, it worked. 

Deep: Yeah. 

Ashwin: Right. Shocked. I'm shocked. So I think, I think there are holes in the paper like that. Plus the method methodology of the paper was, you know, we're gonna ask a bunch of CEOs how many of their, um, how many of their workers are using AI without asking the workers.

Ashwin: Uh, you know, it's, it's shoddy data at best. But, uh, that's, that's just the, the adoption side of things. Um, [00:52:00] I think the big skills that companies are gonna need, if they're gonna be thinking about this stuff properly is they're going to need data skills. Broad bucket, but everything that we talked about today is, uh, you know, we need to think about vectorization of data.

Ashwin: We need to think about rag in, its, you know, broader sense. We need to think about data cleaning, data management, uh, tacit versus explicit, uh, data capture. These are not necessarily skills that organizations currently have, and if they do, they don't, maybe they don't have them in the volume that they need.

Ashwin: We also need to think about the tech. So, you know, we've talked a little bit about, uh, orchestration. We've talked a little bit about, uh, things like agents. This is new tech that's gonna need new skills. Uh, orchestration side isn't new, but we, we need people who can think in processes, think in systems, partition things up.

Ashwin: If you have your processes and you're going to start to make workflows out of them, what's the most efficient way to do it? How do you ensure communication between departments? How do you ensure [00:53:00] data workflows? Again, that's probably not a skillset that is native to some of the non-technical disciplines, hr, learning, marketing, and all of these things.

Ashwin: So redeploying directly, people who used to be employed for content generation and saying, you're going to now be data scientists. You're going to now be orchestrators. It's not a direct lift and shift. So there is a gap there, and that gap needs to somehow be addressed. 

Deep: Well, this has been a, a super interesting conversation.

Deep: I. Maybe let's, let's project out like five or 10 years and we can leave the realm of what we know to be true and like enter the realm of what our instincts tell us, you know, or get us to suspect is gonna happen. But, you know, like five or five or 10 years out, like what do we, what do you really think is happening?

Deep: Like, do you really think that tasks are being eliminated, jobs will be rejiggered and we're gonna have a net loss of a ton of [00:54:00] jobs in the, in the, you know, white collar arena in the very near future? Or are you more optimistic that, you know, we're gonna, there's no shortage of things to do. We're gonna create a whole bunch more projects and maybe there's a migration towards innovation and towards like the data side, like you're saying, but, but overall, there's not, like, it's business as pretty much usual as the last 70 years and there'll be plenty of work for folks.

Deep: Like how, how, where do you sit on that spectrum? 

Ashwin: So I, I think that it depends on timing. So I think in the short term, we are going to see job losses, whether they are. Justified or not is a, you know, separate question. But I think we'll see that. Uh, and then over time we will probably see that some of the tech promise that we have, uh, you know, on our lips, on our fingertips, some of that promise will be realized.

Ashwin: Some of it won't be. And we'll need a refactoring of what white collar work looks like, probably with a human premium in there. And then over time we'll start to see [00:55:00] more jobs with, you know, different descriptions, different skills requirements, uh, being made available and being populated, uh, by people.

Ashwin: But I think there's gonna be a dip. And the size of that dip isn't, is not something I'm, I'm certain about because that dip itself is gonna cause societal problems. You've mentioned recession. I think if lots and lots of people are out of work for a significant period of time, that's very problematic for society.

Ashwin: But I think there will be this kind of, uh, dip effect before we start to see more jobs. Five, 10 years time, uh, you know, 10 years time you might see in that positive, but, but 

Deep: why do you we'll see more jobs after the dip? Like what, why? I mean, like the reasoning models. Are getting more and more powerful.

Deep: They're already incredibly powerful, you know? Mm-hmm. Like you, you can deploy an HR agent who also happens to, you know, be a PhD level, um, physicist, you know, like, at least with respect to Yeah. Existing knowledge. So why, and, and their reasoning capabilities is probably much higher than the majority of [00:56:00] people in hr.

Deep: But, you know, is it because you see a net shift towards human-centric stuff? Nobody wants to, you know, learn flamenco guitar from a robot. It just seems bizarre. Like there's certain things, you know, no one wants to get their drink given to them by a robotic bartender. I mean, it's just dumb. Um, so is that why, like, do you think like the artisanal economy opens up?

Deep: Like what is it that's, you know, that that's gonna lead to that increase? 

Ashwin: Yeah. So I, I think a lot of the things that we know today as work are going to be done by machines. You're exactly right. That, you know, the, the reasoning capability of models and the ability of the tech to do the things that we do now is, it far surpasses, surpasses what most people can do, but the human premium, uh, coupled with human tenacity, right?

Ashwin: I'm not usually the one who believes in, in, in the capability of humans to go off and do, you know, uh, the great things. But, you know, frankly, we usually persevere as a species. So I think there will be, uh, [00:57:00] this kind of human-centric economy that you're talking about, I think the price point's gonna be way, way lower.

Ashwin: So when I said there will be jobs, I'm not necessarily saying, now you go out and you are in a hundred k doing whatever it is, you'll go and find a different job making, I don't know, clay pots and, you know, weaving. Yeah. 

Deep: It'll be like 

Ashwin: a Star Trek 

Deep: episode where we're all, you know, 

Ashwin: yeah. 

Deep: Making 

Ashwin: our, you're not gonna find.

Ashwin: You're not gonna find that job at that price point. There's, it's, it's gonna be a very different landscape of what we can and can't afford. But jobs themselves, governments, generally speaking, are incentivized by having the figures to say that we have this many people at work. So whatever the definition of that is will be somehow, uh, able to justify that.

Ashwin: Lots of people have jobs, whether they are high paying jobs or not. I think most of them are gonna be passing coffee back and forth to each other over. Um, 

Deep: oh, so you do think we're, that we're taking a big step down as far as our role, [00:58:00] I'll cut away be 

Ashwin: jobs, right? 

Deep: Yeah, I mean, like, I think it's a really hard question to answer because, you know, part of me sort of.

Deep: Thinks that, I think that the LLMs in particular, like I feel like we've, we're, we're, we're approaching diminishing returns on what we can squeeze out of, out of this future sequence prediction approach. And I think we need a big breakthrough. Something, you know, along the lines of, um, you know, of, of objective based ai.

Deep: So it's not just predicting feature sequences of text, but predicting like things that actually, you know, achieve objectives. And I think, um, but even with respect to just lms, I think there's easily 20 years worth of run here. I mean, on stuff we can do and, and, and mangle and, and improve, I. I think it's, it's kind of like I'm already seeing, I think that, you know, [00:59:00] Zuckerberg and some of these guys claims that they don't need junior software engineers anymore.

Deep: I think those claims are more hyperbolic and maybe aspirational for somebody as dysfunctional as Zuckerberg. But I think, um, I think that that's, there's like a limit to how many Homer Simpsons we can build that manage the nuclear plant. Like at some point we gotta take our fresh, young, bright minds and reenter them into the workforce.

Deep: And maybe companies have to pay a five or even 10 or maybe even 15 year premium of just getting them up to speed to the point where they can be at the end of that bell curve distribution to keep everything running and making sense. Um, that's on the one end of the spectrum, I think practically speaking, we're, we're gonna every place I most, not every place, most of the places I've been that are.

Deep: That have the attributes of healthy growing companies, I have not seen the problem be lack of ability to [01:00:00] identify new stuff to work on. I've seen it be in abil, like the bottleneck is inability to execute as fast as they need to, to kind of like make it work. So I think that we'll probably keep coming up with more and more stuff.

Deep: Now that said, I think that if you're, you know, at that median and below in from, from whatever skill stack, I think you're very vulnerable. So I think there will probably be a lot more increased pressure on governments. To like up the size of the nonprofit sector and, and that's how you quote, make jobs, you know, on some level.

Deep: So I think capitalism, this kind of raw unbridled capitalism that we've seen in, at least in the US in the last 60, 70 years, I think that pressure point will change with the 50% employment rate or a 40% unemployment rate. And it'll start to go back to the FDR pressures of like, you know, we gotta just get people jobs and, you know, and you can do a lot more with somebody close to the median of that, you know, reasoning dis, you know, [01:01:00] um, on the bell curve position versus somebody at the far left end.

Deep: But even at the far left end, you can have 'em pick up garbage or do something. And up here you could probably come up with stuff so. And also I, I sense that we'll probably start shifting a lot more researches towards open frontier spaces. You know, space, literally, um, um, like bioinformatics, you know, getting people to live a long time, all that kind of stuff.

Deep: Those are places where you can throw an insane amount of government budget, throw people at the problem, throw a lot of AI at the problem, and probably keep society generally satiated. So your 22-year-old males aren't blowing, you know, firebombing buildings and stuff. So I think, um, I think it'll work out, but I tend to generally agree with you.

Deep: Like, I think there will be a dip. I think there's probably a dip happening right now. I think some percentage of it probably has to do with at least the perception of AI solving problems. But I [01:02:00] think how many problems we actually solve with LLMs customer service. Yeah, we'll probably get like 80% of it and the remaining 20% will probably drive us all crazy.

Deep: Um, but we've kind of already done that with customer service in the modern world. Like we already moved it all to the Philippines and like out of country. Um, so I think those places will hurt a lot. Like in the Philippines, those places where, you know, people are, are taking on these kind of call them lower stack like white collar kind of positions.

Deep: So on the whole, I don't know if I would say I'm optimistic, but I'm sort of hopeful that, um, that our societal immune system will like kick in. We'll address the obvious ethical dilemmas over time, there will be very significant dysfunction that comes out of ai. Like I, I shudder to see what the next generation of kids raised by, you know, AI stuffed animals and have [01:03:00] the inability to interact with humans is gonna look like.

Deep: I mean, gen Z's already like, you know, we've seen so many problems with like, that have come from social media impact, like in, in Gen Z and Alpha. So, I don't know. I mean, I guess I'm really hesitant to say, but in general I'm way less techno optimistic than I was 30 years ago. But I'm not a, yeah, I'm probably with you techno negative anymore.

Deep: Like I, you know, I still think we'll figure it out. But it's unfortunately, it's more fake. I mean, like, I don't know, like when I look back, like the first time I saw the internet, 1993, I'm in a lab at the University of Wisconsin and I was like, oh my God, this thing's gonna solve all our problems in democracy.

Deep: Like, yeah, that did not happen. You know, we've got like Agent Orange wreaking destruction across the planet. Like it's, you know, it's, um, it's, 

Ashwin: I mean, it, it turns out that lack of information wasn't the problem. Right. 

Deep: Certainly not [01:04:00] lack of like, poorly curated information that wasn't, that wasn't the problem.

Deep: Awesome. Anyway, Ashwin it was super awesome, uh, having you on the show. Thanks so much for coming on. 

Ashwin: Thank you for having me. It's been amazing.