"The Cognitive Revolution"
00:44:49 · Nathan Labenz
[00:44:49] S0: they do day to day. Yeah. That makes a lot of sense. So there's not a... I mean, this sort of speaks to I think another major question right now, which is like when trying to get value from AI, should we be creating what I traditionally refer to as smart workflows where you have a sort of step by step process that is human designed and each step may or may not be an AI step, but the AI steps, obviously you can optimize your performance and make sure they're working well. One way I characterize those is like, you'll know you're successful when you don't have to check every single output anymore. And I think for a lot of tasks you can get there today. And then there's like this new paradigm of AI that is the more agentic, like the AI chooses its own adventure, you know, maybe within some constraints, but right, like, Claude code, go write this thing or, you know, some of these calling agents like call this person and try to sell them a subscription or whatever. These sort of like open ended, you don't know how many rounds it's going to be, you don't know what the tool calls it's going to choose at any given time are going to be, don't obviously don't know what the environment is going to give How much of that are you seeing deployed in your companies?
00:46:02 · Marc Bhargava
[00:46:02] S1: So on the services side, I think it's going be more of the first case that you outlined where there are seven or eight steps and some of them will be automated and maybe a AI agent or does step one, but a human does step two. And so it's gonna be these hybrid AI enabled workforces where there are multiple steps and some are fully automated, some are not. But in services, think there's still gonna be a huge portion of lawyers, accountants, to your example, doctors, and instead there'll just be different steps that are automated but there'll be key people and it'll be this hybrid AI enabled workforce. That is where I view the world ending with services for many reasons like one, the technology is not there but maybe even more important, many of these spaces are highly regulated so you can't just have an AI doctor even if it was great. And many of them are very human and emotional and it's, you know, a lot of it is talking through learning medical news versus just what's the diagnostic and the prescription, etcetera. So I think in most services industries, it will be the hybrid approach, and that's the one we're employing. I do agree with you though that in certain industries, it could be the full agent automation approach. Go and here's Airbnb, build me Airbnb but for cars instead of for locations, and a swarm of agents will be able to code it up and you'll have sort of an Airbnb for cars that's built for you. So there will be I think some industries that are maybe fully automated. Certainly the holy grail right now is the automation of coding. And you have obviously OpenAI and Anthropic working on that but also Microsoft and Google and then Cognition and Factory and even I think Cursor and others are moving to that trajectory. And so coding is a great example of obviously a massive industry where maybe building websites and products and platforms will be very close to fully automated using agentic swarms.
00:47:45 · Marc Bhargava
But in the industries I'm focused on in applied AI, especially a lot of the services industries which we discussed is really a large part of the global economy, I think it will be a hybrid approach because of the human element, because of the regulatory element, because of the complexity and sort of grey areas of tasks Writing code or doing a math problem generally doesn't have a kind of got to make a gut decision on something while filing your taxes and how you classify something, what's the risk you're willing to take in an audit, etcetera does. And so we're pretty clear on here are the industries where we think this hybrid approach makes sense and we want to incubate companies, maybe they have a role component, maybe they don't. And then also investing in traditional software and agentic companies that we think could maybe fully automate an industry as well.
00:48:33 · Nathan Labenz
[00:48:33] S0: Yeah. One other question I had that prompted is how much do you build versus buy? Because I could imagine like a lot of, you know, another episode coming up with the CEO from Intercom. I could imagine, like, a lot of these companies might ought to just, like, buy Intercom and, you know, have them do, you know, a a big chunk of stuff. Like, it seems like there's a bit of a tension between we want to focus on automating the sort of auxiliary, you know, somewhat less core tasks from the verticalization or vertical specific approach, it feels like the more, you know, true to the vertical you are, the more you're kind of focused in on the core and the more you're kind of trying to do the stuff that, you know, the the pros don't wanna do, the more you're kind of entering into more of a horizontal, like, intercom type scenario. So maybe there's just enough middle ground there that there's still like plenty left to bake, but how do you think about that spectrum?
00:49:31 · Marc Bhargava
[00:49:31] S1: There's a lot of middle ground because there are many point solutions out there in a lot of these industries, legal, accounting, IT, customer support call centers. There are a lot of great technology companies and a lot of solutions. Generally, they're point solutions. Generally, it's not a full platform play. Generally, they're not customized. And generally, they're not trained on your proprietary data. So there's still a lot left and that's why at the core of these transformations is staffing a great... Backing great founders who are great technologists because there's a lot to do in terms of stitching together different point solutions. There's a lot to do in terms of creating custom solutions. There's a lot to do on fine training models and making it unique to your dataset. And so all of those tasks require a really deep understanding of AI and of engineering. And we generally have seen that kind of folks who are coming from Applied AI positions make really good founders in these sort of companies because they absolutely can assess and buy things off the shelf, but then they're also figuring out which model we should use, where, how to stitch things together, do we wanna customize our own model, take an open source one, train it against our unique data. You have to have a lot of really expertise in AI and in building software and actually doing things. It's not just a research question. And we kind of tail off of all the amazing improvements made on the research side by an Anthropic or an OpenAI or a Google or others, whereas they improve their models and can do more, we flow that through and we get the benefit. For example, with legal, at least, has found that, you know, OpenAI is really great at summarization of case text, for example, but the Anthropic model is better at reasoning and asking difficult questions that you might face in a courtroom. And so having kind of both OpenAI and Anthropic every three to six months improving their models, you know, float both of those through at UDIA to deliver value for the end client, which is the in house general counsel at a Fortune 100 firm.
00:51:24 · Marc Bhargava, Nathan Labenz
And so you have to have this technology and AI expertise to understand the different models, where to use which one, and where there are also point solutions that are already built, you know, you have to have the technical expertise to evaluate them and integrate them as well. And so we still think you need a really amazing kind of technical CEO or CTO or founding team to be at the core of these projects. [00:51:47] S0: How much is about customization versus standardization? Like, if you buy 10 businesses, is the idea that you'll have 10 different fine tuned models that, you know, represent the sort of unique flavor of those different local businesses and, you know, a bunch of different custom workflows that kind of recreate the idiosyncratic ways that they did it? Or it seems like maybe more of the economic promise would be in the standardization where you'd kind of say...