The Jagged Frontier of Enterprise AI: Why Adoption Is Failing

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone. I'm Rob May. I'm the co-founder and CEO at Neurometric.

Lauren Slyman: I'm Lauren Slyman. I'm a user researcher at Microsoft.

Anna Kirk: I'm Anna Kirk, sales lead at Thread AI.

Rob May: And this is the AI in NYC show. The next stop. Welcome back everybody for season two of the AI in NYC show. We have a couple new hosts this season. One is here with us tonight, Lauren Slyman. Lauren, do you want to say anything about why you joined the show or?

Lauren Slyman: Well, I talk about AI in my spare time, 100% in my spare time. And so I figured I would do it on a podcast. But I

Rob May: And you're writing a book on AI.

Lauren Slyman: I am writing a book on AI. Meant to simplify complex concepts and teach what a semiconductor is or how neural networks work. And it's what I think is funny. So yep, that'll be out probably within the next 6 months or so hopefully, but we'll see.

Rob May: And then you have Anna and me who you've always had.

Anna Kirk: The usual suspects. We're back for more. We're back.

Rob May: Do you guys know that song, Back in a New York Groove?

Lauren Slyman: Can you sing some for us, please?

Rob May: I can't because we don't want to pay the royalties, but Ace Frehley, the guitarist for Kiss, wrote a song in the in the late '7s called Back in a New York Groove. It's an okay song.

Anna Kirk: Okay. Well, you really fold it. Yeah. Great. Great start.

Rob May: It's good enough to go listen to one time. Anyway, what were you going to say?

Anna Kirk: What did you do for the month of August when we were on break?

Rob May: Ooh. That's a good question. Not much. Mostly worked and you know worked on neurometric and yeah I've been teaching myself this year GPU kernel programming so yeah anyway what did you do

Anna Kirk: I don't I don't even have a good answer that I don't know why I asked you like I worked

Rob May: and you're kind of a homebody but not always

Anna Kirk: that's not true that that is a misconception about me I just I like to concentrate my fun and like bits and spurts. So, yes, I digress.

Rob May: Well, let's start by talking about an article from CNBC and published many other places that New York last quarter pass San Francisco in terms of total tech jobs. Not necessarily all the same jobs. I think a lot of your deep hardcore internet engineers are still on the West Coast. What do you guys think about that? Anybody surprised or not?

Lauren Slyman: No, I feel like with AI, most of the financial industry has been leading a lot of it and that's how the speed of everything and the agility of how it's been implemented was framed from the beginning. I won't make any commentary on that, but I do think that that has had a huge influence on why there is so much tech and interest in tech all of a sudden in New York City.

Anna Kirk: I am surprised that that happened so quickly. I think I would have expected it to take a couple more waves.

Rob May: I mean, we have this big theory that we talk about all the time on the podcast, which is like New York City is the applied AI capital of the world. A bunch of Fortune 500 headquarters here. It's the center of gravity for a whole bunch of industries.

Anna Kirk: But I think it is also kind of telling when you think about like overall tech jobs, which is the metric that this was focusing on. They're not necessarily all AI. And I feel like people are so fixated on like AI as this bubble that you know you forget that like that is you know 2% of the population that is focusing on kind of like these deep like foundational model development development companies and then there are a lot that are focused on you know okay like technology at a large bank that still qualifies as a tech job.

Lauren Slyman: Yeah. I mean, I think it's an interesting headline and I'll see like once that follows by like applied AI jobs, that's when I'll start to get really interested.

Rob May: Yeah. I mean, you probably have a lot of I've got to think this is the home of AI governance people.

Anna Kirk: For sure. For sure.

Rob May: And you have, you know, when when you look at So, we just had OpenRouter acquisition in August by Stripe. Open router was headquartered in New York. 9 billion, nine and a half billion, something like that. 8 billion, somewhere in that range. But really, really good exit for a three and a half, four-year-old company, New York based. We have Reflection who we're hoping to get on the show in the fall when they have their big model release. So, this should be a American open source model on par with the Frontier Labs. A lot of good stuff going on here.

So, cool. Well, let's move on. We have we have a guest today before we talk about some of our other news. We'll we'll have him on so we can he can comment later. But he wrote a recent book that came out called the weight of intelligence which makes some really really interesting points that I think a lot of people aren't paying attention to on the AI side. He's spent time at Accenture and a bunch of other places working with real customers to apply AI. So he's got a lot of credibility. But welcome to the podcast. Satish Viswanathan. Satish, thanks for joining us.

Satish Viswanathan: Hi Rob. Hi Anna and Slyman. Thank you so much for having me today.

Rob May: Yeah, let's get started. Tell us a little bit. First of all, start with like why did you decide to write this book?

Satish Viswanathan: Well, I've been in the field for a long time. I started my my career in 2000 with data warehousing when data warehousing was not so sexy or data was not so sexy and I think since then you know there have been a lot of these questions lingering around in my mind in terms of how else can we approach this whole enterprise transformation piece we never had the time I've been traveling all around and then at one point I decided enough was enough in the rat race and I had to do something different and that was probably the trigger the questions were all there with me just a matter of you know focusing and writing the right kind of narrative for for the for the business leaders.

Rob May: Nice. So there's this concept that we've talked about on the show before called the jagged frontier of AI and for people that haven't seen a previous show came out of some research done by Karim Lakhani and Ethan Mollick at Harvard and at at at Wharton and it pointed out that the frontier of most technologies is kind of predictable, right? When mobile comes around or cloud comes around or database technology something like you can kind of predict how it's going to be deployed and with AI you can't because it spikes in places where it gets really good at some kind of task and it's surprisingly bad at other tasks and so I found it interesting so you argue that the frontier is jagged around AI in part because organizational absorption is jagged not necessarily because the model capability is and so I'm curious what you think about where value a crew. Is it the labs producing capability? Is it whoever solves the absorption? And if this absorption issue proves out to be true, is can you build products around it or or are we just going to see a lot of services businesses around that part of AI?

Satish Viswanathan: Yeah. So, I think that the central tension in the book at least in the first few chapters, Rob, is about is a market oversolving for model capability and undersolving for enterprise metabolization. Okay. And if you see the investment patterns of the past what almost 3 years since now 2022 I think we've been having more headlines on model capability you know starting from foundation to language models to action models to reasoning models but how much have we heard about you know what does it take for some of these rugged you know raw enterprise environments to pick up these models right and run with them.

So I think there's been obviously an imbalance. I've actually taken Ethan's work on the jack frontier and expanded it at an price level. So what I what I what I talk about in the book is the jack frontier of enterprise AI and it actually is a function of two variables. Okay. The first one is the model capability or AI capability frontier and the second one is the enterprise metabization frontier. So, so the way I visualize this is it's not just about one sort of you know playing a part. There are two different waves wave functions which are actually colliding at different levels across you know the maturity of different use cases and that's what is causing this imbalance between the great work being done at the model producers or by by the companies like open aanthropic and the absolute lack of ROI that's actually emerging on the other side.

Anna Kirk: So that that's actually an interesting segue. I mean you we all know that kind of foundation model companies are kind of pricing and being priced on their like promise towards AGI or artificial general intelligence. And you mentioned this. I'm going to quote part of this as well. It's you say artificial general intelligence asks whether machines can become broadly intelligent. Artificial enterprise intelligence asks whether enterprises can become in intelligently organized around AI. So I guess firstly would be really curious to hear a little bit more about you know how you think about that discrepancy between AGI and AEI and if AEI that you alluded to earlier is actually the thing that is determining enterprise ROI. Is the market currently like mispricing a lot of these AI labs or you know also the application layer all of the above?

Satish Viswanathan: I think Anna the answer lies again in a little bit of both right. So you do have for example you look at the restaurant business as an as an analogy you need you need to have the best chefs and the best dishes out there for people to come to the restaurant. At the same time, just because you made the best dish, doesn't mean that you're going to finish every last piece of that dish, right? Little bit of, you know, you do have need to have the best models, which is what we're calling as the AGI equivalent or or let's say the kind of formula, but at the same time, it has absolutely no guarantee to ROI at enterprise level. So you can look very very sexy in this ideal you know glass house kind of environment where conditions are good for asking different questions and answers coming through but in the real world some of these ideal environments don't exist so we do have to sort of factor for that and I think you know the other way of looking at is also some of these models are are challenging human intelligence at a different level so for example

Well, right now most of the prompting the engine or across all kinds of product families is one question is equal to one answer. How often have we been in board meetings or in let's say any kind of leadership meetings where a question is asked and everybody agrees on the same answer? Never. Right? So as as humans we are wired to think differently because of our experiences, our perspectives and so on. So there's also that fundamental challenge in terms of are we trying to oversimplify thinking and execution because we're getting into this formula of one question one answer pretty much as the basis of AGI. So that's why I think there's a lot of I would say differences in terms of how we should approach AGI which we still have we still need to have the best dishes at the restaurant but that doesn't necessarily mean that you're going to have the highest amount of consumption and no food wastage.

Rob May: You know, we've talked about it a couple times on this show. I am not a fan of the belief that you can be that much smarter than a human. And part of it is based on so there I don't know if you've ever read about this satish, but there's a there's a mathematical theory called Ramsey theory. And it basically says that as data sets get bigger, you start to see more and more spurious correlations. And at some point the data set gets so big that you're more likely to add a spurious correlation than any real information to the data set, right?

And so if you think about intelligence as pulling patterns out of data at some point at some level of scale of data, it doesn't that entirely breaks down. And you could see that being an asmtope and when you think about other things. So like I you know before the theory of relativity people assumed we would go to other galaxies in minutes because we didn't know about the speed of light. We thought you could move faster if you just had a bigger engine. And then Einstein showed us like how relativity work and we're like well damn like okay. So we're we have to think about space travel differently because we're we're gated there. And I think we're going to see that there's some limits to intelligence.

Anna Kirk: Yeah. And when you think about it too, like part of the ability to like detect or like create something like Einstein's theory of relativity is that, you know, he was able to interact with kind of a novel observation from what he saw in the physical world. Like an LLM can't do that. And I am so curious. I know we've talked in the podcast before about this like world models whether or not there is kind of any grounding there to having some other kind of relationship to kind of the physical world that is not purely symbolic or linguistic. But yeah, I think to your point there will always be something that that misses or doesn't capture. And certainly in a in a boardroom of a company you're going to have many different viewpoints that all kind of converge on the same question.

Rob May: Yeah. I mean I think we'll I think we'll make machines as smart as humans. My question is just when when when when these lab leaders talk about something 10, 100, 500 times smarter than a human, like I don't know if that's possible. Like it might possible it might only be possible to be 30% smarter than a human and that's the limit of what intelligence could ever be, you know, just like the speed of light or absolute zero or any other physical constraint. So I don't know.

We'll see. Time will tell. What's next? What do we have? Do so you talk about enterprises being given more possibility than they can realistically metabolize and this is something we were talking about before the show and I want to I want to dive into because so I run a tech startup and it works for one and I think probably one of the things that all of us see in the space is a little bit of paralyzation on the part of buyers who like don't know what to make of how fast the techn is moving. And like you know to change and adopt any new technology requires you to change some things about your company and at some point you're like I can't change every week when a new model comes out and so talk a little bit expand on this and talk about some of the implications that it has for the AI industry.

Satish Viswanathan: I think there is a very u deep problem the way we are driving the AI economy which is you know we are trying to build solutions around a model okay now this this is typically how we've done the whole software industry starting from ERP we sort of taken a product and sort of build services around it or integration around it but as you exactly said these models are getting sunset faster than what we would expect right so if the model gets sunset you know and this This is going back to you know 400 500 years back a part of the world thought that earth was the center of the universe and then we later realized it was the sun. We're actually in this pre- era where the model is presumed to be the center of the solution and unfortunately when the models get sunset or you know new capabilities come through it becomes a scope creep. It becomes a surprise. It becomes a kind of frustration that we put all this effort to test this model and sort of make it stable and then I have to do it all over again. Right?

So I think there are some fundamental issues in how the architectures are getting shaped and that is causing also a lot of the frustrations at this at the board level that they can't necessarily see consistent value. They can see it for one quarter or two quarters and then a new model comes in either from the same product family or from a different one and then CEOs are saying why can't we sort of you know increase the capability and goes to ground zero right this actually is a is a serious problem Rob and I think we cannot have the same mindset we had from 30 years back that we s start with one one starting point and that is supposed to be the idle starting point so this is just one of the many issues that we are facing on the ground and that's why I keep sort of looking at enterprise metabilization as being the price and not necessarily model capability.

Sometimes I believe model capability if it goes you know even to whatever we call as AGI it's a glorified quiz master okay you can ask anything it gives you good answers but I think there's been study at at Stanford as well which says that it's no longer about whether the model can produce outputs or outcomes it's more about can the human verify that okay so there's a little bit of that as well so I think we're shifting to the goalpost a bit to even if I get a great answer how much of that answer can I validate as being accurate And all these things are playing out on the ground and business leaders for both technical and nontechnical are struggling with this AI economy wave.

Anna Kirk: Satish there's so we at you know at thread we talk a lot internally about how you know like there has to be there has to be a separation between kind of like your core infrastructure and your core kind of governance processes and the application of any type of new technology. I guess my my question for you like yes I think there is kind of a the application of these foundation models cannot be co-mingled with actual underlying infrastructure and architecture. Both from a technology perspective but also from a process perspective within enterprises. Do you think this is better addressed? Is this a process problem? Is it a governance problem? Are we just so early on into this kind of wave that we're at this point where we're kind of conflating foundation models with infrastructure?

Satish Viswanathan: Yeah. No, I think I think you're absolutely right. This is we're just running in a formal wave where everybody is wanting to do something. We've not sort of taken a step back and thought it through very well. Okay. And we presume that because we've had successes in the past 30 years, you know, starting from ERP to Y2K to sort of the social and of course SaaS and so on, we presume that we will somehow get it right if we keep doing more. But I think what I sort of when I was thinking about it and also researching in the last few months I realized that more AI activity is not necessarily a good thing if if the enterprise is sort of investing more and more without thinking this through in terms of what exactly are we you know having as a final destination. So one of the pivots I'm sort of positioning is instead of doing more AI activity I think what the leader should first understand or sort of think through is what is going to be the enterprise cognitive architecture or orchestration. Okay.

What I what I mean by that is in the preAI era it's been humans making decisions enabled by some technology which sort of gives data points. But now if algorithms are starting to get into this space the whole decision making apparatus and mindset is shifting. So unless we get that clear especially for you know very mission critical industries it's going to be very difficult to bring value if you're just sort of introducing AI solutions here and there just because somebody else is doing it or that's what the market says.

Lauren Slyman: So when comparing that to what you've seen in the past with pod or ERP adoption, would you say that it's primarily a benchmark issue?

Satish Viswanathan: So in the book actually I look at this you know across four stakeholder groups. There's this producers who who actually building models and releasing models. They're implementers who actually building services around models. You know it could be you know McKinsey, Accenture, Deloitte and so on. And then you have the third group which is consumers or enterprises which are actually the ones receiving value. And the last one is shapers and shapers is academia regulators and benchmarking companies.

So I actually looked at this problem in terms of what is different in this wave compared to the previous two waves from all four dimensions. And I can tell you that the incentives are completely all over the place. Right? So nobody has a common incentive of actually making you know enterprise AI ROI workable. Everybody is having different slices of the pie and I think that also is translating to people presuming that if you already have a process for example order to cash or procure to pay all I have to do is sort of either insert a few agents or completely you know pull it off pull the whole process off and build it with you know 5 10 15 20 agents right both of which is not going to solve the problem unless you understand how the decision making through algorithms and a combination of humans is going to translate in the real world. So that's actually one big issue that's happening on the ground.

Hannah and I think the previous the previous eras were a lot more transactional. Okay. The previous eras you could have siloed applications. ERP could be a silo. You know HCM could be a silo, CRM could be a silo because they all had individualized functions but intelligence and especially enterprise cognition is like water. It has to flow. It cannot be puddles of different systems that hold water and somehow miraculously you're going to translate a CEO's decision to give an instant let's say bonus or some kind of a you know a discount to all of these disparate agents who not connected to each other right so I think we are actually looking at a very different paradigm very interesting paradigm of how to not get siloed just like we were in the past and how to build this into a lot more of a river architecture than from what I call earlier as a dam architecture the DAM right this actually this shift is actually going to be very interesting it could actually lead to a lot more you know startups and service providers coming into this enterprise metabolization space which hasn't happened so much

Rob May: how do you guys deal with so you're at a big company how do you deal with the metabolization the adoption of AI like is there pressure from above or is it bottoms up or does nobody care I mean and Microsoft is relatively AI forward so I'm curious what it's like internally culturally

Lauren Slyman: I think culturally you have a bunch of nerds in one big company. They just naturally embrace it. It's something everyone's interested in and talking about and using. And I think there's been a healthy amount of pushing back when you need to push back and question something and then embracing it and knowing how to use it correctly.

Rob May: Cool. That makes a lot of sense.

Anna Kirk: Okay. Question for you, Satish. As you were talking and especially because you're so tenured from your experience at Accenture, there's another potential paradigm that I hear quoted a lot, which is like, well, services are dead. The billable hour is dead. And you know, the more deeply entrenched I get into this space in this industry, the more skeptical I become over that claim. And so I'd be curious to hear your thoughts on that. You know, if knowledge really should flow like water within an enterprise, how does that impact the industry whose entire kind of business model is based on that knowledge?

Satish Viswanathan: Yeah, I think we are we are shifting from the earlier era to a new era in the consulting space. If that's the question Anna I think this whole linear model of growth in terms of more people is equal to more revenue that is completely getting sunset so it's not just going to be more people from some plus some agents it's going to be a different kind of a model there are two theories going around right now and depends on you know whom you talk to in the consulting industry one theory says the bottom of the pyramid is the one that's going to be most impacted because you know all the basic research analysis or say insights gathering could be done better by agents And there are other you know the other school of thought is the middle of the pyramid is what's going to get most impacted. So either way the top of the pyramid is safe which is which is a good part for some of the leaders but I think the middle of the pyramid the reason why they're saying it's going to be the one impacted is because you know we might have a lot more of a flatter organization with the leaders actively working with people at the bottom with some agents.

So, so I think it's one side AI ROI is not coming through as much as it should which means that the consulting organizations are getting challenged from a outcome or output perspective and on the other side you know their own capabilities are getting in a sense eroded because these models are sort of doing some basic things right okay especially on structured data especially on rag models I think the models are having you know as as Rob said some kind of an equivalent or getting to that kind of equal and intelligence though I think this is very relative in terms of different contexts it's not necessarily the same everywhere but I think that there are instances where a good research or a good set of slides that one would take weeks is now take is now sort of getting done in a matter of a few hours so both of these paradigms or theories are circling around but one thing is very sure that the consulting organizations need to be consulted by somebody else or you know to sort of reinvent themselves because they're not going to have the same, you know, long-term plane of high high number of resources anymore.

Anna Kirk: Yeah, it's interesting. I was part of the reason I asked you that is because I was sitting on a plane on a flight recently with a managing director at a big four consulting firm and he was basically saying, well, like all the big four are in this race to productize their expertise fastest. And you know, I tend to lean on the side of the spectrum that says, well, automation in the hands of experts and technology in the hands of experts should go 10x farther. But you know, I'd be curious to hear your thoughts on that. Like, do you think that some of these big four consulting firms will be able to kind of productize and better scale some of that knowledge faster than, you know, the equivalent kind of services firm that doesn't?

Satish Viswanathan: I think it it depends also on how the culture is incentivized. So you know one of the biggest shifts that's going to happen as part of the earlier question as well is we always have appreciated individualized skills or individualized talent as a as a reason for promoting somebody or moving somebody up the you know the corporate ladder. But now if you go to have a model, let's say the model is sort of launched for 100 people who are working in a contact center and they've been all told just follow the model. The model is the best. We spend millions trying to sort of you know get it right. Now all the 100 people are going to use the model diligently 90% 95% of the time. The question is who gets promoted?

So we're going to get into a new wave of what does it mean to excel in the AI economy? And that incentive mechanism and the culture will drive how much of that knowledge they have can be productized because if as a human you're insecure about what you can actually sort of bring to the table because the model is better than you, you're not going to s spend more time and put more of your tacet knowledge you know encoded into the model. So unless you sort of are very sure that you're getting something out of it. So I think the culture and the HR models have to shift. I don't know I've not seen evidence of them shifting yet but once they shift then we will see a lot more confident set of let's say a human training enablements happening on models which could be a new layer of finetuning as well

Rob May: you know I was at last week I was at this AI circle event on data labeling and had some people from scale and some other places and they were talking about this idea that the big problem in pushing the models further now the frontier models is that they need like let's say you want to make the model better at international tax law. Well, like you need somebody that knows something about that data and like that person no matter how much money you pay them doesn't want to sit down and label data and do like it's boring to them. They make enough money like it's not so they're having a hard time getting the kind of data they need to push the models to the next level.

So, it's yeah, so it's interesting to see where that's going to go, but it but it kind of leads into one of one of our articles here that was in the Financial Times yesterday about law firms seeking to make bespoke models in legal AI for what they're doing. So, if you if I look at this really quickly, Harvey, the most prominent legal startup, has made some of their own models. Looks like Kirkland & Ellis has allocated $500 million to building their own AI. Goodwin Procter does a lot of startup law has allocated u millions of dollars to do this. And so is this this canary in the coal mine for the Frontier Labs or is this just like something for law? What what do you what what should we read into this if we're investors or operators? What is this signaling?

Anna Kirk: My my first instinct is that it's a signal that they are particularly riskaverse and not a signal that the Frontier Labs should be threatened.

Rob May: Possible.

Lauren Slyman: It makes me wonder if we should be focusing on if owning it is a real this proprietary proprietary knowledge is advantageous or would it be knowing how to use them properly within a corporate or traditional organizational system? That would be that the advantage there.

Anna Kirk: Yeah, that's a good point because they're probably building they're not building expertise because there's this opportunity cost of like, oh, now we need to go build our own model and like employ the right like technical resources to do that versus actually deploying resources and being able to kind of enable their existing employee base to become more efficient and more effective at their existing practice.

Rob May: Yeah, I think it does make it harder because I think AI sort of raises the bottom for everybody, right? I think I've given this example on the on the show before where if you're an awesome awesome writer, you probably use AI a little less than everybody else, but if you're a terrible writer, AI can make you average to a little above average, like pretty pretty good, right? And I think you see that in a lot of a lot of areas. And so it's interesting to see what that's going to have on these primarily IP driven, knowledge driven jobs like law.

Anna Kirk: I don't know if I don't know like maybe maybe in the like immediate short term it makes the bottom higher but I almost think that like if you are not forced to ever sit down and like just write out your own thoughts and like really think deeply about an issue from a bottoms up perspective there are so many flaws in your argument that you won't catch because you're not holding yourself to the same intellectual standard. So maybe in the short term, but actually like in the long term, I don't know if that's necessarily I don't know if I agree with that.

Lauren Slyman: I think it's interesting because I'm thinking about this lately with using AI myself and I won't use it for writing and I'll use it for I don't know summarizing or turning complex things into things to understand myself. But I do wonder, is everyone going to lose confidence in their own abilities and have to have this buddy to check them every single time with some sort of chat bot to where they won't have creativity, they won't take risks, they'll be slower and they need to have everything double checked and confidence entirely.

Anna Kirk: Yeah, it is very like middling.

Rob May: I don't know. Satish, what do you think? If you care to comment.

Satish Viswanathan: No, exactly. Yeah, I think I think that's the other risk because one side the market is pumping in trillions of dollars to say more AI is great but on the other side I think the question that you know was just being asked is how much AI is too much AI because if you're going to lose your individual ability to think and reason because you're all the time having to validate your own point of view eventually you will plateau down okay and also Rob the point you mentioned about low low achievers or high achievers how they're consuming AI there was a research done on this I think at Harvard and I think this is a couple of years back and they found that you know the high achievers they are confident about what they want to do so they're not worried about it okay so they use AI don't use AI their output is topnotch so but the low low achievers are actually catching up so what's eventually going to happen is instead of having this you know so-called bell curve in terms of value that employees are creating you're probably going to have more of a high-end plateau that everybody's achieving 90% and above and that's again going to cause a lot more problems in terms of who is actually valuable, who is you know important and so on.

So it's a little bit of the other side of AI as well. What does it do on the psyche of employees in terms of their confidence in terms of even trying something new? What if the model says something because it's trained on let's say X trillion or X megabyte or X gigabyte whatever the number is of data of documents but we actually have a gut instinct. So all these are going to be different waves of challenges we're going to face as a human race. If you're going to be constantly compared to a model that's been trained to whatever extent and from an AJ perspective.

Rob May: Yeah. I mean it's interesting right because a lot I mean some of us work in the startup industry and one of the things about the startup industry is sometimes you want somebody who's not an expert in the field to come in and start a company because the conventional wisdom propagates so long that you need somebody to come in with fresh eyes and be like wait why does everybody do it this way? You know there's there's different ideas now. So I wonder if there'll be a similar path for AI be really interesting.

Lauren Slyman: I guess you would have to achieve the higher level of intelligence which would be AGI because at that point what other opinions would it have to give you other than what it's trained on already at a lower level of intelligence.

Rob May: Yeah. We have one more article we want to cover before we wrap up the show. Look it up. There's a there's a programmer out of London. His name is Stephen Steven Deal. D I E H A L. And he wrote an article called the internet is a predatory cesspit and it was very well written clearly not AI-driven or if he did it's like wow he's got a brand new amazing model knowing Stephen he probably doesn't I've actually met Stephen a few times he was a Boston guy before he moved to London and the gist of the article is that the forces that drive the internet and the structure of the internet lend themselves much much more easily to the predators and those types of people rather than I mean you know it's like we had this thing when the internet came out where people this is going to be great because we're everybody's going to be smart and learn so much now and whatever and then we ended up using it for social media and to watch stupid videos and do those kinds of things. And he would argue that like it makes scamming easier and it disproportionately favors the scammers over the, you know, sort of the people trying to fight that kind of stuff. And that AI is just going to make it worse. So curious what you guys thought about that.

Lauren Slyman: He made great points because think about it. I mean like right now if you give your data to any sort of algorithm, it can learn about you. So it can take that, it can take advantage of that. And then you have these models who learn everything about you. These models are going to and telling them their dreams and goals and fears and anything under the sun. You're collecting that data on top of people willingly giving their own thoughts and opinions. It is learning you inside and out. And it is smart enough to be a persuasive and potentially potentially less ethical salesperson. And I'm basing that on just agent research I read recently out of Van Davik. But I think it's I think it's a big risk actually and he made some great points.

Rob May: Are you going to start lying to the models so that they you confuse them off?

Anna Kirk: Yeah. I don't know, but like the fact that like I mean they can track your activity like they're like ingesting a lot of your activity so granularly like I don't know if I could hide or conceal all of my shopping habits for instance.

Rob May: Yeah. So here's some interesting quotes from the from the article. This is an industrial system for manufacturing weakness at scale. A legitimate business can survive a satisfied customer. A griff cannot. It needs the customer frightened, agrieved, lonely, sick or greedy forever. And he talks about that the grift economy is massive, more disturbing still. It is participatory blah blah blah blah blah. So it's and this is the affiliate marketer buys a course about affiliate marketing, then recovers the cost by selling the same course to the next affiliate marketer. The life coach coaches new life coaches. The drop shipper sells tutorials to failed drop shippers. The pyramid is social before it's financial. Everyone stands on someone else insisting they are about to escape.

Lauren Slyman: Wow. You can tell that that was not written by an LLM.

Anna Kirk: Yeah. Yeah. I really like the quote it metabolizes those injuries into new injuries. And Satish to bring this back to your book about metabolizing a framework I am curious what would determine whether an organization would be turning AI into compound compounded cognition versus automated dysfunction at scale.

Satish Viswanathan: I think this this this idea of compounded cognition is something that I really have you know in a sense brought us a fresh concept. I think in essence what I'm trying to say and this also is a reason you know of observing how how CEOs have changed their their goalpost. So in 2024 it was more about you know let's do more AI. We talk about token consumption as a as a symbol of pride and success. In 2026 we're talking about you know I want to see attributable AI impact in my P&L statement. So it's a little bit of the goalpost shifting and no matter which use case you take in enterprise no matter how well researched you think that use case is the best one for a adoption it will never make you know an attributable number on your P&L statement if you are a multi-billion dollar company right it'll always be a decimal of a decimal and that's the reason why I was just reflecting and saying we are doing a good job in local value of that specific use case at that scope that we set out to do but we still don't have the CEO's answer in terms terms of why can't I see the impact in my penal statement and that's where there's this whole concept of compounded cognition comes in.

In essence what it means is that we sort of have to build use cases in a in a little bit of a of a Fibonacci series where each use case the intelligence of each use case leads to the next use case being more intelligent. Okay. And not restricting ourselves to that one use case which we sort of set out to do which is typically how the industry is operating. So that's actually the one side of the story. But the point that you mentioned earlier about the article, I think there's a big risk about the alpha of organizations leaking to the models. No matter what the models are saying that they are not using organizations, you know, core IP and that's also going to become a serious problem. Regulators are going to step in, governments are going to step in terms of how much AI is too much AI and how do we make sure AI is benefiting that enterprise primarily and not resulting in a higher let's say u competitive edge for the other other players in the same industry. So it's a little bit of that nuance that's also you know slowly emerging. But I think we all have to sit and observe how this is going to play out.

Rob May: Cool. And Satish if if people want to check out your book where's the best place? Should they go to Amazon or someplace else? Do you have your own website for it?

Satish Viswanathan: There is obviously it's on Amazon. I do have a you know a paperback hard cover. I'll be releasing a color edition as well because some of the images actually have come out well for for the global market. I'm also working on a Kindle version and an audio book. So it'll be sort of released in phases but essentially Amazon is a go to a go-to location. I do have a website as well in case people want to engage with me on some of these concepts. And I have three more books lined up Rob. So, it's just a matter of, you know, getting the first one through and then start trying to focus the next ones.

Rob May: It's definitely a broad topic. So, well, before we wrap up, we like to do a little bit of NYC trivia.

Lauren Slyman: So, I have a question for you guys tonight. What New York City location has appeared in the most number of movies?

Anna Kirk: Central Park would be the

Lauren Slyman: No. Oh, good. Good guess.

Rob May: Grand Central because it was part of that experiment where people had to like agree on one place in New York and they all chose Grand Central.

Lauren Slyman: Grand Central Terminal. Really?

Rob May: Huh? Yeah. Which makes a lot of sense. Say, say more.

Lauren Slyman: It's appeared in over 250 films, including The Avengers and Madagascar is what the answer says. We can marathon them. Watch order.

Rob May: Anyway, Satish, thanks for being a guest on the podcast today. We will post information where you can find his book and other ways to get in contact with him if you're interested in learning more. Thank you guys for watching. It's good to be back. This is this is fun. And I felt like it's we're only off 5 weeks, but it felt like forever. Got to get back into the rhythm. So please watch, subscribe. We have a newsletter now. Lots of stuff going on. We're going to be doing some events. Send us guests, send us ideas, send us sponsors, and we'll see you guys next week. The next stop.