hey jasper how are you good afternoon it’s dark and cold in berlin the summer is gone now it’s gone oh no like i we like i just had a beautiful sailing weekend out actually in the bay it’s still warm nice weather so yeah silicon valley area right so we have a really cool podcast today we invited eric siegel i know actually eric from a while back where he he has his conference about ai he wrote several books about it he understands ai and all the complications of making ai work getting ai to value not just that he’s first of all a very funny person so we can maybe link a video he has done with ai but he’s especially a very funny person because he’s a very funny person honest person he says maybe he’s a bit negative we just think he’s being transparent of what he thinks is the truth behind ai and what’s possible and that’s what we really want to dive into so it’s it’s great to have him on the show so we talk value we will talk mistakes he will say and you can have the cake and eat it too but we will talk agents we will talk ai and what he thinks about agi we loved it i’m pretty sure you will too so let’s get started let’s listen into it let’s go eric it is so awesome to have you on the call normally before we kick it off we always want to hear a little bit the story like who is eric your life story in the 30 seconds pitch no no problem i’ve been in machine learning for about 33 years i was a columbia university professor i did the grad courses in ai machine learning been an independent consultant for 20 years and i’m the co-founder and ceo of gooder as of a year ago and you always impressed me and we know each other for i don’t know more than 10 years what i really value about you is that you have this amazing ability to make complicated things easy that’s a reason why i really am so glad that you’re here thank you you’ve been in this research field for such a long time i think you’ve been giving hundreds of keynotes you organized conferences you wrote books so and then you moved over to the startup side so you really you must have really known what you were doing going from this kind of secure job into the world of chaos of startups maybe you can shed some light on that that would be very interesting my startup focuses on predictive ai and broadly speaking you know you can divide all things ai which is a very subjective word into those two main categories of generative and predictive predictive ai predictive analytics enterprise machine learning and then you can divide all things ai which is a very subjective so a level of value that the hype conveys or the level of autonomy in general. That’s a very broad sweeping statement, but let me put it this way to frame where I’m coming from. Enterprises in general, depends on the company, should really be investing at least as much into the potential value from predictive use cases as generative. Not only, but right now, it’s becoming kind of a zero-sum game.
A lot of budgets are moving away from predictive, but that’s the technology, predictive, that you turn to, to improve your existing large-scale operations. The holy grail for determining who to investigate, incarcerate, set up on a date or medicate is who’s going to click, buy, lie, or die, commit an act of fraud. What’s the best place to drill for oil? Which satellite is going to run out of battery?
So it’s sort of on a higher macroscopic, still micro from the standpoint of an enterprise, but more macro than generative in terms of using machine learning’s outputs, which are always predictions. But on that organizational, unit rather than on the next pixel as I render an image or as the computer renders an image or the next word or token when it’s generating language. So it’s a very different use of machine learning. They should be apples and oranges.
They’re very different kinds of value propositions. They should compete no more than water parks and ski resorts, but lo and behold, they totally are competing. So I’m going to put my foot down and say, look, let’s not just put all of our bets into generative when the large-scale, existing operations are served with much simpler kind of direct, this is the probability that the customer is going to buy. Okay.
Contact them. This is the probability of the transactions fraudulent. Okay. Block or audit it.
That kind of level of operational detail, large-scale operations are made up of decisions, predictions on that level of detail are the Holy grail for improving those decisions. Okay, folks. So we need to be doing both. Now, one thing that’s really interesting we could get into is that they combine.
You brought in those two. Topics and most of our audience probably is familiar with the difference and you said it actually in your discussion, but give us a better understanding. What is generative and what is the typical predictive AI? So these two terms really refer to the use case of machine learning more than to the particular machine learning technology.
They’re just big categories of the value propositions, how you’re applying the technology application areas. Generative is generating new content items, right? Obviously it’s deep neural networks and transformers and that kind of technology. So it does usually refer to a particular kind of technology, but the word generative just means, Hey, look, it’s going to generate a new image, document, writing, video, music, sound, voice, whatever it is, right?
So you’re generating new content items that for many applications generally must have a human in the loop to review before it goes to an end user, before it gets deployed as code or what have you. Predictive is learning from data to predict. That’s what machine learning is doing. But in this case, the predictions are directly informing operations for targeting credit scoring, fraud detection, marketing, placements of ads, where to drill for oil, which train wheel is likely to fail.
So we should investigate it. Reliability modeling. So it’s that category. So these are not so much about different technologies, but the different uses of that same kind of technology that we know of as machine learning.
Now, in your book, you write a lot about the six steps. And Jasper and I had this many times on this podcast that we said, well, generative AI just created enough hype. So suddenly everybody is looking as well into predictive AI. Now, just having predictive AI in itself is not helpful.
So you created this framework. Can you walk us through this? Yeah, the value propositions for predictive are clear. Perhaps clearer than a lot of noise about generative, right?
Which is more about a panacea or a one size fits all kind of solutionism thing to a certain degree, unless you really focus on a very particular use case. Whereas predictive is already about a use case. It’s what are you predicting and what are you going to do about it? Right?
What’s the outcome or behavior per individual human customer, healthcare patient or organizational unit, like as I mentioned, a train wheel that might go out of commission. Right. Yeah. that’s been created from the core rocket science, right?
It’s only creating value. You’re only capturing it when you actually go to deployment operationalization, going to production. So my new book, the AI Playbook, is focused on those types of predictive use cases and provides a six-step framework that I call BizML, which ushers the project successfully from inception to deployment. Then if we take your framework, Eric, what interests me, I mean, I’m a VC most then, you know so much, you have seen so much, you develop this framework and obviously discuss it with many people.
And there are so many new generative AI companies out there, and especially last year, like very, very early ones. What led you to gooder AI? And maybe you can also allude a little bit what you did not want to do, because you definitely know what’s not going to work, I think. Well, continuing to do things the same way routinely, generally leads to failure.
Maybe 15 or 20% of new machine learning enterprise projects and these predictive use cases actually lead to capturing value because they get deployed. I sort of stayed predictive, but we’re incorporating generative, and there’s a few ways that the two work together in hybrid that I think has the potential to realize a large part of the over-promise of autonomy that goes with today’s hype around generative. So we can come back to that. Predictive use cases, these kinds of predictive AI projects, are always a consulting gig, not a technology install.
You can’t just be like, I’m going to update the backend database and better technology. Now the whole thing is going to run more smoothly and more quickly and less downtime and all that kind of stuff. No, it’s about changing the largest scale business operations. So it’s a business procedure.
And BizML, the six-step framework, it’s not like MLOps or any other kind of technical solution. It’s not the tools and technical procedures. It’s the organizational procedure. It’s a framework, playbook, management protocol, so that the organization collaborates in the right way, bridges that tech-biz divide, which is manifest in so many ways, and that routinely leads to failure, bridging that gap, engendering tight collaboration from end to end, bringing together the business-side stakeholders, the customer of the data scientist who is in charge of the operations meant to be improved with the output of a predictive model, and bring them in, and a semi-technical basis with semi-technical understanding of how these projects work.
The semi-technical is basically what’s predicted, how well, and what’s done about it. And in those three respects, all business-side stakeholders need to get their hands dirty or their feet will get cold, and they’ll balk at deployment. And at the end of the day, they’ll refuse to change existing operations, therefore not improve them, not capture any value. So that has to change.
Talking to you as a VC, you’re probably oftentimes obviously looking for a new business. You’re looking for software solutions. Well, if this is about better consulting or business practices, what are the technology solutions? There’s a few places, but one of the key ones comes down to metrics and saying, hey, look, how good is it?
How good is AI? How much value does it provide? In these predictive projects, typically that question is not answered. So the stakeholder has to make a really difficult decision.
The data scientist is like, this model is awesome at predicting fraud. It has an area, under the receiver operating characteristic curve of 0.88. Isn’t that awesome? And all that tells you is the pure predictive performance.
Abstractly, it’s relative performance comparison to random baseline, and that’s good. It’s good to know that it predicts better than guessing, which means it’s potentially valuable. It doesn’t tell you anything about how valuable it is. That’s what we’re doing at Gooder AI, is navigating that development to value.
If you think about what you said is predictive use case, you always end up in some consulting, consulting gig in order to adjust the organization to suddenly this new tool. And this is true. This is what we see very often in large companies. It is as well what smaller startups come to us and tell us, oh, you know what?
Because you need the complicated org changes, we as the fast about, the more creative group here can actually change it, which comes down to value creation. There is this, like we have always this, this discussion, what’s a feature, what’s a product. If you are a larger corporation, you might say, oh, I use predictive AI, generative AI for me as just a feature. And I own the customer, I own the data.
And now you saying, oh, if you do that, you need consulting. Now, on the other hand, the startups come and saying, actually, I think I can build a product out of it. Can you give for our audience a little bit of frame of mind, how to think about it? Because you clearly, like what you just said, would be the desk for many startups.
All right, guys, it’s a consulting gig anyhow. Now there is the idea of, and this was 10 years ago as data science came all about, right? And we chatted, that was a very similar notion. There is this underlying structure of product versus feature.
Can you talk a little bit about it? How do you think about that? To get these things to work, it does require a consulting gig aspect. And I’m alluding to the main thing that’s generally missing and that therefore leads to a lot of I’m not sure if you’ve heard of it, but it’s a product that’s generally missing and that therefore leads to a lot of lack of value.
You need to have the right practice, framework, collaborative process on a business level. That doesn’t mean that we don’t also need new kinds of software for particular parts of that. I’m just saying that you, one of the necessary conditions is a new kind of practice. One that’s going, that has potential to be implemented by a consultancy, probably one that you’re not interested in investing in, right?
But I’ve been an independent consultant for 20 years. So myself and a lot of our colleagues, we intrinsically do this kind of thing. I’m just trying to brand it with the book. I call it biz ML, five letter buzzword catchy.
We need to deliver the message to business owners in general, to stakeholders that look, you need a very particular kind of practice or framework to get these projects successfully through to deployment. Most business people out there don’t realize you need a very specialized practice, let alone the name of any in particular. The attempts to do so, while the main one was from 30 years ago, business people have never heard of it. In general.
So I’m trying to say, Hey, look, this is business friendly. It’s business relevance, pertinent, accessible, understandable. And Hey, here’s a catchy buzzword. So if we can brand that, then we’ll get the news out there, but I think that’s not very, but not sufficient.
Yeah. If I challenge you on this one, you brought up early on for predictive AI. You said, man, like this could be something like predicting the failure of a device, right? So like GE is using predictive AI to figure out, you know, predictive maintenance reduces cost for maintenance and keeping the actual product more safe.
So that would be one way. Now it becomes a feature and you actually have to change your operations. And people would come to you and say, Eric, we have now this prediction, but that means that we are not maintaining our staff in a regular setup. They’re not going into every flight they’re going into as the machine tells them to.
So that creates org changes. Now you could as well, argue and saying, well, you know what? I’m creating a new model where I rent out airplanes or devices or whatever machines. I can rent them out better than the current financing model because I have a way cheaper way of maintaining them.
Now the AI is ingrained and now I have a completely different business model and a different setup. How do you, in your work differentiate between, at what point, in time is technology so strong that it actually changes as a business model, which is obviously the area where Jasper and I look a lot into versus technology is just a helper. And because it’s a help, but people needed to learn outlook. Oh my God.
Terrible. Yes. And then Google came around and they needed to learn Google, right? It is actually changing the business model.
If you have an existing large scale operations, such as maintenance for whatever you’re renting, you’re renting cars, right? So you’ve got some, some maintenance that’s, it’s covering a hundred thousand fleet of car or a million cars for your rental agency. That’s an existing large scale operation. The sort of business pitch of predictive AI, which was pretty much the main business pitch of any AI before two years ago is one of the last remaining points of differentiation is to improve those large scale operations with per case prediction.
That’s the Holy grail for improving for increasing efficiencies. So, but the only way to actually capture that value is not only to do the number crunching and generate a predictive model that puts those probability with this automobile is much more likely than average to break down in the next hundred miles. And this one’s much less likely than average. And, and then not only general, not only calculate those probabilities, but act on them.
So the acting on them, that’s the part where you’re capturing value. That’s the part where you’re integrating operationalizing deploying, and that’s the part that that’s generally missed. But if you do that, then you, you, and in comparison to not doing it all, whether it’s for targeting marketing and predictive maintenance or anything, generally, there’s this very dramatic improvement on bottom line KPIs related to the process that you’re improving. That’s the gist.
That’s the value proposition. So beyond that, what’s your question? I didn’t quite follow. Like, how does it change a business model?
I mean, I don’t know if I’m a car rental agency and I can drastically cut some costs. That’s a good thing. I don’t know if I’ve changed my business model. Yeah.
That’s so so some decisions around standardization, being opinionated. Basically, you build a product, right? You don’t want to be a full consultancy. However, sometimes consulting work is pretty nice and fruitful.
I was BCG consultant. So I would love to understand your thought process when you set up Goodr and saying, hey, there is some consulting part. There is a standardization piece where we will also charge something for. So how do you balance that for AI?
Because it’s different, right? But Goodr doesn’t address the broad issue my book does. Goodr addresses one particular place where there is a need for a new category of software. And that’s in the metrics.
That’s in establishing the potential value. It’s a business user console for understanding the potential value and making decisions about deployment so that it’s not just a data science decision because it can’t be. It has to involve business users and it has to have to do with business metrics. But stepping back to broader scale, what I’m describing is just simply intriguing.
It’s not intrinsic to any predictive AI project. It needs to be run as a business project first. It’s not a plug and play. It’s not a technology install.
How do you think then about onboarding your customers, customer success, helping them to use it? Or would you say, coming back to my kind of thought about sanitization and being opinionated what they should do, would you say, no, we want to be as self-service as possible because the product is so great? So you’re asking about my product, not about my point about… These projects.
Yeah, your product. The good… Okay. Well, I haven’t even really done the pitch of the product.
I mean, so typically they say, hey, look, this is a technical metric, area under the curve, precision, recall, what have you. Yeah. And it stops there. It’s very, very rare for data scientists to go to the stakeholder and say, hey, look, if you deploy this model to improve your marketing, fraud detection, reliability model for predictive maintenance, whatever it is, whatever the use is, if you deploy this model, it has potential value of this much profit.
Or this much savings. Very straightforward KPIs. Those are just not part of the conversation. Very rarely.
And that needs to change. Now, to make that change and to make this much needed fundamental move from technical metrics to business metrics, you have to open a can of worms. And that’s why here I’ve been in the space for 30 plus years. I’ve been waiting for this to happen.
The reason it doesn’t happen is that you can’t just measure it in a vacuum or in the cubicle of a data scientist. It depends on the… Business context. And it depends on how you’re going to use the model, not just the performance of the model itself.
How are you going to deploy it? Where are you going to draw the line? Am I going to… Now I know which transactions are most likely to be fraudulent.
That’s what the model does. It doesn’t tell me how to use it. You have to decide, okay, well, am I going to then therefore block 0.5% of the transactions most likely to be fraudulent? 1.5%?
That’s a huge difference when you have a large scale operation. And there’s no solution to sort of… Navigate and visualize what the different trade-offs between competing KPIs are going to be, estimating the potential value of doing it one way versus the other. It requires a very specialized UI.
Model training, that’s been the main sort of category of software, but that’s definitely moving to open source. Whereas this is a visualization tool, much less likely to go to open source. And it’s a new category. It needs to be its own category, universally operating across all models, regardless of what tool or solution you use to generate them.
So that’s what we do. Just for me to understand it, Eric, you translate the precision recall, the F1 scores, whatnot, for the business user so they can actually work with it. Yeah, that’s almost right. Translate’s not quite the right word because it’s not like you literally take as input the technical metrics and output the…
It’s just a different calculation. You know, we shouldn’t throw out the technical metrics. They’re useful, but they’re nerdy. The nerds have to talk business speak.
They need to be able to take the model and present it in a way that’s meaningful. For example, in a chart, very simple, where both dimensions are business relevant. They’re going to talk to the business, the stakeholder, and tell them exactly what their options are. One dimension is that decision of where you draw the line.
That’s called the decision threshold. And the other is a business metric like savings or profit. Then there’s other things like certain… Business inputs that are subject to change and uncertainty.
You need to see how big a difference they make. There’s other competing KPIs that interact. So it just becomes a very specialized UI. That’s where the industry is moving.
It’s going to be the only way that we can systematically improve what right now is a devastatingly poor deployment success rate. And I can tell you, Eric, a lot of our startups would love to have something like that. I remember the time nine years ago when we did this accounting automation company. And they had their own precision recall set up.
They tried to translate it into what is for an accountant precision and recall. Actually, what does it mean, right? If I have a certain value there. So I think that’s immensely valuable.
I have seen like in the medical field, you have actually… It’s stunning to understand how well doctors have internalized this understanding from precision recall to what is a clinical actionable. For me, this was one of the things as I went into healthcare that I was like, wow, like, all of this stuff, which I’m trying to explain, they have internalized. They don’t understand why.
Like, they actually do understand why, but they have internalized it. Now, isn’t that problem actually not only a problem which we see in the AI world, but in any toolings world? I used to work for Ericsson and we sold at that time 3G infrastructure. And you had the whole question about what’s the total cost of ownership?
Meaning, if I pay now… The more I have to service less and that kind of changes my setup. Now, you need to break technical features, in this case, servability, down to business features. Haven’t that been a challenge of the business world all along?
Technologies come around and it’s always hard to say, how do I make business out of it? Because at the end, what counts is dollars. You could also just say, we’re talking about forecasting. But this is a very particular type of forecasting because it’s saying, hey, I have a large scale operation.
I’m going to improve it with a predictive model that makes per case predictions that directly inform decisions. So I’m going to integrate it in a certain way. So that’s a very particular kind of operations improvement project, otherwise known as predictive analytics, right? It’s sort of been the main kind of AI before generative, right?
Generative introduces a whole bunch of other capabilities that are very much complementary. Predictive isn’t going anywhere and most of its potential is still on tap. So if we stay on the predictive side for a moment, for those projects where you kind of order, prioritize, triage, and then decide where to draw the line, these most risky patients, some portion, you have to decide that, are going to be retested or reconsidered before you discharge them from the hospital. These train wheels are going to be investigated.
These restaurants are going to be investigated for health code violations. These buildings. These buildings for potential fire. These manhole covers for dangerous incidents.
All large-scale operations are improved with prediction. The act of acting on those predictions is called predictive analytics deployment. It’s one of the main last remaining points of differentiation for improving operations. It’s what everyone used to call AI.
It still needs to be solved. Most of the projects fail. It’s a very particular scenario in which you need to have very specialized solution for visualizing its value. If I’m a founder and you tell me most of these projects fail, why would I start a business?
Obviously, because I want to start a business. But is there anything I can do differently than all the others? Based on your knowledge, why they fail? No, I mean, that’s why I wrote the book, the AI playbook, right?
It’s to run through that. So it runs through the six steps and they mostly involve those three semi-technical ideas that I mentioned, you know, what’s predicted, how well, what’s done about it. Not quite in that order. And it kind of pans out and you’ve got to prep the data, train the model, deploy.
So, but it brings the business user. So the bigger point here, the thing that’s going to make the difference for this field of predictive AI in general is that the stakeholders get their hands dirty so their feet don’t get cold, that they get involved in that semi-technical understanding. It’s like, we’re more excited about the rocket science than the launch of the rocket, right? It’s like, we don’t know how a car engine works, at least not in detail.
Most of us. Including myself, but I’m an expert in driving, right? Momentum, friction, rules of the road, operation of the vehicle and the mutual expectations of drivers. You need the same kind of accessible expertise to drive a machine learning project successfully through to deployment.
Whereas most business stakeholders are going to be like, oh, those details, I’m just going to delegate them to the data scientists. Huge mistake, right? They need to get involved in that very accessible, interesting, fascinating, semi-technical understanding and thereby involve themselves with deep collaboration across the project. Does it even get worse?
Because when we look at many startups from last year, also this year, it feels like people are saying, sorry, it’s not good enough. The large language models are not good enough. We have to push this further. I need money and I want to go for AGI.
I mean, not just open AI, right? There are so many agentic new models out there, startups. So why are we actually doing that? Why do we need AGI?
I mean, is that even a startup? I don’t know. Well, I think you kind of, this is a dramatic change in topic, right? Because everything I’ve been saying is mostly about predictive.
We want to talk about the hype of generative AI, which, you know, this is cognitive dissonance, right? I’m like, I was always really interested in computational linguistics. That was the focus of my also machine learning PhD. I was in the computational linguistics research group for six years.
I think it’s incredible. I think it’s amazing. And it does have lots of particular value. But the hype makes the world 10, 50 times more excited than me.
And I thought I was excited enough. So I’m both at the same time. I’m excited. I’m almost like everybody calm down.
It’s the mismanagement of expectations, the distance between what we’re going to have compared to what the expectations are being set. That’s disillusionment, right? Bad disillusionment. But that’s my point, right?
It’s so weird. We have so many great things. And then you see founders even pushing further and say, this is not good enough. We have to do so much more.
Do you? Yeah. And I think the AGI is essentially a myth. I don’t saying that it’s theoretically impossible.
But what I assert is that nothing in all the advancements of technology represent a concrete, clear step toward AGI as seemingly human like as an uncanny as it is. The distance between what is capable and what humans are is going to become increasingly clear. AGI is just another way of saying artificial human. We’re not creating.
I mean, it’s a ghost story. It’s the novel that Mary Shelley, who wrote Frankenstein, would have written if she knew about algorithms. Funny that you say so. And Eric, I’m so much in line with what you say, right?
There is a saying about AI is whatever was not yet invented. And I think now it’s AGI, whatever is not yet invented. And everybody is hunting this ghost story. Yeah.
And to put it in another way, I’ve quoted that guy. In some of my articles saying that. But to put that another way, the subtext underlying so many of these conversations, if not explicit, is that we’re headed towards AGI. And in fact, it could be very soon, even within a few years or a couple of decades.
And I feel that that’s a false narrative. It always comes down to AGI. Part of the problem is that we shouldn’t be calling it AI. AI is intrinsically hype.
Ascribing the word intelligence, which is specifically about… Humans to a machine always will mean we’re anthropomorphizing. We will always have AI winters so long as we call it AI. The word intelligence will always increase.
There’s no definition other than AGI that will suffice to actually meet the expectations and spirit intended by the very catchy and seductive buzzword AI. So the… I love the fact that you talk about the winter. Yeah.
Yeah. The original sin. This is in the 50s when they… Yes.
No, no. I love the fact because the AI winter, all the dissonance, Eric, this is so true. I was in a board meeting lately and one of the board members says, oh no, we shouldn’t invest. Like very soon, there was an investment proposal.
We shouldn’t invest. Very soon, Gemini will do it all for us. So kind of like the idea of AGI is going to solve it all. So let’s not invest now.
Now, how do you… We don’t have to do anything. We don’t have to worry about the climate or any… Exactly.
World peace. Who cares about world peace? We don’t have to work on anything because it’s going to be the panacea. It would be by definition, right?
If we had AGI, we would have humans and then very quickly better than humans, as many of them as we want operating 24 seven and at infinitum, right? Self-improving, et cetera, right? So it doesn’t… Then we don’t need people anymore.
It’s a wonderful fiction built on a premise that’s whimsical. However, that fiction is out in the world. And I think… We all three here agree that this is fiction and this leads to another winter because it’s overhyped and it’s not what’s real.
Now, how do you, as a founder, as somebody who talks to corporations about this, how do you react to it? Because you talked about the dissonance in your brain. So now here you come that an ROC, like the driving example is super spot on. You know that you need to train driving to the CEO and the CEO says, oh, like, no, We will have self-driving cars.
Now I will have beam me up Scottie technology. Don’t tell me how to drive. That’s a dissonance. How do you deal with that?
The antidote to hype is focused on concrete, credible use cases to focus on value. And there’s plenty of concrete. I’m not saying predictive is better than generative. It’s apples and oranges and generative definitely is valuable.
And it’s also a lot of fun, right? It’s like the coolest thing ever, right? To play with and whatever, if you’ve got any free time. If you’re a customer…
Or one of your customers sort of wants to believe something that’s whimsical, right? They may also believe that if they hop on one leg and say a special poem, you know, that they’re going to have good luck and that’s fine. They can believe whatever. There’s only so much you can do, but if you bring it back to concrete value, what’s the use case, right?
The problem with disillusionment is that the baby gets thrown out with the bath water, right? That it’s very costly. It’s not just an inevitable up and down. First of all, how far up and how far down, right?
Everyone looks at the Gartner thing as if it’s inevitable, but it’s not. We don’t have to be irrational. And even if we do have to be irrational, can we limit that to only being a little irrational? You know, it means that the credible value propositions of both generative and predictive AI will be undernourished at some point when the disillusionment hits too hard, when there’s a reckoning or even a winter or what have you.
It’s just about… It’s just about the fact that we’re not going to be able to do anything about it. And the other thing is that you can have your cake and eat it too. You can have fun with the amazing capabilities, the unprecedented capabilities of, let’s say, a large language model in a way that’s potentially valuable.
And I’ll give you one example is that we’re integrating it into our product because this is exactly the kind of very well-defined, formal, finite scoped arena within which these large language models are very reliable. It’s quite remarkable. If half the purpose of our product is to bridge the gap so the business user can see a new two-dimensional graph, right, or chart that they’ve never seen before, maybe they’re the first-time user, it’s not the rocket science part. It might take a few hours to get you comfortable, what have you.
Wouldn’t it be nice to have a chatbot that you can ask any questions? Hey, why would I want to move this lever? What does this mean? Can you remind me again?
And why does this happen? And it’s like a senior data scientist who’s been well-trained, well-caffeinated, slept well. I mean, it’s incredible, right? I mean, and it’s definitely helping bridge that tech business divide.
So when you find the right scope that we’ve kind of landed on with this in particular, large language models have incredible value and their performance is above the bar. Actually, Eric, thank you very much. I wanted to ask you where the founders, because that’s what we do at the end, should now focus on after debating what’s really, really wrong to do and what to expect from these models. But I think you’re right.
You already gave a very nice summary. The last thing I wanted to ask, this is a bit off script, but since the topic obviously now is a bit more around AI agents with our upcoming podcast, is there anything you can share based on your experience and your thoughts, what for you AI agents are and clearly are not? Yeah, this whole agent, the agentic buzzwords so far is rubbing me the wrong way. I feel like the whole idea of agents is, yeah, we’ve got more than one CPU or GPU out there.
We’ve always known that. We’ve always had the ability to replicate software and have it run in parallel, what have you. Does that really make a qualitative difference in terms of what’s ultimately some kind of ceiling on the capabilities of generative AI, right? How much of the human mind can be reverse engineered just from learning from the written word?
A lot of written words, right? But it’s still just that one arena. We’re not going into our neurons or other aspects of human behavior in general. Well, apparently a lot can be reverse engineered, but there’s going to be a limit, right?
So the idea of the things bouncing ideas off one another and collaborating and all that kind of goes hand in hand with a big part of the mythology, which is that ultimately, once you reach a certain critical mass of quote unquote intelligence, whatever that means, that you’re going to break through the ceiling and it’s ad infinitum. It’s just sort of as if intelligence is a platonic ideal that exists outside of humanity. And we just need to get the formula right. And then suddenly it’s got its own spark of everything.
So I think that agentic narrative partly pulls into that. And that’s a mythology thing. The same time, you know, you have people out there saying that AI is different from other technologies because it’s an agent. It has agency instead of being a tool.
I think that’s a hype narrative that I don’t really know what agency means. I mean, how autonomous it is, is just up to the humans, right? Do you let it make decisions without a human in the loop? It depends on the use case.
Yeah, I think that’s a beautiful summary. And especially for founders. They should really, really listen to that before they start a business and call it AI agent or agentic business. Thank you very much for that, Eric, and for all your thoughts.
Yeah, thanks for having me, guys. We still need to get the song and the dancing in here. Yeah, yeah. Go watch my rap at predictthis.org.
It’s old, but again, predictive, the concepts still apply. So I really appreciate it. And I love getting into the philosophy stuff here about just, you know, about the AGI stuff. That’s a tough one.
That train has left the station. And I don’t, it’s going to slow down eventually, but. You probably didn’t listen to that podcast, but Lutz and I had some fun discussions around AGI because we were basically sitting there trying to explain to ourselves what it would mean. And we also didn’t get to any great conclusion because it’s just, it also doesn’t make any sense, technically.
You won’t get that. I mean, maybe in a few thousand years. I don’t know. Cool.
Thanks so much. Awesome. Thanks, guys. Thank you.
That was really, really cool. I really liked his honesty. Lutz, I mean, usually you read all these AGI and agents are the future, but here’s someone who really understands how technically it works as you, but also has seen it over the years developing and now building his own company around it. So I think for founders, that was something very, very important.
I think for founders, my biggest takeaway from this for founders is actually the fact that many applications will, will be just a prediction. And that prediction will need to be integrated. So if you think you can change one piece of the workflow by having a cool tool, and now you need to raise funding for it, think again, because most likely companies don’t change that fast. So either you change a product or an industry or something completely.
If you just become a piece, then McKinsey is your friend. Or BCG. But we will. No, but.
Don’t worry. We will still speak to some companies who apply AI agents and let’s see what they tell us. It was a pleasure doing the interview with you Lutz and talk soon. Yes.