00:36:55 · Matthew Fitzpatrick
[00:36:55] S0: Look. I I I think this is obviously a topic that gets a ton of coverage in the media at the moment. And and I think one of the reasons it does get a ton of coverage is you've had a lot of announced restructures, etcetera. I would say that with most of those from what I see at least are more what I call AI AI washing than reality, meaning folks that overhire during COVID are using AI as a reason to explain why they're changing their businesses. And so I think that's one reality is that most of the numbers announced to date have not been reality. But I think if you actually look at the next ten years, the main thing I would anchorize what is... Jevons paradox, but the idea that most technologies that roll out actually create a lot more usage. And so as a result of that, a lot more jobs come from that. And so, you know, the example would be, I think, one of the best ones I've heard is, like, bank tellers. So, like, fifty six years ago, everything in a bank was done manually. You then had a bunch of software that ought to be at the vast majority of jobs in a bank branch, but you actually have more bank tellers today than you do back then because people rolled out a lot more bank branches because they're useful. And you could do the same thing with, like, for example, lawyers. Like, the law has been largely digitized the last thirty years, and there's more lawyers today than it was thirty years ago because the law became more complicated and people wanted more sophisticated legal work. And so I think that paradigm has existed in many ways. And I actually think... Here's how I think about it specifically related to AI. Most of the usage, the vast majority of the enterprise usage so far has actually been coding. And if you thought of software as, like, the primary input cost of most businesses, like starting a new company, big obsolescence, can I build a software infrastructure? I think the Jevons Paradox example here is gonna be... It's gonna be much easier to scale technology and software. And so the amount of new businesses, new offerings that scale up and around that is gonna be much, much bigger. And so I think that the nature of work will change quite materially. Like, I think you'll have way more small and urgent fast growing companies. And, you know, I I think that's what... You know, one third of GDP growth last year was AI spent. Right?
00:39:07 · Matthew Fitzpatrick
And so I think you're gonna see a similar paradigm here where we will move from a society where you have bigger companies and a lot of people doing hyper manual work, like gathering information, putting into PowerPoint slide, Two way society where there's a lot more people doing things like selling, designing products, interacting with patients, whatever those topics are, using a lot of the AI tools. And so a lot of the manual administrative work will go down. But, you know, I actually think that's a really good thing for society. Like, the example I I give often is if you take health care, for example, in The US. So, like, our average cost per patient US is, like, 12 to 13,000 per patient. In Germany and Canada, it's, like, 3,000, and about a third of our cost is administrative cost. So, actually, like, bringing it down where where it could be much more about the higher value elements of of health care, it's a really good thing for society, particularly as our aging population means we're gonna have a lot more living people that need care. And so I think all of those evolutions are going to happen. You know, the the often cited example is, like, The US went from a 40% agricultural economy to a 1% agricultural economy, then manufacturing economy, a service economy. And and I think that one of my favorite stats, and if you look at any high school class, it's something like 30% of those people end up in jobs that don't exist when they're in high school. So, like, you just have Interesting. Huge evolutions in the nature of work in every decade. And I think this will be a one where a lot more people will be in creative entrepreneurial roles as a result.
00:40:40 · Guillaume Moubeche, Matthew Fitzpatrick
[00:40:40] S1: And, you know, like, I think Elon Musk was talking about, like, an era of abundance where people don't have to work, etcetera. Do you feel like... Is this a utopia, or do you feel like there is a world where that exists? [00:40:54] S0: I don't have a strong view on that. I mean, in my mind, we're, as I said, about a decade away from even widespread adoption enterprise. So I think... Ask me again if I may. We... I I I think most real dislocations in job markets occur when the change is so sudden that the labor force does not have time to adapt. And I think the reality is if you think this is all changing next year, then that's a real concern. It's not that. We're a long way away. And so I think you're gonna see kids coming out of college now are gonna have five, ten years to adopt and adapt, and that will be that will be. [00:41:28] S1: My feeling is, like, people are actually working more, you know, with AI and not less. Like, it's... Yeah. [00:41:35] S0: Well, the n b... The NBER study I mentioned earlier would suggest that. I mean, if you basically have a conclusion where everyone's doing 15 to 20% more work and then organizational productivity staying the same, it would suggest exactly that, which is a lot of people are doing more, but it's not really changing the nature of work yet.
00:41:35 · Matthew Fitzpatrick, Guillaume Moubeche
[00:41:35] S0: Well, the n b... The NBER study I mentioned earlier would suggest that. I mean, if you basically have a conclusion where everyone's doing 15 to 20% more work and then organizational productivity staying the same, it would suggest exactly that, which is a lot of people are doing more, but it's not really changing the nature of work yet. [00:41:51] S1: Interesting. And I think on the the on the show with Harry, at some point, you pushed back, you know, on the synthetic data thesis. So basically, that human feedback will become, like, irrelevant. Can you basically, like, walk me through why the math doesn't work?
00:42:12 · Matthew Fitzpatrick
[00:42:12] S0: Sure. I I think there's a couple different dimensions for that. So let's take... First of all, you know, I think there's things like commodity labeling, like cat dog, cat dog, where there's enough data around that you can build synthetic data that exists around it. But if you think about the matrix of language, culture, and expertise topics, so, like, seventeenth century French architecture, but have that person be from... Who speak Lebanese, who speak French with Lebanese accent, French Canadian, or France. Like, those are all very different contexts. And very... So that matrix of topic culture language times all the the kind of potential topics on Earth is just a massive amount of data. And so I think what you've seen is actually most of the publicly available data in the world. And, by the way, part of the reason coding has matured so much is coding is probably the most easily accessible data. There's a ton of available data you can train off of similar to video games. But but if you think of all of the other topics that exist, first of all, there's enormous amounts of kind of future training on that. And, you know, there's variety of different podcasts where you... Like, one in particular I can think of where AI researcher was describing. Like, in the age of agents, then if you take hyper specific tasks, human data becomes more important than ever because you need a way to validate that specific task. So it's not just it's not just topic language culture. It's all of that then for a hyper specific task where there's no really good existing data. So an example with... For that would be if you wanted to make a decision around processing a insurance claim for property and casualty insurance in Florida, whatever it is. You know, you you need enough precedent data, enough comparable to, like, what should that 20 page document look like? Does it look accurate? And that's one of, you know, thousands of examples. So I think, you know, it's the topics. It... And then I think the the final thing I'll say is actually the the concept of chain of thought reasoning. So moving to a world where it's not just a simple answer or an annotation, but has the thought process looked accurate, I think you're gonna want human loops validation for that for a long time. And I think that's where the enterprise, you know, is in a complicated place. If you're...
00:44:25 · Matthew Fitzpatrick, Guillaume Moubeche
Let's do medical coding, for example. You can train a a model to do that, but then that's gonna have to go, you know, have regulatory reviews, etcetera. So you actually do need human loop validation. And so I think the answer for this is that humans are gonna stay in the loop of this for a long, long time. Like, I think the same way if you take the mortgage example I gave earlier, They were 20 into mortgage underwriting. About 15% of the... 10 to 15% of those mortgages still have a human loop. And that's where a lot of these tasks will move is you'll move to a kind of a AI driven workflow with human loop validation testing and and triage as needed. And I think that's gonna be persistent. And I think you're gonna want that... Like, if you had a model that worked in health care and then GLP ones came out the next week, you probably need to change that model, and you're gonna want humans to help you fine tune and evolve it. [00:45:18] S1: Yeah. Makes makes a lot of sense. And I think recently, you made an acquisition. Can you elaborate? We