#422 AIはその文書を読めない — Valantor CPO Neil Katz が語る、企業AIに欠けている「文書理解」レイヤー

Let's go. You're listening to Making Data Simple, where we make the world of data effortless, relevant, and yet even fun. All right, podcast listeners, my name is Catherine Nain. I am stepping in for Al Martin today. You may know me as the executive producer of our podcast, Making Data Simple. And today, I am so pleased to be speaking with Neil Katz. He is a chief product officer at Valantor and co-founder of iLevel AI, the company behind GroundX. Before that, he helped lead product and AI at IBM Watson through the weather company, building conversational AI used by hundreds of millions of people. And before that, which I can't wait to have you talk a little bit about, four Emmy Awards as a journalist. Did I get all of that right? You did. I know it's a bit of an unusual background in the tech space, but we can talk about that journey. And actually, before we do, I wanted to say something funny to your listeners, which is you guys are listening on audio. You can't see what I see.

But when I first got on this Zoom or this call with Catherine, there's a caption above your head, and it says Al Martin, but it looks like it says AI Martin. I know. Isn't that hilarious? And I was playing the first 15 seconds. Is this a human being in front of me or not? You might have noticed I was speaking a little bit slowly. and I was trying to enunciate correctly. So if you were AI Martin, you got what I was saying. But I'm glad you're human Catherine, so good to talk to you. I think that's hilarious, and I am actually human Catherine. So tell me a little bit about yourself. Well, you know what? As I said at the top, a bit of an unusual background for the tech space. I actually started my career in technology when I was in college. I created one of the first digital design firms in New York, maybe in the country at that time, for Web1, just to age myself. and built that up over about seven years, sold it, exited that company, and decided I was done with technology. I did not want to spend another night, dark night, until 4 a.m. looking at a computer screen.

I just didn't want to spend my life this way. I've been doing it for a decade, and I want to do something different. So I retooled, went to grad school, became a journalist, started my career in India with my fiancé at the time, my wife today. Thankfully, that worked out 20 years later. and got to travel the world in journalism, reporting from India, obviously Iran, which is now, of course, in the news quite a bit, lots of other interesting places for the New York Times, produced television for CBS News, built the digital news operation at the Weather Channel, which is oddly enough, even though those are impressive news brands before the Weather Channel, the four Emmys come from my time at the Weather Channel where we built a documentary unit that covered climate change all over the world in ways that you would never expect the Weather Channel to tackle the issue. It wasn't about melting icebergs. It was about how, for example, increasing drought and heat was making migrant crossings in America's southern border very dangerous. And several bodies a week would show up in Texas, and they didn't know what to do with these dead migrants or how to save them or how to protect them or even what to do with their bodies, frankly.

So things you wouldn't expect us to cover, deep investigative journalism. And there's the four Emmys, how we got there. And then something unusual happened. IBM purchased the Weather Channel in one of the odder acquisitions in history, but I was an IBMer for about five years. You know, it was actually great because you're laughing. But it was great because we got brought into the Watson division, which was a great place to be inside IBM. And, you know, then I was back to hard tech again, running a large product team at the weather company, responsible for delivering weather data to 200 million global users around the world on web and apps. And we just had an enormous user base at that time. It wasn't just consumers, but we were also delivering weather data to every Apple device in the world, every Samsung device in the world, every Google device in the world, refrigerators that had weather systems, started plants, everything that had weather data in it, we were doing it. And I met my current partner there, Ben Fletcher, who was a long-time IBMer, along with me, who had made several tours as a senior engineer in the Watson Group and in the Research Group. You know, I always say that I would probably say the founders of Allentor have 21 patents between us,

which means Ben has 21 patents and I got zero. That's okay, I got the Emmy. That's why teams are great, man. That's why teams are great. She brings me to the equation. Several years before ChatGPT, you know, blew up and changed the world, you know, at one time, Watson was the bad boy on the block. He was the best model in the world. It was the best model in the world. And we had built something called Watson Weather, which let millions of people every single day, you know, quote, unquote, talk to your weather and get a forecast. Is the hurricane coming? Here's my zip code. What's going on? It wasn't always right. These were very early models. but when it was right the engagement was absolutely skyrocketing through the roof you could see the hockey stick we could see it back then this was you know several years before chat gpt went to the public um and so we left we left actually ibm to build uh i level which then became valentor acquired by valentor and we left it with to solve a really specific problem that was actually happening every day at ibm and then became a much bigger problem when when a when the GPT class models took over the world, which is figuring out basically how to connect a model

with a company's private information. And most of that information is stored in documents. Those documents tend to be visually complex, things like tables and graphics and strange layouts that are very easy for humans to understand and very challenging for computers to understand. You know, every AI project would sort of struggle in the outset. Big dreams, big ideas, good model, okay, and we teach you the things that matter inside our company. And so we built iLevel to solve that problem, train your own vision models to break down documents into small pieces that we can then run through smaller open source models and turn into AI-ready data. You know, what we built eventually became something called RAG, Retrieval Augmented Generation. We built it a couple of years before it had that name. And, you know, ever since, we've been helping all kinds of amazing companies like Air France and Samsung and, you know, a couple thousand devs around the world with great tools to turn their, like the glue that sits between an AI model and your company's data and help you build really accurate and with the IBM Heritage really secure applications because

they run completely on-prem, even air-gapped if you want to from the ground up. So security for us is not an add-on. It's core of the architecture, and I'm kind of grateful to our time at IBM for giving us that mindset from the outset. That was a long answer from my background. That was incredible. So you go off and have this amazing journalistic career, and then you get, it sounds like it. It sounds like you got a lot out of it. I mean, those are heavy stories, though. That's incredible. And then you come back into the technology, and what I'm hearing you say at your core is that you're a sucker for hard problems, and you like using technology to solve them. I think that's a pretty good synopsis, actually. Yeah, yeah, a sucker for hard problems and likes to solve them. So how does the technology work? I mean, I don't want to give away any trade secrets or anything, but especially with the proliferation of the different models we have today and all of the recognition, I think the secure by design is hugely differentiated, and I'm sure there's a bunch that goes into that.

But when we think about how it understands the information that's presented to it, right, how is that different and how does that work? Yeah, so we came at RAG from a first principles approach because it didn't exist at that time. There wasn't patterns in how you're supposed to do it. Great point. And so, you know, the first part of this document, it became clear to us with the document problem, the models just couldn't deal with, at that time, look, we were talking about GPT-3, 3.5. Obviously, it's not a FABO class model. This is years ago. It was just very obvious that, and it's still true today, that the models would just choke on complex information. And the larger the document is, the more dense the document is, and the more visually complex the document is, the more opportunity is for models, even with large context windows, to either make mistakes or drop things out of the middle. And so we took a really practical approach. We said, well, how do we, let's just break the thing into smaller pieces. If the job is too big, let's create small jobs. And so we trained a vision model on nearly a million pages of documents.

That was a really tedious, boring 18 months, I have to say. It's not fun. You know, just drawing boxes around tables and graphics and text blocks on millions of pages of documents. And that helped us build a great unique data set that looks like the corporate data set that you'd see in most companies. We started out actually with training document models, like document parsing models, but they weren't really that good. And we actually wound up hitting on using a video model, which had much higher fidelity. It was kind of an unexpected term. And so we landed on training video models that are designed to detect the soccer ball moving around the screen or the person on the bicycle. And oddly enough, you can train them with document data. They don't really care what they're looking at. and we're able to get a great result out of that. So like a big part of our pipeline is take documents, break them into every, I don't know if there's a video here I could show you, but our vision model basically takes every single page, identifies where's the table, where's the graphic, where are the text blocks. And once we have those locations, those positions, we basically capture them, those objects, run our agentic pipeline with a, as I said, smaller open source VLM.

So you can run them on-prem near your data. We get phenomenal results. And we pair that with, you know, proprietary search that is like holistically synchronized with how we do the ingest so you get better results on the output than folks that just do build a stack piece by piece. I know IBM builds their stack piece by piece. Most of the industry does. Totally fine. We just take a more holistic approach to it. We search by, you know, text, vector, and graph all at the same time, kind of built into the system. And, you know, the end result of all that is a RAG system that gets a really, you know, 98% accurate extraction in your documents if you want to use it for that. accuracy on search that doesn't fail at scale. So most folks don't realize that once you scale a typical vector system to about 10,000 pages, performance degrades. Most people don't know this yet because most of these systems don't get to 10,000 pages. No. So in pilot mode or test mode or whatever. But we've got customers in the insurance space that, you know, a single insurance claim is 10,000 to 20,000 pages, and a corpus of them could be tens of millions of pages. Oh, and when you think about doing due diligence or illegal cases as well,

I can see how that would scale into the thousands and thousands and thousands of pages. That's amazing. Yeah, millions. Like a typical corporate litigation right now is like five million pages of documents. Oh, my gosh. It used to be not that way, but because so much more information is generated in a company now, the litigation corpus has become enormous. You need a way to deal with that. And most AI search systems start to fail at that scale. And the reason for that under the hood is that when you're doing similarities, So when you're in vector search, which is typically a similarity search, you wind up with too many things that are similar. It sounds almost silly, but imagine if I had a jar of 10 marbles, and they're all a different shade of red. And I said, hey, AI computer, find me middle red. It's going to do a good job. If there's a million shades of red, a million marbles, each a different shade of red, and there's a million shades, and I say, find me middle red, there's going to be a lot of results that are pretty similar. And what happens in document corpuses is a lot of blocks of content look pretty similar especially document to document There might be many versions of documents There might be people responding to a document And so if you wind up with things that look too similar

you start to have a lot of failure. It's like the bigger your data set gets, the worse your results are. It's the opposite of what you would think. So we do a lot of work in the chunk creation. We call them semantic objects. Differentiating each of those objects, so they're far more differentiated than a typical chunk because we're stapling a lot of metadata to every chunk. So when we do our search, we just get much better results at scale downstream. There's a lot in there. I could talk about the security piece, which I'd love to get to, but I want to pause there for a minute because I threw a lot at you. That's fascinating. So what I heard you say was that you put all the effort. Now I just have this image of you guys sitting there hunched over day after day of parsing a million documents. That's hilarious, right? But that sounds so tedious and so long. You weren't the only ones. But the upshot of that is that you have a really good understanding of the different elements of a document, which has then been applied to be able to work at scale. So after 10,000 pages, which is a really interesting number to have a cutoff, but then when you go into millions, I mean, that 10,000 is nothing. So the results get worse and worse and worse counterintuitively as you go on

because there's more things that are counted as similar unless you get better at truly understanding what is in that document and what you're searching for, right? You're basically getting higher fidelity at understanding what's in a document. And I think in our show notes when we were preparing, you called it a perception layer. And I think that's – is that what you were saying here? Is that ability to be more understanding of the nature of that document? Yeah, we have a couple of ways of talking about it. We use the term visual intelligence. You're trying to get intelligence on the visual components of the information in your company. This perception or intelligence layer that we're building on top of your document core is another way of describing it for customers. And, you know, on the security front, we just have a really core belief at the company that the world's most important information does not live on the public Internet and never will. And so there's just vast amounts of information that have never touched a model, will never wind up in something from OpenAI or Anthropic or any other streaming model provider.

And it's crazy to ask regulated companies. I know IBM obviously works almost exclusively with regulated industries. It's crazy talk to ask a regulated company to send their data across the Internet, the stuff that they've been protecting, they've been building security around for the last 30 years. And because you want some cool AI features, you're asking them to transmit that data across the Internet. That's just crazy talk. Hopefully that's not going to happen to your banking records or your payroll records or your health care records or all the important information in the world. It's fine for Joe's flower shop. It's fine for a million startups. But for larger companies or regulated companies or secure-focused companies, we think you've got to bring the AI to the data. And so we built the whole architecture so you could run it next to your data. You can run it, obviously, in your managed cloud. You can run it in your colo data center. You could run it in the basement with the old IBM iron mainframes if you want. You could run an air gap and unplug the Internet. And it all works just as well from the ground up. And you can do it at scale for these documents. So do you essentially retrain the model? You talked about using open source models. Do you retrain the model on that company's corpus every time

using the proprietary visual intelligence that you talked about? Is that the general idea? You actually don't have to. First of all, you can use a whole wide variety of VLMs and architecture of your choice. At the moment, we put a lot of customers on Gemma 4. Some customers have preferences. Some people love Chinese models. Some folks, for security reasons, don't want a Chinese model. You know, we take a neutral position on all of that. Customers get to decide. Because we've trained our vision model, which sits in front of the LLM or the VLM, We actually don't need, it's very unusual that we need to retrain things to understand a company's documents. Just something that's super specialized, yeah, we can come in and we can retrain our vision model on your industry. Like maybe you've got, you know, airline engineering diagrams, which are really specific to what you guys do or something like that. It's a good use case that we can go in and do a fine tune of our own model and give our customer a private version of the model. In most cases, the system really works out of the box. what we do wind up doing with customers is what we call custom workflows and all that really is

is our whole ingest platform you can twist every dial in the whole system if you want to about 100 parameters of that ingest that happens and if there's very specific things you want to pull out of your document specific extractions a custom workflow just lets you define what those fields are but the cool part is that we don't care what the documents look like So we've got a lot of customers, but half our customers use us for like end-to-end rag, right, AI search on complex documents. But half our customers use us for what they would call AI OCR, and I kind of hate that term. I kind of hate that term because it sounds pretty old-fashioned. We're doing OCR, we're not. Yes, it does. It doesn't sound like it's OCR at all, but, yeah, I take your point. But a lot of customers go, hey, I've got a lot of documents, and they have what we would call document diversity. It's a tip of props. It's a problem you see all the time. Let's take contracts, for example. Whenever I onboard a customer, this is something we see a lot. I onboard a new customer, and I have to take in, you know, a couple thousand pages of contracts from that customer to onboard them.

Well, all those contracts look different customer to customer, but the information I have to pull out usually looks the same. It's usually the information about the services or information about the legal aspects of the services or the liability of things or it's going to be industry-specific. Our whole thing is that we train a custom workflow, and it doesn't matter how diverse those documents are. You don't need templates. We don't care what the documents look like. You're going to get the same super high quality results downstream no matter what. How does it not lose the thread? I mean, when I think about my work with when I use AI, it will lose the thread. It will get off track. It will forget. How does that work? And so that's one question I have. I'm going to stack questions, which you're never supposed to do in a podcast. I'm going to stack questions to make it absolutely terrible for editing in the background. But that's one of my questions. Another one of my questions is how much energy does it take to run this? Like what kind of infrastructure does it take? Everything like that, because it sits alongside the data. And this is a huge corpus of data. So how much energy and effort does it take to do this?

And then what's the overhead for training people to use this software? And trusting the results, right? Because one of the big inhibitors to, I think, everybody using AI is that they have to trust the answers that they're given. So how does that work with people? So I apologize to my post-production team right now, but I'm going to stack those three questions together because you get into a flow and it's beautiful to listen to. So go. There might have been even four or five questions in there. So starting with long context. So look, underneath the hood, this is fundamentally a RAG system, right? So as we are extracting information from documents or converting documents into A-ready data, that gets stored in our systems. So when people then have a downstream question or an agent has a downstream question, you're basically doing a search across your focus to get the correct information out of the system. So that typically isn't a problem with what we're doing. In terms of energy usage, I don't actually know what the energy bill is. We haven't tracked that. But you can run this on regular NVIDIA chips.

I mean, you can run it on, you know, they don't even have to be state-of-the-art. H-100 is totally fine. you're going to get great speed and great results. Depending on how much speed and how much scale you need, you may have to stack them, starting with literally one chip, and you're going to do fine for most applications. We even have folks that run on, and actually this is how we used to run our cloud, on Tesla chips, which are the generation before the H series, older chips, stacking a couple of those. And, you know, we make it run in part because, well, we've been doing this a while, so we're efficient at how we run the chips. But also we try to promote the use of, you know, smaller models. like a Gemma 4, 31 billion parameter model. You can run that on a lot of devices these days. And so, yeah, what your electricity bill is, I'm not sure at the end of the month. Well, I guess I didn't put that together. That's a really good point. You're using a small model, so it's not going to be that intensive, but you still do need to use a GPU, but it's not having to be state-of-the-art because you've got the efficiency in there and you've got the ability to make use of it, right? That's right. It still has to be industrial-strength GPU,

so we're not at the point where you can run this on your phone or your laptop. You do need to run it in a server somewhere. But if you're a regulated enterprise, you know, you've got machines or you're provisioning them from a hyperscaler. Sure. So it tends to not be a problem. There was a question. It was four. What was the question for you after that? Human adoption. Oh, how to train people. So on the – and then how to trust the results. We have two really, really important topics. So I'll do them one by one so you're ever showing up. So in terms of how do engineers learn to use GroundX, our core technology, honestly, it used to take a bit. But what we've done in the last six months is we've built a harness. A lot of people are building harnesses. We've built a harness for GroundX that connects to any of your AI coding tools, Claude, Codex, Cursor, whatever you like. And that combined with an MCP, and it completely understands how to build, test, and deploy GroundX-based applications.

And so we've got medium technical or even some non-technical people in our own company that are building, testing, and deploying applications now with the GroundX harness connected to the Claude or the Codex. So a lot of that complexity has just been sort of melted away, and that's why easy ad-coding tools are fantastic. And we just kind of sit between them and your documents to make it actually work on your information side, your company. So that's the thing. Yeah, so what I hear you saying with that is that you use Claude or whatever you'd like to use to be able to build the workflows and do the coding that's necessary, and then you go implement in your AirGap or on-premise system, right, to preserve that security? Yeah. Okay. Yeah, that makes sense. So a recent customer, the job was an onboarding team. We're spending an onboarding team at a big regulated company in the financial services sector. They had an onboarding team that should be spending their time doing warm handshakes. Hey, how are you? welcome to the company. We've got these great services. I'm your representative. I'm going to be your partner. We're going to walk through this. Maybe I'll help you upsell some things later on. The nice hug, human hug of a company.

But they were spending 80% of their time, this team, literally onboarding the contracts from the customer and trying to extract data, about 150 fields out of every contract. So they had 80% of their actual time was spent doing what I would call kind of document grunt work. And 20% was doing the work they were actually hired for, which was the highest value human work, which was where all the value was in their job to the company, but they had to spend all their time doing the other stuff so they could get to the high value part of their job. They had to do the homework so they could get to the part that mattered. We see this a lot, actually, with people almost doing the wrong job inside a company, not what they think they should be doing, but what they have to do to prepare for their actual real human work. And so working with them we said this is actually how it runs We say look give us five sample documents Give us a golden data set So you have to have the answers prepared already What are those 150 fields that have to be extracted from those documents Give us a description, literally in JSON. We work with you. If you have

descriptions, great. We work with you. Give us descriptions of what these fields are, what you're trying to extract out of the documents, and we will basically, or you can do it, or we will basically put it into Claude, run it in a loop with our harness, and say, Claude, talk to GroundX. use these sample documents use this golden data set to test your work and use these descriptions of the fields they want to extract and run it in a loop until you get to at least 90% accuracy on this extraction set and it just churns for like an hour and then boom, you're there so that's kind of like how easy it can be with these tools to get where you want to go I will say, the bad news is the last 10%, the last mile you cannot use LLM as a judge to create its own homework. Everyone uses LLM as a judge and it's okay as far as it goes, but that last 10% is pretty shaky on. And so that's why we don't generally say, hey, keep looping until you get to 99% because it's not going to actually be accurate. Loops, you get to 90% and then we have to do a human layer a little bit.

Us and the humans of the customer have to review the final ones that, that have low confidence and put in those corrections to make sure things are just right. So it can be increasingly a pretty automated process, even to use the tools to build the software that you want on the far side. How is shaky defined on that last 10%? What do we mean by shaky? Yeah, so if the results don't match the golden data set, right? So you've got, you know, your golden data set is the answer key. Okay. And it's returning things that don't match the answer key, and we would expect it would. Correct, yeah. And so you have to make the manual changes, the fine-tuning to be able to do that. You do, just because the LMs just aren't that good at that final piece of that. No, that makes sense. I just wanted to make sure I was clear on what shaking meant in this case, but that sounds totally reasonable. And now, going back to our original persona, who was spending 80% of their time trying to parse that document, they now have evidence to see that we got 90% of the way there using this brilliant automation, using all this powerful technology. We do the human intervention because humans matter.

Like, I keep coming back to everybody. Humans matter. We still got to do the human things to not only capture that last 10%, but do the handshaking and do the onboarding, right? That's so powerful. I was talking at an event a couple weeks ago, and it was about AI and the future of work. And there's about 200 people in the room in Atlanta, great room. And everybody was really scared. And it makes sense why they're scared, because the leaders of this industry, of our biggest LLM companies, keep saying crazy things. You know, they oscillate between – their basic sales pitch is, buy our products, we'll fire half of humanity and kill the rest. Win, right? Everybody wants to go on board with that. Sounds great. And they're confused why people don't want a data center in their backyard. Right. What on earth is this pitch? But I think it's BS, actually, because what we've seen now, customer by customer by customer, is not people getting fired because we've automated something. It's people actually doing the part of their job that has value for the company.

So I gave you the contract, the onboarding team example. Let me give you another one from insurance. So we have a product called Fraudex, which we built on top of GroundX. And Fraudex is a point solution that helps insurance carriers identify evidence for fraud inside their claim files about 40 times faster than pure human review. So, you know, what is an insurance claim? Well, if you're dealing with general liability claims, let's take people, someone fall down on a construction site. you know, that claim is going to be 5,000 to 10,000 pages of documents. A typical investigative team at a carrier is like 12, 15, maybe 20, but it's like not that many people, dozen to dozens. And each one of them may have to, in theory, read more than a million pages a year. In reality, they cannot. No. And so you're spot-checking, you're scanning, you're scraping, you're, you know, traditionally. You're using history, experience, intuition. Yeah, maybe doing keyword search. You're just like, you're typing in the dark, basically. It can take a, to review those 5,000 pages, you know, that's like a week of work, right?

It's like 40 human hours. It's spread out over different times, but it's like 40 human hours. So we've now taken that down to an hour, an hour of human review. Our AI, basically our AI agents go and search for things that investigators usually search for. We then, you know, build the report for you. And importantly, this gets to the trust question you asked a minute ago. We show you exactly where every one of those claims comes from in your document set. We are huge believers of humans judge. AI searches and finds and presents. We think humans still need to judge at the end of that flow for critical things. And an insurance claim is very critical. Getting it right could be life or death for people. Getting it right could be the difference between getting paid out on a terrible accident immediately or being tied up in court. And for the insurance companies, if a claim is fraudulent, they don't find it. That's millions of dollars every single time. So these are important decisions. But we find is nobody's like, hey, this is great. We can automate 80, 90 percent of this work now. Let's fire our investigative team.

It's wow. The investigative team can finally actually cover the amount of documents we've given them in the first place. They can do the job without hiring like another hundred people, which the carrier is not going to do. Investigators love it because they've got like robotic like our, you know, like Mr. gadget hands now, and they can spend all of their time judging this evidence we've found. Is this valid evidence? What is it indicative of? Do we believe in what should we do next? Should this be paid out? Should this be sent to a litigation? Should we call the FBI? Those are human decisions, not computer decisions. And they can do it now across the whole claim corpus instead of just a fraction of them. We see the same thing in audit. What are audits? The whole audit universe is spot checking. If you do, if you're a typical audit company, you'll audit 5% of whatever the document load is. So they're just getting lucky or unlucky. Like that's crazy. Let's audit 100% of everything and let the humans judge whether the audit's good or not. We're strong believers in the future of work is human, human, human, but all of us, you know, doing the part of your job that actually brings the value to your company. I was looking on the site. I was trying to find

the MIT article that came out earlier this year in March, 2026, I believe it was. It was just brilliant. And it was on the idea of new work and what it entails. And I think this is a beautiful example that there is no finite amount of work. No company has ever said that there is a finite amount of work. So it's not like it's a fixed pie that we're trying to scramble after shares for. The pie keeps growing. The nature of that work keeps changing. I was talking to a friend of mine And we were talking about like a database engineer, right? And to your point of what you're talking about here, we're offloading the parts that doesn't need that human intervention, that human touch and really elevating and giving the space to the humans to not only have more volume, potentially more volume of work that only humans can do, right? So that it may be 10% of what it was, but now it can go 100% of the time because you get 10 more cases that you can work on. But I think that's a really great example. And I don't have fear that there won't be enough work to do.

My concern is making sure that people are aware of the work that they can be done and have the education and skills to be able to participate in that economy and participate in that market. I agree 100% with that. I mean, if you, you know, look, the computer was introduced into the office, what, 40 years ago now? Yeah. Did it reduce the amount of work? It created work. It just, like, generated. Like, all the work we're worried that's going to go away today literally didn't exist. got created because the computer went into the office. And now you've got people who are like data analysts and spreadsheet jockeys and a million jobs that exist because of the information revolution from computers. The bottlenecks just moved to some new place, and the humans wound up filling in where the bottleneck was now. We just see the same thing happening. The bottleneck just moves inside a company, and humans have to fill in the gap where the new bottlenecks are. I think one of the things that is unique, though, is because AI is evolving really quickly because it's mostly at the software layer right now, not the hardware layer. Obviously, hardware is the ultimate speed limit of everything,

but so much innovation is happening on model development. It might require a lot of flexibility by humans and organizations in general to be like, let's just keep chasing the bottlenecks and dispatching humans to go work on the next bottleneck that we have in the company. The most obvious place you see that is in software development, which has been ironically hit first by, you know, AI displacement. Yes. And it's generated some panic in some corners, but it creates no panic for me because, you know, for a couple reasons. You know, number one, the thing you see about the bottleneck is you see the software engineers wore the bottleneck forever. So the product people had a ton of time on their hands to design screenshots and then discuss, you know, focus group screenshots and then make prototypes and then focus group the prototypes and make plans. And there's a lot of stuff happening over here on the non-software engineering side. And the reason was they knew whatever they came up with was six months away from the engineering team. So they better get that decision right, first of all.

So you better do all this preparatory work about user surveys and maybe testing and just getting that right so once you make that decision, engineering can go build a thing. But you also knew you were six months away. Now that's compacted. That might be compacted to six weeks or six days. Sure. And now you've got a bottleneck on the product side of, like, what to make and how to figure out if this is what users want. And, you know, that sort of hasn't been solved yet, actually. AI hasn't really gone in there and solved what we should make for a living. That's still a pretty human job. So the bottleneck just completely moved. And now you've got to surge people into that part of the problem while the engineers are, you know, have this, like, you know, spinning wheel tool. Everyone else doesn't quite have it yet. It's going to be our imagination is the limit. So I'm glad you said that. because that is actually the constraint that I think is creating all the fear. It's very easy to see what parts of your job you won't be able to do anymore. You won't have to do anymore. You won't be doing anymore, whether you want to or not. Yeah, it's not going to be there. Right.

Imagine how hard it was in the Industrial Revolution for, I think it was like 90% of Americans worked in agriculture one way or the other prior to the Industrial Revolution. now some of those people were slaves so they weren't exactly working under their own behalf but the rest of america slave or not was also working everyone was everyone was a farmer industrial revolution comes and obviously industrializes over 100 years how farming is produced today two percent of people are working agriculture in america we produce god only know 100x more food than we used to i don't know what the number is some massive amount more food than we ever did before the thing is though if you were a farmer in you know 1800 or 1850, there's no way for you to imagine what the jobs the Industrial Revolution would create The negative part which I think you brought up we have to be careful about is in the beginning of the Industrial Revolution some of those jobs were terrible People went from working outside, still tough back-breaking work, to working 80 hours a week in factories, repetitive, like just move the hammer, click the dial, switch, move on the factory floor.

Insanely repetitive, difficult, super long hours, no laws and regulations about the safety of any of those workers. Like there was a period in that eventually our social safety net laws got created and we got a much better economy out of it. But there was that crunch period there that not just people in companies, but like law, the citizens and law creators have to be very focused, I think, on figuring out that transition period. So when we get to the new jobs of the new economy, you know, we're not swimming in a lot of pain to get there. Liminal phases are tough for people. And we are in one right now. This is and there's so much sensationalism around it. And it's really tough to imagine what you don't know, just like we were saying. And so I have a lot of concern that we be mindful of how it's framed and keep repeating and advancing the message that this is huge. It is amazing technology. There's a ton of risks. I mean, we started off and have threaded security throughout this entire conversation. I understand from we wouldn't want sensitive documents to be able to get out.

But that said, on the other hand, I mean, the article just came out about how their models are not yet able to use your Roomba robot vacuum to self-propagate, but it's perhaps not far from the IoT, right, because of all the advances that we're making in AI. So I think there's a lot of... Be nice to your Roomba, by the way. Yeah, I have to say please and thank you and name it. Yes, remember that. So there's a lot... But hopefully my AI overlords who are listening to this recording will know and be benevolent and bestow on me great favor in the future. I have to tell my room, but don't worry about it. I see my room. I got it. I got it. I think my room doesn't mess with it. It's okay. I got this. It's okay. You're doing a great job. You're doing a great job. Good job. Yeah. Take a break. Oh, my God. I lost my train of thought because that was a funny image. So I completely agree that we've got to worry on this dystopian narrative. But on the other hand, like you said, we've got great examples of how humans have been resilient. This is going to make even more painfully clear how necessary it is to have humans in what part of the decision-making process to what degree.

So I'm very optimistic with caution, especially around the cybersecurity aspect of what opportunities this will provide, right? Like everything else, we've got to be careful as we do it. There will be dangers along the way. I'm sure there will be screw-ups that are huge. no one wants to hear that, but I think it's the more realistic version of it, rather than, you know, it's going to kill everybody. I think there's going to be big accidents that are painful, in the same way that, you know, like, companies used to spew toxic waste into the water. That was one of the... That's a great one. ...exogenous effects of the Industrial Revolution. And they didn't care. And some of the best brands in the world, they were like, you know, I think it was GE that polluted the Hudson. They were like, yeah, whatever. We just destroyed the Hudson River. Hey, sorry, man. That's how it goes. They walked away. You know, that was the exogenous, you know, blast radius, I guess, of the industrial revolution. That's going to happen with AI, too. But it doesn't mean that the end result of it isn't fantastic for all of us. Oh, man, you know you're talking to a journalist when they haul out exogenous during a Making Data Simple podcast conversation. That was beautifully said.

So I have a few more questions in our last couple of minutes here. We were talking about the imagination is your limit. The sky is the limit, but it's your imagination. Three to five years out, what are you thinking? your chief product officer, you're the guy who's sitting there imagining what the future is going to be for us. What's happening in the next three to five years? Well, I mean, my sci-fi mind and brain wants to think about the really cool stuff, maybe cooler than document automation, which is a lot of what we do at our company. But, you know, I'd love to see us be able to build models around animal language. Okay, now that's fun. I haven't heard that before. That's amazing. that would be that would change pretty much probably just change our understanding of everything in the world and it might even change religions you know just like just having a completely different perspective if you really could talk to the smarter animals out there um could be amazing um i think we're we're literally just the beginning of of models that are that are

that can advance physics and the physical sciences and architecture obviously we just saw some big news out of open AI and controversial news. They solved some major math problems. Mathematicians are pretty panicked about that and upset. But, you know, like, we think what would happen between the pre-industrial and industrial evolution, it was like magic. I mean, cars were invented. Airplanes were invented. I mean, this is like outrageous stuff. And I think we've gone through a period of lagging of the physical sciences because so much energy and mind power has gone into the information sciences and computer science. Oh, that's a great point. That's a great point. Like the time between, this is a crazy stat, the time between the Wright brothers inventing flight and the 747 is the same amount of time between the invention of the 747 and today. The difference is the 747 and the 747. Oh, that is wild. It hasn't changed. We went from like two guys flying a paper mache plane,

like bicycle pedals that were definitely going to crash, to a 747, like, what was it, 50, 60 years? Now, 60 years later, we're still flying in 747s. They still fly the same amount, relatively the same speed, a little bit more efficient on gasoline. The seats are worse. Like, the physical sciences are really not pushed far enough. I'm super excited about models that can move beyond answering my email and take us to a new magical era where we're inventing the physical world, reinventing the physical world. I guess that's Elon's vision to go into space and these things. You can't imagine going to space. You can't do it right now. but if you could have an age of invention where new unbelievable things are created every like five years instead of every 50, that could be a very exciting place to live in. Okay, new podcast, AI Optimism. And we're just going to haul everybody on to brainstorm and figure out ideas of how to put this to work to dial down the existential terror that people have and dial up the opportunity for great that this will lead to. That's wonderful. I love it. I actually think that's a great idea for a podcast.

I do, too. I do, too. Okay, so offline, we'll go figure this out, and maybe I will have it be AI Martin or AI Main, but I'll still be the human co-host, maybe. What didn't I cover from a technology perspective, company perspective? What didn't I cover that you want people to know? I think what we'd love people to know is that we're at a moment where companies like ours and others can help automate things that could never be automated before in work. A lot of hard data or structured data work kind of got automated over the last 20 years in the machine learning space. And now we're at the point where unstructured components of a company can be automated. and like complex workflows. Like when we think about the stuff we're doing in insurance, like it'd be hard to imagine a few years ago that you could automate 90% of the work of reading millions of pages of insurance claims to try to understand what's in them,

collect the evidence, write a report, and then give that to a human being to judge. So that's an exciting moment for almost any business. And it's a moment of reinvention. We're trying to figure out, oh, which things can be automated in our company that we never thought, almost you don't even think about automating them because like, oh, Joe does that or Susan does that or just like, that's just the job. It doesn't have to be the job. The job could be the interesting part, the judging, the thinking, the talking, the relationships, the building, the human part that only humans can do. And so, I don't know, whenever we talk to companies, they're excited about saving money. We're actually excited about reinventing how they do work inside a company. And, you know, I think it's a special moment for that. I like to talk about it in terms of assure that the humans will be there. What is the work that needs to be done? And it's really hard for people to identify and kind of decouple that because you get so involved in that I'm doing the thing. But just like you said, if there's two people who have been doing it, it's let's articulate the work that needs to be done. And I really like let's keep the interesting judgment based, discernment based, human based elements of it and think about the work that you're doing,

which requires a lot of critical thinking to decompose the end to end project into individual tasks and think about how we can leverage technology to offload that. But that's the skill I think companies need to continue to develop is articulating what is the work that needs to be done, what are the tasks that compose that work, and then how can we leverage technology to our advantage, AI to our advantage to be able to do that so that humans are doing the interesting part of the job. And how do we design a new workflow where the human is in the loop in the right position? That's right. That's the parts that have to be verified or judged or analyzed or, you know, communicated. And that's just a reinvention of almost every workflow in a company. It is. Which is fun. Very cool. All right. Neil, this has just been delightful to speak with you today. Where can people go to find you and where can they go to find out more about Valantor and GroundX and all that you bring to the table? Where should people go? Yeah, so it's pretty simple. Go to valantor.com, V-A-L-A-N-T-O-R.

Honestly, I have no idea what that means. It sounds a word of, like, medieval battle language. Maybe it was created by some pharmacists. It almost sounds like what you use for, like, a pharma drug. But, anyway, valantor.com. It completely does. Like, I just think, yes, you're charging out there in your... Yeah, we're not an erectile dysfunction pill. We are an AI company that can help you greatly. valantor.com. In a secure way. So, yeah, come to us and see us there and talk to us. We love talking to companies, especially in the regulated space, that have big document problems and have big AI dreams, and that's where we feel most comfortable. Wonderful. Okay, we'll make sure that gets in our show notes. Listeners, you heard it here first. I do have one closing question. Do you say please and thank you to your AI chatbot or your various devices? Well, it depends on what it is up to. Sometimes I'll be nice, but if it's screwing up, I'll be pretty mean. and we're like, what is this slop garbage you just sent me? Do not send me more of this. Send me good answers, not these answers.

None of it matters, of course. It's just a human thing to do. We are what we are. So it's strange enough that we built machines that understand us. We built machines that understand us far faster than we were ever going to change ourselves to understand machines. I'm pretty sure that's going to be the opening line for our new podcast, AI Optimist. Well said. I love it. Thank you so much. to it. I will. Neil, thank you so much for your time today. And podcast listeners, you know where to find us. If you have any good ideas or want to be on our show, please send us an email and we'll get back to you on that. Neil, thank you so much. Thank you.