Earley AI Podcast - Episode 30: Busting AI Myths and Building Enterprise Intelligence with Alex Babin

From Misconceptions to Mastery: How Ontology, Taxonomy, and AI Infrastructure Unlock Real Enterprise Value

Guest: Alex Babin, Co-Founder and CEO at ZERO Systems 

Hosts: Seth Earley, CEO at Earley Information Science 

             Chris Featherstone, Sr. Director of AI/Data Product/Program Management at Salesforce 

Published on: October 16, 2023

 

 

 

In this episode, Seth Earley and Chris Featherstone speak with Alex Babin, Co-Founder and CEO at ZERO Systems, a Silicon Valley AI infrastructure company building cognitive automation tools for enterprise knowledge workers. Alex unpacks the two biggest misconceptions holding organizations back from AI success — that AI works out of the box and that dumping data into a model solves everything — and explains why the real power of enterprise AI lives at the intersection of ontology, taxonomy, and purpose-built orchestration infrastructure. 

 

Key Takeaways:

  • AI cannot work out of the box; organizations must define ROI first and reverse-engineer the steps needed to achieve it.
  • Throwing data at AI without structured ontology and taxonomy produces unreliable outputs that fail enterprise requirements.
  • User-generated feedback loops create a new layer of metadata that continuously reinforces and improves AI model performance over time.
  • Hallucination in AI outputs is a critical reliability risk in regulated industries, making traceable and referenceable models non-negotiable.
  • Taxonomy answers what and how data is structured, while ontology captures the why, the meaning and relationships that unlock true AI power.
  • LLMs have become commodities faster than any prior technology, making AI infrastructure, orchestration, and data quality the real competitive differentiators.
  • Skilled AI Modules (SAMs) combine modular capabilities like entity extraction and summarization into chainable, end-to-end enterprise solutions that accelerate time to value.

 

Insightful Quotes:

"Throwing ChatGPT on top of your problems will not solve it. If you want to have something effective and working to solve your problems, you have to have all the right ingredients and components inside." - Alex Babin

"The holy grail, the Narnia of AI and large language models lives in the realm of ontology." - Alex Babin

"There still has to be human judgment in here. There still have to be services to help organizations understand those processes, understand where they want to have an intervention, understand what data is required." - Seth Earley

Tune in to discover how enterprise organizations can cut through AI hype, build the right data foundations, and deploy modular AI infrastructure that delivers reliable, measurable results.

 

Links:



Ways to Tune In:



Thanks to our sponsors:

 

Podcast Transcript: AI Misconceptions, Ontology, and Enterprise Infrastructure

Transcript introduction

This transcript captures a conversation between Seth Earley, Chris Featherstone, and Alex Babin about the two biggest misconceptions organizations hold about AI, the critical role of ontology and taxonomy in building reliable enterprise models, and how ZERO Systems' modular AI infrastructure - including its Hercules orchestration layer and Skilled AI Modules (SAMs) - is helping Fortune 1000 companies move from fragmented AI experiments to end-to-end solutions.

Transcript

Seth Earley: Good morning, good afternoon, good evening, depending upon your time zone and when you're listening to this. This is a welcome to another edition of our podcast, Early AI. My name is Seth Early, and I'm Chris Featherstone, and today's guest is a tech-savvy individual. He leads a growing, high-performing team. He stays on top of both what's going on in the legal world and the AI world. His goal is to help legal professionals do the work better and faster. I'm really excited about today's podcast. This is going to be tremendous. We have CEO and co-founder at Xero, Alex Baben. Welcome to the show, Alex. Thank you.

Alex Babin: Thank you, Seth. Thank you, Chris. Pleasure to be here. You give me too much credit. Like, no one can be on top of everything in AI right now. It's moving so fast. You're on top of some things.

Seth Earley: You're on top of something. We'll leave it at that. Something. Correct. You're a one-eyed man in the kingdom of the blind. That's how I- So I wanted to start with a question about what are some of the popular misconceptions that you are coming across about AI and AI in the legal field? What are people operating under that's just not correct?

Alex Babin: Yeah. It's such a big can of worms you just touched. So I'm even scared of opening it up. But there are so many misconceptions and how people understand AI and what they think about AI. And not just in the legal industry that we serve, but also in Fortune 1000 and insurance and management consulting everywhere. So we are not focusing on legal alone. Legal is just one vertical that we're working with. And those misconceptions, they are very universal. Number one, the biggest one, the prize goes to: AI can work out of the box. No, it can't. And the reason that that is the biggest misconception right now is that OpenAI did an amazing job showing everyone what AI is capable of. And we as humans, we have this kind of ability to extrapolate what we see and what we can touch to what we believe those things can do and how they can do things. So if you go to ChatGPT and you can ask it to do something for you, provide information and so on and so forth, you extrapolate - well, in a sort of a way, you extrapolate like, oh, it can do everything for me. And the second misconception is: okay, if I believe that it can do anything or everything that I require, I just need to throw a bunch of my data in it and it's gonna be El Dorado, Holy Grail, Narnia, you name it. That's the second biggest misconception that we see today. I would focus on those two because those are the biggest ones, especially if we talk about enterprise.

Seth Earley: And what do you recommend? What's the best practice - what should our listeners walk away with? What do you do? What's the simple argument for dealing with that?

Alex Babin: The simplest argument is to start not from the end result, from the very beginning. What is the ROI you are looking for? What is your return on investment? What do you want to do? And then basically reconstruct or reverse engineer all the steps required to getting there. And in many cases, our clients and partners see: wait a second, we don't need all those fancy new tools. What we want to achieve can be achieved with much less effort with existing solutions if applied the right way. And most of those conversations start with the data. And how the data is structured, its ontology, and all these other elements of data.

Seth Earley: Topic near and dear to my heart. I know, we're dealing with that all the time.

Alex Babin: Because garbage in, garbage out. And if you want to have something effective and working to solve your problems, you have to have all the right ingredients and components inside. Just throwing ChatGPT on top of it will not solve the problem.

Seth Earley: Say more about that. Why - what are organizations- and I know I have my approaches and my theories, and again, generative AI is generative. It means it's creating stuff. It's not a retrieval mechanism, but we still need our knowledge and we still need to be able to deal with that knowledge and put guardrails around the generative AI so that it's giving accurate answers. So what are your thoughts around that? How is that dealt with today?

Alex Babin: Yeah, it is such a big piece of conversation - it can be a whole series around that. But let's try to summarize it into something. So first of all, generative AI is again a pretty big concept. And a lot of people think about generative AI like: I asked a chatbot or ChatGPT or any other solution, I ask it something and it generates content that never existed before. And I get it, right? It's not correct. The content is already somewhere. It's just like a combination - chopping into pieces, putting it together. There are so many components in there. But generative AI also works really well with data and data retrieval. So let's say I have a 50-page document. On a top level of your ontology, you probably do labeling and add metadata - what client it belongs to, what project it belongs to. But that's just scratching the surface. What is inside the document? What are those clauses, due dates, components, all of those elements inside? Generative AI can actually generate the metadata by analyzing the document, extracting the important components, summarizing them into something searchable.

Seth Earley: Right. So you're saying that you can use generative AI - you can feed it your content and then leverage its ability to do things like entity extraction so you can have additional signals on that piece of information, and then you can break it into components and understand what the elements are. And then how is that metadata used? What's the next step?

Alex Babin: That's one piece of what generative AI can do for the data - basically extracting the data. The next layer of data or metadata that never existed before, but is being generated, is the user-generated data, this feedback loop. And actually, this piece is the most important one because let's say you have a value-add product - an automation tool based on AI or whatever it is - and a user interacts with the tool. During this interaction, a new type of metadata is born. This kind of reinforcement learning. And the machine now has more data to be fed back into the model, reinforcing and providing better and better results. So it's kind of a flywheel. And this is how ChatGPT was born out of GPT-3, because of the human-in-the-loop reinforcement learning. Otherwise, it's just the model. We know what the difference is between GPT-3 and ChatGPT. It's the conversational piece, and it required a lot of human work to interact with the model to get it to that level. So that kind of data is much smaller in volume, but the quality and importance of that data is absolutely critical for the continuous improvement of AI solutions. This is the goldmine.

Seth Earley: The thing I have talked about in the past is being able to semantically deconstruct the utterance - the thing people are asking for. Generative AI is very good at identifying intents. It's going to be able to identify elements of those intents and entities in those intents, and those can be used to boost the signal. That metadata can be used in order to retrieve. So we still have to have some retrieval mechanism. That's where the balance is right now - this is generative. Now we also need to retrieve because we can't have it hallucinate, especially in a regulated industry.

Alex Babin: Absolutely. Hallucinations and the reliability of the output - I would say it's probably the most critical component, especially in regulated industries, for large enterprises. Being able to get reliable output is one of the most critical components. And whenever I talk to people and they're like, okay, whatever, it might hallucinate - let me give you an example. Let's say you go to a bank, you put all your money in a bank, and you know that the bank is using AI models that can hallucinate. So instead of your bank account, your money, there is a slight chance - just a 10% chance - of your money going into Seth's account or Chris's account. Would you trust that bank even if it's a slight possibility? No. And that's of course an exaggeration, but sometimes this reliability issue can be critical for systems that are dependent on that quality.

Chris Featherstone: Alex, let me ask you something because we have a number of managed services out there with predetermined models. We've got now gen AI that's generating content. At what point do you see - because I feel like there are organizations that will pivot over with the pendulum swing to: hey, we need to do everything generative. And I feel like it's almost a little bit of paranoia moving them over. Whereas, hey, I've got my models, I know exactly what the outcomes are, I've trained them, I've got moderation put on top of them, I've got governance. So where do you- what's your recommendation centered around that?

Alex Babin: That's an amazing question, Chris. Sometimes I call it self-supporting hysteria, because everyone is running this way and then everyone else says, oh, I should be running this way. And literally, generative AI and AI in general is the topic number one for everyone, including CEOs of large organizations. I've been having those conversations with Fortune 1000 companies where even six months ago, I was finding it hard sometimes to explain what a large language model is even to technical folks. Now every CEO and their mom knows what a large language model is. And in this case, it's again very important to trace it back to the original intent: why you need this, what do you want to achieve. And there are so many components that are not there yet - ethical walls, for example. A lot of folks think: we'll just set up a model, source the data in, and it will give us superpowers. Then they realize data proliferation happens inside the organization and some sales manager in Nebraska gets all the financial statements of the Fortune 1000. I'm exaggerating, but it happens. Recently, Samsung employees were using ChatGPT and putting the source code of the chip design they were working on in, asking ChatGPT to fix errors in the code - and it got exposed. So we're going to see normalization of AI strategies for enterprises, but they will have to have components to set up AI the right way. And that's what we do. We provide our clients with an infrastructure layer, an orchestration layer, to actually work with AI - both outside models and internal models - and all the components required to do it effectively, safely, and without shooting themselves in both feet.

Seth Earley: When you said everybody and their mother knows what a large language model is - I think they know at a certain level. What are the things that people need to be aware of about large language models? My premise on this is they're generalized, so they cover a broad swath of topics and content. They're trained on enormous amounts of information, but they may not be specific enough for a highly technical industry. It's getting better all the time. But then there's company-specific IP - you just mentioned Samsung. That's where the differentiation is. If everybody's going to use a standardized language model, that's great for efficiency, but it's not going to give you competitive advantage. We need differentiation based on our knowledge assets, our data, our content, and our customers. So talk a little bit about what large language models are and then this idea of specificity for industry and company.

Alex Babin: Yeah. Such a big topic. I'll start with a very basic thing. A large language model is a type of neural network that is being trained on enormous amounts of data - large corpora of data. One of the critical components in LLMs and how they've been developing is the transformer. It started with a paper published by Google folks - which surprisingly they didn't take advantage of - it was probably 2017, called "Attention Is All You Need." And transformers actually allowed large language models to get to the next level because they existed before - we've been working with the BERT model and training it on data. But to get to the level of sophistication we see today, the attention mechanism - not just predicting the next token in the sequence but also understanding the importance of elements inside that particular sequence of tokens - that's what made the big difference. When that happened, applied to a large corpus of data, OpenAI made a bet: now we have all those components, what if we throw in an enormous amount of data like the internet? And they did. And I know they didn't expect those results at the very beginning themselves. It was basically an experiment. But the results were there. And that's how this whole thing started. So a large language model in a nutshell is a neural network that can operate with a large amount of data and, if trained well enough with specific mechanisms, can provide specific results. The second piece: "large" is a very broad, ambiguous concept. For specific tasks, 100 million parameters is a pretty big model. For other tasks, even 100 billion parameters would not be enough. And I believe - and this is what we see with our clients - that it's always going to be a combination. There's not going to be one large language model that rules them all. Never. It's going to be a combination of large language models, task-specific models, external models, internal models, and everything in between. It's going to be an ecosystem. It's already an ecosystem. LLMs have become a commodity faster than any other technology I've ever seen - four months in and everyone has a large language model or claims they do. Some models can be trained on a specific corpus of data - maybe a small subset, but very high-quality data with a lot of ontology efforts on top of it, making sure it's absolutely crystal clear and high quality. And that model will provide absolutely amazing results that even GPT-3 or Cohere won't be able to match.

Seth Earley: So I was actually going to ask that question about ontology. Let's define ontology for our audience who may not be as familiar. My definition is it's a knowledge scaffolding. It has conceptual relationships. Imagine if you had all the taxonomies in an organization for different entities and elements, and then all the relationships between those different entities and elements - including non-preferred terms and the thesaurus structure. What it does is help you understand how things are conceptually related. It becomes the knowledge scaffolding for the organization. It can contextualize information and knowledge. Some of the ontologies I've seen for life sciences are too big - the language model is too big - so they have to be fine-tuned. So talk about the relationship of ontology to language models.

Chris Featherstone: Alex, one thing to keep in mind - Seth's middle name is ontology. So it's Seth Ontology Earley.

Alex Babin: Do you have it? Do you have it? Ontology is all we need, right?

Chris Featherstone: I'll send him an email. He's like, what's the ontology for this email? I want to make sure the structure is correctly done. Anyway, keep going.

Alex Babin: No, no, no. It's great. So I wanted to get it down to the level of a bit more simplification. It's important to start not from the ontology, but actually taxonomy. And because those two things are - I'm trying to bring it up as I would explain it to my mom, because I like this level of abstraction and explanation. Taxonomy is something that has a structure. We are talking about the data, of course, like a catalog. There is a relationship between the data. Like if we are talking about consumer goods and I'm a manufacturer, I have men's clothes, I have women's clothes, and then pants, shirts, socks, and so on and so forth under each category. That's a very simplistic way our brain works - we catalog things. This is a taxonomy. It's a very simple but straightforward way of doing things.

Seth Earley: Hierarchies are for humans. Exactly. Because machines don't care. Exactly. This is hierarchy. But we have to be able to understand how things fit together.

Alex Babin: Exactly. This is hierarchy. And we have our own model - it's much smaller than big models, but it's designed just for classification and labeling of data. So let's say your email comes into the inbox and it's not classified. It doesn't have its place in the taxonomy of your data. But suddenly the model says, wait a second - this is client A, project B, stage C. Boom. You have a taxonomy and this piece of data has its place. And it's a reference model. So we're matching the unstructured data to the taxonomy of your organization. Because if you don't have that classification label in your taxonomy, no system would classify by matter and project or client. This is a simple way of structuring the data. But then we go into ontology. And this is a much more important concept because if taxonomy answers the question what and how, ontology answers the question why. And this is where we get into the realm of large language models, because the interconnectivity between the data - having that meaning and context captured and labeled, this metadata around the data surfaced - then you can apply models on top of it and get maximum power out of it. So here's another example related to the email. Email comes into your inbox. Taxonomy first, of course - what is the client, what is the project, what is the stage - all those surface-level metadata types. But what is inside the email? Is it a due date? Is the client asking for a report? What is this report? Have I been doing those reports before? Where does the data for that report live? Is it in Salesforce? Is it in my document management system, content management system? There are so many - it's like a door to Narnia. You open it up and it just goes beyond belief. So this is the next level. And a lot of organizations - most of them, as far as I know - don't go beyond advanced taxonomy.

Seth Earley: The holy grail of enterprise ontology. I'm sorry, I didn't mean to cut you off. Go ahead.

Alex Babin: The holy grail, the Narnia of AI and large language models lives in the realm of ontology.

Seth Earley: That's great. And when you were talking about enterprises having taxonomies - many of them don't have enterprise taxonomies, or they have 50 or 60 taxonomies. Those do need to be reconciled at the right level of granularity, and you do need to build those relationships because that is the contextualization of the data and the content.

Chris Featherstone: Let me ask you too, and then I want your example, Alex - the other piece of this is, what happens - because nobody gets it right the first time. How do you embrace the evolution of the taxonomy over time because of changing needs?

Seth Earley: We always like to say that a taxonomy is a living, breathing thing. When you're done with your taxonomy, you're done with sales, you know? Because there'll be new products, new opportunities, new topics, new issues. And so we always have to have change management, and we have to be able to build domain models so that they're broad enough to include extensions. The other thing is not every application cares about the detailed granularity. I like to say there are two types of taxonomists: lumpers and splitters - people that group things together and people that divide them into finer points. The idea is you're going to have some situations that require lumping and some that require splitting. A telecommunications firm had routes to market on the website - they'd only have about six categories of industries. In the CRM they'd have about 50 categories because there were different routes to market, different conferences, different titles, much deeper granularity. So when you have those models - they might be called industry - it might be the same vocabulary, but you have to add additional metadata on the terminology to tell it what the right context is. That has to be done with ontology.

Alex Babin: I absolutely agree. It's a living, breathing organism. And if it stops breathing and moving, it's dead. More data comes in - data creates, with the help of AI, more data. It's a self-supporting flywheel. If it's not providing more data, more results, it's just stagnant, and then it's not AI - it's rule-based something. So I was going to bring an example of policy enforcement, where two different ontologies that are not typically connected together should be connected to provide extra results. Let's say you have a set of rules or guidelines or policies that you need to enforce inside your organization, and you have basically an ontology of it - what should be done, what can be done, how it can be done. Typically, it's like 50-page documents. No one reads them - I've never seen a person except for those who create those reading those documents. They sit somewhere. And then there are other people who actually should be doing things according to those guidelines or policies, and they have their own ontologies of the data they work with. And those two worlds never or rarely overlap. But AI can actually match those and bring them together - understanding the ontology on one side, the way it should be, and understanding the ontology of the data, and then matching those: wait a second, this should not go there, or you can't expose that here, and this information should go this way. So in this case, AI can be a stitching mechanism between those ontologies. It's not quite there yet, but it's getting there. One of our modules actually does that. Because if we talk about a Fortune 1000 organization, there would be thousands and thousands of different ontologies related to different data sets, different lines of business, different departments. And those are not interconnected. It's like trying to boil the ocean if you start with one. That's where AI can help.

Chris Featherstone: But that's why it's a science and not something solved. Because I may have a master taxonomy and an ontology to go with it that maps to a catalog. But the beauty is that catalog may be multiple catalogs, and I may have to conjoin those catalogs in order to get better results. And so we get into this notion of all the data governance for system, people, and process. Then we filter that up to give access to the data science teams, and now they're doing feature development on top of that and adding new features and almost generating completely new taxonomies and catalogs at the same time while testing everything.

Alex Babin: You're right. But, Chris, let's take it a level higher. That's the foundation. Without that, nothing would work. But what organizations actually need, they need end-to-end solutions - solutions to the problems. Without having the right data, right ontology on top of taxonomy, it would be possible. But that's not what they're looking for. They're looking for solutions that will solve the problems. And that's another thing that doesn't exist on the market right now, at least at enterprise scale. There are so many fragmented solutions that work here and there, but there is no way for enterprises to say: hey, this is the layer I'm going to apply on top of my data. This is the level of APIs, integrations, protocols that will help me get those modules or solutions to work on top of my data, providing me with end-to-end solutions, and do it quickly instead of a two-year development cycle. We see this as probably the biggest problem because it's the Wild West right now. Data sources, lack of instruments and infrastructure and protocols - it's just such a big mess. We're trying to bring structure into it. We have our own orchestration and infrastructure layer, Hercules, that sits inside the security perimeter of an organization, prepares the data, helps get the ontology in place, interconnects data, depersonalizes it, and then the organization can use external models, internal models, old models, whatever it is to get end-to-end solutions. And they can actually build those end-to-end solutions quickly. We call them SAMs - Skilled AI Modules. Each SAM can provide a small specific subset of the value. But if you multiply those modules by the number of use cases you apply them to, your organization can immediately get to the next level. This is the future of enterprise and how AI is going to be applied in enterprise. All about time to value.

Seth Earley: So what does the S stand for again? Skilled AI Module. Okay. So talk about the skill.

Alex Babin: Skill is a small subset - basically a function that this module can tap into. There is a library of those skills. For example, summarization is a skill. Data extraction or entity extraction is a skill. And once you have those many, many skills - there might be a small model that provides high-level accuracy for entity extraction specific to your content and your reference data. So for example, there is a Skilled AI Module that needs to do invoice processing and then connect it to an internal system. It can tap into multiple skills: one is entity extraction, another is API-level connection to the system, summarization, and so on and so forth. So this provides basically Lego blocks on top of your infrastructure.

Seth Earley: We are a little more than halfway through. We're speaking with Alex Babin and ZERO Systems - zerosystems.com. He's the CEO and the founder. I wanted to ask about the business and your role - what is keeping you up at night in this field? It's moving very quickly, there are a lot of unknowns, and the competition is moving quickly. What are you concerned about?

Alex Babin: Well, the main concern for everyone on the market right now is where things go. The comforting thing, though - no one knows. So we're all in the same boat together, we'll figure it out. I would not say we predicted what is going to be happening, but we knew that enterprises would require AI to be part of their business processes. We started about seven years ago - just over seven years ago - and we knew it was coming. We've been building our building blocks. It's like in a gold rush when most of the money is made by guys who sell shovels and picks and jeans. We were building a shovel, picks, and jeans factory for ourselves. We didn't know where this gold rush would come from. And when in November, OpenAI released ChatGPT, I was telling my co-founder: this is the best marketing ever done for us without us spending a dollar. We realized early on that everyone would need infrastructure, everyone would need end-to-end solutions. And if we worked hard enough and long enough to create it preemptively, when the gold rush started and everyone rushed for AI, they would need those components no matter what, and we would be the first supplier in town to get those components in. Eventually there were the large language models and the popularization of AI across all industries. And now we have those components and we're focusing on Fortune 1000 to provide them with those elements.

Seth Earley: There still has to be human judgment in here. There still have to be services to help organizations understand those processes, understand where they want to have an intervention, understand what data is required - fine-tuning taxonomies and ontologies, choosing specific language models, maybe building their own language model based on the content they have. So there's still a lot of work that needs to be done to connect these pieces together and have those solutions. I certainly don't see my company going out of business soon.

Alex Babin: Oh no, you guys are going to be blooming, because that's what I started with - the biggest misconception is that AI can work outside the box. We can provide all the components and the building blocks, but we are not a services company - we are a technology company. And guys like you and Bain and Deloitte and others - they provide the value-add services helping organizations get to ROI faster. I see your goal and our goal as the same: to get clients to ROI as fast as possible without mistakes in the process.

Seth Earley: Totally agree. Totally agree. Accelerate that time to value. Let me ask you quickly - what kinds of problems has your team solved recently? Can you give us some examples of either a client story or something you've done recently that is notable?

Alex Babin: Absolutely. One example I want to bring up - can't name the client, but they are on our website. It's a Fortune 100 company. This example is the financial analyst process where private equity notices are being processed. Private equity notices - capital calls and distribution notices - coming in, and the company's business division is processing those notices, providing fund management services to their clients. Those are thousands and thousands of documents from different sources coming in each week. And there are 80 people basically opening up a PDF, looking for information, extracting it, and putting it somewhere. Typically, IDP can deal with that - you open up the document, extract stuff. But there is a second component that was never done before because there is business logic after you extract the documents. They're all interconnected. That's where the new approach with large language models kicks in - because you need to understand the logic. You don't just take the data and put it somewhere. You understand where it came from, what the amount is, you do the math, you apply business logic, and only then you provide it somewhere. And that logical block - that reasoning block - was missing until now. Our SAM, for example, not just takes the data out beautifully with zero-shot learning, it also applies business logic. It mimics the cognitive process of that particular financial analyst. So now they don't have to do the napkin math, and the error rate just drops. And so that process needs to be defined - you need to extract those rules, understand that judgment.

Alex Babin: Sometimes those aren't even rules. Sometimes people actually can't even explain how they do it - that's the way I do it, and another person might do it differently. So this reasoning module needs to really understand what the end goal is, what the user wants to achieve, and reconstruct back - basically do the reconstruction of the whole process knowing how to replicate it. And it was not possible before. Now it is. And those examples exist everywhere. You look at any organization in the world - there are thousands of processes like this where the cognitive processes of knowledge workers should be augmented. And that's where the power of AI comes in. But it wouldn't be possible without large language models and the infrastructure layer that we provide.

Chris Featherstone: You mentioned in the beginning that the orchestration of these SAMs is super important to string them together in the right way to get the outputs. And I'm assuming your technology stack does something like that?

Alex Babin: Yes. Because like knowledge workers - real humans - do not work in isolation. The same goes for the modules. Modules need to talk, they need to send data, they need to share. That's where it's important because otherwise it's just isolated items without providing each other with important information. Chain of tasks, chain of events, chain of data points that are combined together - that's where the real power is.

Chris Featherstone: It's super easy to understand an ASR model talking to an NLP model, right? But it's when you pull in classification or something like that - you can string however many of these together, I'm assuming, and then generate these outputs.

Alex Babin: Exactly.

Seth Earley: What are your biggest challenges today? And then what excites you about the future?

Alex Babin: I hope those are the same things. The things that are exciting us also make us sometimes be very careful. I would say the ethics of AI. We are very laser-focused on AI that is 100% reliable and accurate. That's our nature. We work with enterprises. Our models need to be traceable and referenceable. But the advancements of AI on the other side go this way - it's literally an arms race. It's good for us because better models existing outside means better for us - we'll just provide our clients access to those models in a safe and secure way. But they are advancing so fast. It's an arms race - bigger, faster, more powerful. There is no controlled governance layer that helps everyone understand what's going on. And because we don't understand what's going on - by "we" I mean the market, everyone - we don't understand where it goes. It's less predictable, less controllable. So this is a shaky situation, but we'll collectively, I believe, figure it out.

Chris Featherstone: It feels a bit like the transaction arms race of the '90s with databases. What in the hell is the real business value of doing four billion transactions in a single second for a business that has long-running transactions?

Seth Earley: Let me switch gears just a little bit. Tell us about who you are, where you're from, how you got to this space, how you got into AI. Just give us a little rundown.

Alex Babin: I've been into AI for about - well, as we call it right now. My previous company was actually doing interactive video. We were trying to understand the elements on the video, automatically tracing it and then applying interactivity to it. It was a pretty interesting concept. No tools and no processing power existed at that time. If you've seen recently Facebook released the segment-all model which does amazingly what I was trying to achieve twelve years ago. That's where I got this AI bug. My co-founder and I started the company and wanted to build cognitive automation for knowledge workers. That's what we still do - just more tools, more powerful elements. We basically spent a lot of time building those building blocks that now pay off. And now everyone needs what we were trying to explain to people why it was cool - five years ago no one could understand. I live in Silicon Valley, that's where our headquarters is. Father of two beautiful children and the husband of a beautiful wife. And I have a crazy hobby. I don't know anyone here in the Valley who does the same thing. I do Japanese blacksmithing - Japanese swords.

Seth Earley: Japanese blacksmithing! Japanese swords! That is fascinating. How long have you been doing that?

Alex Babin: About five years now. And I'm still pretty bad at it. I'm getting better. To me it's just more of a meditation rather than a craft. We're here - headquarters in Campbell, Silicon Valley, right between San Jose and Palo Alto. About 100 people now.

Seth Earley: And let me ask you another reflective question. If you were able to go back in time and give yourself some advice as you were getting out of college - what would that advice be?

Alex Babin: These "go back in time" advice can actually backfire because what we know right now might not be that helpful if we go back. But I would say very simplistically: if I go back to 1999, I would tell myself to buy Apple stock. Forget about everything else, you'll figure it out, just buy Apple stock as much as you can.

Alex Babin: So tell me where people can find you. Zero Systems - zerosystems.com. That's our website. Or on LinkedIn - Alex Babin on LinkedIn as well.

Seth Earley: Great, great. And I think this has been wonderful. I'm really excited. I love when people tie together the things that I'm passionate about, such as ontologies and knowledge and knowledge work - taxonomies and large language models and AI and cognitive systems. So this has been really wonderful. Thank you, Alex, so much for being here.

Chris Featherstone: Thank you. I'm just more excited that I now know a sword maker. I mean, AI aside, let's dig in.

Alex Babin: Maybe we should do another hour just on the sword making. You know how my wife puts it? She says, when AI will take over the world finally - that's my wife's word - when AI will take over the world and we go back to the Stone Age, your hobby is gonna be a very important profession.

Chris Featherstone: Exactly. I agree with you 100%. That's awesome.

Seth Earley: I want to also thank our audience for listening. If you've enjoyed this podcast, if you've learned something, certainly share it - tell somebody about it. And again, thank you, Alex. I really appreciate your being here. It's been a pleasure.

Chris Featherstone: All feedback is welcome. Thanks, Alex. Appreciate the time.

Seth Earley: This has been another episode of the Earley AI Podcast, and we will see you next time.
Meet the Author
Earley Information Science Team

We're passionate about managing data, content, and organizational knowledge. For 25 years, we've supported business outcomes by making information findable, usable, and valuable.