Welcome to Cybersecurity Insights, the podcast for the CyberEd.io learning community. Our goal is to bring cybersecurity practitioners the latest and most relevant education and training to upskill and dive deeper into topics that matter in today's modern cybersecurity world. Good day, everyone. This is Steve King.
I'm the Managing Director at CyberEd.io. With me today is Rodrigo Liang, Co-Founder and CEO of SambaNova Systems. SambaNova is a leader in enterprise-scale AI solutions. Rodrigo has deep roots in Stanford and the university and the ecosystem there.
And he and his co-founders had designed a full-stack hardware software platform optimized for AI workloads, which set the company on a mission to enable the future of AI for the enterprise. Contrary, I just spent a week at RSA and talked to a lot of people and there were a lot of differing views about AI. So it'll be great to hear what you have to say, Rodrigo, today, because I think you and I agree that we're witnessing probably the fastest industrial revolution in history, at least in my lifetime. And it's happening right before us, the way that I've never seen product release cycles, if you can call them that, happen the way that they've happened in the last four months.
So I'm delighted to have you here today and thanks for joining us. Thank you for having me, Steve. Wonderful to be here. And yeah, I agree.
It's the fastest industrial revolution we've seen. And I've been telling people this and people kind of, you know, usually pause and this question has like, yeah, it sounds about right, you know, but we're excited to be in the middle of all this and watching the technology enable people to do these just amazing things. And so really excited to be here to share some of those with your audience today. Yeah.
And if this is if the last four months are an example of the future scale, I can only dream of what the next four years are going to look like. So it's very exciting. Tell us a little bit about SambaNova and how you fit into, I guess we'll call it the AI space. Right.
Yeah. So I started this company back in 2017 with two Stanford professors. And at the time we were thinking about how these models were progressing, how these artificial intelligence models were evolving, because, you know, I mean, machine learning is not something that's been just recently invented. It's been around for decades.
But we've recently over the past 10 years gone to the point where the capabilities of these models are doing these amazing things now. And yet, you know, what you're seeing people do with them today is still just the beginning of a multi-decade journey that you're going to see these technologies evolve. Right. And so we started looking at this from the angle of how can we make technology available to everybody?
Right. Because the trend were that the largest companies in the world, the Googles, the Facebooks, you know, the biggest companies in the world were already investing significant amounts of money and people and energy into developing these amazing capabilities that artificial intelligence was enabling. But how does everybody else do it? Right.
How does everybody, every other company in the world, every other individual in the world who do not have that level of investment ability keep up with such a transformative technology shift, one that's akin to the Internet, right, where we saw this back in the late 90s. If you did not embrace the Internet within a few years, you were irrelevant. Right. The economy became 100 percent Internet driven.
I believe the economy is going to be 100 percent artificial intelligence driven. Right. Not to say that that's going to be the product, but it will be the means by which we do commerce. It will be the means by which we do business engagement in everything in the world.
Right. If you don't have it, you will be at a significant disadvantage. And so some of that we came in with the technology stack that ultimately was supposed to solve. But we thought about how do we solve that for most of the companies?
The answer here, the answer to how to do that is threefold. One, you have to simplify the technology stack. Sure, there are chips involved, there are models involved, there's software, all sorts of things that people have to become experts in if you're going to do it yourself. And so some of that we embrace this idea of a full stack so that your point of interface is as a user.
You don't need to become a large language model expert or a computer vision expert or a kind series expert. What you have to do is just bring your data and allow artificial intelligence to help you gain insights. So that's one. So just the simplification, simplifications of the interface to what's ultimately a fairly sophisticated stack.
Right. So we build that whole interface for you. The second piece of it is we've come to realize over the first few years of some of Nova that the access to machine learning expertise to these engineers that are getting hired by the Googles and Microsofts and to some extent some of Nova. That skill set is extremely difficult to find.
And so if you're a Fortune 2000 or global 2000 company looking for these machine learning experts, it's extremely difficult to hire them. And so what we decided to do was offer that technology stack as a service, as a service to companies that do not want to or are finding it hard to hire all those people. And we come in and we do it on your behalf to allow our expertise to become an extension of the company. And then finally, the third piece of it is being able to actually create privacy and ownership and security, which is probably the most interesting audience today, which is we fundamentally believe that data is the crown jewels of every company and models trained on your data are ultimately the most powerful vehicle to gain access to that information.
Most companies don't know what they know. Right. But with these large language models, as you're seeing today, just train on public data on Internet open domain data is already incredibly powerful. Imagine what it can do if it was trained on your personal data.
Right. And so what Samanova does is we'll bring the entire tech stack to wherever your data is, especially to your most sensitive data, and allow you to actually create a model that's your own owning the model. Owning the model will become a critical item to most companies because they do not want to actually give away that insight and give away that data and knowledge to other people. And so all these companies are trying to use shared models, use open vendor models.
It becomes something that they cannot own in the long term. And so what we at Samanova do is we actually come and help you build your own so you can own your own models based on your data in perpetuity. Yeah, that certainly makes sense from a business model point of view. And in particular, given what you point out is the biggest or at least one of the biggest challenges is to find people that are adequately trained in how to build things here with knowledge of large language models.
And from a cybersecurity point of view, I've been talking with several companies now that have product readiness that are focused on the what you would not be surprised to find are the labor intensive portions of the incident lifecycle where you've got SOC analysts looking at a lot of data over a long period of time and trying to pick out, you know, separate wheat from chaff relative to false positives and real threats. You know, and then we had the thousand signature, I think it was more like 2000 letter, which honorably suggests that we should slow this down. Of course, nobody else in the world is going to. Are we exposing ourselves to a massive cybersecurity threat?
And are we currently protected at all against the bad actors who are going to use this stuff against us? You know, Steve, you know, I started working in the early 90s and, you know, by the time the Internet was really in full force, people were talking about the fact that, look, we've now cracked open the doors for all sorts of people to hack into our systems, our data. You know, we've got, you know, all these security issues now. We've got data being exposed.
We have fraud being committed through all these different access points now and all sorts of issues. And guess what? 20, 20 some years later, they all became true. Right.
And yet there's probably not a single person on this planet. Maybe there might be one or two, but the large, large, large majority of the world would not go back to a world without the Internet. Right. And so I think we're facing the same challenges that the technology transition from pre Internet to post Internet was so significant, not just from a technology perspective, but it lifted entire economies.
Right. That people could out of their garage, you suddenly have a global business and we forget that prior to that, you know, you couldn't do that. Right. They have a global business just selling out of your garage.
And so we have these technologies that have completely transformed the way that we operate. And I'm specifically focused on the on the on the business world. Right. But that the technology has lifted that and has come with challenges, would then create an ecosystem to go and try to safeguard against it.
Right. Which many of your audiences today part of that part of that ecosystem to help us protect against the bad actors because the bad actors will always be there. They're only figuring out what tools are available to them to be bad actors. Right.
And so artificial intelligence for sure will follow a similar path where it's going to create entire economies that we have in. And we have yet to imagine it's going to open markets to people who never thought they could ever get access to those markets because maybe they thought they don't have the knowledge or the expertise or whatever. And suddenly that knowledge and expertise is delivered to them in the form of artificial intelligence. And it's going to open up these ways that we have to be able to protect ourselves and There are three tiers that, you know, I really hope the audience can understand, right?
There are three tiers of exposures that come with these models. And again, we fundamentally believe that these models will transform your business. Full stop, right? You'll have to start on this journey because these models are so powerful and they're doing so much that it will collapse the cost structures that you have for decades by tenants, which frees those resources and investment to actually generate the things for your business, creative things for your business, versus the administrative things that the robots are now much better at doing.
And so 100%, you have to embrace the technology that's coming. But with that, I start with the question that, look, you have to own the model. And the reason I say that is there's three exposures that businesses have all the time when it comes to this type of thing. One, you saw it with very recently, the Samsung leak.
OK, so if your audience hasn't hasn't seen that, what it is, is every time you're engaging with ChatGPT or public model, what people don't understand is every question becomes in perpetuity part of the corpus for that model, right? And so if somebody comes and says, hey, what is, you know, what's happening with this company? Let me understand more about that company. In perpetuity, they will associate those queries and those questions with that individual.
It's almost like, you know, disclosing your Google searches, right? And so people don't understand that that's becoming part of the shared corpus that the entire world has access to. Right. This is what that the engineers were trying to, you know, do some stuff in their code and they posted the code to actually generate some stuff off of the code base, which was private.
But then that code base is now in perpetuity part of the model. Right. Anybody can pull that code down now. Right.
Because like, oh, what is that? You know, so so that's leak number one, that people aren't understanding just the use model of what a shared, a shared model, a shared backbone means. A shared backbone means every engagement with the model becomes part of the global model that everybody has access to. So we have to educate people on exactly how to engage with that.
That's one. The second one that I think people have to understand is, and this is what I touched on, those models are trained on public web crawl. Right. This is every everything that's on the web.
It's not it's not biased one way or the other. It's just whatever is out there on the web. Right. And so when people say, hey, this thing hallucinates, sometimes it says inappropriate things, sometimes it gets racist, whatever.
Well, it's whatever is on the web. Right. And so the qualification of how this model responds needs to actually be done carefully by industry, especially industries that actually have to be certain on the responses that the machine gives. We advocate that you have to be able to open the model so you can test it, verify it, make sure you can exclude certain data sets from places that you don't believe you'll follow your business.
Right. Or you're tied to kind of the values that your business has or, you know, I mean, you you want to be able to actually verify that the behavior of this model, which ultimately becomes a representation of your business if you're using it as part of your workflow, truly does represent your business. Most people don't understand how those models got generated, exactly what it's going to say, exactly what it's going to do. Right.
And so that's the second thing that you've got to make sure you understand how this model was created and does do you stand behind it. Finally, the third thing to tell people is these models are generated again, open domain. So it puts as much value on the Jay-Z and Miley Cyrus lyrics and Shakespeare's background as it does for a bank's business. They all treated equally.
If you're that bank, you care about your business, your products, your customers, your process, your people a thousand times more than you care about some, you know, social media influencer. And so what we promote is this idea of you have to train these models on your data. Your data is more important than the public domain data. And so you want to be able to gather insights off of your data at a higher level of importance than Jay-Z and Miley Cyrus.
Right. Because you run your business. So in order to do that, you need to own the model because your data, your should be your model. Right.
And so those are three things that we focus our clients to understand when you're if you're serious, if you're serious on this AI transition, you have to protect yourself. You have to understand what's gone into it and you have to be able to to come in and protect yourself on the security side. So you don't have data IP leaks into the domain. You have to protect yourself by not having Trojan force come in and so understand all the data sources that came in to train this model.
And then finally, making sure that if you train a model with the highest value, that means that you incorporate your data into it. You want to protect that model. Right. So the more risk averse industries and businesses among us, like banking, like insurance, et cetera, are going to be laggards, I assume, in early adoption of this particular technology.
What's that going to do? That's going to create a startup environment that should be very rich and very aggressive and leave a lot of a lot of these companies in the lurch, you know, at the end, with the additional vectors that are kind of flying around in the West now with banking being so vulnerable to closure and as a result of not operating best practices. There seems like there'll be a lot of disruption in that space. What's your what's your view of that?
Yeah, that's right. I actually think, you know, highly regulated environments are rightfully most conservative and most careful in the space. And that's fundamentally where we find the most traction as well today. As a startup in AI, we're the most deployed in the U.S.
government. Right. And so governments are very careful about this because their information in many, many cases is extremely sensitive. Right.
And so we've deployed. Yeah, but most of that is already stolen and owned by our adversaries. Well, you know, all that said, you know, I think our governments are very careful, you know, but yeah, that's probably true. I think the banks, you know, and we've deployed a number of banks now as well, where the banks, here's the interesting part.
It's actually today one of the more aggressive industries when it comes to artificial intelligence adoption. Do you think that's a function of its labor intensity as well? Yeah, exactly. Because the processes that banks have to operate in are labor intensive, they're administrative, and fundamentally they have to be right.
Right. So if you think about risk, you think about risk. New GDPR rules come in and you have to go and scan a million contracts or whatever. Those are very labor intensive and you have to be right because the penalties of being wrong are significant.
And so the machines actually doing a pretty good job of actually helping them with that. Think about like all the data management and tying together various data into all these different systems and making sure they're right. These systems are actually really good at what we call name entity extraction or name entity recognition extraction, which is pulling information out of all of this unstructured data, emails and text messages and, you know, whatever. Right.
You know, just Slack messages these days. Right. And all these unstructured messages and pulling the data and putting in places. Machines are really good at these things now.
So where where you have industries and companies where you have a lot of processes where manual translation is actually a very significant part. Right. These machines are incredibly, incredibly effective at being tools and parallel assistants to everybody. And so when someone, what we say, the way that you can, you know, you can bridge the situation of privacy and yet adoption of the new technology for the step function benefits that AI brings is why we've chosen to go full behind firewall type of solution.
Most of these environments until they figure out kind of what all the risks when I expose my data over to a third party vendor or expose it to the cloud or expose it to these all these different places. You know, some of them, we actually, because we build a full stack, we bring the entire stack behind your firewalls where you're most secure, where you already have all the policies and protections. And we actually let you train your models behind your firewalls and use them behind the firewalls. That's the safest way today for you to start building your knowledge base, building the processes, building your understanding of artificial intelligence and how it works for your business before you expose your data into an environment that you don't know if the tools and policies are there to protect you.
That's why we chose to do it. And we're agnostic. We run in the cloud. We run on on-prem.
We run wherever you want to run it. But I would say today, the large majority of our clients are running behind one firewalls because it's the safest and they want to see how the market evolves to protect them so that they don't have leaks like we've seen in these very public cases where people's data got leaked through the models. Yeah, sure. What's to prevent, you know, if we go back to the idea that a company owning its own data model, if you will, what's to prevent new bad guys from replicating that data model and acting as that company?
I suppose that that can happen here. Like, you know, it's the same as what prevents somebody from actually replicating your database. Right. You know, if you're a large company, you know what prevents them from replicating your database and your websites and being able to operate.
Well, How data is distributed, how data is organized is also how data will be integrated and generated by these machines. So why not just now replicate that same compartmentalization of data with an artificial intelligence model above it to give you easier access to what information you already have, right? You have that information. You just want to interpret it.
You want to generate stuff from it. You want to be able to edit it, modify it, be able to actually generate value out of those things. And if you actually are able to do it from a database that already exists, we'll keep it within that. Those structures already exist.
You don't have to go invent everything. Right, right. How about we wrap up with a question about the issue around waiting? I don't have a commercial dog in this hunt, so I will say right off the top that there is no benefit to waiting.
If you're not in the game immediately, then you might as well participate in a different event because you're not going to compete. What is your view about that? And do you think there's any real threat here from a global threat actor point of view? And the question there is, does Congress understand that the Chinese language, for example, is very, very different than the similar expression in Latin and would require just tons of compute capability to do what we've already done here?
A hundred percent. Look, what we're doing, what we're talking about today with GPT models, and it's the tip of the spear, right? This is the tip of the spear. Two scary things that people, companies are thinking about, organizations are thinking about, should I get into AI or not?
What we see as kind of like these little blips and kind of announcements that people are doing, oh, I have this offering now with GPT. I have this thing that's AI enabled, et cetera, et cetera. What they don't see is the amount of work under the water beneath the iceberg that's happening every single day in some of the largest companies in the world. We see it because we engage with them all the time, right?
And so over the coming 12 months, you're going to see this explosion of services and products that will attack the market in every single industry that will completely change the order of those companies within those industries. Because what we see today, I mean, that's one of the things with AI is, it's not entirely visible how it's actually getting deployed because it's so silent. We already see what Alexa is in our life. Certainly Alexa got a little smarter.
Suddenly Siri got a little smarter. You don't see what's happening, and yet we know because we're working with a lot of some of the largest companies, incredible amounts of work is going in every single day with thousands of people being invested into it to build products and services in every industry using AI. And so, yeah, so your first question is, is there a reason to wait? No, you'll be behind already because these companies are investing.
Even though they only give you a little glimpse of the things that they're doing, the amount of work, all the investments that's happening in AI today is beneath the water in the iceberg. You know, in a sense, right? And it's coming. It's coming over the next 12 months.
It's going to be a full-on attack on all these services that for years and years and years, we thought that, you know, the market share will not move. And yet AI is going to represent an opportunity for companies to reorder their market share within the industry in a significant way because it's certainly going to capture, you know, things that you couldn't capture before. Here's a quick example. In banking and financials, the market share for a long, long time was localized to language barriers.
By how many companies, how many banks were basically protected because they had the language advantage and kept some of the other foreign banks out. Well, what if suddenly every bank has access to 170 languages and produce products and services by, you know, a snap of a finger and they can get into the market now. Word before they couldn't do it, right? And so these are things that are happening right now.
So I would, I a hundred percent believe that you have to start. It's a journey. You're not going to get it overnight. You have to get going.
We accelerate people by as much as 18 to 20 months when it comes to getting into the game. And so that's what we do. If people are feeling behind, we give them a hot start. When it comes to languages, like I said, this will find places for our time.
I agree. But look, you know, one of our clients is OTP Bank in Hungary. Hungarian, which we now train, have trained and we actually do a really good job helping that bank understand, comprehend, and generate documents in Hungarian, it's an incredibly difficult language to get right. So we have experience with these complex languages.
It's one thing to train in English. It's one thing to train in Spanish, but it's something else to train in some of these more sophisticated languages. And yet you're going to need them as you go global because AI is going to touch every company on this planet in every industry, in every segment, in every single department within those companies. And so hopefully this is helpful to your audience to start getting a sense of the journey that we're all going to be going through over the next 10, 15, 20 years, right?
And like you said, there's no reason to wait. In fact, if you haven't started yet, you know, let us help you because it's a race with every industry and we've got to get going fast. Indeed. Yeah.
Thank you, Rodrigo, for taking time out of your day to share a lot of these ideas and discussion with us and our audience. And in particular, Jay-Z and Miley Cyrus, who I'm sure are listening as well. I hope to catch up with you again in say, I don't know, three to four months to see what's progressed and, you know, what we've discussed here and see how much of that has become reality or gone well beyond it. Wonderful.
Looking forward to it, Steve. I'm excited to have the conversation again. Great. We'll stay in touch and thank you to our listeners too for attending another session here.
And hopefully it was as interesting to you as it was to us. And until next time, I'm signing off, your host, Steve King. Take care. Thank you for joining us for another episode of Cybersecurity Insights.
You can connect with us on LinkedIn or Facebook or send us an email at [email protected]. For more information about the podcast, visit cybered.io forward slash podcast. Until next week, stay safe and secure. And we'll see you on the next episode of Cybersecurity Insights.