Lior Strugach is the co-founder & CTO of Argu AI, an Israeli-founded startup backed by a16z Speedrun

I want to talk about technical foundations. You and your founders come from elite Israeli military backgrounds. I want to talk about how does that background shape argu's architecture, your security standards and risk mindset.

So as you know, every citizen in Israel serves in the military and I personally come from the intelligence sector. I started my career in the ground technology R and D division and then made the shift to intelligence. So first of all, we take security very highly because we know the threats out there. We know that if your software has even the slightest hiccup, it can affect, especially in mission critical operations, it can affect the way our clients are able to use the software. Security is something we take very dearly because we in those mission critical environments we have to give our clients 100% reliability or else. Besides that, accuracy in the mission critical sense is something that we cannot. Coming from a military background, you cannot compromise in accuracy. And having all of the levers to influence the way the model and the system works is something that you have to give to the client. We cannot give them a 95% solution here because in those mission critical environments where you have to know if somebody has fallen, if you have to know if somebody breached the perimeter, you can't compromise on that. And those are the two values that we hold dearly throughout the development of the product.

Your system interprets the world in real time. Argo, your company is next generation AI operating system for cameras. How do you define accuracy when the real world is constantly changing?

Excellent question. So the term accuracy has a very concrete definition and we benchmark excessively against real world scenarios to always test test ourselves and make sure that the accuracy that we provide our clients is state of the art. Now having said that, the definition inside our system is not set in stone. Our goal is to allow any non technical person in the organization to be able to shift the model's understanding on the fly and by proxy also influence the accuracy of the model without needing to train a model. So what our system allows our clientele is to approach the problem not in the sense of I want to detect this and this is how it detects it from now on. But in the sense of I get this initial result, now I'm now I am able to influence it using different parameters, for example giving it false annotations, adding negative prompts, or just shifting the equation. Because I personally as the product manager, now that this is up and running, understood better what I want to detect. So our goal is to shift this equation from how well does this model perform to how well can I Define the problem to make it perform better.

You're building one of the fastest growing startups in the AI vision landscape. ARGU is an AI operating system that connects to physical cameras and real time video analytics. I want to talk about surveillance versus safety because that's a gray zone. How do you ensure your technology is is empowering people rather than watching them?

Amazing question. So the focus of the product in the traditional sense is to empower the individual in the organization that is currently the head of anything related to security. So when we install the system, we install it based on permissions. We want to give the power only to the individuals that need this power.

So.

So we will not create a situation where an unauthorized individual is able to access data that was otherwise inaccessible to them, for example tracking an individual. So when we install with different clients, we first of all understand the current limitations. For example, in Europe, facial recognition and LPR is a very sensitive subject and we adhere deeply to the constraints in that field. And besides that, managing inside of an organization, the way people can interact with it, which cameras they can interact and which system capabilities is a deep concern that we try to mitigate as much as possible in the development and the integration with clients.

I want to talk about the future of AI and vision intelligence. You're building one of the fastest growing companies in the AI vision landscape. You're connecting to physical cameras and real time video analytics. What's one application of vision language AI that will shock the world in the next three years?

Most surprising application I think that people are not expecting from not even VLMs vision language models, but from the world of multimodal language models in general. Is the modality fusion. Think about searching when did the altitude of an airplane change? And the model will be able to read the gyroscope data and search through it without needing any translation in the middle. Not language, but a totally different modality of understanding this fusion I think can unlock so many use cases that are currently just require either expertise from an individual that is able to read this data or just infeasible because of the volume of data. So this is the use case that I'm personally most excited about because when we will reach this modality fusion and then we'll be able to not only interpret the world visually but but also using sound and different sensors. This is the point where we will truly unlock the potential of AI.

You are building one of the fastest growing companies in the vision AI world. Your company is a real time AI operating system that connects to physical cameras and real time Video analytics. Cameras are everywhere. Are we heading toward a world where every camera becomes a real time reasoning AI agent?

Definitely yes. Cameras are powerful tools that are currently just being underutilized. Think about how little we knew about what we can do with a computer when it was first invented. Even though we already had the basics, the programming language, we already could program things. Even though it wasn't that powerful, we could already, having the knowledge we have today, create programs that do 10 times more than what they did back then. This is where we are at with cameras. They are there monitoring, collecting data, but still less than 1% of their potential is currently being tapped into. And this data just sits out there waiting for somebody to create something that will be able to take it to the next level.

You're at the forefront of AI vision. You're building an AI operating system that connects to physical cameras and real time video analytics. 10 years from now. What does the word camera even mean?

So this one can be taken to so many different extremes, but I think that in 10 years from now, the generation that will look at the camera and think this is a camera will probably be more accustomed to it doing everything for them. I think the manual settings, the manual operations that you're currently doing for either, searching for either, defining for, interacting for modifying the camera will all be simple down to a process where you just want what you need and you get it without any of the barriers in between. Being specialized in this field will still be something that is present. You will have people who know how to operate a manual camera, but it will be a lot less common and that we and it will be a lot less necessary to achieve the same results.

AI started with text to text and now it's going towards vision and audio. And the intersection of AI and the physical world is just beginning. What happens when every object, not just cameras, has eyes and a brain?

That's an interesting one. In such a world, possibilities can truly be unlocked for a variety of use cases we can't even think about. Right now I'm the most excited about a time when all of those different objects will share their data with each other and will allow us to get insights about every aspect of our lives, helping each of us personally improve and see the full picture of anything we want. That can be, for example, understanding our health habits, from just the passive monitoring of different devices around our house, to us being able to back propagate different security incidents because we just have the data everywhere. So this is what personally I envision will happen when we will start integrating AI into so many different mediums that we'll be able to just recollect any event in a matter of few seconds.

You're at the forefront of AI when it comes to vision. You're building an AI operating system that actually connects to physical cameras and real time video analytics. The thing is, is that the public's fear of AI is so how do you earn and maintain trust in a space that people are fearing?

We focus highly on giving our clients the confidence that their data will remain as private and secure as possible. To mitigate this fear, we have two main points where we prioritize the client to help him mitigate the risks of the system being used for inappropriate purposes. The first one is us segregating the system by permission on every level, down to the camera, so the client can decide who can access what from system functionality to the individual camera stream. This allows him to segregate each one of the functions by the priority and by the needs of individual people in the organization and by that prevent abuse. The second part is that the system is deployed fully on prem. It does not send any data outside of the organization. This allows our clients to have the confidence of the fact that their data will not be used to any by any third party for any purpose and will remain secure and encrypted on their hardware.

You're the founder of one of the fastest growing companies in AI vision. You're backed by Andreessen Horowitz, A16Z, and you are building an AI operating system that connects to the physical cameras and real time video analytics. Right now in the world, defense is big. It's always been big. But technology when it comes to defense is getting very popular, let's say. How do you think AI vision is going to impact defense?

So it will impact it in a very broad manner. Defense is a sector where vision can be used heavily throughout the entire pipeline from intelligence gathering, interrogation, to active combatants in robotics, or even to assess risks in different scenarios to logistics. Every base still has logistics behind the scenes, so defense will be impacted heavily in all of those sectors. But my personal thoughts is that defense will be less human oriented and will be more robotics, intelligence and cyber oriented in the upcoming future. And because of that shift, vision will play a crucial part in the way intelligence and decisions are being made without the need to involve active combatants in the field to act on those decisions later on.

Lior, you're building argu. It's one of the fastest growing startups in the AI world. You're backed by Andreessen Horowitz, a16z and you're building AI for vision. And you know, we are familiar with large language models and how they affect text to text. We're now, you know, seeing a lot of applications when it's voice related. But vision, this is, I think less popular. So I'm curious, what are some of the most prominent applications that are going to come about that everyday users are going to be able to engage with and get value from with AI vision?

Excellent questions. The consumer use case for AI vision is immense, but it is currently highly limited by the fact that you need to have something that actively monitors your day to day to get that value. And my personal thought is that we will not see a shift of day to day video recording in the next 10 years. Having said that, I do think that day to day consumers will be able to live in safer cities, to operate in highly optimized factories and to just in their day to day life, feel safer, more secure and know more about their environment thanks to vision Technology. Technology because the way it will monitor the public spaces and allow us to create a society that can act upon use cases a lot faster and a lot more deliberate manner than we had before.