While every company wants in on the AI party, the hurdle of turning thousands of gigabytes of data into an app is steep, not to mention costly. That’s where Labelbox comes into play.
The company starts with a standard foundational AI model, then injects unique data — what co-founder Brian Rieger calls the “real signal” — to build the system a brand wants, from certifying underwater welds to tracking trains arriving at a station. Companies no longer need to program to build an AI model for their business, they just need to tap the know-how they already have.
“The most exciting part of AI is that you don’t need to be an expert in AI anymore,” he says. “You need to be an expert in your domain.”
Hear more from Rieger in the video below or read our lightly-edited transcript.
Transcript:
Labelbox co-founder and Chief Product Officer, Brian Rieger
How did we come up with Labelbox? It’s kind of funny. For many of the companies have started, I’ve worked with my current other co-founder, Manu (Sharma), and we would always think so hard about the name, what are we going to name this thing? Just so many anxiety filled conversations about that, because we thought it was so important. With Labelbox we’re, “Why don’t we just work on the company and stop thinking about the name?” So obviously, when we started the business, it was all about labeling. A box is kind of like a nomenclature in Silicon Valley for a place to do things, store things, Labelbox, and we thought about it for about 30 minutes, and we got back to working on the company. Maybe that, three hours of not thinking about the name was the difference. But it was just that simple. There isn’t any super elegant rationale behind it.
Labelbox is a data-centric AI platform for building intelligent applications. What that means is the way we teach AI how to understand the world and how to make decisions on behalf of our companies and our teams, and the particular applications that we’re applying it to is through examples and through data. So Labelbox is that data-centric system that allows teams and AI developers and engineers to teach these AI what to know and how to make decisions and ultimately, how to perform for the business applications that are being used for.
One of the cool things about these new AI models that are available, the GPTs and these foundation models, you can say from lots of different companies and open source communities, is that they do have general understanding. What that does is it allows you to start somewhere. Five years ago, you would have had to start with a bunch of data and a bunch of labeled information and all of these things in order to just get that model to work in and of itself. But now we have these foundation models and these generative AI models that are generally knowledgeable, and they can kind of figure things out to some extent. So the place to start and the opportunity is, we can take a business problem. Let’s say it’s figuring out where in a legal document there are risky clauses for my business. We can apply one of these New Age machine learning models to it, and we can start to work with it in terms of, where are the risky clauses, and it’ll come back with some answers. From there, we can begin to teach this model from that point about our particular application, our particular way of understanding legal clauses and the peculiarities of our business. We do that through labeling or reinforcement learning, or fine tuning or retrieval augmentation system, there’s lots of different things out there to do this. But the great part is there’s sort of a foundation to start with. Even if you’re not an AI expert, you can start there because these are fairly simple models to use, to apply to your problem and kind of see what you get.
As we’ve always believed at Labelbox, the most exciting part of AI is that you don’t need to be an expert in AI anymore, you need to be an expert in your domain. That’s the real hardy information, the real signal and the noise that the AI can use to understand your business. Because you no longer really have to program the models anymore. The models or algorithms are sort of programmed already, there sort of setup already. The real programming is done through example. So if you’re an expert in a particular thing, like manufacturing of a particular shampoo, and you’re trying to do a quality assurance process on that, your expert understanding of how to understand the quality of a manufactured shampoo product is the exact data and expertise that an AI model needs from its foundation to become useful and perform it in your business domain. If you’re an expert, I think the huge opportunity in AI is you have the knowledge in your head to really take an AI model that you find online and turn it into something really compelling for the business.
We started Labelbox in 2018. The challenge that my founding team and I encountered was we were trying to build a computer vision system for satellite imagery to understand certain aspects about the world. We had lots of camera images from the satellites, and we were trying to provide intelligent insights and actionable data for businesses, from banking to the government to agriculture, applications. Things like counting the number of cars in a parking lot on Black Friday to understand whether or not BestBuy is going to do well in their earnings report or understanding the level of oil in oil tank reserves in different parts of the country. These are insights that businesses can use to more intelligently run their companies and make decisions about the future, and to understand the logistics of the world: the nature of crop health, and the nature of the Amazon’s health and how much of the Amazon is being cut down versus how much is growing, and how, the climate science understanding in terms of what we should expect from, you know, the future of global warming so far. So lots of different applications.
Our insight was that the process of taking data and turning that into AI ready, training data used to train a machine learning model was very difficult. You would have to use all these different services, third party services, and what companies really need to do is use their own companies expertise to draw on these images or input their own expertise, like this is an oil level, that’s a certain level, or this is a, area of the rain forest that has been cut down. Those are all computer vision, sort of problems that you need to teach an AI, an artificial intelligence model, how to understand. So the process of taking raw satellite imagery and turning that into data that a machine learning system can use to understand the world is actually fairly complex. It’s sort of like developing software. There really wasn’t a tool available to do this data-centric workflow. We built Labelbox to serve that purpose for all industries.
The challenge right now is that the AI capability and just the pace of innovation is moving so quickly, that it’s hard to put a pole in the sand and say, “Okay, we’ve got the tech, how do we build testing and best practices around that?” Because the second, we put that pole down we’ve got a GPT5 that can do a whole new myriad of things. We’re sort of in that world right now. It’s one of the fastest technology waves that certainly I’ve seen.
What really excites me about working at Labelbox, I would say in general, because our product is directly involved with the data part of building AI and machine learning technology, we get to interact with companies like Procter and Gamble, that are working on really interesting problems. We get to see those problems manifest in data, which means we get to see the visual information or the textual information. Our company is built of people that love understanding the world and are curious about how things work and that sort of thing.
For us, the most exciting thing that we get up in the morning for is, we get to see how the whole world kind of does things. Like how does a system track which trains go down the track. It turns out, they actually use sound to understand which trains are at which stations. You’d never think of it that way. But there are these new AI systems that can track trains with sound and with the whistles and so forth. It’s stuff like that. Or we work with a company that creates custom nails that scans your whole nail and can give you a perfectly fit nail to stylize with your own design. We just get to see these amazing use cases, underwater welding where you can where someone can certify their weld for a structure using a vision model that’s right next to them. And we get to understand how a good weld looks like and what it means for a weld to be certifiable for 100 meters underwater.
We just love that we just get to see the world and understand how it works and see what companies are trying to do to improve their businesses and improve the products that they provide to their customers. So that’s what’s what we love. We talk about it every week we get together to show different use cases that are part of our user base customer base.
