- Cerebral Valley
- Posts
- Labelbox is AI's high-quality data factory 🌐
Labelbox is AI's high-quality data factory 🌐
Plus: Founder/CEO Manu on Boost, evals and alignment...
CV Deep Dive
Today, we’re talking with Manu Sharma, Co-founder and CEO of Labelbox.
Labelbox is the data factory that gives AI teams an unprecedented level of data quality and control over the labeling process. Founded in 2018 by Manu, Brian, and Daniel, Labelbox was born out of their experiences at companies like Planet Labs, where the need for high-quality labeled data to power deep learning models became evident. Labelbox offers a comprehensive suite of tools and services for data labeling, evaluation, and alignment, catering to both frontier AI labs and mainstream enterprises.
Today, Labelbox serves a wide range of customers, including AI labs, autonomous vehicle companies, and major enterprises. Their Boost labeling services, powered by the Alignerr platform, connects skilled domain experts with companies needing precise AI alignment, further enhancing their offerings.
Labelbox last raised a $110 million Series D in 2022, led by the Softbank Vision Fund, with participation from Andreesen Horowitz, Catherine Woods (founder of Ark Invest), Databricks Ventures, Snowpoint Ventures, and B Capital.
In this conversation, Manu discusses the founding story of Labelbox, the competitive landscape of data labeling, and what sets Labelbox apart in the market.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with Manu 💬
Manu - welcome to Cerebral Valley! First off, give us a bit about your background and what led you to found Labelbox?
Hey! Thank you very much for having me. I've been running Labelbox for about six years, and I co-founded the company with Brian and Daniel in 2018. The idea came from insights we uncovered while working at our respective companies. I was at Planet Labs, one of the leading geospatial satellite imaging companies. Planet Labs operates like a Starlink constellation but with satellites taking images.
At the time, deep learning was taking off, and there was a lot of excitement around self-driving cars. However, we saw a broader opportunity: deep learning models were going to power many facets of life. The only way to create those models at the time was by labeling data with human input. We saw an opportunity to build a product for this.
Initially, we thought Labelbox would be a small business, not a venture-backed company. This was our third business together with Brian. We moved to a low-cost place outside the Bay Area and built a product on nights and weekends, launching it on Reddit. It took off quickly. Every week, more people signed up, and we were excited about solving their problems and learning about new ones. Within a couple of months, we were generating revenue and had a list of features our customers wanted. At that point, we knew we were onto something and decided to raise capital.
Give us an overview of Labelbox for the uninitiated AI developer.
It's quite simple: we focus on high-quality data labeling. We offer tools for customers to do it themselves, and we provide an end-to-end service. All of these capabilities are part of the same platform. This is what makes Labelbox unique. Other players might provide labels in a CSV file or an API call, but it's up to the customer to vet the quality. With Labelbox, we offer all the tools needed for this process.
Users can make comments, track label quality, and access performance dashboards. Both the labelers and our customers use the same tools on the same platform. This integrated approach streamlines the process and makes life easier for our customers.
We see human evaluations (evals) as part of the same problem as data labeling. Whether it's instruction fine-tuning (SFT) or preference ranking, it's all about asking humans for feedback. This feedback can take various forms, from supervised experiences to semi-supervised techniques like clustering or pre-labeling with foundation models.
Developing AI models, whether frontier models like LLMs and multimodal models or task-specific models for mainstream enterprises, requires two key things: compute and data. Labelbox is the data provider. We are essentially a data factory that produces the data needed for these models. We specialize in creating data touched by humans.
In summary, Labelbox offers a comprehensive platform for all forms of data labeling and alignment, providing a unified experience for generating high-quality data for AI models.
Who are your users today? Who’s finding the most value in what you’re building with Labelbox?
It depends on the products and services our customers are using.
For frontier AI labs, when we provide the data, it is typically a fully managed experience. Here, it's usually the data operations and data labeling teams that run a shared service inside the organization. They work with us to kick off new campaigns, and we deliver the labels. Ultimately, these labels are sent to the researchers and machine learning teams who use the data to improve their models.
For companies using our software platform, there are quite a few personas involved. 1) Machine learning engineers, 2) data labeling managers, and 3) the actual data labelers interact with our product on a daily basis.
Machine learning engineers focus on the quality of data, pre-labeling, and making the entire pipeline more efficient or automated.
Data labeling managers are concerned with throughput, reporting, and ensuring that the data factory produces the right kind of labels at the right price and throughput.
SMEs (subject matter experts) or data labelers prioritize having a great experience while labeling data, ensuring the tools don't get in the way and that they get compensated fairly.
Our products are used by these three groups quite frequently, catering to their specific needs and workflows.
Any customer success stories you’d like to share?
A vast majority of generative AI consumer apps that people use today likely have models trained with data powered by Labelbox. While I can't name all of them, many top LLM apps rely on us for their data.
One public example is Google Cloud, which introduced a human LLM evaluation service for all its customers powered by Labelbox. This highlights the level of trust and success we've achieved.
Another example from the mainstream enterprise sector is John Deere. Their autonomous tractors can detect and spray weeds with herbicides, reducing herbicide use by 90%. The entire data engine behind this precision is powered by Labelbox.
For frontier AI Labs, which is our current focus, there's a lot of excitement. Many generative AI startups working on text-to-image, text-to-video, and text-to-audio use cases are using Labelbox today.
When most people think of data labeling for AI, they first think of Scale. Could you help us understand the competitive landscape of this space? What sets Labelbox apart?
What sets Labelbox apart is the unprecedented data quality and level of control. Let's dig deeper into what those two things really mean.
First, Labelbox is the only company that provides a guarantee on our data quality. This means customers will only pay us if it matches the quality thresholds they've set, in any configuration and at any scale. We stand behind that, not only for our managed labeling services but also for our platform and tooling.
Labelbox offers both labeling services and a comprehensive platform that allows even Frontier AI labs and researchers to log in, QA the data, provide feedback, and access the labeling teams directly. This direct access to labeling teams is crucial for creating the edge in AI right now. Our platform helps customers achieve the highest quality data, which is vital because, in generative AI, quality is often subjective and linked to the model's reasoning capabilities or other key dimensions.
Regarding control, we are not a black box service. Our product was born as a tools company, providing a full set of tools, techniques, and workflows that allow customers to configure and fine-tune their data pipelines as they see fit. This flexibility ensures the most optimized data production for their models.
For example, Frontier AI labs might want to label data with an internal team of experts for a new idea they're testing. They can use our tools for this, or they can outsource to our experts through our platform called Aligner (alignerr.com). This platform is similar to Fiverr but focuses on AI alignment, with domain experts and PhDs labeling data for our customers. This level of flexibility and control is unprecedented, and Labelbox is the only company currently offering it.
What trends in AI currently excite you the most? How do upcoming trends such as agentic and multimodal AI affect your roadmap for Labelbox?
Well, I think the most exciting trend I see is the merging of AI tasks into a unified architecture, where transformers can handle everything. Initially, this started with LLMs, but now we're seeing models like Microsoft's Florence, which can perform segmentation, bounding box detection, and chat-like text experiences all within the same architecture. This convergence is thrilling because it hints at the development of a general-purpose system capable of interacting across various modalities without requiring extensive engineering. We're likely a few quarters away from fully realizing this, but it's on the horizon.
Another exciting trend is the shift towards enhancing the utility of these models in practical contexts. While current LLM leaderboards showcase impressive capabilities, they often represent vanity metrics. The real challenge now is to make these models more useful for specific tasks and jobs. This ties into the concept of agentic AI, where models perform more complex tasks autonomously. I believe we'll see significant breakthroughs in this area soon.
On a broader scale, the paradigm of reinforcement learning from human feedback (RLHF) is proving to be more substantial than many anticipated. Essentially, any task a human can demonstrate can be learned by a model, given enough data. We're seeing this with Tesla's imitation learning from drivers and with LLMs learning from human feedback. This framework, where models learn from human-performed tasks, is incredibly promising.
While this may not necessarily lead to AGI or superhuman intelligence immediately, the fact that we have a solid framework for models to learn a wide range of human activities is very exciting. It opens up numerous possibilities for making AI more integrated and effective in everyday applications.
Can you tell us about the culture at Labelbox? Are you hiring, and what do you look for in prospective team members?
Absolutely. Our culture at Labelbox is built around three core values.
First, craftsmanship. Our team members love the tools they work with and take pride in building the best possible experiences. We're not always the first to market, but we strive to be the best. This ethos drives us to create the highest quality products, like our RLHF offering, which we believe is the best available.
Second, curiosity. Our people are always learning and staying humble in the face of rapid changes in the industry. This fosters a spirit of collaboration and a constant drive to seek understanding and improve.
Third, a customer-first mindset. Every decision we make is aimed at helping our customers succeed. This is easier said than done, as it requires hundreds of daily decisions to be made with the customer's best interest in mind.
We're a down-to-earth, hardworking team, passionate about the field and committed to building the best products. We are hiring across various roles, particularly in engineering, product management, and customer success. We're looking for individuals who share our values of craftsmanship, curiosity, and a customer-first mindset, and who are excited about contributing to the advancement of AI.
Conclusion
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.