In the rapidly evolving landscape of artificial intelligence, startups are increasingly recognizing the value of proprietary data as a cornerstone for innovation and market differentiation. This shift marks a departure from traditional practices of sourcing data from public domains or relying on third-party providers. Instead, companies are investing in the collection and curation of unique datasets tailored to their specific needs.
The Case of Turing: A Hands-On Approach to Data Collection
Consider the experience of Taylor, an artist who, along with her roommate, participated in an intensive data collection project for Turing, an AI company focused on developing advanced vision models. For a week, they wore GoPro cameras on their foreheads, capturing synchronized footage of daily activities such as painting, sculpting, and household chores. This initiative aimed to provide the AI system with diverse perspectives on various tasks, enhancing its ability to understand and replicate human actions.
Taylor described the process: We woke up, did our regular routine, and then strapped the cameras on our head and synced the times together. Then we would make our breakfast and clean the dishes. Then we’d go our separate ways and work on art.
Despite the physical discomfort, including headaches from prolonged camera use, Taylor found the work rewarding, both financially and creatively. She noted the challenges: It would give you headaches. You take it off and there’s just a red square on your forehead.
Turing’s strategy involves contracting individuals from various professions—chefs, construction workers, electricians—to gather a wide array of data. Sudarshan Sivaraman, Turing’s Chief AGI Officer, emphasized the importance of this approach: We are doing it for so many different kinds of blue-collar work, so that we have a diversity of data in the pre-training phase. After we capture all this information, the models will be able to understand how a certain task is performed.
Fyxer: Leveraging Specialized Data for Enhanced AI Performance
Another example is Fyxer, an email management company utilizing AI to sort emails and draft replies. Founder Richard Hollingsworth discovered that the effectiveness of their AI models was significantly influenced by the quality of the training data. This realization led Fyxer to employ experienced executive assistants to train the models, ensuring a deep understanding of email management nuances.
Hollingsworth explained: We realized that the quality of the data, not the quantity, is the thing that really defines the performance.
This approach required a substantial investment in human resources, with executive assistants sometimes outnumbering engineers four to one. However, the focus on high-quality, proprietary data has proven to be a key differentiator in Fyxer’s AI capabilities.
The Competitive Advantage of Proprietary Data
The emphasis on proprietary data collection is not merely about improving AI performance; it also serves as a strategic move to establish a competitive edge. By owning unique datasets, companies can create barriers to entry for competitors and offer more tailored solutions to their customers.
Hollingsworth highlighted this advantage: We believe that the best way to do it is through data, through building custom models, through high-quality, human-led data training.
Synthetic Data: Amplifying the Importance of Quality
The use of synthetic data—artificially generated data that mimics real-world scenarios—has become prevalent in AI training. While it allows for the expansion of training datasets, the quality of the original data remains paramount. Sivaraman noted: If the pre-training data itself is not of good quality, then whatever you do with synthetic data is also not going to be of good quality.
Investor Perspectives: The Value of Proprietary Data
Venture capitalists are also recognizing the significance of proprietary data in AI startups. Paul Drews, managing partner at Salesforce Ventures, stated: It’s really hard for AI startups to have a moat because the landscape is changing so quickly. He emphasized the importance of differentiated data, technical research innovation, and compelling user experiences.
Similarly, Jason Mendel, a venture investor at Battery Ventures, pointed out: I’m looking for companies that have deep data and workflow moats. Access to unique, proprietary data enables companies to deliver better products than their competitors.
Bria: Ethical AI Training with Licensed Data
Bria, an AI company specializing in image generation, exemplifies the trend of using licensed data for training models. By partnering with entities like Getty Images, Bria ensures that its AI models are trained on legally obtained content, addressing ethical and legal concerns associated with data usage.
CEO Yair Adato explained: Bria has mitigated biases that can sometimes emerge in AI-generated visual content by training its models on globally representative datasets.
Conclusion
The shift towards proprietary data collection and curation reflects a broader understanding within the AI industry: the quality and uniqueness of data are critical drivers of success. By investing in their own datasets, AI startups are not only enhancing the performance of their models but also building sustainable competitive advantages in an increasingly crowded market.
