Turing Labs is an innovative AI company that creates new ways to collect data. Their visionary efforts are making this new model of artificial intelligence development possible. As a company focused on vision models, we are constantly on the lookout for high quality datasets. In order to do this, they hire a creative inter-disciplinary team of skilled professionals who work with artists and electricians. This strategy is indicative of a broader trend in the nascent AI industry. Organizations are getting a handle on their data inputs to improve model accuracy.
Taylor, a data freelancer, who was recently contracted by Turing Labs to work on these efforts. She needed to create five hours of synced content per day. This commitment required seven hours of her energy to provide enough time for breaks and recuperation. To capture data in a new way, Taylor strapped GoPro cameras to her forehead. This creative methodology allowed her to document real-time spaces and activities as she lived through them.
Richard Hollingsworth, Turing’s founder and CEO, shared the reasoning for taking the process manual and low-tech. What he found is that using more smaller models trained on very narrow, very specific datasets work better. Hollingsworth noted that the richness of that data is what really makes performance true performance. It’s not the quantity of data that matters, instead, it’s the quality. This philosophy carries over into Turing’s overall approach of custom models delivered via top-tier, human-led data training.
The company’s Chief AGI Officer, Sudarshan Sivaraman, explained why data diversity is crucial in the pre-training stage. “We’re trying to place primary emphasis on all types of blue-collar work. As long as we can’t prevent it, we need to address it,” Sivaraman said. This strategy helps us obtain varied data in the pre-training phase. Turing goes a radical route, building its own models from scratch. Players such as Fyxer rely on established foundation models.
Hollingsworth identified the difficulties of gathering a specialized workforce to help with data collection. “It’s a very people-oriented problem. Great people are very hard to find,” he noted. This is a critical barrier that Turing is addressing by hiring effective executive assistants. They’re most interested in ironing out the nuts and bolts of data processing, such as determining when to open an email.
Taylor recounted her experience working on this innovative project: “We woke up, did our regular routine, and then strapped the cameras on our head and synced the times together.” This type of hands-on approach is a great example of Turing’s devotion to collecting diverse and high-quality datasets that further their AI mission.
Sivaraman made the case for collecting better, more robust data. He emphasized that if we discredit bad pre-training data, we would make all future attempts to create synthetic data a failure. He warned that shoddy pre-training data results in shoddy results even with synthetic data. If the premise is not of a solid foundation, then nothing established on top of it will be worthwhile.
Turing Labs is quickly inserting itself deeper into the operational fold and fine-tuning its approach. Clearly, this massive growth will be felt throughout the already-booming AI industry. The company’s focus on manual data collection and diverse professional backgrounds may set a new standard for how AI startups approach their data needs. To Turing, quality trumps quantity. By encouraging partnership among industry, academia, and government, it is firmly establishing itself as a leader in the rapidly changing space of artificial intelligence.