Rapidata, a new startup, has developed a platform that significantly shortens AI model development timelines compared to traditional methods. The company’s approach effectively gamifies reinforcement learning from human feedback (RLHF) by leveraging a global network of nearly 20 million users from popular apps like Duolingo or Candy Crush. This allows AI labs to iterate on models in near-real-time, eliminating the need to wait weeks or months for a single batch of feedback.
The platform converts digital footprints into training data by offering users the choice to provide feedback for AI models instead of watching traditional ads. Rapidata’s crowd intelligence approach reaches between 15 and 20 million people, processes 1.5 million human annotations per hour, and ensures quality control and anonymity for users.
One of the key advancements introduced by Rapidata is ‘online RLHF,’ which integrates human judgment directly into the training loop by partnering with GPUs running the model. This real-time feedback mechanism prevents reward model hacking and enhances the training process with human nuance.
By providing a scalable network for AI teams to access human judgment at a global scale, Rapidata aims to redefine the AI development landscape. The company’s $8.5 million seed round investment signifies the growing importance of incorporating human feedback efficiently into AI model training.
Source: VentureBeat