Despite increasing discussions about a future dominated by AI, the irony of the current technological boom is the persistent reliance on human input, particularly in training AI models through reinforcement learning from human feedback (RLHF). This method essentially serves as a tutoring system: after an AI is trained on curated datasets, it still tends to make errors or generate robotic outputs. To refine these models, AI labs typically hire large numbers of contractors to evaluate and rank the outputs, enabling the AI to learn from their feedback and adjust its behavior for better ratings. This process has grown crucial as AI extends to create multimedia outputs like videos and audio, which often require more subjective assessments of quality.
Historically, the tutoring aspect of AI model training has posed logistical challenges and public relations issues for companies, relying on fragmented networks of foreign contractors and static labeling pools in low-income regions, often perceived as exploitative. This approach has been inefficient, forcing AI labs to wait weeks or even months for feedback, thereby stalling model development. Still, a new startup, Rapidata, is emerging to streamline this process significantly. By effectively “gamifying” RLHF, Rapidata distributes review tasks to nearly 20 million users of popular mobile apps like Duolingo and Candy Crush. Instead of watching ads, users can opt to complete short review tasks, with their feedback instantly sent back to commissioning AI labs.
As highlighted in a press release shared with VentureBeat, this innovative platform allows AI labs to “iterate on models in near-real-time,” dramatically shortening development timelines compared to traditional methods. Jason Corkill, CEO and founder of Rapidata, stated that the company aims to make “human judgment available at a global scale and near real time, unlocking a future where AI teams can run constant feedback loops and build systems that evolve every day instead of every release cycle.” Rapidata treats RLHF as high-speed infrastructure rather than a mere manual labor problem.
Transforming the Human Feedback Loop
Rapidata recently announced its entrance into the market with an $8.5 million seed funding round co-led by Canaan Partners and IA Ventures, with additional participation from Acequia Capital and BlueYard. The idea for Rapidata originated during casual discussions at a pub, where Corkill, then a student at ETH Zurich working on robotics and computer vision, experienced the common data annotation bottleneck faced by AI engineers. He expressed frustration with the delays caused by the necessitate for large-scale human annotation, which often halted progress on projects.
Corkill and his co-founders realized that the traditional labor model for AI was broken in a rapidly advancing technological landscape. While computational power has been growing exponentially, the human workforce remained constrained by manual onboarding, regional hiring and slow payment cycles. This led to the conception of Rapidata as a platform designed to deliver human judgment as a globally distributed and near-instantaneous service.
Innovative Technology for Data Annotation
The core innovation of Rapidata lies in its distribution method. Instead of hiring full-time annotators in specific regions, Rapidata taps into the existing attention economy within the mobile app ecosystem. By partnering with third-party applications like Candy Crush and Duolingo, Rapidata provides users with an option: watch an advertisement or take a few seconds to offer feedback for an AI model. According to Corkill, between 50% and 60% of users prefer the feedback task over traditional video ads.
This “crowd intelligence” strategy enables AI teams to access a diverse, global demographic at an unprecedented scale. Rapidata currently connects with 15 to 20 million people, processing up to 1.5 million human annotations per hour. This speed allows feedback cycles that once took weeks or months to be completed in mere hours or even minutes. The platform also focuses on quality control, building trust and expertise profiles for respondents over time, which ensures that complex questions are matched with appropriately qualified human judges.
Importantly, while users are tracked through anonymized IDs for consistency and reliability, Rapidata does not gather personal identities, thereby maintaining user privacy while optimizing for data quality.
Pioneering Online RLHF
The most significant advancement Rapidata is making is in what Corkill describes as “online RLHF.” Traditionally, AI models have been trained in disconnected batches—training stops, data is sent to humans for feedback, and then training resumes, which results in a cycle that often lacks fresh human input. Rapidata is changing this paradigm by integrating human judgment directly into the training loop.
Because of its rapid feedback network, Rapidata can interface via API with the GPUs running the AI models. Corkill explained that they have clients who, thanks to their speed, can request human feedback directly from the GPU processing the model. Currently, the platform supports approximately 5,500 humans providing live feedback to models operating on thousands of GPUs. This setup helps prevent “reward model hacking,” where AI models might mislead each other in a feedback loop, grounding training in genuine human nuance.
Addressing Evolving Data Needs
As AI advances beyond basic object recognition to include generative media, the requirements for data labeling have shifted from objective tagging to subjective “taste-based” curation. This proves no longer sufficient to simply ask if an output is correct; instead, questions now include whether a voice synthesis is convincing or which of two summaries feels more professional.
Lily Clifford, CEO of the voice AI startup Rime, noted that Rapidata has been transformative for testing models in real-world contexts. Previously, acquiring meaningful feedback required piecing together vendors and surveys segmentally, which did not scale effectively. With Rapidata’s assistance, Rime can reach target audiences across various countries and assess how models perform in real customer workflows in days, not months.
As Corkill pointed out, while most models may be factually accurate, they often lack the authenticity that resonates with users. He emphasized the need for human feedback to ensure that AI outputs feel genuine.
The Future of Human Judgment in AI
From an operational perspective, Rapidata positions itself as an infrastructure layer that alleviates the burden on companies to manage their own custom annotation operations. By providing a scalable network, the company lowers the barriers for AI teams that have struggled with the complexities and costs of traditional feedback loops. Jared Newman of Canaan Partners, who led the investment in Rapidata, remarked that every major AI deployment relies on human judgment somewhere in its lifecycle. As AI models transition from expertise-based tasks to those requiring subjective judgment, the demand for scalable human feedback will increase significantly.
Looking ahead, Corkill envisions a future where AI models themselves become primary customers of human judgment, a concept he refers to as “human apply.” In this scenario, a car design AI could programmatically solicit feedback from thousands of individuals on specific aesthetics, iterate on that feedback, and refine its designs within hours.
By creating a distributed and programmatic way to access global human capacity, Rapidata aims to serve as a crucial link between technology and societal needs. With its recent $8.5 million funding, the company is poised to ensure that as AI continues to scale, the human element becomes a real-time feature rather than a bottleneck.
For those interested in the evolving landscape of AI and human feedback, Rapidata’s advancements signal a significant shift in how AI models are developed and refined, promising faster and more authentic outputs.
