AI’s promise of automation is still being held back by one stubbornly human problem: training. Even the most advanced models rely on human judgment to refine their outputs—yet that process remains stuck in the slow lane, dependent on fragmented labor pools and static feedback cycles. Now, a startup called Rapidata is rewriting the rules by turning millions of mobile app users into an instant, global feedback network, shrinking what once took months into near real-time iterations.

The core idea is simple but radical: instead of waiting weeks for contractors to label data in bulk, AI labs can now tap into a distributed pool of 15–20 million people—already engaged in apps like Duolingo or Candy Crush—to provide feedback on demand. Users opt in to short review tasks as an alternative to watching ads, and their responses flow back to the model instantly. This isn’t just faster; it’s a fundamental shift from batch processing to what Rapidata calls online reinforcement learning from human feedback (RLHF), where human judgment is woven directly into the training loop.

The implications are immediate. Where traditional RLHF requires stopping a model, sending it to human annotators, waiting for feedback, and then resuming training, Rapidata’s system integrates feedback in real time. A model can now request human input mid-training, adjust its weights, and continue—eliminating the lag that once turned weeks into months.

How It Works: From Ads to AI

Rapidata’s approach leverages the existing attention economy. Users of partner apps are presented with a choice: watch a 30-second ad or spend 10 seconds reviewing AI-generated content (e.g., summarizing a document, rating voice synthesis, or comparing image quality). Roughly 50–60% opt for the feedback task, which earns them micro-rewards or in-app perks instead of ad revenue. The platform processes up to 1.5 million annotations per hour, with feedback cycles reduced from weeks to minutes.

Key to this speed is the platform’s ability to match tasks to the most relevant reviewers. Over time, users build expertise profiles—meaning complex judgments (e.g., evaluating nuanced voice tones or cultural appropriateness in generated text) are directed to those with demonstrated proficiency. Anonymized tracking ensures consistency without compromising privacy.

Why This Matters: The End of the AI Bottleneck

The bottleneck isn’t just about speed; it’s about scale and adaptability. As AI moves beyond text into multimedia—video, audio, and imagery—the need for subjective, context-aware feedback grows. Traditional annotation teams struggle to keep up with the diversity of global audiences or the rapid iteration cycles demanded by modern models. Rapidata’s model addresses this by

  • Global reach: Feedback from users in 190+ countries, enabling models to adapt to regional preferences (e.g., tone, humor, or cultural references) without delay.
  • Real-time integration: API connections to GPUs allow models to pause, request feedback, and resume—preventing reward-hacking loops where two AIs manipulate each other’s outputs.
  • Programmatic access: AI teams can now treat human feedback as an on-demand service, scaling up or down based on need, rather than managing fixed labor pools.
  • Taste curation: Beyond factual accuracy, models can now be fine-tuned for authenticity—whether an AI-generated email sounds human, a voice assistant’s tone feels natural, or a generated image aligns with cultural aesthetics.

For companies like Rime, a voice AI startup, this means testing models in real-world workflows across languages and regions in days, not months. Previously, gathering feedback required stitching together vendors by country or segment—a process that didn’t scale. Rapidata’s platform turns this into a seamless, automated pipeline.

A Vision for ‘Human Use’

Looking ahead, Rapidata’s founder, Jason Corkill, envisions a future where AI doesn’t just consume human feedback but programmatically requests it. Imagine an AI designing a car interior: instead of guessing what users prefer, it could dynamically poll 25,000 people in France, iterate on their feedback, and refine the design in hours. This human use model—where AI acts as a conduit for real-time societal input—could redefine product development across industries.

The $8.5 million seed round (led by Canaan Partners and IA Ventures) signals confidence in this vision. With the funding, Rapidata aims to expand its network and infrastructure, ensuring that as AI models grow more capable, the human element doesn’t become a bottleneck—but a real-time accelerator.

Takeaway: Rapidata’s platform isn’t just about faster AI training; it’s about making human judgment as scalable and instantaneous as compute power. For labs racing to build more nuanced, culturally aware models, this could be the missing link between silicon and society.