Why We’re Building Poseidon

Unlocking the most valuable and overlooked form of IP: real-world data.

Real-World Data is the New Frontier

We’ve exhausted internet-scale data. As AI systems become more capable and general, they also become hungrier for data that reflects the richness and edge cases of the world we actually inhabit. Public datasets, many of which form the backbone of today’s foundational models, have been used and reused to the point of diminishing returns. Common Crawl, Wikipedia, and piles of scraped forum content still have their place, but they weren’t built to power the next, physical wave of AI: robots that move through kitchens, autonomous vehicles that drive at night, and voice-to-text features that analyze different accents and dialects.

I learned this firsthand while building models during my PhD at Stanford. Most of the footage I recorded was mundane, but the 1% that captured actual edge cases (such as Waymo vehicles making unexpected turns or failing to parallel park) drove the most significant performance gains. That long-tail data is where robustness lives.

This shift from screen-based intelligence to embodied AI changes the entire equation. The data that matters most for what’s next isn’t sitting on web pages. It’s hiding in the long-tail: stereo video of someone navigating a messy home, audio files with whispered speech, or motion data from hands folding laundry. This data is the most valuable subset of intellectual property, as it is used to train AI to understand the subtle chaos of the real world, and it’s exactly the kind of data that’s scarce, unstructured, and legally fraught to collect.

What Poseidon Is

Poseidon is a full-stack, decentralized data layer for AI, purpose-built on Story to bring real-world data into the next wave of AI innovation. It combines structure, licensing, and incentive alignment in a way that turns scattered, high-value edge-case data into something composable and accessible for builders. 

At its core, Poseidon is designed to make the hardest-to-source data usable at scale. That means robust metadata and standardized formats for multi-modal input. It means ingestion pipelines that work for sensor data, not just scraped text. And just as critically, it means IP clarity from day one so that contributors can share data confidently and model developers can use it commercially, without risk. Poseidon is built on top of Story’s licensing infrastructure, which means every dataset can carry its own terms, provenance, and logic for how rights and revenue flow.

​​Scaling this requires automation without compromising on quality. My research lab developed uncertainty-driven validation systems that flag only the ambiguous edge cases for human review. We’ve baked those learnings into Poseidon’s pipeline, allowing us to scale high-signal data collection while keeping quality intact, even at network scale.

Building a Better Data Economy 

Poseidon isn’t just infrastructure, it’s a new economic layer for AI, one where long-tail, high-signal data can be shared safely and fairly without being trapped in institutional silos or lost in legal ambiguity. 

By tying together licensing on Story, metadata, and incentives, Poseidon allows datasets to behave more like software – versioned, traceable, and legally enforceable – while still remaining contributor-driven. This opens the door to new kinds of datasets that evolve over time, improve through network participation, and carry clear commercial rights from the moment they are created.

This is what was missing when I tried to crowdsource robotics datasets in academia. Even the data I collected myself couldn’t be cleanly shared or reused. There simply was no infrastructure for IP-cleared data collaboration, even though it was mission-critical for advancing robotics and other areas of AI.

What’s Next

Poseidon is just getting started, but the foundation is here: structured ingestion for complex, real-world data; built-in licensing and traceability only possible on Story; and an incentive model that rewards the contributors who power AI’s next wave. We’re proud to be backed by a $15m seed round from a16z crypto and to be working with the world’s leading AI companies that understand the magnitude of what’s ahead.

If you’re training AI for the physical world, we’d love to talk. If you’re collecting unique datasets – sensor data, voice recordings, edge-case video – we’d love to help you bring that data into the ecosystem safely while ensuring you’re recognized and rewarded for your contributions.