Project Mossly — Initial Concept

This page outlines the initial concept for Project Mossly.

The core idea is to launch a background initiative to collect internet data for potential use in model training. This effort would begin even before we have clear picture of what kind of data we’ll eventually need. As the AI startup space grows more saturated, the only lasting competitive edge will come from data ownership. In that sense, this may seem like a move driven by FOMO, but it’s better understood as a prudent investment in long-term defensibility.

Good data compounds in value over time. Compute continues to get cheaper each year, but data—especially exclusive, high-quality data—becomes more critical and less replaceable. Major tech companies have already poured millions, sometimes billions, into their foundational datasets, and they guard them closely.

While new state-of-the-art models keep emerging, many of them are just iterations—small improvements over previous ones, often driven more by better datasets than breakthrough architectures. You can fine-tune these models, but that comes at a cost: fine-tuning tends to skew the model’s understanding of the original data distribution. It might make the model extremely good at one task while degrading its performance across many others.

If your goal is to make a model that improves in a general way, you need a dataset that’s at least as comprehensive and well-distributed as the original. This makes data—especially original or hard-to-find data—a core asset for any serious AI company.

Public datasets are available, but they’re limited in both scope and uniqueness. To build something truly differentiated, we’ll need to go beyond what’s easily downloadable. That’s where Project Mossly comes in. The mission is simple: use web crawling to collect wide range of internet data quietly and safely. Most servers sit idle during the night; we can use that time to run background jobs that steadily build our internal dataset.

— Sprited Dev 🌱