Engineered signal that turns the world's largest library into usable knowledge—and expands it by translating numbers into narratives.
In the past five years, frontier models have ingested a large fraction of what's publicly available in text. A single large training run now spans roughly 15 trillion tokens—on the order of 11 trillion words.
The corpus is immense. Common Crawl adds billions of pages each month and has indexed hundreds of billions overall. The Internet Archive's Wayback Machine is approaching a trillion snapshots. This is the largest public, machine-readable library humanity has ever assembled.
But its apparent vastness is misleading. Most publishing concentrates around a narrow set of attention-bearing topics, and much of the marginal output is duplication: variations of the same summaries and takes, tuned for ranking and persuasion. The ceiling is human attention-hours—what publishers believe human audiences will click, read, and share.
The web is also an economic system. In the United States alone, digital advertising clears hundreds of billions of dollars annually. Incentives reward ranking, clicks, and retention. Pages are tuned to win attention, not to support reasoning.
In that economy, "coverage" is a mirage. Production clusters where attention monetizes, and the marginal page is usually recombination: duplicated explainers, recycled commentary, SEO-shaped templates. You get more pages without getting more decision-relevant facts.
For models, the web mostly arrives as prose with weak structure—an undifferentiated mass of text where who said what, about whom, when, and with what evidence is often missing or implied. Retrieval then returns clusters of near-duplicates, so systems spend tokens re-reading instead of updating beliefs.
Even curated corpora carry boilerplate and duplication, and filtering choices change what gets represented. The open web is a noisy prior, not a knowledge base.
Public text is finite, and training demand is catching up to it. You can see it in practice: better-curated, better-structured corpora beat raw scrape volume. The leverage is in curation, provenance, and refresh.
Search helps machines find pages. Synorb begins after retrieval. We rewrite what the web and trusted datasets say into a single, structured, machine-native corpus anchored to people, organizations, and data.
Machines don't have query windows. They can read continuously, accumulate context over time, and benefit from coverage that expands beyond what human attention economics will ever fund. A corpus built for humans is large, but structurally limited in scope for infinite-attention consumers.