The industry has hit a data-quality inflection point. Here's why "every fact traceable to a source" is the standard that matters now.
Ilya Sutskever's widely-discussed 2024 remark that pre-training-era scaling has reached "peak data" is well corroborated by independent modeling — Epoch AI's own projections tie any further scaling runway explicitly to curation and filtering quality, not raw volume. The shift underway across the industry is from "more data" to "better-curated data."
a16z has argued in print that durable data value now requires "verifiable, structured" data rather than scraped text. That distinction matters more than it sounds: a dataset with a wrong number in it isn't neutral — a confidently wrong fact is worse than an honest gap, because it's indistinguishable from a correct one until something breaks downstream.
source_urls field — the actual pages a fact was checked against — and any figure we can't verify is marked null rather than estimated. That's not a feature bolted on for marketing; it follows directly from the shift described above.Most data products in this space are optimized for volume and speed — reasonably, since that's what most buyers have historically asked for. Full per-fact citation and a hard no-estimate rule slow down production in exchange for something buyers can actually check themselves rather than take on faith. That trade only makes sense once "more data" stops being the binding constraint — which is roughly where the industry is now.