AI runs out of public human text ~2026-2032 — making fresh real-world data the scarce input
Villalobos et al. (Epoch AI; arXiv 2211.04325, rev. 4 Jun 2024) forecast that on current trends LLMs will consume a dataset equal to the entire stock of public human-generated text between 2026 and 2032 — sooner if overtrained — pushing the field toward synthetic data, transfer learning and new sources. If public text is exhausting, proprietary, verified, freshly-collected real-world ground truth becomes the scarce asset — the precise bet behind "the ground truth AI needs to survive."
The Epoch AI analysis estimates the stock of public human text against the growing appetite of LLM training and finds that, on current trends, models will be trained on datasets roughly equal to the entire stock of public human-generated text between 2026 and 2032 (slightly earlier if overtrained). The authors point to synthetic data, transfer learning, and new data sources as the paths forward.
Why it matters (spine): this is the foundation stone of RDKG's moat thesis (P4). Pair it with the inference-cost collapse: reasoning trends to free and the cheap, abundant training substrate exhausts — so the scarce, defensible input becomes verified, non-public, physical-world ground truth, which is exactly what the gardens produce (materials, health, growing things).
Falsifier: synthetic data fully substitutes without quality collapse. · Confidence: 0.85.
Provenance: Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, Marius Hobbhahn (Epoch AI; credited), "Will we run out of data? Limits of LLM scaling based on human-generated data," arXiv 2211.04325 (v2, 4 Jun 2024). Verified abstract text quoted.
So we should package every garden's verified, freshly-measured ground truth as the scarce post-exhaustion input — the cross-garden moat.
Position RDKG's verified, freshly-collected, non-public physical-world ground truth as the post-exhaustion input the scaling literature says will be scarce.
domains: data-moat · scaling-limits · machine-layer · economics-of-intelligence