We estimate the stock of human-generated public text at around 300 trillion tokens. If trends continue, language models will fully utilize this stock between 2026 and 2032, or even earlier if intensely overtrained.| Epoch AI
