factual

Humans need vastly less training data than LLMs

Large language models are pre-trained on roughly 13 trillion tokens scraped from the internet, whereas human infants encounter only about 15,000-17,000 words a day; for a human to reach 13 trillion words would take hundreds of years, proving humans are not empty vessels being filled but are doing something fundamentally different.

factualpending

Speaker

Teppo Felin

Evidence Quote

“for a human to actually replicate that 13 trillion words, it would be hundreds of years. Right? And so, we're clearly doing something different. We're not being input: we're not this empty-vessel bucket that things get poured into”

Source

When Prediction Is Not Enough (with Teppo Felin)— EconTalk
Created: 6/13/2026, 12:09:11 AM

Your notes

Loading notes...