Скачать или смотреть AI Trained on the Internet. Now It's Destroying It.

Sign up to our newsletter 👉 https://absolutelyagentic.com/?modal=...

In 2023, StackOverflow was the largest programming knowledge base on Earth, built over 15 years by millions of developers.

By 2025, question volume had collapsed by 78%, gutted by the very AI models trained on its content.
It's not just StackOverflow - Chegg lost 99% of its stock value, and publishers lost a third of their search traffic.

But the real crisis runs deeper: researchers have discovered that when AI trains on AI-generated content, its outputs degrade rapidly in a process called model collapse.

AI may be devouring the very ecosystem it needs to survive, and high-quality human data could be exhausted by 2028.

This video explores what happens when the internet starts eating itself.

Chapters
00:00 - Intro
01:58 - The Substrate
04:54 - Model Collapse
08:00 - Ouroboros and the Casualties
12:25 - The Counter Case

Sources to Google
Ilia Shumailov - AI Models Collapse When Trained on Recursively Generated Data (Nature, 2024)
Rice University - Model Autophagy Disorder (MAD) Study
Epoch AI - Will We Run Out of Data? Limits of LLM Scaling (2024)
Emily Bender - On the Dangers of Stochastic Parrots
Press Gazette - Publishers Lost a Third of Search Traffic to AI Summaries
University of Maryland - Reliability of AI-Generated Text Detection Tools
University of Waterloo - Growth of Synthetic Content on the Web

#AI #DeadInternetTheory #ModelCollapse #AITraining #ArtificialIntelligence #AbsolutelyAgentic