What happens when AI eats its own tail?

The big LLMs (ChatGPT, Midjourney, Stable Diffusion, etc.) are trained on massive piles of scraped “public” data. 

(Btw: “Paging the Copyright Cops.” I mean, wtf!) 

A couple of new studies show that when AI starts training itself on its own AI-generated output, results degrade until it’s just craptastic gobbledygook. 

It only takes a few generations for this to happen. And it’s worse for larger (more general) LLMs. 

Uh oh.

WHEREAS most AI output is currently not tagged as AI-generated; and

WHEREAS there’s an explosion of “synthetic data” now polluting the internet; and 

WHEREAS LLMs also want to stay up-to-date; 

THEREFORE it looks like we’re headed to Craplandia.

Or maybe we’ll just be stuck in a pre-2023 world, forever believing Drake & Doja Cat are the hottest acts in the world and Spain hasn’t won the Women’s World Cup yet.

Some ways out, perhaps:

🔸  Only use tightly-trained (perhaps private) LLMs. 

🔸  Look into “classifier-free guidance”. 

🔸  Be vewwy vewwy caweful with synthetic data.