“Like taking a photocopy of a photocopy, we lose more and more information and connection to reality,” This is the Achilles heel of LLMs. With the proliferati

“Like taking a photocopy of a photocopy, we lose more and more information and connection to reality,” This is the Achilles heel of LLMs. With the proliferation of models, training new LLMs on the output from older LLMs becomes all but impossible to avoid. It's the AI equivalent of inbreeding --- where defects are inevitably magnified over time.