#mode-collapse
#mode-collapse

[ follow ]

#chatgpt #user-experience #ai-evaluation #creative-taste #human-judgment #benchmarking #chinese-language

Contra Labs - Human Creativity Benchmark

Evaluator agreement reflects shared best practices, while evaluator disagreement reflects legitimate taste and intent; separating these signals shows no current model is both reliably correct and steerable.

ChatGPT Is Saying VWeird Things in Chinese

ChatGPT’s Chinese responses often reuse repetitive, culturally awkward phrases due to training biases and mode collapse, making outputs feel meme-worthy to users.

ChatGPT Has 'Goblin' Mania in the US. In China It Will 'Catch You Steadily'

ChatGPT overuses specific Chinese phrases, driven by mode collapse from post-training feedback, annoying native users and reflecting rigid language patterns.

ChatGPT Is Saying VWeird Things in Chinese

ChatGPT’s Chinese responses often reuse repetitive, culturally awkward phrases due to training biases and mode collapse, making outputs feel meme-worthy to users.

ChatGPT Has 'Goblin' Mania in the US. In China It Will 'Catch You Steadily'

ChatGPT overuses specific Chinese phrases, driven by mode collapse from post-training feedback, annoying native users and reflecting rigid language patterns.

You Are Not Immune To Mode Collapse - LessWrong

Mode collapse is an issue where AI models often produce outputs from their training distribution, leading to repetitive patterns.

Artificial intelligence

fromFast Company

How AI-induced cultural stagnation is already happening

Autonomous generative AI systems trained on their own outputs converge to narrow, generic visual themes and lose initial prompt diversity.

[ Load more ]