LLMs are fundamentally text predictors. There are many high-probability replies to “Tell me a funny joke”, so you wouldn’t necessarily expect them all to tell the same one. But perhaps, somewhere in the training data, someone published their conversation with an LLM in which they said “Tell me a funny joke” and it replied with the joke about atoms. Next-gen LLMs learn from this training data that if an LLM is asked to tell a joke, the probability-maximizing answer is that particular joke. So now they all start telling the same joke.
Why can’t the mode-collapse just be from convergent evolution in terms of what the lowest-common denominator rater will find funny? If there are only a few top candidates, then you’d expect a lot of overlap. And then there’s the very incestuous nature of LLM training these days: everyone is distilling and using LLM judges and publishing the same datasets to Hugging Face and training on them. That’s why you’ll ask Grok or Llama or DeepSeek-R1 a question and hear “As an AI model trained by OpenAI...”.
An untested hypothesis:
LLMs are fundamentally text predictors. There are many high-probability replies to “Tell me a funny joke”, so you wouldn’t necessarily expect them all to tell the same one. But perhaps, somewhere in the training data, someone published their conversation with an LLM in which they said “Tell me a funny joke” and it replied with the joke about atoms. Next-gen LLMs learn from this training data that if an LLM is asked to tell a joke, the probability-maximizing answer is that particular joke. So now they all start telling the same joke.
Why can’t the mode-collapse just be from convergent evolution in terms of what the lowest-common denominator rater will find funny? If there are only a few top candidates, then you’d expect a lot of overlap. And then there’s the very incestuous nature of LLM training these days: everyone is distilling and using LLM judges and publishing the same datasets to Hugging Face and training on them. That’s why you’ll ask Grok or Llama or DeepSeek-R1 a question and hear “As an AI model trained by OpenAI...”.