Re:Taste Is All That's Left
by red75prime on Thursday August 06, 2026 @05:09PM
(#49202602)
> the LLMs were trained to mimic how people speak (write)
It's a foundational model (fresh from autoregressive pretraining) that approximates the probability distribution of human texts. And, no, it's not the statistical average of how people speak. It approximates how a person who could have written a text in its context would have written the next words.
Fine-tuning, RLHF, reinforcement learning change this probability distribution. I guess, it's mostly RLHF that shapes the way LLMs write. The similarity of style is due to common providers of RLHF data.
It's a foundational model (fresh from autoregressive pretraining) that approximates the probability distribution of human texts. And, no, it's not the statistical average of how people speak. It approximates how a person who could have written a text in its context would have written the next words.
Fine-tuning, RLHF, reinforcement learning change this probability distribution. I guess, it's mostly RLHF that shapes the way LLMs write. The similarity of style is due to common providers of RLHF data.