He posed an open question on training corpus design.
An LLM fintuned on a book does not act like a human who read a book; it doesn't have consistent episodic memories about the experience, etc.
What does the finetuning corpus that produces an LLM that acts like a human who read book look like?
Many users express amusement at creative approaches to fine-tuning LLMs for human-like book reading, such as suggesting agent book clubs that gossip instead of analyzing plots.
@universeinanegg There’s so many funny ways to approach this problem.
An LLM fintuned on a book does not act like a human who read a book; it doesn't have consistent episodic memories about the experience, etc.
What does the finetuning corpus that produces an LLM that acts like a human who read book look like?
@universeinanegg Does any kind of training or context engineering ever produce an LLM that acts like a human?
@universeinanegg Their whole life's tokenized thoughts and verbal output. Presently impossible, so at least everything they've ever written or read. But I'm not sure even that version would really understand what Gravity's Rainbow means to me.
@universeinanegg An agent book club where they just gossip instead of analyze the plot…
@universeinanegg margins