alt Hacker News> The book isn't in the LLM.
An LLM is essentially a lossy compression of the training data. The book absolutely is in there, it’s just mangled to the point of unrecognizability.
> The book isn't in the LLM.
An LLM is essentially a lossy compression of the training data. The book absolutely is in there, it’s just mangled to the point of unrecognizability.
The wood tends to have an impression of the hammer that hits it. The book isn't in there, the weights are just shaped by what tools were used to form it.
When large quantities of source material are replicable by prompting its a bug not a feature.