|
|
|
|
|
by potatopatch
922 days ago
|
|
The difference seems too small not to rule out all sorts of things, but the general idea would be that you can't predict shorter/longer for one minor change, so the average of samplings of each should be similar unless the thing modified relates to terseness in the training set. |
|
This feels like the original author is over anthropomorphizing LLMs, and expecting them to interpret prompts the way humans would, but it seems obvious that changing the prompt results in a slightly different context window, which results in a slightly different response distribution? Similarly, if you changed whether the bit about time of year was at the beginning or the end of the prompt, I would expect a statistically different distribution of response lengths.