|
|
|
|
|
|
by jbellis
384 days ago
|
|
|
I don't see anything here to indicate it's "actually" Sonnet under the hood Possibly it was intentionally trained on some of Sonnet's outputs, but given that this only happens in thinking mode and Sonnet 3.5 did not have a thinking mode, I think the most likely explanation is just that LLMs are at their core a next-token predictor and sometimes that gives you weird artifacts when you slurp in a bunch of data from the web, which increasingly includes other LLMs' outputs |
|
|