3 ms·
> This along with a few tokenizer related tests people ran, made people suspect that we are just serving Claude with post-processing where we filter out words l
by thorum 2y ago
> This along with a few tokenizer related tests people ran, made people suspect that we are just serving Claude with post-processing where we filter out words like Claude.
Didn't these "few tokenizer related tests" prove the API was using Claude's tokenizer instead of Llama's, based on how words were being divided into tokens?
That's a hard one to explain (it doesn't appear they're even trying to).
- refulgentis 2y agoPeople keep asserting that but, really, it was just people pointing to setting max tokens to a certain value and getting a certain # of words out. They didn't actually have tokens. Perfectly possible to have collisions, I'd wager even likely in the scenarios they tested, simple question, < 10 tokens, in English.