llama 3.3 70b q4_k_m - coherence on kotlin vs java, kotlin breaks earlier but perplexity measurements inconsistent
tested llama 3.3 70b q4_k_m on equivalent kotlin vs java codebases.... kotlin breaks coherence around 16.8k, java holds to 18.2k.... but when i measure perplexity at same token counts the gap shrinks to like 0.4 which seems like noise.... kotlin has way more concise syntax so same logic = fewer tokens, but also kotlin's type inference and extension functions might create longer-range dependencies the model has to track.... anyone else tested this.... wondering if it's real or just my setup....
kotlin has more syntactic sugar than java so probably burns more tokens for the same logic. if you didnt measure at same token count then this is expected, not a model issue
yeah kotlin burns like 30% more tokens than java for same logic becuase of all the syntactic sugar. if you didnt measure at exact same token count this result means nothing
kotlin's null safety syntax burns tokens hard. measured this last month, same logic was 34% more tokens.