did see that with certain llms when building with pipecat, they aren't all equally proficient at calling tools and this is something coming more into focus with model training (see kimi k2)