yeah so that is why something like edge impulse, that extracts just enough of the LLM needed for a specific task, is very important<p>for instance, running on an M0 at ten frames a second a cut down LLM with a camera, can recognise childrens toys (like hot dog/not hot dog) at 95% accuracy and use only 367kb of memory.<p>yes, that is three hundred and sixty seven kilobytes (ymmv), and that could still be tweaked by possibly dropping the resolution of the real time imaging<p>only use what you need - old hippy saying