Most of the "magic" of large models are really just function calls, so as long as the small models have access to the same functions they work well. They fixed

Most of the "magic" of large models are really just function calls, so as long as the small models have access to the same functions they work well. They fixed the "how many R's in Strawberry" issue by offloading the question to a function, not spending a godly amount of money/energy on training another model. Oops, sorry "Tools". Gotta maintain the grift these statistic based lossy text compression cool bar tricks are "thinking".