Small AI models and specialized vertical AI models are very brittle. Any unusual situation or out-of-distribution issue and they break down. You also won’t get emergent leaps or good problem solving. They still have uses, but benchmarks don’t do a good job of showing weaknesses