The frontier of how good models are also shifts and will remain ahead of local models unless we hit some dead end limitation in the algorithms themselves. A ceiling so to speak on how good LLM can get before the law of diminishing returns starts to apply.