I forget where I first saw this analogy, but describing LLMs as a broken clock just makes so much sense. You cannot scale up or improve LLMs to address fundamental limitations with the technology any more than you can make a hyperscale broken-clock data center to more accurately tell time.