This is a base model, just because it's 90% there on its own doesn't mean you can't improve on it by adding extra safe guards. For example you can get LLMs to be more accurate by asking another LLM to proofread the work. I am frankly amazed that the base models are this good to begin with. I was totally expecting to need way more safeguarda from the get go, but we're getting a lot even without them. But I fully expect there to be AI tools that are specialized to identify where the base model messes up and then corrects it.