LLM needs to be dumb or we screwed
No really, with all due respect to people working in AI. AI is smart when the LLMs are dumb.
You see LLM is practically a zip file (quite big file) which has vectors for indexing and querying token based on the previous token(s).
And there is a practical hardware limit to how much size,index and query. And that's why you are now seeing optimization of last great models like opus 4.1 to 4.5. Yes opus 4.5 is an optimisation and not an improvement. It sucks in compare to 4.1 both in performance and cost.
Now with that as a context if you were to make LLM smart you will basically keep making the model havier and costlier in a never ending spiral. Let's say you want a trained model on your companies data and have a personalised AI for your company.. the data within your company moves in gigabytes based on new data generation every month.. let's say 10k new users onboards monthly and then your team on slack + jira + GitHub generates 200 tickets etc.
You will very easily have an information overload problem.. and the decision for the data being right or wrong let's say that comes in six months(user feedback or churn or lost deals for company to improve). The smart LLM you are trying to make now goes bad at such an intense magnitute that it's fucking impossible to recover and get back to a same stage.
Hence, keep the LLM dumb, more like a natural language parser.
The intelligence needs to come slow and steady in terms of operational data and its output impact in its due time and it needs to live seperately and not as part of the LLM.