You missed the part where deep seek uses a separate inference engine to take the LLM output and reason through it to see if it makes sense.

No it’s not perfect. But it isn’t just predicting text like how AI was a couple of years ago.