AI is in the end a glorified next-word prediction function.

When you try to optimize this in a way that usually results in reasoning, the path it takes is still biased by the base model