Mohitm

Most modern LLMs correct or bypass the "strawberry" counting issue through specific engineering changes, prompt workarounds, or execution tools rather than a fundamental leap in character-level perception.

you're right that "magnitude carries no meaning" is too broad. It holds for the static lookup table under cosine similarity, but whether norm matters really depends on the training objective and similarity function, and the retrieval case proves it: L2-normalizing a dot-product-trained representation can reorder results, not just speed up the same ranking.