Ask HN: How does GPT-4V work?

Does it really understand images or it is using "old" technologies to interpretate images and convert them to a text description to then reason over it and provide answers to questions?