The title is very misleading: GPT-3 generated the titles, but as usual, they were manually curated on the "picked the title I liked the most" criteria.
There's more to getting to the front of Hacker News than just a good title (although it helps); ironically this framing undersells the writing capabilities of the author by attributing success to the AI only.
There are ways in theory for AI models to generate "better" titles, which is something I'm actively researching (rough example: https://github.com/minimaxir/gpt-3-experiments/tree/master/e...).
