[saltythrowaway]

Embedding-similarity based retrieval for one. Probably shouldn't get into the details here but Google has made incredible advances in the retrieval step of your typical retrieval / ranking pipeline and for a variety of reasons meta has not implemented them in production, despite having open sourced tools in this area. There are a few other advances like this Google has been using for long enough to publish on, meta engineers have working prototypes, but nothing is in production and everything breaks all the time.

[tomaskafka]

Oh, I have a similar story from a local country-leading job portal. I tried to persuade them to invest in embedding-based search instead of the category (which is often wrong and ambiguous by both advertisers and searchers) + keywords, but '2 people for 1 quarter to demonstrate a prototype able to be tested in production' seemed too risky for the management, despite having really cool results from the offline prototype :/.

The risk of missed opportunity is harder to visualize and thus rarely considered.