1) Current patent search systems use boolean and keyword searches dominated by metadata, but new systems use ranked retrieval and move away from metadata. 2) Evaluating search requires moving from anecdotes to statistical benchmarks like TAPAS using real queries against patent corpora. 3) To improve search, we need to go beyond proximity queries that approximate meaning to true semantic search connecting mentions to normalized entities, though coverage exists primarily in biomedical domains. 4) Future search systems will integrate technologies like natural language processing, topic modeling, knowledge bases, and visual search of images and diagrams to build a true understanding of applications and prior art.