I generally don't waste time with patents. I think most patents in deep learning can be overturned by prior art.

My current approach to IP is trade secrets. If we publish, we are careful to avoid details that would make the techniques easy to productionize.

As a founder building in the AI video space (Genve.ai), I've grappled with this exact question. Our conclusion? Speed and execution are the only real moats. In the time it takes to file a patent, the SOTA (state-of-the-art) models change twice. For us, instead of spending $20k on legal fees for a patent that might be obsolete in 18 months, we chose to invest that capital into our engineering—specifically developing our parallel multi-language translation engine. In AI, by the time someone tries to 'copy' your patented method, a new research paper or a model update has likely made that method inefficient anyway. We prefer keeping our specific lip-syncing logic as a trade secret while focusing on scaling the product as fast as possible.