Nvidia

Five AI & Machine Learning patents cover a wide range of applications, from content generation to content moderation to conversational interfaces. One filing describes a system that trains an auto-regressive auto-encoder to compress 3D meshes into fixed-length latent codes, allowing the model to produce meshes with more than 8,000 faces that match the kind of topology an artist would create manually. A separate patent takes a different approach to multimedia generation, decoupling image retrieval entirely from a language model by pre-indexing text-to-image associations and using vector similarity search to match a chatbot's generated responses with relevant images, avoiding the computational cost of running a full multimodal model. 2 of the remaining patents focus on real-time communication: one applies deep neural networks to filter abusive language during live voice and text chat at low enough latency to avoid disrupting the exchange, while another uses natural language processing on game-generated text logs to pull out event data for highlight reels and activity feeds, sidestepping the need for frame-by-frame video analysis entirely. The fifth patent in this category describes conversational agents that convert multichannel Audio into a fixed B-format spatial representation, allowing a large language model to process not just what a sound is but where it comes from in a 3D environment.

The sole Graphics patent addresses a long-standing constraint in GPU programming. Rather than requiring developers to wait for Hardware-level additions whenever they need a new atomic memory operation, the filing describes a system that lets programmers define those operations in software, executed by specialized processors sitting within the memory hierarchy itself. This moves atomic operations from a fixed set baked into silicon to an open-ended, programmable layer, removing the ceiling on complexity that lockless concurrent programming has historically run into.