Cartesia
Architecture for Real-Time AI
Cartesia is an AI research company on a mission to enable pervasive, real-time intelligence with long-term memory. Their first product is a fast, hyperrealistic voice generation model, Sonic. We are privileged to back Karan, Albert, Brandon and Arjun in the pre-seed and seed, in partnership with the brilliant Chris Re.
Technical creativity is in short supply. At a time when many lament a stalling of true innovation in AI—beyond merely scaling Transformers on vast GPU clusters—the Cartesia team stands apart. They bring unparalleled technical creativity, vision, and velocity to a field too often resigned to incremental improvements. Their breakthrough work promises to reshape the direction of the industry.
Cartesia’s approaches, inspired by systems theory, move beyond the conventional architectures that have dominated recent years. S4 and Mamba are novel state space models (SSMs) that treat data as signals flowing through structured systems. Rather than relying on attention mechanisms (as in Transformers) or explicit, step-by-step state updates (as in RNNs), these architectures use mathematical structure to capture both short- and long-term patterns with remarkable efficiency and precision. This design is especially powerful for tasks involving time-series data and audio processing, or anywhere that understanding long-range dependencies is essential.
It’s becoming obvious that realizing the full value of AI will require further breakthroughs—particularly in scaling, speed, and efficiency. Cartesia’s models meet these challenges head-on. Their computational cost scales linearly, not quadratically, with sequence length. They learn to compress extended sequences into fixed-size states and achieve fast, cost-effective inference. The vision: real-time, multimodal intelligence accessible on every device. They’re hiring!
(On a personal note, before diving into state space models, we first tried to convince Brandon Yang, whom we knew from Snorkel, to start a different company. To his credit, he refused to work on anything less than truly transformative).