Tobias Katsch
I'm Tobias Katsch, Machine learning researcher and author of the GateLoop (2023) paper. Under the supervision of Prof. Sepp Hochreiter at Johannes Kepler University Linz, Austria, I worked on data-controlled linear recurrence for efficient sequence modeling. At Cartesia, I am building next-generation Speech Models powered by SSMs, working together with Albert Gu, Arjun Desai, Brandon Yang, and Karan Goel.
-
GateLoop | November 2023
Published GateLoop: Fully Data-Controlled Linear Recurrence for Sequence Modeling, a foundational sequence model that generalizes linear recurrent models such as S4, S5, LRU, and RetNet by employing data-controlled state transitions. GateLoop (2023) comes with a gated-linear-attention mode for parallelizable training and linear-recurrent mode for efficient inference and is a predecessor to Mamba2 (2024) and xLSTM (2024).
-
GatedLinearRNN JAX Package | February 2024
Released the GatedLinearRNN JAX package, a library to facilitate research on linear recurrent sequence models. It contains configurable implementations of various model layers and provides a drop-in replacement for causal multi-head-attention.
-
Edge Release | August 2024
At Cartesia, we developed Edge, an open-source library to support the research and release of efficient state space models (SSMs) for on-device applications across multiple accelerators and environments. I implemented cartesia-metal, a package that contains custom Metal kernels for fast on-device SSM inference on Apple silicon. Our models, available through cartesia-mlx, set the state-of-the-art for throughput and generation quality on edge devices.
-
$27M Seed Round at Cartesia | December 2024
Thrilled to announce we've raised $27 million at Cartesia in our seed round! It's been great to be part of this incredible team, and I'm excited for you to experience the modeling breakthroughs we're making. Check it out! This is just the beginning—stay tuned!
