At Tilde, we believe achieving true progress in interpretability requires innovation across the entire stack. From uncovering the fundamental building blocks of model computation to enabling precise steering of their behavior, we're driven by the fascinating scientific and mathematical problems that underlie true understanding of intelligence.
We are backed by Khosla Ventures, researchers at OpenAI and Anthropic, Yufei Zhao, and others.
If coming from an ML background, ideally has a strong track record publishing novel results at top-tier conferences (NeurIPS, ICML, ICLR).
If coming from more of a mathematics or physics backgrounds, ideally has done interesting research or strong competition experience and cares about ML/AI (and has demonstrated this in some way).
(Ideally) Has high-effort blog where they discuss deeply technical concepts (e.g. something like this or this).
Extensive experience writing and optimizing GPU kernels, particularly using Triton, PTX, CUDA, or low-level systems languages (C/C++).
Track record of practical kernel optimizations, demonstrated improvements in training efficiency, and involvement in experimental hardware-software integration.
Strong familiarity with training pipelines, debugging, scaling, and performance optimization of open-source models.
Ability to independently design, run, and iterate large-scale ML experiments.
(Ideally) Has made impactful open-source projects or contributions.
The Tilde Fellowship supports research in reverse-engineering important phenomena in pretraining science (arch, optimizers, etc).
We will provide GPU compute, direct mentorship, and collaboration with Tilde staff, along with opportunities to match with other fellows.