Tufalabs

Tufa Labs is a small, independent research group based in Zurich, Switzerland, dedicated to fundamental AI research. Since its founding about a year ago, the lab has been fueled by a passion for AI since around 2008, when the founder always believed that AI research would be the most impactful technology. After initial attempts in ML/AI research, the lab was born from applying AI to quant finance and securing funding. Their research priorities include investigating o-style systems (post-training, system2 reasoning), wrapping up ARC-related research, and learning to train and fine-tune LLMs. They focus on LLM-based reasoning, self-improving loops, and scaling trends. The lab is involved in ARC-AGI, a benchmark designed by François Chollet (creator of Keras) to measure intelligence by skill-acquisition efficiency. Tufa Labs has published research on topics like AlphaWrite (inference-time compute scaling for writing), self-rewarding self-improving models, LLMs for engineering, and LADDER (self-improving LLMs through recursive decomposition). They also have a paper on Deep Learning for ARC. The lab values intellectual curiosity, high intrinsic motivation, perseverance, and fast experimentation. They are funded for at least the next 3 years and aim to elevate from GPU-poor to GPU-rich. They are looking for obsessive individuals who share their research interests.