Research that moves into Matilda and onto MC-2
We’re an applied lab, so when a result holds up we publish it and build it into Matilda, which runs on MC-2.
Selected work
Research areas
Six research areas, applied and long-range.
Some of this work goes straight into Matilda, and some tests ideas that may take years to reach a product.
-
Agents
We study how models use tools, plan and keep memory across long tasks, and the results go into Matilda.
-
Safety and alignment
We test when a capable system can be relied on, through interpretability, oversight and red-teaming.
-
Training and algorithms
We work on how models learn at scale, with new objectives, optimisers and reinforcement learning (RL) for reasoning.
-
Data
The data a model trains on sets what it can do, so we work on how that data is curated, generated and mixed.
-
Multimodal
We train vision-language models that tie their answers to what is in the image.
-
Efficiency
We train and serve on our own GPUs, so speed matters from the kernel to the model.
Research residency · 3 to 6 months
A paid residency for
your own research question.
It’s open to late-stage PhD students and standout early-career researchers, who set their own direction and work on our GPU cluster with our team.
From the lab