Pensera is an independent research group studying the safety of AI systems that improve AI systems. We build the evaluation, interpretability, and control systems needed for the point at which models begin to participate in their own development.

For most of the field’s history, the people who improve AI systems and the systems being improved have been clearly separate. That separation is eroding. Models now help write training code, propose architectures, filter and label data, generate evaluations, and review the work of other models. As this continues, more of the loop that turns compute into capability runs through systems we do not yet fully understand.

Future AI systems may increasingly participate in the process of building stronger AI systems. Pensera studies this transition: when models begin improving models, how those loops can be measured, and how they can be kept safe. We think it deserves dedicated study before it is complete, not after.

The scientific understanding of frontier systems already lags their capabilities. A world in which models meaningfully accelerate their own improvement widens that gap — unless the tools to observe and constrain the process keep pace. Our work is an attempt to build those tools. We organize it around three questions:

These are research problems, not solved ones. We are early, and we are deliberately narrow: we would rather understand one part of this transition well than gesture at all of it. We keep our claims close to our evidence, and we aim to make our work measured, reproducible, and useful to the wider community working on the same questions.