Research

Recursive self-improvement starts small: automate one piece of research, use the freed time for something harder, then automate the next piece. What each system learns feeds into the next. AI research is where this loop should close first, as agents take over training code, data curation and experiments.

More scale on a static recipe won’t get there: a model frozen at deployment stops learning. We want systems that keep improving from their own experience, which puts post-training at the centre. Pretraining mostly consolidates what is known; post-training on long-horizon experience is where models discover something new.

We build the machinery and the measurement: benchmarks for whether agents can post-train models or invent faster algorithms, dense signals that tell an agent which steps helped, and forecasting that turns research hunches into checkable predictions. We evaluate at the level of individual samples, so no gain or regression is averaged away, and every benchmark, codebase and dataset we build is open.

Models, tools and agents change every few months, faster than workflows can adapt. Using them well means understanding them: what a model has learned, where in its data that came from, and what shifts with each update.

We study neural networks from this angle: mapping what models forget during post-training, tracing how capabilities follow from data, and building large open datasets for images, language and video so anyone can ask these questions, not only a few companies. We also test whether agents can do real research work, such as finding the source behind a scientific claim.

The goal: AI that keeps improving in the hands of its users, and stays understandable as it changes.

Foundation models now predict behaviour and neural activity better than hand-built models. We use them as scientific instruments: turning vision-language models into models of human eye movements, predicting human choices across many cognitive experiments, and training retina models on recordings from many labs and species.

We train digital twins, study what they appear to learn, and build open, shared tools so results can be compared and built upon across labs.

The aim: a cumulative, quantitative account of biological vision, and what it teaches us about systems that keep learning.