Models that keep learning

Over the past few years, scaling a single model toward broader knowledge and capabilities has delivered remarkable progress. We believe this paradigm is approaching an inflection point, with diminishing returns from further scale alone. There are other paths to AGI. At Sous Labs, we pursue a different training objective: a general ability to learn.

Such a model could be much smaller. Placed in an environment, it should acquire knowledge, develop skills, and grow into a specialist through experience. Its expertise becomes specific to where it works; its capacity to learn remains general.

We want this learning to be internalized in the model’s weights. As people turn experience into lasting knowledge and skill, models should become more capable through changes within themselves. Harnesses and agent memory are valuable complements. Our focus is on learning in the model itself, not only in the systems around it.

For us, continual learning means exploring an environment, choosing what is worth learning, testing understanding, and improving through feedback. New experience should build on what has already been learned, not erase it. Learning how to learn is central to our research.

We imagine people and organizations developing intelligence of their own: models that learn the context of their work and become increasingly capable within it. This matters wherever knowledge, tasks, and conditions keep changing, from science and software to decisions in the physical world.

Our vision of AGI is collective: countless smaller models, each with a general ability to learn, developing different expertise across different environments. Rather than concentrating every capability in one model, general intelligence can emerge across this population of continually learning specialists.