Our mission
Our mission is to understand the mechanisms that make artificial intelligence more capable, adaptable and useful over time. We are interested in systems that can build on experience, make reliable discoveries and carry what they learn into new situations.
That inquiry spans individual models and the larger systems around them: tools, memory, other agents, environments and institutions. We want to understand which improvements come from a better learner, which come from better conditions for learning, and how those contributions fit together. Our long-term aim is to contribute principles and methods for cumulative, broadly useful intelligence.
Why this matters
We believe that improving the ability to learn and discover could have value across science, engineering and many other fields. A useful advance can solve a problem; an advance in how problems are investigated could help people make many further discoveries.
Capability also changes what systems can influence. The rewards, resources and institutions around AI shape how it is developed and used. We therefore treat reliability, incentives and human direction as part of the research problem. Understanding where improvement fails, plateaus or pursues the wrong objective matters alongside understanding what makes it work.
What we focus on
Three connected priorities organize the work:
- Learning and discovery. How training, reasoning, search and interaction help agents solve unfamiliar problems, develop reusable knowledge and retain earlier competence.
- Evolution and collective intelligence. How diverse approaches are generated and preserved, and when communication, shared memory and specialization help populations accumulate more than individuals could alone.
- Improving the process of improvement. How systems might adapt their own learning and search methods, what evidence would demonstrate progress, and which computational, evaluative and safety constraints limit it.
These priorities connect our foundational questions with current work on frontier language models and agent systems. They leave room for several approaches, including reinforcement learning, evolutionary search and combinations of learned models with external tools. The frontier research agenda develops the current investigations; the open questions give the broader context.
How we work
Rubedo is at an early stage. Our work currently consists of conceptual analysis, literature review, two exploratory notebook experiments and proposed controlled investigations. We have not selected a general architecture or established validated research findings.
We aim to turn broad ideas into narrow comparisons that can change our minds. That means stating assumptions, using meaningful baselines, counting resource costs, checking transfer and retention, and documenting failures. Claims about discovery need independent checks; claims about improving intelligence need evidence beyond a higher score on the tasks used for development.
We intend to share useful methods, reproducible artifacts and negative results as the work develops. The experiments page records what has been tried and what remains proposed. The reading map and notebook separate external evidence from our working interpretations.
Our roots
Rubedo began with questions about decentralized AI: could many interacting agents learn, specialize and build something more capable together? Those discussions led us toward evolution, developmental learning, cumulative knowledge and the organization of adaptive systems.
That background still shapes our interest in populations and distributed intelligence. It is a source of research questions, rather than a commitment that one architecture must be best. Our scope now includes both individual learners and the wider systems that support their development. The inquiry traces that progression in more detail.
Why Rubedo
The name comes from alchemy: rubedo, or “reddening,” is associated with the completion of the Great Work. We chose it for the idea of transformation: developing greater capability by refining and combining what is already there.
For us, that offers a metaphor for AI research: understanding how learning, experience and organization can give rise to more capable intelligence. The name expresses the ambition behind the project; the work itself proceeds through computation, experimentation and evidence.