Exploring intelligence and open-ended optimization
Rubedo explores how adaptive systems learn, discover, evolve and organize—from individual agents to populations, institutions and civilizations—and what enables, limits and steers their capacity to improve.
What conditions allow an adaptive system to discover, preserve, compose and value instrumental stepping stones, and progressively better ways of learning, without humans prescribing those discoveries?
Rubedo explores how adaptive systems, from single agents to populations, markets and perhaps whole civilizations, might discover their own stepping stones, hold on to them, and learn to optimize better over time.
This is a working question, not a settled theory. It started with agents in a shared environment and now includes pretrained models, experimental feedback and evolutionary search. Rubedo has no validated research findings; we distinguish external evidence, our hypotheses and proposed experiments throughout the site.
From frontier capabilities to testable questions
Can learning, search and verification make a system a better discoverer? Our current agenda connects frontier AI to Rubedo's older questions about stepping stones, evolvability and collective knowledge.
What changes after pretraining?
Reasoning and agentic RL, inference-time search and tool harnesses. Distinguishing reliable sampling, added information and transferable learning.
Questions and evidence →Which alternatives should survive?
Quality-diversity archives, curricula, multi-agent coordination and shared memory, tested against their compute and coordination costs.
Complementary tracks →Can discovery improve itself?
Verifiable artifacts, novelty and attribution, meta-optimization and the unresolved limits of recursive improvement.
What would count →Three layers we keep returning to
Most of our discussions end up at one of three scales. They aren't separate projects. They look like the same problem at different magnifications.
Optimization theory
What does optimizing mean when the landscape is enormous, changing, deceptive and only partly knowable? Local optima, exploration, stepping stones, evolvability, and learning how to learn.
Can a learner find the path to a distant goal when nothing along the path looks like progress?
Optimizer ecology
What happens when many optimizers interact? Competition, cooperation, specialization, communication, trust and coalitions generate selection pressures, and those pressures can shape what each optimizer ends up pursuing.
When does a population of agents start behaving like a higher-level agent?
Civilizational meta-optimization
Markets, firms, laboratories, infrastructure and AI are becoming coupled into larger feedback loops. If civilization is itself a kind of optimizer, what is it optimizing, and who decided?
What is the objective of the optimizer composed of all optimizers?
Lines of inquiry
Ten foundational questions and a frontier learning agenda that connects them to concrete investigations. Each page sets out what we think, what the literature shows, and what we don't know.
Stepping stones
Distant objectives rarely reveal their prerequisites. How does a learner value what doesn't pay yet?
Preservation
Keeping unusual discoveries alive before anyone knows they matter, with finite memory.
Curiosity
Self-directed learning that is more than chasing surprise, and drives that might themselves evolve.
Abstraction
How predictive structure, schemas and replay might give rise to reusable concepts.
Evolvability
Evolving the capacity to learn, including the machinery that decides how to search.
Heredity
Genomes as developmental programs, safe recombination, and the error catastrophe.
Compute
Biology had billions of years. What general shortcuts make artificial search tractable?
Agents
Memory, huge action spaces, priced thinking, and communication that costs something.
Collectives
Populations, culture, institutions, and the emergence of higher-order optimizers.
Incentives
Evaluators, Goodhart's law, and rewarding real usefulness in decentralized systems.
We expected selection pressure alone to produce genome-like heredity. Instead the population went extinct in 68 steps.
Our approach is theory first, then tiny controlled experiments. The handful of toys we've run so far, an evolved weight generator and a digital-heredity world, mostly showed us where ideas break. We publish them anyway, failures included.
Philosophy of optimization
Longer essays on the bigger picture: what a sufficiently general optimizer would do, whether society already is one, and where the analogy to nature holds and where it breaks.
The patient optimizer
The paperclip problem without the absurdity. Why a long-horizon optimizer might make no paperclips for a very long time, and when it should stop waiting.
Read essay →Civilization as an optimizer
Optimization isn't arriving with AI. AI is being inserted into loops that already exist. What objective emerges when no one wrote it down?
Read essay →The physical substrate of power
From mines to machine tools to capital markets: why recursive intelligence improvement may really be recursive industrial improvement.
Read essay →Generate, evaluate, select, amplify
Evolution, science and markets share one search pattern. A careful look at how far the metaphor of reality as optimization can be pushed.
Read essay →Currently thinking about
Recent entries from the research notebook. These are working notes, not results.
From better answers to better discoverers
Where reasoning RL, search, evolution and collective memory might complement each other, and what evidence would distinguish progress from a stronger proxy score.
Mapping the industrial stack
Walking from extraction through refining, materials, components and capital to see what any physical objective eventually depends on.
Engineering a system that learns how to search
The middle ground between blind evolution and full engineering: supply the capacity to adapt, and leave the discoveries and the search methods open.
How to read this site
Rubedo's proposals are unvalidated. External research is cited with its scope and limitations; claim labels distinguish it from our own hypotheses and working notes.
Something we're trying to understand. No answer implied.
A plausible idea we take seriously but haven't tested.
Our current distillation of a discussion. Likely to be revised.
What a published paper shows, bounded by its actual setting and evidence.
An illustration chosen to make a problem concrete, not a plan.
Philosophical analogy with no scientific evidence behind it. Interesting, not established.