We're asking how optimization becomes open-ended. We haven't answered it.

The question

Our working question is this:

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?

Most machine learning is told, in one way or another, what intermediate progress looks like. A designer rewards the agent for collecting wood, then stone, then iron. A curriculum orders the tasks. A dataset encodes what humans already know. Those shortcuts work, and they're often the right engineering choice. But they hide the problem we care most about: how does a system figure out the path for itself?

That covers more than learning a good policy for one task. It covers how a system might develop its own representations, drives, skills, memories and abstractions, and eventually better ways of adapting. It also covers how populations, cultures and institutions accumulate what individuals discover, since that may matter as much as any one learner's ability.

Why this is exploratory

We want to be direct about where things stand. Rubedo has no selected architecture, no benchmark, no compute budget and no results. Several proposed architectures have come and gone, because committing to one too early would defeat the point of the inquiry. We've run a couple of notebook-scale toy experiments, and one of them failed in an instructive way (see Experiments). Most of the work is conceptual: clarifying the problem, mapping the literature, and finding which narrow questions a small experiment could actually distinguish.

The method we've settled on is theory → tiny controlled experiments → falsifiable results → progressively richer systems. We're near the start of that sequence.

Earlier versions of this site described "programs" and "active experiments". That overstated things. The ideas behind them (developmental genomes, error thresholds, gestation learning) are still part of the history of this project, and you can read about them in the notebook. But none is a running experiment, and we'd rather this site show what we're actually thinking than imply results we don't have.

Much of the thinking here was developed through long-form working discussions, including extended dialogue with AI research assistants. We treat those discussions as a way to generate and test ideas, not as evidence. Every page on this site labels what kind of claim it's making.

How the question has changed

Rubedo has moved a long way from where it started. Compressed, the arc looks like this:

It began as a plan to spread one neural network across many untrusted machines. That ran into latency, synchronization, adversarial nodes and proof overhead, and it risked becoming a cryptography project instead of an intelligence project. The pivot that followed still defines us: don't decentralize one mind; ask whether intelligence can emerge from a decentralized ecology of minds. (More on that in Incentives.) Each step since has been a response to something the previous step taught us. The notebook records how.

Looking back, the inquiry now runs across a stack of questions, each built on the one before:

1. How does an agent learn?
2. How does evolution produce better learners?
3. How do populations of learners cooperate, compete and specialize?
4. How do populations preserve discoveries and accumulate knowledge?
5. How do economic and institutional structures direct resources toward productive strategies?
6. How can the whole optimization process improve its own ability to search, learn, coordinate, acquire resources and create new optimization machinery?

And beneath all of them, the question we'd most like to answer:

What minimal set of mechanisms is sufficient for a system to begin discovering the rest for itself?

From open-ended intelligence to open-ended optimization capacity

Rubedo began with multi-agent systems: many adaptive agents sharing an environment with scarce resources, communication, competition, cooperation, trust, deception and specialization. The hope was that rich enough pressure would produce interesting forms of intelligence and social organization.

Two things have shifted since then.

First, the object of study widened. Intelligence is one way optimization capacity grows. Others include capital accumulation, communication, specialization, memory, institutions, tools, scientific knowledge, energy infrastructure and coordination. Ten moderately capable agents that invent an effective institution might outperform one brilliant agent. So the deeper object may be open-ended optimization capacity, with intelligence as one of its mechanisms.

Second, the environment got real. We used to ask how to design an artificial world rich enough for meaningful pressures to arise. But civilization already contains agents, scarcity, trade, property, capital, institutions and technological change, and AI is being inserted into every one of those loops. Rubedo doesn't need to invent the world where open-ended optimization happens. That world exists. The question is how to understand it, and how increasingly capable artificial agents will take part in it.

That doesn't make synthetic environments obsolete. We have exactly one civilization, so n = 1. We can't rerun the Industrial Revolution ten thousand times with different property rights. A simulated world plays the role of a wind tunnel: we don't build one because the atmosphere isn't real, but because reality has too many interacting variables to isolate a cause. Real civilization generates hypotheses, and controlled worlds test mechanisms. More on this in Collectives and Civilization as an optimizer.

How frontier AI fits the inquiry

Recent discussions ask whether pretrained language models, reinforced by interaction and external checks, can become better discoverers. This connects directly to Rubedo's existing concerns: exploration that survives premature convergence, preserved stepping stones, learned search controls, cumulative memory and useful cooperation.

Our frontier learning agenda separates changes to model weights from inference-time search, tool-use harnesses and evolutionary outer loops. These mechanisms can complement each other, but their contributions need matched comparisons. A correct answer, a novel result and a process that improves future discovery require different evidence.

The concrete directions are small: compare reasoning and agentic RL with fixed-model search; test diverse archives and curricula against simpler selection; measure whether shared memory and coordination repay their costs; and allow bounded changes to search procedures while keeping evaluation independent. The proposed pilots specify controls, failure conditions and transfer checks. None has been run by Rubedo.

External demonstrations justify investigating these mechanisms. They do not establish that agentic RL is sufficient for open-ended discovery or that recursive self-improvement will continue. Reward hacking, contamination, specification errors, forgetting and physical resource limits are part of the question.

What we're committed to, and what we're not

A few preferences shape how we approach the problem. They're commitments about method, not claims about how intelligence works.

• Account for supplied structure. The diamond testbed asks for discovery without prescribed milestone chains. A separate LLM track can use pretrained knowledge, demonstrations and curricula, provided we document their contribution and avoid conflating the two settings.
• No premature component diagrams. It's tempting to declare that intelligence is a world model plus a goal generator plus a skill library plus a planner. Maybe it is. But we want to understand the principles before fixing the boxes.
• Theory before implementation. We look to evolution, development, neuroscience, ecology and cultural evolution for insight before choosing an architecture, without pretending any one formal model is the whole answer.
• Take compute seriously. Biological intelligence rests on billions of years of physical search. A small training run won't recapitulate that, so we need to know which general principles make search tractable. See Compute.
• Preserve what's been learned. Repeatedly destroying competence and rediscovering it is the failure mode we most want to avoid.
• Learn through interaction. Experimentation, search, self-play and changing environments, not only static data.
• Treat incentive problems as design problems. Evaluator manipulation and Goodhart's law are things to solve, not reasons to stop.

Equally, there are things we are not claiming: that an assumption-free learner is possible, that everything must emerge from scratch, that biology must be simulated literally, or that language models and existing tools have no place anywhere in this work.

What progress would mean

Working synthesis

Progress would look like a system that discovers useful abstractions without labels, preserves unusual discoveries before their value is obvious, transfers capabilities to new situations, improves its own learning efficiency, and accumulates knowledge across time or across agents.

Those are different outcomes. Higher reward, novelty, diversity, complexity, generality and evolvability can each go up while the others don't. We haven't found a single metric that captures all of them, and we're wary of anyone who claims to have one. Claims of exponential improvement or sustained open-endedness stay hypotheses until someone measures them, and no finite run proves indefinite innovation.

For the next controlled experiment, the question we'd want it to answer is modest:

Under a limited budget, can a system improve its own way of generating and evaluating alternatives, retain useful competence, and become a better learner in unfamiliar conditions?

Why "Rubedo"

Our name draws on the alchemical idea of transformation. The About page explains that connection briefly, alongside Rubedo's mission, research priorities and roots in decentralized AI.