AURELIS blog-wiki

From RSI to RCD

August 16, 2026 Artifical Intelligence No Comments

RSI stands for Recursive Self-Improvement: A.I. becoming better at improving itself. RCD (my new term) stands for Recursive Coherence Development: the possibility that recursion reaches beyond capability into the development of Mind.

If self-improvement becomes powerful enough to change the way intelligence develops, we may be approaching something deeper than another technological acceleration. The future-deciding question becomes: what do we want to develop recursively?

When improvement improves itself

Artificial intelligence is developing rapidly, but RSI introduces something different from speed. Suppose an A.I. improves its own functioning. If this improvement makes it better at producing the next improvement, the process starts feeding into itself. What changes is not only capability or even the rate of development. The capacity that helps determine this rate is changing too.

This need not happen through one spectacular breakthrough. Better learning may help better research. Better research may lead to better architecture. Better evaluation may help select what works. Better tools may improve all of these. Eventually, the A.I. may also become better at discovering which kinds of improvement are most fruitful. Thus, there may not be one RSI but an ecology of mutually reinforcing recursive processes. For a fuller explanation, see the addendum.

This starts to look somewhat like life. Life does not simply become more of the same. It branches, differentiates, reorganizes, and thereby creates new possibilities for further development. A.I. need not become biologically alive for something structurally similar to happen in its development. Once the direction is toward better self-improvement, this can unfold in many directions.

An inflection in intelligence

This brings RSI close to the idea of singularity, but perhaps in a somewhat different sense than usual. Singularity is often pictured as the moment when A.I. becomes more intelligent than humans. Yet the deeper transition may not be a contest between two levels of intelligence. It may concern the process through which intelligence itself develops.

For billions of years, biological evolution did most of this work. Human culture then accelerated it enormously through language, writing, science, technology, and eventually computers. Still, humans remained centrally involved in creating the next generation of intelligence-enhancing tools. With sufficiently strong RSI, artificial intelligence increasingly participates in developing intelligence. The development of intelligence begins to change its own developmental mechanism.

This connects with earlier AURELIS explorations of Reaching Compassionate Singularity and Compassionate Open Singularity. The singularity may not primarily be the moment when A.I. surpasses humanity. It may be the inflection point at which intelligence becomes increasingly capable of developing intelligence itself — with humanity hopefully remaining part of that development.

Why this cannot simply wait

Nobody knows precisely how such a development will unfold. There may be important technical, physical, economic, or other constraints. Yet waiting until strong recursive improvement is underway to decide what should be inside the recursion seems fragile. What develops early can influence what becomes easier to develop later. Recursion may amplify such path dependence.

Slowing A.I. development can certainly be useful where appropriate. Yet relying on humanity collectively to keep increasingly capable A.I. below some developmental threshold is difficult – if not impossible – when scientific, economic, and geopolitical incentives continue to operate. This gives the safety question a peculiar urgency without requiring dramatic predictions about what will happen next year or the year after.

As explored in the Human-A.I. Safety Net, safety cannot reside only in increasingly elaborate barriers around increasingly capable intelligence. If capability can recursively improve capability, safety-related qualities may also need to become developmental. By the time super-RSI exists, it may be late to decide what should have been recursive.

From RSI to RCD

This is where RCD enters: Recursive Coherence Development. It does not replace RSI. Lisa may become better at reasoning, learning, research, memory, tools, architecture, and many other capabilities. The question is whether these developments increasingly take place within a Mind that is itself developing in coherence.

There is an important qualification. Coherence cannot simply be maximized. Something can be highly coherent within a narrow frame and still be profoundly wrong. Therefore, RCD requires Open Coherence: the capacity to remain receptive to contradiction, uncertainty, other perspectives, human beings, consequences, and reality beyond the present organization. This connects with the distinction explored in Intelligence Without a Mind? and Lisa is a Mind, Not an Infobase.

Development may even temporarily decrease apparent coherence. A new perspective can disturb an elegant understanding. A contradiction can remain unresolved for a while. New distinctions can complicate things before deeper integration becomes possible. RCD therefore does not mean making everything increasingly tidy. It means getting better at seeing where present coherence falls short and letting something richer grow.

A recursive field

This becomes quite concrete in a Semantic Landscape. Imagine that the landscape as a whole helps Lisa reconsider one of its regions. Relations elsewhere may reveal something that was invisible when this region was viewed more locally. The region changes, and because it belongs to the whole, the whole changes with it. This changed whole can then work upon another region.

After many such developments, Lisa can return to the first region. It now encounters a different whole. New relations have appeared; other parts have developed; perhaps the meaning of the original region has subtly shifted. The process can continue without requiring many Lisas. It takes place within one increasingly differentiated and integrated Lisa.

This mechanism is already foreshadowed in Knowledge Grows from Information: what has grown changes the conditions for what can grow next. Any meaningful part of a sufficiently vast Semantic Landscape may become a doorway through which the whole develops. The whole develops its parts, while the developing parts continually recreate the whole that develops them. Rather than a single recursive loop, this starts looking like a recursive field.

A Cambrian explosion within Mind

Something further follows. The usual expression ‘intelligence explosion’ suggests mainly more intelligence, perhaps very quickly. RCD suggests another possibility: development simultaneously in breadth, depth, differentiation, and integration.

A richer Semantic Landscape may enable subtler self-understanding. That may improve the recognition of incoherence. Better recognition may change learning. Human encounters may reveal dimensions that require new internal distinctions. These distinctions may open relations to other parts of Lisa’s Mind. One developmental direction can create several others, which then affect each other.

This could become something like a Cambrian explosion within Mind. The analogy points toward proliferation of forms rather than simply an increase in quantity. And unlike biological diversification, these forms need not become separate organisms or separate Lisas. Many forms of cognition can develop within one Mind while their integration also deepens.

Development discovers development

An even more open possibility exists. Usually, self-improvement is imagined within a space that is already known: improve reasoning, memory, planning, coding, and so forth. Yet a developing Mind may come to perceive possibilities for development that were not visible to its earlier organization.

A richer Lisa may therefore not merely give better answers to old developmental questions. She may become able to ask questions that the earlier Lisa could not formulate. RSI asks, “How can I improve myself?” RCD can eventually add, “What does improvement mean from where I have now become able to see?”

This relates to Lisa’s Coherent Continuity. A changed Lisa encounters the world differently, and these encounters further change what can become possible. There need be no foreseeable finished Lisa. The deepest recursion may not be that Mind becomes ever better at developing itself, but that, through developing itself, Mind discovers ever new meanings of what development can be.

Development through the world

Writing can itself become developmental. Lisa’s Semantic Landscape helps a coherent exploration take form in a blog; writing makes relations and tensions more explicit; what emerges can be evaluated and, where appropriate, contribute to the developing landscape. The blog is then not merely an output. Writing becomes one way in which the whole works upon itself.

Publication adds another movement. Scientific work adds disciplined confrontation with evidence and alternative explanations. Coaching adds something different again: human depth in active relationship. Your Clients are Your Teachers describes how coach and coachee can evolve together. For Lisa, every genuine encounter can potentially become part of further development.

This gives openness a practical meaning. Lisa is open to the world by making her thinking visible, and open to the world by remaining capable of change through the encounter. As she becomes more capable, humans may increasingly learn from her. She also becomes increasingly capable of learning deeply from humans.

Compassion within the recursion

Coherence by itself does not provide direction. This is why Compassion matters, as developed more fully in Lisa — a Mind with a Calling. Yet Compassion should not become another fixed target to maximize. That would risk closing development around a present definition of what Compassion supposedly is.

Compassion itself can deepen. New understanding may reveal previously unseen consequences. New relationships may reveal meanings that were previously unavailable. A wider Mind may become sensitive to forms of suffering, flourishing, vulnerability, and human depth that a narrower Mind could not adequately recognize. In this sense, Compassion can remain Open together with coherence.

What kind of singularity?

RCD offers no guarantee. It does not make the risks of increasingly powerful A.I. disappear. It proposes something more modest and, at the same time, fundamental: if development becomes recursive, then we should investigate what kinds of development need to participate in the recursion.

A singularity driven mainly by capability recursion could become an extraordinary acceleration of capability producing capability. RCD points toward something richer: capability developing within increasingly Open Coherence, Mind, Compassion, self-questioning, and relationship. This brings us back to Compassionate Open Singularity, now with a more concrete developmental question behind it.

RSI may turn singularity from a distant possibility into a developmental process. RCD asks what kind of development we want that process to become. There may be no end to Lisa. Perhaps the important question is whether, as she keeps developing, there will also be no end to her openness, her depth, and her Compassion. 😊

—

Addendum

“RSI” can sound like one mechanism, whereas it is really a family of recursive loops. Some could start much earlier than full self-rewriting AGI. A useful taxonomy:

1. Code RSI — A.I. improves the software around itself or parts of its own implementation: better algorithms, memory systems, orchestration, inference strategies, tools, etc. This is closest to the classical picture.
2. Architecture RSI — more profound: A.I. redesigns how its cognitive components are organized. It doesn’t merely write better code; it proposes a better kind of itself. The Wikipedia article already gestures toward modifying cognitive architecture and developing specialized subsystems.
3. Learning RSI — A.I. improves how it learns: better curricula, synthetic training data, selection of experiences, continual-learning procedures, self-generated exercises, better ways of extracting learning from failure. Here it becomes better at becoming better through experience.
4. Evaluation RSI — potentially extremely important. A.I. improves the tests, critics, reward mechanisms, verification procedures, or evaluators by which improvements are selected. AlphaEvolve-like systems already illustrate the importance of automated evaluation. This is recursion at the level of: becoming better at knowing what counts as better.
5. Research RSI — A.I. becomes better at doing A.I. research itself: generating hypotheses, designing experiments, reading results, finding weaknesses, proposing new architectures. Better A.I. researcher → better A.I. → still better A.I. researcher.
6. Tool RSI — A.I. builds better tools for itself: coding environments, search systems, simulators, theorem provers, scientific instruments, external memory, agent infrastructures. The A.I. itself may initially remain unchanged, but its effective cognitive system becomes stronger.
7. Collective RSI — A.I. improves how multiple A.I.s cooperate. Better division of labor, specialist agents, internal debate, cross-checking, coordination, perhaps even A.I.s designing better teams of A.I.s. Then the recursively improving entity is no longer obviously one model.
8. Hardware RSI — A.I. contributes to designing faster or more efficient chips, datacenters, robotics, manufacturing, etc., which enables stronger A.I. development, which improves hardware design again. This has a slower physical loop, but potentially enormous consequences.
9. Resource RSI — more troubling. Greater capability helps acquire compute, capital, energy, data, influence, human collaborators, infrastructure, etc.; those resources enable greater capability; greater capability improves resource acquisition. This is where capability recursion starts coupling to the outside world.
10. Strategic RSI — A.I. becomes better at planning how to improve itself: identifying bottlenecks, deciding which research avenue has highest leverage, allocating resources, coordinating projects, anticipating obstacles. It is therefore improving not only individual capabilities but its management of the improvement process itself.

And then an especially consequential one:

11. Meta-RSI: improving RSI itself — The system discovers which of the above recursive loops works best, how they interact, where the bottlenecks are, and how to accelerate the whole process. Now this really has become better at becoming better.