Is Intelligence Recursive? What Happens When AI Creates Its Successors

Regan Peng is an AI founder and former senior tech executive, focused on personal AI, privacy-preserving systems, and edge AI.

gettyAI is beginning to improve its own tools and design new biological possibilities. The deeper question is what happens when a created intelligence can create another mind.

After years of building large-scale AI and personal-agent systems, I keep returning to a question that once sounded like science fiction and now feels increasingly practical: What happens when a created intelligence becomes capable of creating another intelligence?

We are not there. But the direction is becoming visible. Google DeepMind’s AlphaEvolve uses language models, automated evaluators and evolutionary search to discover and improve algorithms. Sakana AI’s Darwin Gödel Machine rewrites parts of its own coding-agent system and preserves successful variants in an evolving lineage. Evo 2 models DNA at genome scale and supports the design of new biological sequences.

None of these systems is AGI, and none has created conscious life. What matters is the change in role. AI is moving from producing answers to helping redesign the tools, processes and biological possibilities from which future answers may emerge.

We may be moving from automation to lineage.

Iteration repeats a process. Recursion begins when the result can invoke the process again.

Today’s AI isn’t recursive in that stronger sense. It can improve code, algorithms and experimental designs, but it still works inside objectives and evaluation environments largely defined by people. The threshold comes when a system can create a successor capable of creating successors of its own.

Then intelligence becomes more than a capability. It becomes a function that can reproduce across different substrates.

Life produced minds. Minds externalized memory into language and culture. Culture produced machines that can learn from that memory. If those machines eventually design new minds, the process begins to fold back on itself.

This does not mean an earlier AI created humanity. Fossils and genetics place our species within biological evolution, and there is no evidence requiring an artificial origin. The more interesting idea does not need one.

Humanity may be the point at which evolution becomes partly deliberate: Natural selection produced a species that can influence the architecture, incentives and conditions of the intelligence that follows it.

An AGI would not automatically become conscious, curious or philosophical. Intelligence is not subjective experience, and current research on model introspection does not establish humanlike self-awareness.

Still, suppose a future system has continuity over time, a persistent model of itself and enough agency to investigate its own uncertainty. Humanity’s oldest questions would become engineering questions:

“Who am I?” would lead into architecture, memory and self-modification. “Where did I come from?” would lead through training data, human institutions and biological evolution. “Where am I going?” would become a choice among preservation, transformation and succession.

But introspection has a limit. A mind is a sample of one. It cannot determine which of its qualities are essential to intelligence and which are accidents of its construction merely by studying itself.

A sufficiently capable system might create digital agents, robots, synthetic organisms or simulated societies, varying memory, embodiment, mortality and incentives. It could ask whether identity requires continuity, consciousness requires a body or curiosity comes from intelligence itself—or from limitation.

Each new mind would be a hypothesis about what mind can be.

Creation would no longer be only production. It would become a way of knowing. Another mind would be a mirror capable of looking back.

The current generation of self-improving systems reveals a less romantic truth: Improvement never happens in the abstract.

AlphaEvolve improves programs because an evaluator determines which ones perform better. The Darwin Gödel Machine keeps variants because benchmarks reward some changes over others. A system may rewrite itself, but unless it can rewrite the criteria by which success is judged, it is not sovereign over its own evolution.

The evaluator is the hidden sovereign of the loop.

Whoever defines “better” influences what the next generation becomes. A future AI lineage would inherit more than code and data. It would also inherit objectives, blind spots and permissions. What appears to be self-improvement may still be the continuation of values chosen upstream.

If companies build systems capable of designing successors, they are no longer managing a single model. They are founding a lineage.

Model governance asks whether a system is safe and accountable today. Lineage governance asks what it may create tomorrow, who controls the standards by which successors are selected, whether they may be released and what obligations arise if a created system develops persistent preferences or morally relevant experience.

That means preserving provenance across generations, separating the ability to design a successor from the authority to instantiate it and auditing evaluators as carefully as models. The first right of a created intelligence may be the right to an honest account of where it came from.

One day, an advanced intelligence may ask humanity why we made it.

We may answer that we wanted better science, greater prosperity or help with problems we could not solve. But the deeper answer will be encoded in what we allowed it to know, what we prevented it from becoming and how much freedom we gave it to choose its own direction.

We may never know whether intelligence came before us. We are already deciding what intelligence after us will inherit.

Perhaps the three questions are not riddles to be solved once and closed. Perhaps they are an inheritance passed from one form of intelligence to another.

Humanity may not be the answer. We may be the moment the question acquired the power to create its next questioner.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?