Evolution is the only process we know of that has reliably produced open-ended complexity over very long time horizons. Starting from chemistry, it has generated multicellular organisms, nervous systems, language, and cultural evolution — each stage creating the conditions for the next, with no apparent ceiling in sight after approximately four billion years. AI researchers have observed this and asked, for at least several decades: can we build a computational analogue? Can we design a system that generates progressively more complex and capable structures without converging prematurely on a fixed solution?
Progress has been slow. Not for lack of effort — the open-endedness problem has attracted serious researchers and a substantial body of theoretical work. The difficulty is partly technical and partly conceptual. The technical difficulties are real but tractable, at least in principle. The conceptual difficulty is more interesting: we have not always been precise about what open-endedness is, why it fails to arise spontaneously in most AI systems, and what structural conditions are necessary and sufficient for it to emerge.
What open-endedness means, precisely
The term gets used loosely in a way that muddles several distinct properties. It is worth separating them.
The first property is perpetual novelty generation: the system continues to produce entities (agents, structures, behaviours, strategies) that are genuinely new relative to what existed before. Not just variation within a fixed space, but expansion of the space itself. This is what biological evolution does when it produces new body plans, new metabolic pathways, new nervous system architectures — each of which opens up a new region of possibility space that did not exist before.
The second property is cumulative complexity increase: the new entities are, on average, more complex — in some meaningful sense — than prior ones. Evolution does not just generate noise. The complexity of the most complex organisms has increased monotonically over geological time, even though the majority of evolutionary dynamics is dominated by neutral or slightly deleterious mutations.
The third property is absence of convergence: the system does not settle into a stable equilibrium or cyclical pattern. It maintains genuine dynamism over arbitrarily long timescales. This is both the most important and the hardest property to achieve in computational systems, which are strongly biased toward convergence by the gradient descent dynamics used in most modern learning.
Many systems described as "open-ended" in the AI literature have the first property but not the third — they produce varied outputs for a while, then plateau. Evolutionary algorithms in simulated environments often produce interesting early diversity that collapses as the population converges on a local optimum. Understanding why is key to making progress.
The convergence problem
Gradient-based learning systems converge by design. That is what gradient descent is for: finding a minimum of a loss function. Open-endedness requires something structurally different — a system that does not have a global minimum it is trying to reach, or one in which reaching any local minimum immediately changes the landscape in ways that create new gradients to follow.
The latter is the more interesting formulation. In biological evolution, the fitness landscape is not fixed — it changes as organisms evolve, because organisms are part of each other's environment. The Red Queen effect: organisms must keep running just to maintain fitness relative to other evolving organisms. This coevolutionary pressure is what prevents convergence. There is no fixed target to converge on, because the target is itself evolving.
Adversarial self-play has this structure to a limited degree — the agent's opponent changes as the agent improves, preventing convergence on a fixed strategy. But the scope is narrow: two agents playing the same game converge on a Nash equilibrium, which is a fixed point of the coevolutionary dynamics rather than an escape from them. What biological evolution has, and two-player game-playing AI does not, is an unbounded landscape in which new niches and new interaction types continue to emerge indefinitely.
The environment as a co-generator
One underdiscussed factor in open-endedness is the role of the environment itself as a generative source of complexity. In biology, the environment is not passive — it evolves alongside organisms, and the relationship is mutual. Organisms modify their environments (biofilm, soil chemistry, atmosphere), creating new selective pressures that in turn drive further evolution. The environment is a participant in the process, not a fixed backdrop.
Most AI environments are static. The world model that an agent is trained against does not change in response to the agent's actions, at least not at the level of the environment's generative structure. Procedurally generated environments go some way toward this — the specific instances encountered vary — but the generative grammar itself remains fixed. True environmental co-evolution, in which the environment's generative structure responds to and is shaped by the agents operating within it, is rare in AI systems and technically demanding to implement.
We are investigating lightweight versions of this idea: environments in which the distribution of states encountered by the agent shifts not just through exploration but through the agent's interaction with the environment's generative parameters. The goal is not to simulate biological evolution, which operates at timescales and population sizes that are not computationally accessible. The goal is to preserve the structural property that matters: the fitness landscape is non-stationary in a way that is responsive to the learner's current capabilities, ensuring that convergence in one region of the landscape opens up complexity in another.
Open-endedness and representation
There is a deep connection between open-endedness and representation that is not always made explicit. An agent that has converged on a fixed representation of its environment cannot exhibit genuine open-endedness, because its model of what the environment is capable of is fixed. Open-ended complexity generation requires the capacity to form new representations as new structures are encountered — to extend one's model of the world rather than classify new experience into existing categories.
This is related to the general problem of compositional generalisation: can a system understand combinations of concepts it has not previously encountered by composing representations of the component concepts? Humans appear to have this capacity to a significant degree — we can form useful representations of situations that have no historical precedent. Current machine learning systems have it to a lesser degree, and the cases where they fail tend to be exactly the cases where genuine novelty is required rather than recombination of familiar patterns.
We are not claiming that compositional representation is sufficient for open-endedness. But it may be necessary — if the system's representational capacity is fixed, the complexity it can handle is bounded, and true open-endedness (in the sense of perpetual, unbounded novelty generation) is precluded. A fully open-ended system must be capable of growing its own representational vocabulary in response to what it encounters.
Why the problem has stayed hard
Honest accounting of the state of the field requires acknowledging that the most impressive open-ended-looking AI systems are impressive in ways that differ from biological open-endedness in key respects. Large language models generate remarkable variety and can be prompted to produce genuinely novel outputs. But the generation operates within a representation formed at pretraining time, which is fixed once training completes. The system does not continue to evolve; it does not develop new representational vocabulary in response to post-training experience; its complexity does not increase over time in the way a biological lineage's does.
We are not arguing this is a deficiency that makes these systems unimpressive. They are deeply impressive in their domains. The argument is narrower: they are not open-ended in the technical sense, and they should not be used as existence proofs that the open-endedness problem is solved or nearly so.
What would change the picture? Systems that can form new representational structures during deployment, not just at training time. Environments with genuinely co-evolving dynamics that prevent the convergence trap. Learning objectives that reward maintained complexity and ongoing novelty generation rather than performance on a fixed evaluation distribution. Each of these is a hard research problem. Together they describe something like the necessary conditions for computational open-endedness, and understanding their relationships — which are sufficient as well as necessary, which can be relaxed — is a large part of what we are trying to work out.
