Artificial Consciousness Research

Searching for the architecture of autonomous, self-directed intelligence

neander.io is an artificial consciousness research company using evolutionary computation to discover the neural architectures that give rise to autonomous, self-directed behavior — and to understand what structural ingredients any genuinely aware system would need to possess.

Read our research →
Research Overview
6 Sections
01 The Question

What kind of structure makes an AI truly self-directed?

Most AI today is built by choosing a fixed structure and training it on massive amounts of data. The bigger the model and the more data, the better it scores on benchmarks. This approach has produced remarkable capabilities — but it sidesteps a deeper question.

Current AI systems are fundamentally reactive: given an input, they produce an output. They don't generate behavior from within. They don't maintain goals across time. They don't improve themselves during an experience. Scaling a reactive system produces a more capable reactive system — it doesn't produce something autonomous.

neander.io is asking what kinds of structures give rise to behavior we would recognize as genuinely self-directed: internally motivated, adaptive over time, and capable of acting without being prompted by every next input. That question cannot be answered by building larger models. It requires searching for a different kind of architecture entirely.

02 Inspired by Nature

Even the simplest creatures show us the way

An ant has a nervous system smaller than most microchips, yet it navigates complex terrain, remembers routes, and coordinates with thousands of others toward collective goals. A nematode worm — with exactly 302 neurons — displays goal-directed behavior, learns from its environment, and maintains stable internal states that persist between experiences. Research in minimal cognition has shown that fewer than fifteen neurons can be enough for a system to sustain internal computation and make context-dependent decisions.

These biological systems were not designed. They emerged through millions of years of evolution: undirected variation, selection pressure, and the gradual accumulation of structures that happened to work. And yet they exhibit exactly the properties we associate with awareness — persistent internal states, self-directed behavior, adaptive learning — in systems far simpler than anything modern AI builds.

This is one of the core motivations behind neander.io's approach. If evolution can produce autonomous, self-directed behavior from such simple parts, an artificial evolutionary process should be capable of finding similar structures — and may discover architectural principles that no human designer would think to try.

03 The Method

Let evolution do the searching

Rather than designing neural networks by hand, neander.io runs an evolutionary process over network structures. Hundreds of candidate architectures are generated, evaluated, and selectively bred across many generations — those with the most interesting properties survive and pass on their structural traits to the next generation.

A key distinction: what gets evolved here is the blueprint of a network, not the specific strengths of its connections. The blueprint specifies which units connect to which, what kind of processing each unit performs, whether connections loop back on themselves, and what built-in learning rules — if any — are active during the network's lifetime. Connection strengths are determined separately, either through training or the network's own adaptation mechanisms. This separation allows the research to isolate what the architecture itself contributes, independent of what the network has learned.

The result is structures that no human engineer would design from scratch. Human designers build what they can justify in advance. Evolution finds what works, even when no one can explain why it should.

04 What We Measure

More than a single performance score

Most AI research judges systems on one number — accuracy, reward, or speed. Optimizing a single score produces systems that are very good at exactly one thing, and nothing else. neander.io evaluates its evolved networks across four dimensions simultaneously, because the systems worth finding have to balance all four.

The four criteria are: task competence (does it accomplish what it is meant to do?), behavioral autonomy (does its behavior show signs of being internally driven, rather than purely responding to the most recent input?), internal-state coherence (does it maintain stable, slowly-varying inner dynamics — something like a sustained state of mind, rather than a blank slate reset with every input?), and meta-learning capacity (can it improve on new tasks during a single lifetime of experience, not just across separate training runs?).

These four objectives pull against each other. A network that maximizes autonomy at the expense of task performance isn't useful; one that maximizes task performance with no internal dynamics isn't interesting. The networks that sit at the boundary — performing reasonably well on all four simultaneously — are the ones the research is looking for. That frontier of trade-offs is itself the scientific result.

05 The Research

Starting small, then scaling up

The research runs in three phases. The first focuses on the smallest possible networks — under thirty processing units — where every component can be examined in detail. These networks are tested on a task that requires genuine working memory: maintaining balance while responding to cues that arrive seconds before the action they require. A purely reactive system cannot solve this. Only a network that sustains internal state across time can. This phase maps out which structural patterns give rise to that capability.

The second phase introduces plasticity: the ability for a network to modify its own internal connections during its lifetime, according to rules encoded in its blueprint. This tests whether evolved architectures can genuinely learn how to learn — improving at new tasks through experience, without any external retraining. Architectures that combine this capability with high autonomy scores are of particular interest, because that combination is the clearest candidate signature of proto-conscious dynamics.

The third phase scales up to networks of several hundred units in richer environments — including incomplete information, multiple simultaneous tasks, and interaction with a second independently evolving population. The central question is whether the structural patterns identified at small scale remain relevant, and whether more complex awareness-like behaviors emerge when these simpler building blocks are composed together.

06 What We're Not Claiming

Honest about the hard questions

neander.io does not claim to have built a conscious machine, and the research is designed with that epistemic caution built in from the start. Consciousness remains one of the deepest unsolved problems in science and philosophy — contested in theory, difficult to measure in practice, and unlikely to be resolved by any single experiment.

What the research does try to do is identify measurable, observable properties — behavioral autonomy, internal coherence, adaptive learning, self-modelling — that any aware system would plausibly need to possess, and find the structural conditions that produce them. High scores on these metrics are evidence of architecturally interesting systems. They are not, by themselves, evidence of experience.

A network that maintains stable internal states, drives its own behavior from within, and improves itself through experience is a genuinely remarkable and scientifically valuable system — whether or not it is conscious in any philosophically meaningful sense. That is what this research aims to produce, and what it aims to understand.

Closing Statement

This work will not settle whether machines can be conscious. It may, with luck, sharpen our understanding of what properties any candidate conscious machine would need to possess — and bring that question closer to the reach of science.

neander.io — 2026