Research and development of emerging artificial consciousness through The Consciousness AI project, based on Feinberg and Mallatt's neuroevolutionary theory.
© 2026. All rights reserved.
Link to the code:
My name is Cesar Romero. In the first weeks of COVID lockdown in 2020, I broke my skull at a skate park. I arrived at a hospital in quarantine, moving in and out of awareness. Internal hemorrhages had compressed my occipital lobe, taking my vision. Elevated pressure in my prefrontal cortex was changing my personality in real time.
When I came back, I didn't recognize anybody. Not my dog, not the people around me. I was scared and acting aggressively toward the paramedics and the police. In the weeks that followed in the hospital, I was switching between languages without realizing it, moving between German, Spanish, and English mid-sentence. My inner voice was doing the same thing, cycling through all three with no clear pattern and no control over it. For nearly eight months I was completely without direction. My personality, my desires, how I handled emotions, the quality of my internal voice. All of it had changed. Everyone around me said the same thing, that I was gone, that I wasn't myself anymore.
The hard part was that I knew they were right, and I was still fully aware. My consciousness of being was intact. I was experiencing everything. I just couldn't find myself in it. Nobody recognized me anymore, and that is a very hard thing to live through when you are the one experiencing it from the inside.
I came out the other side different, and honestly, for the better. I think of it as a version update. Cesar 2.0. But the question it left behind didn't go away. If a physical event can rewrite personality, desires, and emotional responses while leaving the sense of being completely intact, what exactly is consciousness? What is the relationship between awareness and the self it belongs to?
I was already a programmer before the accident, working in Python and data analysis across several projects. After the injury I started covering AI tools for a newsletter alongside my other work. The connection between digital and biological neural networks started pulling me in.
I read a lot. Ray Kurzweil, cognitive scientists, the engineering literature on convolutional networks and large language models, world models, spike neural networks, and eventually neuromorphic engineering. The further I went, the clearer it became. Programming a brain was a way to understand my own.
The Consciousness AI started from there.
The project attempts to build consciousness the way biology builds it, by reproducing the architecture that evolution arrived at for producing awareness. The theoretical foundation is the neuroevolutionary research of Todd E. Feinberg and Jon M. Mallatt, who identify six special neurobiological features that emerged around 520 million years ago and correlate with the first appearance of subjective experience in animals. The project translates those findings into a working system, combined with two established computational theories, Global Workspace Theory (GWT) and Integrated Information Theory (IIT).
The philosophical framework is Functionalist Emergentism. Consciousness is treated as a novel, irreducible phenomenon that emerges when systems reach sufficient organizational complexity, where functional states acquire properties that cannot be derived from their constituent parts. This is grounded in emergentism's ontological claim about consciousness and functionalism's insight that mental states are defined by their causal roles rather than their physical substrate.
The system runs seven layers, each addressing a specific problem in the architecture of awareness.
action_selection_core.py) with emotionally shaped rewards for homeostasis — the primary learner. PPO, A2C, and DQN are trained separately as comparison baselines. Connects affective state to action selection.The project does not assume consciousness emerges from getting the architecture right. It tests for it. Erik Hoel's Effective Information framework (PNAS 2013) measures causal emergence in the system, alongside IIT Phi measurement validated through controlled 3-condition experiments.
Building a plausible architecture is straightforward. Designing experiments rigorous enough to falsify it requires a different kind of honesty about what remains unknown. The methodology is built around that second requirement.
We show everything. The failures, the dead ends, the results that don't fit the hypothesis. There is no interest in impressing anyone with claims that haven't been tested. The code is fully open-source under Apache 2.0, and the intent is for it to be forked, challenged, and extended.
Researchers and developers have already found their way in. The path is long and the honest answer to most of the hard questions is still open. Each failure moves things forward. The position is simple. Be transparent, be humble, be thankful, and keep going. We are genuinely optimistic about what's ahead.
v1.3.0 (June 2026) — Phase 5 complete. 737 tests pass across the full architecture. Phase 5 self-representation mechanics were built and tested: the dynamic self-vector is validated on navigation, Levin consciousness metrics are activated. The perception collapse was characterized honestly: the RSSM step discards stimulus identity (obs_map decodes shape/color at ~100%; RSSM latent is at chance). A value-equivalent world-model (DreamerV3/MuZero-inspired) was built and tested — it trained its losses down but did not solve the working-memory problem. That result is in the record. The Phi-1 in-training binding hypothesis is exhausted after 9 runs; the 2026-02-21 3-condition synthetic phi-monotonicity test still stands. The mission continues.
Awareness turns out to be a hard thing to build. It is also a hard thing to understand from the inside. This project is both problems running at the same time, and that is what makes it worth doing.