In this blog post, I explore the topic of self-consciousness. I'm not an expert or a neuroscientist, but I have a strong interest in information theory and machine learning, and I often think about this topic. First, I will explore how consciousness could emerge in a complex dynamic system, supporting the hypothesis with some weak evidence. Then, I will speculate on the characteristics a system should have for self-consciousness to arise within a simulated brain.

I wrote this post despite my limited knowledge of the topic and will convey my low confidence in the hypothesis. I believe it's worth being speculative for the sake of telling a compelling story. The goal is not to be exhaustive; I have purposely oversimplified and linked to oversimplified external material that I hope you find interesting.

What’s self-consciousness?

Everyone experiences consciousness. Access consciousness allows us to retrieve information for reasoning. Phenomenal consciousness involves subjective sensory experiences, emotions, and thoughts. Self-consciousness provides a sense of individual identity, rather than viewing ourselves as just a collection of cells. For the rest of the post I will focus on self-consciousness.

Not all beings require consciousness to survive and reproduce. For example, ants mainly respond to their external environment, reproduce quickly, and, if they live long enough, store information in their genetic code. Their internal state exists but holds minimal information; for instance, some bacteria use chemical gradients, while multicellular beings can use hormones. More complex beings, which reproduce more slowly and have fewer offspring, need to learn from experiences, maintain complex internal states, react based on emotions, and engage in complex problem-solving.

Self-consciousness is intriguing, as it involves experiencing our brain's functioning while reflecting on itself. In the next section, we will explore how consciousness might emerge from a complex dynamic system with numerous feedback loops.

Consciousness as emergent behavior

Consciousness may arise in the brain through a process that combines past learning, problem-solving skills, recent observations, and current sensory inputs to produce sophisticated behavior. Evolution likely pressured humans to simulate the world in their minds. For instance, early humans could track prey for days without constant visual contact, using subtle environmental clues, predicting animal behavior based on past learning or stories, and inferring the mental states of fellow hunters. We evolved from choosing actions based solely on immediate observations to complex reasoning based on generalized past experiences, learned directly or indirectly through stories told by others.

Can consciousness, particularly self-consciousness, be explained as behavior emerging from a complex dynamic system? Can it be simulated? Is it possible to distinguish a truly self-aware system from one that merely imitates it, like a Large Language Model?

Those are significant philosophical questions. We are far from answering them, especially since we lack a precise definition of consciousness and a way to test it. Here, I aim to reason through analogies and speculate a bit. Next, let's explore how a single consciousness could arise from two brains.

Single individual consciousness

It's fascinating how humans perceive themselves as individuals despite having brains with two distinct, connected hemispheres. These dynamic systems cooperate in such synchrony that the feeling of a single "self" arises.

Your brain isn't aware it's split in half, suggesting that self-consciousness might be linked to our senses. Since we can't sense having two distinct hemispheres, and because the two systems communicate effectively, we don't experience two separate consciousnesses.

This holds true even after they are physically separated. Some people experience disrupted connections due to surgery or trauma, resulting in two distinct brains that must coordinate without directly communicating.

Experiments have been conducted on people with split brains, such as showing images to one eye and asking the hemisphere responsible for speech to describe what is seen. While perception is split, consciousness does not appear to be. The two hemispheres find alternative ways to communicate, and a single consciousness still emerges from these two separate dynamic systems functioning in close proximity.

Next let’s ask if it’s possible for multiple individuals to share the same self-consciousness.

Multi-individual consciousness

Can a multi-person consciousness emerge? Can a group experience a single self-consciousness? How should the brains be connected for this to occur? What is the minimum bandwidth required?

In the past three decades, the Internet has served as a connection medium. Now, billions of people are connected at near-light speed. However, the bandwidth remains relatively low, limited by typing, reading, talking, and listening. It's comparable to offline communication, just on a larger scale.

Thanks to Brain-Computer Interfaces (BCIs), communication bandwidth is set to increase exponentially in the coming years. We are now at a point where we can begin addressing these questions.

What happens when neurons firing in one person's brain directly influence neurons in another's? Neuralink has demonstrated the ability to read individual neuron firings and has a working prototype in two humans. BCI technology is gaining momentum, funding, and media attention. Increasing BCI communication bandwidth is now an engineering challenge. I predict it will scale exponentially, similar to the growth of computational power in silicon-based transistors.

I'm particularly excited about BCIs because, for the first time, we can read and write signals from individual neurons. This allows us to probe the brain at its smallest meaningful resolution: the single neuron, the brain's "atom". We now have the tools to shift these questions from philosophy to testable, falsifiable hypotheses.

I recommend this interesting Neuralink immersion, you know, when you have 9 hours.

Or just summarize it with Fabric.

yt --transcript https://www.youtube.com/watch?v=Kbk9BiPhm7o | fabric -sp summarize

In the next section I want to step back from philosophical questions and compare complex biological brains with artificial neural networks and see what they have in common.

Simulated consciousness

Artificial neural networks (ANNs) vaguely resemble the brain's biological neural circuits. Extensive research is being conducted in this field, and we are building increasingly complex networks, exploring various wiring methods (a.k.a. ANN architectures). They are an ideal candidate for a complex dynamic system where consciousness could arise. I propose the hypothesis that, for consciousness to emerge, having the right architecture (the wiring of neurons) is more important than a high-fidelity simulation of the actual biological system.

Will a sufficiently complex ANN with the right architecture truly begin to experience self-consciousness as humans do?

In recent weeks, I've been exploring the field of mechanistic interpretability of artificial neural networks. It's fascinating how exponentially more facts can be stored in a network compared to the number of parameters, thanks to the Johnson-Lindenstrauss lemma. Loosely speaking, the number of parameters in an ANN is analogous to the number of neurons. With the right architecture (e.g., transformer architecture) and sufficient parameters (enough neurons in our analogy, enough complexity), the network exhibits emergent reasoning behavior. For example, it becomes capable of simple arithmetic or deducing novel ideas from memorized facts, rather than merely regurgitating training set text. It develops a compressed, complex internal representation of the world, and the reasoning behavior emerges. What other aspects of consciousness might emerge?

Having explored the similarities in reasoning and fact storage between brains and ANNs, let's examine the differences and where ANNs fall short in simulating consciousness.

Most complex ANNs are decoupled from the environment. For example, LLM parameters remain fixed after training. The ANN doesn't have the chance to continuously probe the environment and discern the boundary between "self" and the environment. If self-consciousness arises from sensory inputs, removing this tight feedback loop might disrupt it. Some experiments involve agents interacting with a simulated environment. I'm very curious to discover the emerging properties of increasingly complex environment-probing agents.

Another important missing element is the drive to survive. If self-consciousness arises from evolutionary pressure for self-preservation, removing it during training suggests that resources for self-preservation could be allocated elsewhere. Thus, self-preservation might be a building block of self-consciousness. Without incorporating self-preservation into the objective function, creating self-consciousness may be challenging.

Conclusion

In this post, we explored self-consciousness: the feeling of being a single entity, distinct from the environment, capable of reasoning and introspecting our thoughts. We discussed how self-consciousness might emerge when complex dynamic systems are connected with high bandwidth, using the example of the two brain hemispheres to support this hypothesis. We also speculated about the possibility of multi-person consciousness and how Brain-Computer Interfaces could enable it in the near future.

I included some YouTube videos that I found interesting on the topic. While I didn't answer any of the questions I posed, my main goal was to highlight and connect intriguing ideas I found online.