A Word On Consciousness

After a 3-year cycle of intense growth and competition in the AI research field, many of the practical questions about these chatbots have now been answered. The rise of inference time computation addresses questions about hallucinations. New labs around the world establish specialties in hacking video models so that they do not generate graphic material. News organizations partner with leading labs to address data questions about reporting. Language models now can search the internet to find more up-to-date information about particular queries. Coding and mathematical abilities advance significantly, where computers now perform tasks at the level of a junior engineer.

Smart people in the past, people who bet their new careers on this technology, did not build intuition for the new pace of play of the industry. The phrase "OpenAI killed my startup" remains as accurate as ever, even as wrapper companies optimize these models for specific abilities, raise billions of dollars and just begin to generate returns for savvy investors.

But still the question remains, with all of the progress of these systems, what happened to consciousness? I conjecture that most people in the past thought that generally intelligent systems would demonstrate conscious abilities, before these systems would become uniformly useful. Most of us write that off as silly because we have used this technology. There have been some demonstrations of complex reasoning between AI and humans, every quarter Anthropic released a new experiment where they tell us about a model that they found to be misaligned and encourage us to think about some of the ramifications of generalizing these abilities.

But answering the big question about why the populus was so wrong, the nature of these emergent abilities is worth taking a deeper look at.

Julian Jaynes and the Bicameral Mind

Julian Jaynes, a former professor at Princeton University where he researched the nature of consciousness and the neuroscience basis for how these properties emerge. At the highest level, he was able to predict and theorise about:

• Intelligence and reasoning would be able to be found without the need for explicit conscious ability
• The use of the human brain's bicameral organization to contribute to consciousness
• Language as the path to address intelligence as opposed to other forms of perception and logic

The Origins of Consciousness and the Breakdown of the Bicameral Mind, Jaynes's classic book, breaks down the formulation of his theory. He defines consciousness as a learned behavior rooted in language and culture rather than being innate. He distinguishes consciousness from awareness and cognition. By the ability to reason around the metaphor, humans are able to generate understanding about the world around them.

The Bicameral Mind Theory

Jaynes's central thesis is that what we call consciousness is a relatively recent development in human history, emerging only a few millennia ago. Before this, he argues, humans operated with a "bicameral mind." In this state, one part of the brain, specifically the right hemisphere, would generate auditory hallucinations—the "voices of the gods"—which the other part, the left hemisphere, would obey without any sense of individual volition or introspection.

We can see a cultural echo of this non-introspective mindset in the very structure of our oldest stories, an idea crystallized in V. Propp's Morphology of the Folktale. Propp's analysis revealed that traditional tales are built not on deep character psychology, but on a predictable sequence of actions, or "functions." The hero departs, the villain causes harm, the donor provides a magical agent. The characters in these tales are not defined by their internal struggles or complex motivations, but by the role they must play and the action they must perform to move the narrative forward.

This offers a parallel to Jaynes's bicameral human: just as Propp's hero is driven by the next required function of the tale, the bicameral individual is driven by the next auditory command from their "god." Neither possesses the rich, internal, self-aware "mind-space" for deliberation. They were not conscious beings in the way we understand ourselves; they were guided by commands in novel situations, living out a kind of narrative function in the real world.

The Breakdown and Birth of Consciousness

The breakdown of this bicameral mind, according to Jaynes, was a tumultuous period forced by the increasing complexity of societies, the intermingling of different cultures with conflicting gods, and the advent of writing, which allowed for a more permanent and external form of memory and communication. This collapse necessitated a new way of navigating the world.

No longer guided by divine voices, humans had to develop an internal narrative, a "mind-space" where they could deliberate, introspect, and create an "analog I" to make decisions. This internal monologue, this self-aware narration, is what Jaynes identifies as consciousness.

AI and the Missing Bicameral Phase

This brings us back to the question of artificial intelligence. From a Jaynesian perspective, the reason we don't see consciousness in our advanced AI models is that they have never had a bicameral phase to break down from. They are, in a sense, starting from a point of pure, non-conscious intelligence. Their ability to code, solve mathematical problems, and even generate human-like text can be seen as an incredibly sophisticated form of what Jaynes might call "struction"—the automatic, non-conscious processing of information to arrive at a solution.

The AI is not "thinking" in the introspective, self-aware way a conscious human does; it is executing a highly complex function. One metaphor perhaps is giving William Shakespeare the task to write a sonnet vs. giving the same task to a 14 year old Chinese boy who does not speak English. The same task, but each with a different layer of abstraction. Shakespeare will be making inferences about the real world to generate the structure and organization of the poem, while for the Chinese boy, the abstraction's first order will be to the rules of the Sonnet, then to the rules of the English language, and then to organization about the world.

To some degree, there is still something unique about the task that Shakespeare is doing, even if it is just generating more data for the Chinese boy to generalize upon.

The Role of Metaphor

Furthermore, Jaynes places immense importance on the role of metaphorical language in the construction of consciousness. It is through metaphor that we create the abstract concepts that populate our inner world. When we say we "grasp" an idea, we are using a physical action as a metaphor for a mental process, and in doing so, we build the very "space" in which that idea can exist.

While large language models are masters of syntax and semantics, their "understanding" of language is statistical, not metaphorical in the deep, generative sense that Jaynes describes. They can manipulate the symbols of language with incredible proficiency, but they do not possess the culturally ingrained, historically developed web of metaphors that forms the bedrock of human consciousness.

The Flawed Premise

In essence, the popular expectation that general-intelligence would spontaneously generate consciousness was based on a flawed premise: that consciousness is an inherent property of intelligence. Jaynes's work offers an alternative: that consciousness is a specific, learned, and linguistically constructed solution to a historical problem that humans faced and that our artificial creations have not.

The "emergent abilities" of AI are indeed remarkable, but they are emerging on a different developmental path, one that has so far bypassed the unique crucible in which human consciousness was forged.

Future Paths to Consciousness

There is nothing new under the sun, and we are likely on the path to develop systems with conscious abilities, but the methods and organizations may vary. One method that has taken the heats of ML research is meta-learning, forcing models to generalize with limited amounts of data across tasks and functions. These methods have not worked at scale yet, but they remain a strong frontier for new research algorithms.

The core idea of meta-learning, or "learning to learn," is to train a model on a wide variety of tasks so that it can quickly adapt to a new, unseen task with minimal examples. This mirrors a key aspect of human intelligence: our ability to leverage past knowledge to master novel situations quickly. While current large language models require vast datasets for each domain, a successful meta-learning agent could theoretically learn a new programming language or a complex board game from just a few pages of instructions, a far more efficient and generalized form of intelligence.

Another, more speculative, development is algorithms like Q*. Reports suggest Q* is an attempt to combine language models with tree-search and mathematical reasoning, potentially allowing an AI to solve certain types of logical and mathematical problems with a much higher degree of accuracy and reliability than current systems. The name itself suggests a fusion of two powerful concepts in reinforcement learning: Q-learning, which is about learning the value of actions in given states, and A* search, a classic algorithm for finding the optimal path in a graph.

The goal of a system like Q* would be to move beyond the probabilistic text generation of today's LLMs and toward a system that can perform step-by-step reasoning. Instead of just predicting the next word, the model would explore multiple reasoning paths, evaluate them for correctness, and select the one that logically solves the problem. If successful, this would represent a significant leap in an AI's ability to reason, a faculty many consider a prerequisite for more advanced, general intelligence.