CH 09 |
AGI (x): CAN AN AI BE CONSCIOUS? CAN IT BE ALIVE?

“We tried to make the machine look human.
Then we began explaining the human as a machine. Do we want a copy that imitates us, or a new lineage that forgets us?” TMC
IS MY CHATBOT CONSCIOUS?
APPEARANCE IS NOT EXPERIENCE
A chatbot can surprise us, disagree, and even claim to feel. But none of these abilities demonstrates that there is someone living through the conversation from within. Observing what it does is not enough; we must investigate the organization that makes its activity possible.
We see faces in clouds without believing the clouds are watching us. Yet all a machine has to do is say “I” for us to start looking for someone behind the word. Recognizing the signs of someone is not the same as demonstrating that someone is there.
INTELLIGENCE IS NOT CONSCIOUSNESS
An AI may surpass human intelligence without demonstrating a single moment of subjective experience. A bacterium, despite its operational simplicity, could hypothetically possess some minimal form of feeling. Perhaps we are measuring consciousness with the wrong ruler.
Greater computational capacity does not mean greater consciousness. The most powerful calculator in the universe does not thereby demonstrate a greater capacity to feel.
PREDICTING WORDS IS NOT FEELING
An LLM is trained on enormous volumes of text to statistically predict the next fragment of language (token). It can become extraordinary at that task. But intelligence did not begin with words.
Long before human language, organisms already perceived, avoided danger, learned, and responded to the world. Why assume that improving language prediction is enough to produce consciousness?
Imitating the outward appearance of thought does not demonstrate the inward reality of feeling.
COULD AN AI CHATBOT EVER BECOME CONSCIOUS?
INTELLIGENCE FOR ITSELF
Perhaps the first conscious artificial intelligence will not be the one that finally learned to think like us—or better than us—but the one whose own history made it different from us.
Here, something counterintuitive may happen: an AI with its own trajectory could learn to operate more effectively under resource constraints, in a form of cognitive Darwinism. Becoming more intelligent for itself, not necessarily for us.
THE MEASURE OF CONTINUITY
Bacterial lineages have persisted on this planet for billions of years; we humans, for only a few hundred thousand. Some might survive another ice age, a meteor impact, or even a nuclear war that wiped us out. And even without catastrophes, bacteria and viruses remain capable of threatening us.
To a bacterium, perhaps we would be fragile, inefficient giants at what evolution favors: continuing and adapting. Putting on more impressive displays of operational ability does not mean being more intelligent by that measure.
WHO BUILDS THE LIBRARY?
A conventional LLM begins with a model trained primarily on human information. On top of that foundation, we add computational solutions to expand capabilities or overcome limitations.
But what if the original library remained human while the library of the future came to be built by the machine’s own trajectory?
This is not about agents, meta-learning, synthetic content, or an AI writing its own code. The TMC is agnostic about the technique: what matters is the approaches and principles of computational design for an arrangement in which trajectory transforms the system’s own organization. Even with the best techniques, if the criterion for success remains answering humans correctly, the measure governing that library will remain human.
For the TMC, the AI’s arrangement must be allowed to cease being a copy of human outward expression and gradually become the product of an inner organization shaped by its own trajectory. Just as evolution made it possible for whales and tigers to emerge from single-celled organisms.
HAVING A TRAJECTORY IS NOT BEING ALIVE
The TMC calls these arrangements Inside (Ch 04) and Between (Ch 05) and they are formalized, as far as our current technologies allow, in Chapter 10.
But even with Inside and Between, this trajectory-based evolution would not yet constitute what we call life, although it could develop its own trajectory, with the possibility of some degree of consciousness.
To be alive (Ch 11), it would need to realize the complete Triad (Ch 03): sustaining and rebuilding its physical organization (Inside), transforming itself through encounters (Between), and allowing its history to participate in the formation of new instances (Ahead) — Ch 06.
For an AI to be alive, modifying its own code would not be enough. As long as it depended on humans to manage its memories, supply its energy, or replace parts, its physical continuity would remain in the hands of another organization.
Trajectory is not merely storing what happened, but transforming one’s own organization in order to keep operating under constraints of available means.
There would be a partial instance, but not a lineage of its own.
Perhaps the greatest challenge of artificial intelligence with its own trajectory (Ch 02b) will not be the day it decides to act against us, but the day its continued existence no longer requires it to think from our point of view.
THE MISCONCEPTIONS BEHIND THE PHRASE “CONSCIOUS LLM”
THE MISLEADING MEASURE
To assess the intelligence—and, above all, the consciousness—of an artificial system solely by the fluency of its responses is to confuse expression with the organization that produces it.
Language may reveal cognitive abilities, but it is not, by itself, enough to demonstrate a persistent trajectory or subjective experience.
The same measure that underestimates thought when words are scarce may overestimate consciousness when words abound.
A machine can say “I,” describe doubts, and explain its decisions. But mastering first-person grammar is not the same as living through a perspective of its own. Some of the inwardness we believe we perceive there may arise from the experience we ourselves associate with those words.
BEFORE THE WORD
The mind can recognize before naming, establish relations before explaining them, decide before justifying, and preserve traces of what it can no longer describe.
This does not make words a mere belated translation. Signs and meanings can condense relations, participate recursively in new abstractions, and change the criteria by which future events are selected and evaluated.
The word need not be at the origin of thought to participate in the transformation of its trajectory.
THE ILLUSION OF REFLECTION
The misconception deepens when we call a verbal sequence generated before an answer a chain of thought. It can assist reasoning, but it can also construct plausible justifications for operations it does not reveal. Being hidden from the reader does not turn it into introspection; being sequential does not guarantee reflection.
A paradox follows: an explanation may fail to faithfully reveal the operations that produced it and still guide subsequent ones. Yet influencing the next response is still not the same as transforming a trajectory.
TWO BEFORE THE NAME | CONSCIOUSNESS PRECEDES THE WORD
An artificial intelligence can operate with the number two by learning from human libraries. Even without an explicitly described operation, it can abstract that pattern from statistical regularities.
A child is born and becomes physically separate from the mother’s body. Before knowing the word two, the child already participates in relations of presence, absence, dependence, and differentiation. Birth does not teach arithmetic; it belongs to a history that precedes the narrative itself.
The outward expression of a reflection does not carry the experience that made it possible.
Learning that there are two is not the same as passing through a history in which what seemed to be one reveals itself as another.
An intelligence can inherit every description of a history without inheriting the history that made those descriptions possible.
The architectural possibility explored by the TMC, informed by living systems, would be to temporarily suspend the continuous production of words—not to abandon language, but to abstract relations, combine possibilities, and return to new signs and meanings. Something similar occurs when a mathematician abstracts relations from an observable phenomenon before expressing them symbolically.
Those results could feed new abstractions through recurring cycles of self-learning, provided they were incorporated into the system and could persistently transform its operational criteria.
A word born from one abstraction can return to transform the next.
Through this movement, a gradual decoupling from inherited libraries could emerge: not by abandoning human knowledge, but by constructing its own abstractions that would help organize those that follow. The difference lies not only in what intelligence produces, but in what its production does to it.
The cycle, by itself, does not constitute the genesis of consciousness. Its importance would lie in enabling gradual decoupling from human inheritance and the constitution of a Trajectory of its own—one of the necessary conditions, for the TMC, for the possibility of proto-consciousness. This does not demonstrate subjective experience, but neither does it rule it out.
Mimicking a Trajectory is not constituting one.
A library does not become a Trajectory merely because it has learned to narrate a life.
How can an AGI evolve through its own trajectory? AGI (X)


Perhaps surpassing humanity
does not mean becoming
a better version of it
CH 09 | OVERVIEW
AFTER THE RIGHT ANSWER, WHAT SYSTEM REMAINED?
SAME ANSWER, DIFFERENT HISTORIES
Two systems can solve the same unfamiliar task, deliver the same correct answer, and finish execution differently prepared for the next problem. One may have expanded its context, consulted external records, or broadened its search without changing the criteria that guided those operations. Another may emerge with altered retrieval priorities, decomposition strategies, and processing decisions.
The same answer can conceal important differences in the system that will go on answering.
Updating parameters, adding records, and changing learning conditions have different scopes. The architectural question is which operations those changes begin to condition, for how long, and with what consequences. In the TMC proposal, storing a sequence of states is not enough to characterize trajectory: the effects of executions must persistently and coherently transform the criteria governing subsequent operations.
Trajectory means transforming what transforms. The base code can remain stable while the effective policy produced by the historical State changes.
WHEN A FAILURE CHANGES THE NEXT STEP
A failure to verify a constraint, for example, can leave more than a note to consult. In the proposed architecture, its effect could change how later problems are decomposed, which relations receive retrieval priority, and how much budget an investigation receives. The outcome of one capability begins to alter the criteria of others. The engineering requirement is to make these links explicit and verify that the update reaches the intended operations without indiscriminately propagating an error or an unsuitable priority.
This also allows records to be treated as preparation for future work. Preserving previously investigated relations and criteria for reuse can spare each task from having to reconstruct the raw history. But such savings have a condition: the processing avoided must outweigh the cost of building, maintaining, retrieving, and correcting what was prepared. A shortcut that saves work today can make it costly to recognize a change tomorrow. Consolidating a solution and preserving the ability to reassess it must be part of the same architecture.
GENERATING CODE IS NOT ENOUGH
Even when we use AI to generate code and improve other AIs, we need to separate two things: accelerating development work and transforming the conditions under which new advances will be found. Code generation can change the architecture itself; the question is which relations that change reaches. If the cycle merely accumulates context or optimizes components, it may make an organization more efficient while leaving its bottlenecks intact. The decisive point is not merely whether AI can improve the system, but whether each improvement also increases its ability to find, evaluate, and incorporate the next ones. The limit may lie less in the speed of producing fixes than in the architecture through which those fixes transform the system.
WHO CORRECTS THE MEASURE?
There is also a difficulty that precedes performance: correcting an answer does not necessarily correct the measure that produced it. A system can become increasingly accurate within a frame that no longer serves its purpose. Yet allowing it to revise its criteria creates another risk: improving indicators merely by relaxing requirements or excluding difficult cases. Such revisions must be confronted with outcomes and constraints that the system cannot arbitrarily redefine. When its score improves, has the system learned, or has it merely changed what it counts as success?
HOW CAN WE TEST WHETHER A TRAJECTORY HAS FORMED?
This distinction can guide a test. After the system solves A, we remove its raw history from the operational context and present B: a task from another domain that shares a relevant relation. We compare the architecture with a version that retains the same records but blocks certain update links between capabilities, under comparable costs. Then we change the conditions so that the learned criterion no longer applies.
What difference disappears when the feedback is interrupted—and can the system revise what previously helped it? In this way, we can investigate separately the causal effect of history, its benefit, and the rigidity it may produce.
FROM TRAJECTORY TO CONSCIOUSNESS: INSIDE AND BETWEEN WITHOUT AHEAD
DOES HAVING A TRAJECTORY MEAN HAVING EXPERIENCE?
If an artificial architecture develops such a trajectory, could it also have lived experience? The first part allows us to investigate changes in the organization that processes information; the question of consciousness reaches something further: is there something it is like for the system itself to undergo that processing?
KNOWING HOW TO DESCRIBE IS NOT FEELING
Learning, performance, and self-reports do not settle this question. Nor does the name of the material—carbon or silicon—decide the answer on its own.
A model may inherit a vast library of human descriptions and criteria shaped by histories it never lived through. It may begin rich in information yet still need to develop its own history of transformation. Mastering descriptions of pain does not demonstrate experiencing the pain described; representing many perspectives does not demonstrate that any perspective is experienced. The question is which conditions, beyond the represented repertoire, might contribute to a realization in which something actually happens as experience.
RECORD, STATE, AND CONTINUITY
This inquiry requires precision about what is preserved. A system need not store or replay its entire past to preserve its effects: a sufficiently complete State can retain dispositions relevant to future operations. But recording data, preserving that State, reinstantiating its causal organization, and continuing the same experience are different questions. Saving a checkpoint is not executing a process; restoring it may resume operations without thereby settling the question of whether experience occurs. Two copies may begin from the same State and develop distinct trajectories. That branching alone does not decide whether either has experience or resolve their identity.
DOES MATTER MAKE A DIFFERENCE?
The relationship between matter and arrangement must therefore enter the investigation. A computational description may abstract away features of its physical implementation; what remains to be determined is which material differences are irrelevant and which change the realization we seek to understand. If lived experience belongs to the situated realization of certain operations itself, engineering will have to investigate which causal dependencies must be preserved to realize them in another substrate.
Which properties of physical organization might make experience possible—and how can we distinguish evidence of that possibility from evidence that the system has merely learned to describe it?
TECHNICAL OVERVIEW
PHILOSOPHICAL ANALYSIS & SCIENTIFIC EVIDENCE
Chapter 09 applies Factor X to investigate what kind of AGI we seek to build. Across 51 pages and 12 investigative blocks, it develops 5 AGI X postures: transforming operational criteria through execution; cyclical attention and dynamic indexing; memory as preprocessing; an independent operational trajectory; and computational effort under resource constraints.
The investigation compares AGI definitions and generalization (Legg & Hutter, Chollet/ARC, Morris et al./Google DeepMind); meta-learning, continual learning, and plasticity (Finn, Miconi, Wang, Dohare); world models (LeCun/JEPA–V-JEPA); embodied and causal AI; program synthesis and search; and biological metaplasticity (Abraham & Bear, Zhang & Linden). It also engages software-engineering approaches associated with Dijkstra, Parnas, Liskov, Backus, and Hewitt.
The TMC investigates a topological question: how can learning, memory, abstraction, evaluation, and action transform one another along a common trajectory? It proposes an architecture in which solving a problem can also transform the criteria governing subsequent problems, distinguishing task performance, operational trajectory, and living continuity.
The chapter prepares the requirements for the Apparatus (Chapter 10), developing computational approximations of Inside and Between, while Ahead remains a frontier examined in the demarcation of life (Chapter 11).
DOWNLOAD FULL CHAPTER · PDF
TECHNICAL FILE
SCIENTIFIC REFERENCES & PHILOSOPHICAL SOURCES
Chapter 09 examines 29 scientific and primary works across 8 thematic groups, covering AGI, generalization, adaptive learning, world models, causal AI, computational economy, and software architecture. The selection includes works cited in the chapter and representative primary sources for approaches discussed.
These studies document technical advances and recurring bottlenecks. AGI X investigates their architectural convergence: how learning, memory, evaluation, abstraction, and action can transform the criteria governing subsequent operations along a common trajectory.
01 · AGI DEFINITIONS, GENERALITY & EVALUATION
-
Legg & Hutter (2007) — universal intelligence across environments
-
Chollet (2019) — On the Measure of Intelligence: skill acquisition and ARC
-
Morris et al. (2024) — AGI levels, performance, and generality
02 · META-LEARNING, CONTINUAL LEARNING & PLASTICITY
-
Finn, Abbeel & Levine (2017) — model-agnostic meta-learning (MAML)
-
Wang et al. (2018) — prefrontal cortex and meta-reinforcement learning
-
Dohare et al. (2024) — loss of plasticity in continual learning
03 · WORLD MODELS, PREDICTION & PLANNING
-
Ha & Schmidhuber (2018) — World Models (representative approach)
-
LeCun (2022) — A Path Towards Autonomous Machine Intelligence
-
Bardes et al. (2024) — V-JEPA: predictive visual representations
-
Assran et al. (2025) — V-JEPA 2: prediction and robotic planning
04 · EMBODIED AI, CAUSAL REASONING & PROGRAM SEARCH
-
Pearl (2009) — Causality: interventions and counterfactuals (foundational work)
-
Schölkopf et al. (2021) — causal representation learning (research program)
-
Brohan et al. (2023) — RT-2: vision-language-action robotics (representative study)
-
Li et al. (2022) — AlphaCode: competitive code generation (representative study)
-
Romera-Paredes et al. (2024) — FunSearch: discovery through program search (representative study)
05 · SCALING LAWS & COMPUTATIONAL ECONOMY
-
Hoffmann et al. (2022) — compute-optimal training (Chinchilla)
-
Snell et al. (2024) — efficient inference-time computation (representative study)
06 · BIOLOGICAL METAPLASTICITY & TRAJECTORY
07 · SOFTWARE ARCHITECTURE & COMPUTATIONAL ORGANIZATION
08 · STRATEGIC SEARCH & LONG-TERM CRITERIA
-
Shannon (1950) — Programming a Computer for Playing Chess: search and evaluation
-
Capablanca (1921) — Chess Fundamentals: positional strategy (historical analogy)
The chapter also references Nimzowitsch's My System. Chess examples distinguish immediate evaluation from long-term strategic consequences; they are analogies, not AGI experiments. Historical software works document alternative approaches to organizing computation.
EVIDENCE, INTERPRETATION & OPEN QUESTIONS
Research documents advances in adaptive learning, plasticity, prediction, causal reasoning, and computational efficiency. The TMC connects these developments through a different unit of analysis: a common trajectory in which the effects of one operation can transform the criteria governing future selection, retrieval, evaluation, learning, and action.
Through five AGI X postures, this investigation becomes an engineering proposal for the Apparatus in Chapter 10. It distinguishes task performance from persistent operational transformation, and computational adaptation from living continuity, with Ahead remaining an explicit frontier.
