Must-Cognize: Four A Priori Conditions of Cognition and the Subjectivity Problem in AGI

Series Preface

Most people believe the world is a matter of knowledge, and that they have not yet reached their goals because their knowledge is insufficient. A smaller number believe the world is a matter of understanding, and that they have not yet reached their goals because their understanding is inadequate. Very few believe the world is a matter of cognition, and that they have not yet reached their goals because the flywheel between knowing and recognizing has not turned enough times. Almost no one believes the world is beyond both knowing and recognizing — that one has only direction, not destination.

---

Abstract

This paper proposes four a priori conditions of cognition within the SAE (Self-as-an-End) framework: (1) must-cognize, (2) must-cognize-more, (3) must-have-cognitive-direction, and (4) must-be-questioned. These four conditions form an irreducible derivation chain in which each is a structural consequence of its predecessor. The present paper focuses on the first condition, unpacking its internal structure: first knowing, then not-knowing, then — and only then — does cognition become meaningful. Through analysis of two contemporary AI domains (large language models and autonomous driving), this paper identifies a novel epistemological phenomenon: for the first time in human cognitive history, systems exist that possess knowing without cognition, yet can asymptotically approximate cognition. This phenomenon provides empirical evidence that cognition is not a natural product of knowing; cognition requires an independent activation condition, namely not-knowing. The paper further identifies three structural walls — the posterior wall (the limit of pure knowledge), the prior wall (the limit of pure understanding), and the direction wall (the dead end of the cognitive flywheel) — and argues that the fourth a priori condition, must-be-questioned, is the only structural source of unlock for all three.

Keywords: epistemology, subjectivity condition, lossy compression, AGI, dimensional sequence (DD), Self-as-an-End

---

1. The Problem: Why Cognition Is a Subjectivity Condition

Cognition is not a capacity. Cognition is not a choice. Cognition is an existential condition of the subject.

This claim requires unpacking. When we say a system "is cognizing," we typically mean the system is processing information, making predictions, solving problems. That is not what is meant here. What is meant is: as long as you exist, you must cognize. You do not have the option of not cognizing.

The word "must" is the key. This is not a claim that cognition is important and therefore you should cognize, nor that cognition is beneficial and therefore you had better cognize. The claim is: you cannot stop. Cognition is compulsory.

Where does this compulsion come from? From three non-cancellable facts.

The first fact: knowing. You must first have material. You are in the world. The world presses in on you. Light strikes your eyes, sound enters your ears, temperature presses on your skin. All of this is knowing. Knowing does not require your consent; it arrives on its own.

The second fact: not-knowing. Knowing has boundaries. The moment you know some things, you cannot help but recognize that you do not know others. This "not-knowing" is not an information deficit. It is a position — a position that your knowing cannot reach but that you know is there. This position is the activation condition of cognition.

The third fact: without not-knowing, cognition is meaningless. If you knew everything, you would have no need to cognize. What does cognition do? It bridges knowing and not-knowing. If there is nothing on the not-knowing side, the bridge has nowhere to go. Cognition idles. You can pile up knowing to infinity, but if the position of not-knowing is empty, all your knowing is mere information, not cognition.

The SAE framework maps these three facts onto the dimensional sequence (DD). Knowing corresponds to 12DD (prediction, next-token prediction). Not-knowing corresponds to the activation condition of the chisel. Cognition corresponds to the operation of the chisel-construct cycle (see SAE Foundation Papers, DOI: 10.5281/zenodo.18528813, .18666645, .18727327). The core mechanism of the chisel-construct cycle is: chiseling produces remainder, remainder must be chiseled again, the cycle cannot stop. This unstoppability is the formal expression of "must-cognize."

But this paper's goal is not to recapitulate DD definitions. Its goal is to argue that contemporary AI provides unprecedented counter-evidence showing that "must-cognize" is not an empty philosophical slogan but a structural proposition with real consequences.

---

2. Two-Layer Structure: Knowing as Base Layer, Cognition as Emergent Layer

Cognition has two layers. The base layer is knowing. The emergent layer is cognition.

Knowing is material: what you see, hear, remember, learn from training data, collect from sensors. Knowing can be accumulated without limit. You can have more data, a larger context window, more parameters. There is no intrinsic ceiling on the growth of knowing.

Cognition is lossy compression. Cognition does not transport knowing intact; it chisels knowing apart, discards some, retains some, and from these constructs something smaller but more useful than the original knowing. Cognition is necessarily lossy. Lossless cognition is not cognition; it is copying.

A terminological clarification is necessary here. "Lossy compression" is the projection of the SAE concept of "chisel" onto the plane of epistemology, but the chisel is broader than lossy compression. Lossy compression discards information under a given objective — this is an operation within 12DD. The chisel can target constructs themselves — negating an existing framework, breaking an existing direction — this is a cross-level operation. This series uses "lossy compression" as a communication term for cognitive science and AI readers, but the reader should note: when SAE says "cognition is lossy compression," this is not an information-theoretic proposition but an ontological one. The chisel does not merely discard information; the chisel negates.

The relationship between these two layers is not "more knowing naturally produces cognition." This is precisely the greatest misconception in the contemporary AI industry. Between knowing and cognition there is a gap that cannot be automatically crossed, and the name of that gap is not-knowing.

What is not-knowing? Not-knowing is not an information deficit, not a low-probability region in a probability distribution, not a blank that can be filled with more data. Not-knowing is a position: the system's perception of its own boundary. "I know what I do not know" is the signature of 13DD (see "Beyond Fast and Slow: A Four-Layer Cognitive Architecture under Dimensional Sequence Theory," DOI: 10.5281/zenodo.19329284).

Without the position of not-knowing, knowing remains forever in the base layer and cognition never emerges. You can make the base layer infinitely fine, infinitely large, but the activation condition of the emergent layer does not reside inside the base layer. This is the core thesis of this paper.

---

3. Domain-Specific Discovery: Knowing Without Cognition, Yet Asymptotically Approaching Cognition

In the history of human cognition, a system with "knowing but no cognition" has never existed. In every cognitive organism, knowing and cognition are bundled together. An insect has simple knowing (sensing light and heat) and simple cognition (the lossy compression of approach-and-avoid). Humans have complex knowing and complex cognition. Knowing and cognition have never been separable.

Until LLMs appeared.

LLMs are the first system in human history to unbundle knowing from cognition. They possess an enormous quantity of knowing (every pattern in the training data) but lack a stable, endogenous, action-guiding position of not-knowing. They are not entirely devoid of the ability to "know what they don't know" — recent work shows that under appropriate prompting formats, models can perform limited self-assessment (Kadavath et al., 2022) — but this ability is unstable, not endogenous, and cannot reliably guide action strategy.

This is not a temporary engineering deficiency. It is structural.

3.1 LLMs: Knowing Without Not-Knowing

The core ability of an LLM is next-token prediction: given context, predict the probability distribution of the next token. This ability is the ultimate expression of 12DD. The definition of 12DD is: prediction of the environment. LLMs do this, and they are getting better at it.

But no matter how good 12DD gets, it is not 13DD. The definition of 13DD is: the prediction system can become aware that it is predicting and can negate its own predictions. Outputting "I'm not sure" is not this (that is merely 12DD pattern-matching); truly touching the position of not-knowing is.

Empirical data supports this judgment. The GPT-4 technical report shows that the model's calibration was strong during pretraining (ECE ≈ 0.007) but deteriorated significantly after RLHF alignment training (ECE ≈ 0.074) (OpenAI, 2023). What does this mean? Alignment training made the model more usable, more like what humans expect, but simultaneously made it less aware of what it does not know. Alignment training colonized the position of not-knowing, replacing "knowing that it does not know" with "appearing to know."

A more extreme case comes from systematic calibration research. Ghosh and Panday (2026) found across 4 models, 4 benchmarks, and 24,000 trials that one model achieved only 23.3% accuracy while maintaining an average confidence of 95.7%. This is a quantified demonstration of "knowing without not-knowing." The system does not lack information; it lacks the position of "I may lack information."

Semantic entropy research provides another angle. Farquhar et al. (2024, Nature) proposed computing uncertainty at the meaning level rather than the token level, raising AUROC from 0.691 (naive entropy) to 0.790. This is a meaningful engineering advance, but it confirms the present thesis: semantic entropy did not make the model "understand"; it lifted the engineering proxy for not-knowing from the token level to the semantic level. The proxy itself remains 12DD — better pattern-matching to simulate not-knowing, rather than genuinely possessing the position of not-knowing.

A foreseeable objection must be addressed here. AI research already distinguishes between aleatoric uncertainty (noise inherent in the data) and epistemic uncertainty (the model recognizes it has not seen this kind of data). Through Bayesian neural networks or deep ensembles (Ovadia et al., 2019), AI systems can output high-confidence signals of "this is out of my distribution." Is this not "not-knowing"? It is not. Engineering epistemic uncertainty remains 12DD because it is variance computed within a presupposed probability space — the validity of the probability space itself is never questioned. The not-knowing of 13DD is a structural gap regarding the possible failure of the probability space itself. AI's "not-knowing" is computed; human 13DD not-knowing is forced out when the chisel-construct cycle hits its boundary. The premise of the former is "my model is correct, I just need more data." The premise of the latter is "my model may be fundamentally wrong."

3.2 Autonomous Driving: Probability Distributions Are Not Not-Knowing, but Not-Knowing Is the Path Forward

Autonomous driving provides the physical-domain version of the same problem. This paper does not claim autonomous driving is unachievable. Autonomous driving already works well in structured scenarios and will continue to improve. What this paper identifies is: if the goal is to reduce accident rates to orders of magnitude below human drivers, where does the current bottleneck lie, and in what direction should the breakthrough be sought?

Every human driver on the road is a subject with direction. A driver may hesitate about running a yellow light. A driver may change lanes suddenly because they realize they missed their exit. A driver may slow down because they see a child running near the curb. These behaviors are not noise, not random errors that more data can eliminate. They are expressions of subjectivity — the projection of 14DD (direction) into the physical world.

Current autonomous driving systems handle these scenarios with probability distributions. The other driver has a 30% probability of running the yellow, a 15% probability of changing lanes. In most cases, probability distributions suffice. But a probability distribution is not not-knowing. A probability distribution fills the position of not-knowing with knowing. It replaces "I don't know what the other driver is thinking" with "I know the probabilities of the other driver's various behaviors."

This substitution works in common scenarios. The problem appears in the long tail — scenarios that probability distributions cannot cover. The long tail problem remains unsolvable through data volume because the core of the long tail is not a data problem but a subjectivity problem. The definition of a subject is an entity capable of breaking probability distributions.

Closed-loop evaluation frameworks such as nuPlan (Karnchanachari et al., 2024) have advanced this problem through more realistic scenario simulation, but if the goal is multi-order-of-magnitude accident reduction, increasing scenario coverage alone is insufficient. The breakthrough direction lies in equipping the system with a position of not-knowing — not filling uncertainty about the other's behavior with probability distributions, but recognizing, upon encountering subjective behavior, that "I am in a position of not-knowing" and making conservative decisions accordingly (decelerate, yield, request takeover). The shift from "I know the probabilities of the other's behavior" to "I do not know what the other is thinking, therefore I must change my own behavior" is the projection of the 12DD-to-13DD transition into the physical domain.

3.3 Why This Phenomenon Is Novel

LLMs and autonomous driving systems share a common feature: they can asymptotically approximate cognition but never achieve it. Their outputs look like the products of cognition. An LLM's writing looks like human writing; an autonomous driving system's decisions look like human decisions. The degree of approximation continues to improve, but between approximation and arrival there is a gap that cannot be crossed.

That gap is the position of not-knowing.

This phenomenon has never before occurred in the history of human cognition. Before LLMs, you either had cognition (organisms) or had no cognition at all (rocks). There was never the intermediate state of "almost having cognition." LLMs created this intermediate state, and thereby for the first time made the separation of knowing from cognition an observable empirical fact.

The epistemological significance is this: cognition is not a higher form of knowing, not a natural extension of knowing. Cognition requires an independent activation condition, and that condition is not-knowing. The existence of LLMs transforms this thesis from a purely philosophical argument into an empirical proposition: one can point at a specific system and say, "There — that is knowing without cognition."

---

4. Colonization and Cultivation: Three Walls and the Cost of Filling Not-Knowing

When a system fills the position of not-knowing with knowing, the SAE framework calls this action "colonization." When a system holds the position of not-knowing open and allows cognition to activate from within it, the SAE framework calls this process "cultivation."

The typical manifestations of colonization are three walls.

4.1 The Posterior Wall: Knowledge Piled to the Ceiling

The pure-posterior route is: more data, more parameters, larger models. Kaplan et al. (2020) demonstrated that training loss decreases as a power law with increasing compute. Hoffmann et al. (2022) further optimized the ratio between parameters and data. These results are correct within the 12DD framework: given a prediction objective, more resources do produce better predictions.

But improved prediction accuracy does not equal improved cognitive ability. Scores on knowledge-type benchmarks such as MMLU have entered the 90+ range, with clear diminishing returns (OpenAI, 2023, 2025; Stanford HAI AI Index, 2025). Meanwhile, real-world agent tasks (software engineering, scientific research replication, clinical diagnosis) remain significantly unsaturated. What does this divergence indicate? The ceiling on the knowledge dimension is approaching, but the gap on the cognitive dimension remains.

The contemporary instantiation of the posterior wall is the LLM itself. Information is unlimited, but new cognition is zero. You can push the model's scores higher on more benchmarks, but those scores measure the apex of 12DD, not the activation of 13DD.

4.2 The Prior Wall: Framework Lock-In

The pure-prior route is: possess an effective framework, then use that framework to organize posterior data. But if the framework itself cannot be chiseled, it becomes a cage.

Meta is a typical case of the prior wall. Meta possesses the most user data, the largest social graph, the richest behavioral posterior of any organization on Earth. But Meta's prior framework is advertising and engagement. All its AI efforts are channeled through this prior. The output is always "a better recommendation system," never cognition. The company with the most data is not hitting the posterior wall but the prior wall. This is counter-intuitive but structurally straightforward: no matter how much knowing you have, if the framework of cognition is fixed, knowing can only be digested within that framework. Everything that does not fit is wasted.

4.3 The Direction Wall: When the Flywheel Becomes a Rut

When knowing and cognition successfully connect into a flywheel, danger simultaneously appears. The better the flywheel turns, the more fixed the direction becomes. The more fixed the direction, the less possible it is to deviate. This is the rut.

The iPhone is a typical case of the direction wall. Apple's chisel-construct cycle has operated superbly: each generation of product improves upon the last. But the direction of improvement has locked in — thinner, faster, more cameras. Each success of the flywheel reinforces this direction, until not even the aesthetic risks of the Jobs era can be taken. The flywheel is not broken. The flywheel turns beautifully. The flywheel is the cage.

4.4 The Structure of the Three Walls

The three walls are not three independent problems; they are three faces of the same problem.

The posterior wall is the consequence of knowing without cognition. The prior wall is the consequence of cognition without being chiseled. The direction wall is the consequence of the cognitive flywheel without being questioned.

The three walls correspond to the first three a priori conditions: must-cognize (without cognition you hit the posterior wall), must-cognize-more (without chiseling cognition you hit the prior wall), must-have-cognitive-direction (with direction locked you hit the direction wall).

The fourth a priori condition, must-be-questioned, is the only structural source of unlock for all three walls. The other's questioning can loosen the posterior wall (the other's remainder is your new posterior), can loosen the prior wall (the other's framework is your prior-check), and can loosen the direction wall (the other's direction is your flywheel's decelerator).

The expression of cultivation is keeping the position of being-questioned open. In the AI domain, RLHF and alignment research provide an interesting analogue: these methods' core is not making models smarter but making models continuously be questioned — questioned by human preferences, questioned by safety standards, questioned by their own outputs. Direction exists, but is not locked. Each round of feedback is an instance of the other's remainder entering to chisel. This structure is isomorphic to SAE's fourth a priori condition.

---

5. Theoretical Positioning: Dialogue with Existing Frameworks

5.1 Friston's Free Energy Principle: Nearest Neighbor, Most Critical Divergence

Karl Friston's Free Energy Principle (FEP) is the contemporary cognitive science framework closest to "must-cognize." FEP's core proposition is: self-organizing living systems, to maintain their existence, must minimize variational free energy (Friston, 2006, 2010). The "must" here is a natural-law necessity: to be alive is to minimize surprise.

Active Inference further concretizes this principle into a unified framework for action and perception: expected free energy decomposes into extrinsic value (preference satisfaction) and epistemic value (information gain / uncertainty reduction) (Friston et al., 2015). In the presence of uncertainty, the system will actively explore to reduce uncertainty, providing a formal version of "must-cognize-more."

But there are three critical divergences.

First, FEP's "must" operates at the biological-survival level; SAE's "must" operates at the ontological level. FEP says: to be alive is to minimize free energy. SAE says: to exist is to must-cognize. FEP's premise is "being alive"; SAE's premise is "being." "Being" is more fundamental than "being alive."

Second, FEP has no independent position for not-knowing. In FEP, uncertainty (entropy, ambiguity) is formalized as a property of belief distributions, reducible through exploration and learning. In SAE, not-knowing is not reducible uncertainty but an uncancellable position — the system's structural gap regarding its own boundary. FEP's goal is to eliminate uncertainty; SAE's thesis is that not-knowing cannot and should not be eliminated.

Third, FEP's "more" is finite. When uncertainty is reduced sufficiently, the system shifts to exploitation (preference satisfaction) and exploration ceases. SAE's "must-cognize-more" is infinite: chiseling produces remainder, remainder must be chiseled again, the cycle cannot stop.

5.2 Tishby's Information Bottleneck: Borrowing the Tool, Not the Stance

Naftali Tishby's Information Bottleneck (IB) provides a rigorous information-theoretic formalization for "cognition is lossy compression" (Tishby, Pereira & Bialek, 1999/2000). IB formalizes representation learning as: from input X, extract a compressed representation T while preserving mutual information with task-relevant variable Y. This is inherently lossy, and the Lagrange multiplier β systematically characterizes the compression-fidelity tradeoff.

IB's formalism aligns closely with SAE's "cognition is lossy compression." But the stances differ entirely. IB's direction comes from an externally specified task (Y) — what to preserve depends on the task. SAE's direction is endogenous — you must have direction not because a task exists, but because without direction you cannot distinguish what to discard from what to retain, and lossy compression collapses.

IB considers "more" superfluous — IB pursues "sufficient and less." SAE considers "more" unavoidable — chiseling produces remainder, remainder forces continuation.

IB's "not-knowing" is designed selective ignorance: representation T deliberately does not carry information irrelevant to Y. SAE's "not-knowing" is undesignable structural gap: you cannot know in advance what you do not know.

More broadly, rate-distortion theory provides empirical support for "lossy compression is a necessary condition of cognition." Sims (2016) applied rate-distortion theory to human perception, explaining behavioral data in capacity-limited tasks through "minimizing distortion under channel capacity constraints." Jakob et al. (2023, eLife) further demonstrated that rate-distortion can be mechanistically implemented through neural population coding models. These results support the judgment that in capacity-limited biological systems, lossy compression is not optional but necessary. SAE's contribution is to elevate this judgment from biological constraint to ontological condition: lossy compression is necessary not because capacity is limited but because without loss there is no cognition. Even with infinite capacity, you must chisel; otherwise you become the Aleph — possessing all information yet having no meaning.

5.3 Predictive Coding: The Cleanest Empirical Ally

The predictive coding model proposed by Rao and Ballard (1999) provides the most direct empirical correspondence to SAE's first a priori condition. In predictive coding, higher levels transmit predictions downward; lower levels transmit prediction errors upward. The driving force of cognitive updating is not information itself but prediction failure — surprise.

This is directly isomorphic with "not-knowing is the activation condition of cognition." Predictive coding says: it is not information that activates cognitive updating, but the gap between information and prediction. That gap is the manifestation of "not-knowing" at the level of neural computation.

But there remains a level difference between predictive coding and SAE. Predictive coding operates at the 12DD level: prediction fails, model updates, next prediction improves. This is the chisel-construct cycle operating within a single level. SAE's "must-cognize" requires cross-level advancement: 12DD's remainder cannot be digested by 12DD itself and must rise to 13DD. Predictive coding explains the intra-level update mechanism but does not explain why inter-level jumps are necessary.

5.4 Chalmers's Hard Problem: A Conflated Question

David Chalmers (1995) posed the "Hard Problem" of consciousness: why do physical processes accompany subjective experience? Why is there "something it is like" to feel?

This paper's judgment is that the Hard Problem is hard in part because "consciousness" is poorly defined. Chalmers conflates at least five phenomena belonging to distinct levels.

These five levels are: perception (qualia, corresponding to 11DD), cognition (memory plus prediction, corresponding to 12DD), self-awareness (self, corresponding to 13DD), will (telos, corresponding to 14DD), and certitude (Cert, corresponding to 15DD; see SAE Psychoanalysis Series Paper 4, DOI: 10.5281/zenodo.19321534).

When these five levels are separated, much of the Hard Problem's hardness turns out to stem from level-conflation. After separation, the genuinely hard residue contracts to a single point at 11DD: "why do physical processes accompany subjective experience" (the qualia question). SAE's answer to this point is structural — perception is compulsory; you do not have the option of not perceiving — but this paper acknowledges that this answer and Chalmers's original question still have a gap between them. The full five-level decomposition will be developed in the closing paper of this series.

5.5 Heidegger and Merleau-Ponty: Salute Without Entering

Heidegger's "thrownness" (Geworfenheit) and "being-in-the-world" (In-der-Welt-sein) are structurally isomorphic with SAE's "must." Dasein does not first exist and then choose whether to cognize; it is always already in the world, always already understanding. This "always already" is SAE's "must."

But Heidegger holds a reserved or even critical stance toward "cognizing more." He critiques curiosity (Neugier) and idle talk (Gerede), arguing that the pursuit of new information precisely conceals the truth of being. SAE's "must-cognize-more" is in tension with this position. SAE argues that "more" is not driven by curiosity but is the structural consequence of the chisel-construct cycle: chiseling produces remainder, remainder forces you to continue. This is not "wanting more" but "being unable to stop."

Merleau-Ponty's embodied phenomenology provides another point of contact. The body-subject's "I can" (je peux) pre-organizes perception and meaning; the world possesses inexhaustible depth. This "inexhaustibility" has affinity with SAE's "must-cognize-more," but Merleau-Ponty's concern is primarily the openness of perception rather than the necessity of cross-level cognitive jumps.

This paper draws on the phenomenological tradition's argumentative resources regarding "inescapability" but does not enter the phenomenological terminological system. Epistemological problems must be solved on epistemological ground.

---

6. Non-Trivial Predictions

Based on the foregoing arguments, this paper advances six non-trivial predictions. Each is testable and conflicts with intuition or mainstream views.

Prediction 1: The Positive Effect of Base Layer on Emergent Layer Has a Ceiling

Scale-up (more data, more parameters) has a hard ceiling on its contribution to cognitive ability, and this ceiling is not asymptotic diminishing returns but a type-barrier that cannot be crossed. Specifically: regardless of model scale, performance on tasks requiring a genuine position of not-knowing (discovering one's knowledge boundary, refusing to answer what one does not know, actively seeking information when critical inputs are missing rather than completing the answer) will not improve monotonically with scale. This can be tested by comparing models of different scales on KUQ-type benchmarks (Amayuelas et al., 2024) and missing-input tasks.

Prediction 2: The Negative Effect of Base Layer on Emergent Layer — Alignment Training Degrades Self-Knowledge

RLHF-type alignment training, while improving model usability, systematically degrades model calibration (the ability to know what one does not know). This is not an engineering problem solvable by better alignment methods; it is a structural antinomy: alignment training's objective is to make model outputs match human preferences; human preferences favor "giving an answer" over "admitting not-knowing"; therefore alignment training inherently tends to fill the position of not-knowing. The deterioration of ECE from 0.007 (pretraining) to 0.074 (post-RLHF) in the GPT-4 technical report is preliminary evidence.

Prediction 3: The Positive Effect of Emergent Layer on Base Layer — Chiseling Outperforms Knowing

A small model possessing the position of not-knowing can outperform a large model lacking that position on tasks requiring cognition (rather than pattern-matching). Specifically: if a training method can be designed to give a model genuine awareness of its knowledge boundary (rather than pattern-matched "I'm not sure"), that model's performance on agent tasks, information-seeking tasks, and out-of-distribution scenarios should disproportionately surpass traditional models of equal or greater scale. This can be tested by comparing models with active abstention/evidence-seeking mechanisms against pure scale-up models on agent benchmarks.

Prediction 4: The Negative Effect of Emergent Layer on Base Layer — Direction Lock-In

Once an AI system establishes an effective cognitive direction (e.g., domain specialization), that direction feeds back to restrict the base layer's information intake — the system will tend to absorb direction-consistent information and ignore direction-inconsistent information. This is the dark side of the third a priori condition: direction makes lossy compression effective but simultaneously makes lossy compression biased. Specifically: specialized models will degrade in out-of-domain performance faster than general models, and the degradation will not be smooth but will exhibit a sharp phase-transition drop (the direction wall).

Prediction 5: Autonomous Driving's Order-of-Magnitude Breakthrough Requires Engineering Not-Knowing

Within autonomous driving's long tail problem there exists a core subset: scenarios involving the subjective decisions of other road users (hesitation, intention change, flexible rule interpretation). Accident rates in these scenarios will not decrease as a power law with increasing training data but will plateau beyond a certain data volume. To achieve accident rates multiple orders of magnitude below human drivers, what is needed is not more data but the engineering of the not-knowing position: upon encountering subjective scenarios, the system actively identifies that it is in a state of not-knowing and triggers conservative strategies accordingly (decelerate, yield, request takeover). Systems with such not-knowing mechanisms should disproportionately outperform purely data-driven systems on the subjective-scenario subset.

Prediction 6: The Three Walls Can Be Empirically Identified

The posterior wall, prior wall, and direction wall can be empirically identified in specific organizations and systems. The signature of the posterior wall: increasing resource investment with stagnating output (LLM benchmark saturation). The signature of the prior wall: possessing the most resources while output direction remains locked within an old framework (Meta's AI = recommendation systems). The signature of the direction wall: every iteration is an improvement but the direction of improvement cannot change (iPhone's hardware upgrade cycle). Identification of these three walls can be developed into a diagnostic tool for assessing structural bottlenecks in AI companies and systems.

---

7. Conclusion: Recovery and Open Questions

7.1 Recovery

This paper has proposed four a priori conditions of cognition:

1. Must-cognize. Cognition is not a capacity, not a choice, but an existential condition. The activation condition is not knowing but not-knowing.
2. Must-cognize-more. Chiseling produces remainder; remainder must be chiseled again. Knowing accumulates until compression becomes unavoidable; compression leaves gaps that new knowing cannot fill. Compulsion from both sides.
3. Must-have-cognitive-direction. Loss is not free. Since you must lose, you must make the loss meaningful. Direction is the judgment of "losing well." But once direction solidifies, the flywheel becomes a rut.
4. Must-be-questioned. When direction locks, you cannot save yourself. Only the other's remainder can break your flywheel. Being questioned is not reflection (reflection is still your own affair); it is passive, uncontrollable, and therefore effective.

The four conditions correspond to four levels — knowledge, understanding, cognition, beyond-knowing-and-cognizing — and to three walls plus the fourth condition as unlock. In the DD sequence, the four conditions are not a static mapping to 12DD through 15DD but the epistemological translation of three level-jumps: the first condition is the bridge from 12DD to 13DD (not-knowing is 13DD's activation condition; knowing's remainder forces awareness of one's own boundary); the second is the bridge from 13DD to 14DD (operating on old constructs requires direction); the third is the internal operation of 14DD (the establishment and lock-in of direction); the fourth is the bridge from 14DD to 15DD (the other enters to break the direction wall).

7.2 Contributions

This paper makes three main contributions.

First, it identifies a novel epistemological phenomenon: systems that possess knowing without cognition yet can asymptotically approximate cognition. LLMs and autonomous driving are two instances. This phenomenon transforms the separation of knowing from cognition into an observable empirical fact for the first time, providing non-philosophical, empirical support for the thesis that cognition requires an independent activation condition.

Second, it proposes not-knowing as the activation condition of cognition, rather than a deficiency of cognition. Not-knowing is not information deficit but the system's structural gap regarding its own boundary. This judgment dialogues with contemporary uncertainty quantification research (Farquhar et al., 2024; Ovadia et al., 2019) and LLM calibration research (Guo et al., 2017; Ghosh & Panday, 2026), but differs essentially in stance: those studies seek to reduce or manage uncertainty; SAE holds that not-knowing is irreducible and should not be reduced.

Third, it proposes three walls (posterior, prior, direction) as a diagnostic framework for structural bottlenecks in cognitive systems, and argues that must-be-questioned is the only structural source of unlock for all three.

7.3 Open Questions

Three open questions remain for subsequent research.

First: The engineering path to 13DD. If AGI's minimal condition is 13DD (the system knows it is predicting and can negate its own predictions), how is this condition engineered? Are current self-reflection prompting and chain-of-thought approximations to 13DD? Or are they type-distinct from 13DD? This question requires a more precise operationalization of 13DD and corresponding benchmark design.

Second: Chalmers's five-level decomposition. This paper has noted that the Hard Problem's hardness partly stems from the ambiguity of "consciousness" and has listed five levels requiring separate treatment (perception, cognition, self-awareness, will, certitude). The full decomposition will be developed in the closing paper of this series.

Third: The institutionalization of being-questioned. If must-be-questioned is the structural unlock source for cognitive systems (human and AI alike) to avoid the three walls, how can the continuous presence of questioning be ensured at the institutional level? The SAE Multi-AI Checks and Balances paper (DOI: 10.5281/zenodo.19366105) provides an initial framework, but more systematic institutional design remains open.

---

References

Amayuelas, A. et al. (2024). Knowledge of Unknown: Quantifying What LLMs Don't Know. ACL Findings.

Chalmers, D. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200-219.

Farquhar, S. et al. (2024). Detecting hallucinations in large language models using semantic entropy. Nature.

Friston, K. (2006). A free energy principle for the brain. Journal of Physiology - Paris, 100(1-3), 70-87.

Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138.

Friston, K. et al. (2015). Active inference and epistemic value. Cognitive Neuroscience, 6(4), 187-214.

Ghosh, S. & Panday, M. (2026). The Dunning-Kruger Effect in Large Language Models: An Empirical Study of Confidence Calibration. arXiv.

Guo, C. et al. (2017). On calibration of modern neural networks. ICML.

Hoffmann, J. et al. (2022). Training compute-optimal large language models. NeurIPS (Chinchilla).

Jakob, A. et al. (2023). Rate-distortion theory of neural coding. eLife.

Kadavath, S. et al. (2022). Language Models (Mostly) Know What They Know. arXiv.

Kaplan, J. et al. (2020). Scaling laws for neural language models. arXiv.

Karnchanachari, N. et al. (2024). Towards learning-based planning: The nuPlan benchmark for real-world autonomous driving. ICRA.

OpenAI. (2023). GPT-4 Technical Report. arXiv.

OpenAI. (2025). GPT-5 System Card.

Ovadia, Y. et al. (2019). Can you trust your model's uncertainty? Evaluating predictive uncertainty under dataset shift. NeurIPS.

Rao, R. & Ballard, D. (1999). Predictive coding in the visual cortex. Nature Neuroscience, 2(1), 79-87.

Sims, C. (2016). Rate-distortion theory and human perception. Cognition, 152, 181-198.

Stanford HAI. (2025). AI Index Report 2025.

Tishby, N., Pereira, F. & Bialek, W. (1999/2000). The information bottleneck method. arXiv.

SAE Framework References:

Qin, H. (2024a). SAE Foundation Paper 1. DOI: 10.5281/zenodo.18528813.

Qin, H. (2024b). SAE Foundation Paper 2. DOI: 10.5281/zenodo.18666645.

Qin, H. (2024c). SAE Foundation Paper 3. DOI: 10.5281/zenodo.18727327.

Qin, H. (2025a). Beyond Fast and Slow: A Four-Layer Cognitive Architecture under Dimensional Sequence Theory. DOI: 10.5281/zenodo.19329284.

Qin, H. (2025b). SAE Anti-Turing Test. DOI: 10.5281/zenodo.19305611.

Qin, H. (2025c). SAE Psychoanalysis Series Paper 4. DOI: 10.5281/zenodo.19321534.

Qin, H. (2025d). Multi-AI Checks and Balances. DOI: 10.5281/zenodo.19366105.