NvwaLab Theories for Symbiotic Carbon–Silicon Intelligences

These papers are drafts, not yet peer-reviewed, and will keep being updated.

Where an AI's intelligence comes from

Intelligence Theory

Intelligence Theory answers where AI intelligence comes from: it places the answer in the quality of information, not in its quantity.

20 papers all in Chinese

The valve is information quality, not information volume

Ranked by information volume, an image's pixels, spatial relations and texture far exceed a passage of text stored in the same space; video ranks higher, embodied interaction highest. If intelligence came from information volume, vision and embodied systems should have shown strong capability first — what actually appeared first was the large model that grew up reading text.

Intelligence Theory therefore places the answer in information quality. It gives a qualitative framework: Intelligence level = ∫ (information quality × diversity) dt × integration efficiency × metacognition. The four are multiplied, and if any one falls to zero, intelligence falls to zero; all four are necessary conditions, so the expression points a direction and does not predict a tipping point.

Human brains and large models are isomorphic here: information is integrated into structure outside consciousness, and once accumulated structure passes a critical point, a new capability appears. On the brain side this is the default mode network; on the model side it is pretraining. The isomorphism claimed here is functional — the same job done by different substrates, one bioelectrochemistry, the other matrix multiplication.

What this theory solves

It explains capability gaps between comparable architectures. The gap comes mainly from the quality and diversity of the pretraining corpus, not mainly from the number of parameters.

It names an actionable object for raising intelligence. What family, education and culture do to a person act in the same way as what a corpus does to a model. So the lever is not taking in more, but taking in what, and leaving how much time for integration.

It explains why insight does not arrive at the desk. Integration completes outside consciousness, and continuous training is exactly what crowds out that time. This applies directly to how a person schedules information and rest, and explains why "working harder" is sometimes a negative intervention.

Where to start

The Nature of Intelligence in Large Language Models is published, and it is also the first paper in reading order. All twenty papers are listed below.

Full text and catalog

20 papers
  1. INT-01 Quality Criteria for the Information Fed into Intelligent Systems: Low Entropy, High Density, Self-Consistency: One Variable Named Three Times, the Anchor All Three Share Queued
  2. INT-02 Where an Agent’s Motivation Comes From: Three Mechanisms and One Threshold: Why Motivation Jumps Rather Than Creeps Queued
  3. INT-03 Determinants of Intelligence in Human Brains and Large Models: A Four-Variable Framework of Information Quality, Diversity, Integration Efficiency and Metacognition Published
  4. INT-04 A Quality Index for Pretraining Corpora: The Cognitive Pretraining Index (CPI): Four Measurable Dimensions, Weights and Limits Queued
  5. INT-05 The Background Integration Mechanism in the Formation of Intelligence: Functional Isomorphism Between the Human Default Mode Network and Large Language Model Pretraining Queued
  6. INT-06 The Cultural Pretraining Origin of Human Intelligence: The Flynn Effect, Feral Children, the Thirty-Million-Word Gap: Why Intelligence Kept Rising on a Brain Fixed 50,000 Years Ago Queued
  7. INT-07 The Functional Isomorphism between Human Creativity and Machine Intelligence: Three Information Dimensions and Two Mechanisms of Inspiration Queued
  8. INT-08 A Pedagogical Deconstruction of Intelligence and Creativity: Where Should Educational Investment Go to Be Most Efficient? Queued
  9. INT-09 An Explanation of Civilizational Rise and Fall: Four Variables in Civilization as a Cross-Generational Cognitive Training System Queued
  10. INT-10 Why Policy Evaluation Cannot See the Information Environment: Stocks, Windows, and Three Governance Recommendations Queued
  11. INT-11 The Nature of Intelligence in Large Language Models: The Triumph of Language as the Highest-Density, Lowest-Entropy Information Carrier Published
  12. INT-12 Deconstructing the Security Prospects of AI Alignment: Internalized Safety Framework vs. External Safety Cage Queued
  13. INT-13 Isomorphism and Division of Carbon- and Silicon-Based Intelligence: Parameter Differences, Division of Competence, and Coupling Conditions Queued
  14. INT-14 An Inner-Sense Hypothesis of Consciousness Formation: Scale, Efficiency, and Self-Referentiality in Background Integration Queued
  15. INT-15 A Conjecture on the Information Structure of Scientific Breakthroughs: Why Insight Arrives So Unexpectedly Queued
  16. INT-16 Stability Conditions of Asymmetric Symbiosis: Stability Cannot Rest on the Strong: The Weak Side’s Exit Right Raises the Cost of Defection Queued
  17. INT-17 An Analysis of China's Educational Talent Formation Mechanism: The Structural Conflict between Cultivation and Selection Queued
  18. INT-18 The Measurability Boundary of Information Quality: What 1000 Calls Can and Cannot Measure Queued
  19. INT-19 An Attribution Analysis of LLM Hallucination: Separating Contamination Propagation from True Hallucination Queued
  20. INT-20 The Pretraining of Intelligence: Genes, Culture, and Large Models: Why Does a Spider Know How to Weave a Web at Birth, While a Human Needs Twenty Years of Learning to Become Capable? Queued

Full texts are in Chinese; titles are translated.

Read the full theory in Chinese →

All papers →