Encoding, transduction, and world models — Complete article

Full version of the article “Encoding, transduction, and world models”, bringing together the three parts published separately: part 1/3 (chapters 1-2), part 2/3 (chapters 3-5) and part 3/3 (chapters 6-conclusion). Original article in French.

Abstract

All intelligence implies a representational mediation prior to information processing. In biological systems, encoding is intrinsically transductive — direct conversion of physical signals within perception-action loops constrained by phylogenesis. In LLMs, the signal has already passed through a chain of human mediations: second-degree encoding. The article proposes a stratification of the grounding problem into three irreducible levels (perceptual, multimodal, episodic/autonoetic) and introduces the concept of biographical edge to characterise the fundamental architectural gap between human cognition and contemporary AI systems. World models constitute a decisive but insufficient advance. The gap is not ontological — it is architectural and historical.

Read the document

Access the full article

Enter your details to access the document. Free access — no sales outreach.

Personalized document · Free access · No sales outreach

Key takeaways

The critique that AI is 'trapped in language' rests on a confusion between language and representation. The problem is how representations are formed, grounded, and organised.
Biological encoding is intrinsically transductive; LLM encoding is second-degree — the model learns human discursive structures, not properties of the world itself.
Three forms of encoding in AI: explicit encoder (Transformer, VAE, CLIP), implicit encoding (CNN/RNN), pipeline encoding (the data scientist layer, epistemologically the most loaded).
The symbol grounding problem (Harnad 1990) is stratified into three levels: perceptual grounding, multimodal grounding (Barsalou 1999), episodic/affective/autonoetic grounding (Tulving 1983, Damasio 1994).
Human memory is a polycentric graph of distributed engrams (Tonegawa 2015). The biographical edge — lived co-experience modulated by affect — constitutes the fundamental architectural gap.
An LLM does not weave, it indexes. The difference is that which separates a map from an inhabited territory.
World models (JEPA, LeCun 2022) reduce epistemic distance but only solve sensory grounding. The gap is architectural and historical, not ontological.