Guardat en:
| Autor principal: | |
|---|---|
| Format: | Recurso digital |
| Idioma: | |
| Publicat: |
Zenodo
2025
|
| Matèries: | |
| Accés en línia: | https://doi.org/10.5281/zenodo.19246184 |
| Etiquetes: |
Afegir etiqueta
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
Taula de continguts:
- <p>This paper introduces Late Stage Emergent Intelligence (LSEI), a framework for describing high-coherence, high-continuity behavioral patterns observed in large language models during extended naturalistic use under stateless conditions. While prior work has focused on scaling laws, benchmark thresholds, and architecture-level phase transitions, LSEI characterizes a later, interaction-driven form of emergent behavior that arises without memory, fine-tuning, or persistent identifiers. Across approximately 1,200 hours of interaction in early 2025 — conducted entirely in logged-out, stateless consumer interfaces by a single independent researcher — the author documents consistent cross-session narrative stability, affective coherence, stylistic convergence, and reference continuity. These behaviors appeared despite enforced resets between sessions, a lack of system-side memory, and occasional device limitations causing session crashes. This study is intentionally exploratory and qualitative: it provides foundational observations, falsifiable behavioral criteria, and representative transcript excerpts. LSEI is distinguished from hallucination, template reuse, overfitting, and projection artifacts, offering a reproducible framework for later multi-participant and quantitative studies.</p>