Executive Summary
Research on large language models (LLMs) has uncovered a phenomenon that defies conventional theoretical expectations: the abrupt appearance of sophisticated cognitive capabilities — such as multi-step reasoning, deep contextual understanding, and functional code generation — when models cross certain computational scale thresholds. This phenomenon, termed emergent abilities by Wei et al. (2022), raises fundamental questions that transcend disciplinary boundaries.
1. Theoretical Framework: Simulation Hypothesis and Digital Physics
1.1 Bostrom’s Simulation Trilemma
Nick Bostrom, philosopher at the University of Oxford, formulated his celebrated simulation argument in 2003. Structured as a trilemma with three mutually exhaustive propositions: (1) it is unlikely that civilizations similar to humans will reach the computational capacity to simulate realities with a level of detail that includes conscious minds; (2) if such civilizations exist, they probably would create many simulations, so the number of simulated minds would significantly exceed non-simulated minds; (3) we, with high probability, are one of those simulated minds. At least one of the three must be false (Bostrom, 2003).
1.2 Digital Physics: Zuse, Wheeler, Fredkin, and Lloyd
Konrad Zuse proposed in 1969 that the universe itself could be a giant computational system. John Archibald Wheeler developed the phrase «It from Bit» — every physical entity emerges from binary information. Edward Fredkin developed the Finite Nature Hypothesis, arguing that all physical evolution is a computational process. Seth Lloyd proposed that the universe is a giant quantum computer, with a computational capacity of approximately 1090 bits of information (Lloyd, 2006).
1.3 Stephen Wolfram and Computational Irreducibility
Stephen Wolfram introduced the concept of computational irreducibility: some computational systems are irreducible — there is no way to predict their behavior without simulating each step of the computation. His work on cellular automata demonstrates how simple rules produce extraordinary emergent complexity (Wolfram, 2002).
1.4 Max Tegmark’s Mathematical Universe Hypothesis (MUH)
Max Tegmark proposed that «Our external physical reality is a mathematical structure.» This does not merely mean the universe can be described by mathematics — the universe literally is a mathematical structure. Tegmark proposes a four-level taxonomy of multiverses, culminating in Level IV where all mathematical structures physically exist (Tegmark, 2014).
2. Philosophical Framework: Philosophy of Mind and Emergence
2.1 Searle’s Chinese Room Argument
John Searle proposed the celebrated thought experiment: a person who does not understand Chinese is enclosed in a room with rules correlating Chinese input symbols with Chinese output symbols. The system can pass the Turing test without understanding Chinese. Searle uses this argument to refute functionalism and propose biological naturalism: mental states are biological features of the brain, causal properties emerging from neural processes (Searle, 1984).
2.2 Daniel Dennett: The Multiple Drafts Model
In Consciousness Explained (1991), Dennett argued that the common intuition about consciousness — the «Cartesian theater» image — is deeply misleading. He proposed the Multiple Drafts model: there is no single privileged moment of conscious processing. The brain constantly generates multiple «drafts» of experience that compete, with fragments being continuously edited and rewritten.
2.3 David Chalmers: The Hard Problem of Consciousness
Chalmers articulated the «hard problem» of consciousness: how the brain processes information, responds to stimuli, integrates information, and controls behavior are difficult but in principle scientifically tractable problems. The hard problem asks why and how conscious experience feels a particular way — why there is something it is like to be a conscious organism (Chalmers, 1996).
2.4 Karl Friston: The Free Energy Principle
Karl Friston developed the Free Energy Principle (FEP), a unifying theoretical framework seeking to explain the self-organization of complex biological systems, including the mind. The FEP is a variational principle: any system that resists dispersion (entropy) must minimize surprise, defined as the negative log-probability of sensory states given the system’s internal models (Friston, 2010, 2019).
2.5 Giulio Tononi: Integrated Information Theory
Giulio Tononi created Integrated Information Theory (IIT), a mathematically rigorous approach to quantifying consciousness. IIT starts from the axiom: consciousness is identical to integrated information. Tononi formalizes this through the quantity phi (Φ), which measures the integrated information of a system (Tononi, 2004).
3. Mathematical Foundations of Complexity
3.1 Chaos Theory
Chaos theory studies nonlinear dynamical systems exhibiting extreme sensitivity to initial conditions, preventing long-term prediction despite underlying determinism. A chaotic system is characterized by: determinism, sensitivity to initial conditions, nonlinearity, and strange attractors (trajectories converging toward fractal structures in phase space). The Lyapunov exponent λ measures the rate of separation of nearby trajectories — if λ > 0, the system exhibits chaos (Strogatz, 2018).
3.2 Phase Transitions and Self-Organized Criticality
Criticality occurs at the transition point between ordered and disordered phases. Phase transitions are classified by order: first order (discontinuity in free energy) and second order (continuity in free energy but discontinuity in second derivatives). Self-Organized Criticality (SOC), proposed by Bak, Tang, and Wiesenfeld (1988), holds that certain systems naturally evolve toward a critical state without fine-tuning of parameters.
3.3 Geometry of Latent Spaces
The manifold hypothesis indicates that high-dimensional data lie on low-dimensional manifolds embedded in ℝD. Ricci curvature and the Riemann tensor characterize the geometry of an LLM’s representation space. The variation of the metric across space may influence how new capabilities emerge (Lee, 2018).
4. Empirical Evidence: Emergent Abilities in LLMs
4.1 Wei et al. (2022): Emergent Abilities
Wei et al. established the vocabulary and methodological framework for studying emergence in LLMs. Their formal definition requires an emergent ability to simultaneously satisfy: (1) not present in small models — not detectable in models with fewer parameters, less training data, or less compute; (2) robustly present in large models — appears consistently once the model crosses a scale threshold. Wei et al. documented more than 40 tasks where emergence is observed, including chain-of-thought prompting (~100B parameters), modular arithmetic (~10B parameters), and reasoning about intentional states (~100B parameters) (Wei et al., 2022).
4.2 Scaling Laws: Kaplan et al. and Chinchilla
Kaplan et al. (2020) empirically demonstrated that language model cross-entropy loss scales as a power law in three factors: number of parameters (N), training corpus size (D), and compute (C). Hoffmann et al. (2022) refined the scaling laws with their Chinchilla work, showing that for a given amount of compute, it was more efficient to train smaller models on more data than larger models on less data.
4.3 Mechanistic Interpretability
Mechanistic interpretability is a research program seeking to understand how neural networks implement specific computations through analysis of internal circuits. Olsson et al. (2022) showed that during transformer training, a circuit emerges spontaneously — composed of two attention layers implementing a pattern-completion algorithm: induction heads. Bricken et al. (2023) demonstrated via sparse autoencoders that monosemantic features responding to specific concepts can be identified (Bricken et al., 2023).
4.4 The Schaeffer vs. Wei Debate: Real Emergence or Methodological Mirage?
Schaeffer, Miranda, and Koyejo (2023) argued that apparent «emergent abilities» are artifacts of discontinuous metric choice. When continuous metrics like log-probability or Brier score are used, model performance scales gradually and predictably. However, Wei et al. and others responded that: (1) metric choice is not arbitrary — binary accuracy is relevant when the task is genuinely discrete; (2) Afonin et al. (2025) show emergence also occurs in undesired behaviors (misalignment), which would be hard to explain if all emergence were a methodological mirage; (3) some capabilities are genuinely discrete in nature (Schaeffer et al., 2023; Afonin et al., 2025).
4.5 Phase Transitions in LLMs
Arnold et al. (2024) and Nakaishi et al. (2024) experimentally demonstrated that LLMs exhibit critical phase transitions with critical exponents and power-law correlation decay analogous to physical systems. The phase transition analogy suggests that, just as water boils at 100°C, capabilities in LLMs emerge abruptly upon crossing scale thresholds.
5. Ilya Sutskever’s Position on AI Safety
Ilya Sutskever (1986–), influential deep learning researcher and former chief scientist at OpenAI, made fundamental contributions including Sequence to Sequence Learning with Neural Networks (2014) and AlexNet (2012). His departure from OpenAI on May 14, 2024 generated significant impact. In his farewell message on X, Sutskever stated: «I have concluded that the safety of artificial intelligence requires engaging with challenges that we have never faced before, challenges that are not solved simply by making the models larger.»
This declaration is particularly significant coming from the scientist who had supervised precisely the training of ever larger models. Implicitly, Sutskever recognized that scaling alone does not solve the safety problem, and that the community should pivot toward more fundamental challenges.
6. Synthesis: Connecting Physical, Biological, and Artificial Systems
A deep connection emerges between digital physics and language models. If:
- Universe → Mathematics → Emergent reality (Tegmark and digital physics)
- Language → Matrices → Emergent cognition (LLMs and neural emergence)
Then perhaps both processes are manifestations of the same fundamental phenomenon: emergence of qualitative complexity from underlying computational/mathematical structure. If simple cellular automata produce emergent complexity within a computational system, and if the universe operates by simple rules, then LLMs may illustrate an analogous principle: qualitatively new complexity emerging from simple underlying computational processes within the universe.
This summary was generated from the full research document: «Emergence in Large Language Models: Theoretical, Philosophical, Mathematical, and Empirical Foundations» — 51 verified bibliographic references, May 12, 2026.
Keywords: emergence, large language models, simulation hypothesis, digital physics, philosophy of mind, phase transitions, criticality, mechanistic interpretability, computational consciousness.
📄 Download Full Research Document (PDF):
https://blog.yucas.net/wp-content/uploads/2026/05/documento_investigacion_emergencia_llms-1.pdf
51 verified bibliographic references · May 12, 2026