The quest for artificial intelligence has long been characterized by a shifting battleground of cognitive theories, structural architectures, and terminology. Over the past several decades, two primary theoretical frameworks have competed to define how machines might achieve human-level cognitive faculties: connectionism (the foundation of neural networks) and computationalism (often referred to as symbolic AI or classicism). As connectionist models evolved, they merged with a modern ideology known as "expansionism" or the "scaling law hypothesis"—the belief that continuously increasing compute, parameter count, and data volume will eventually yield emergent, generalized intelligence.
This article provides an objective, academically grounded assessment of the theoretical evolution of AI, historical terminologies, and the current debate surrounding whether Artificial General Intelligence (AGI) can truly be achieved via contemporary autoregressive "next-token prediction" methods, or if the current AGI narrative is a sociotechnical fiction designed to mobilize capital.
The War of Paradigms: Connectionism vs. Computationalism
For the first four decades of artificial intelligence research, the dominant paradigm was computationalism (also called the symbolic approach or Good Old-Fashioned AI - GOFAI). Rooted in the physical symbol system hypothesis, computationalism posits that human cognition is a form of digital computation where meaningless symbolic tokens are manipulated according to explicit syntactic rules (Davenport, 2013). This architecture excels at structured, logical reasoning, mathematical calculations, and formal language rules, but it struggles with noisy, real-world sensory inputs such as computer vision and speech recognition.
Its historical rival, connectionism, rejects rigid symbol manipulation in favor of parallel, distributed processing (PDP) modeled roughly on biological neural networks. In connectionist networks, representations are not discrete symbols; rather, they are represented as activation patterns across highly interconnected, virtual "neurons" (Can, 2013). While connectionism was historically criticized for lacking logical compositionality and its inability to perform simple algebraic generalizations (such as the notorious XOR problem highlighted by Minsky and Papert in 1969), the advent of deep learning and multi-layered neural networks in the 21st century dramatically reversed the paradigm's fortunes (Zhang, 2023).
"The connectionist design of artificial neural networks cannot be comparable to biological neural networks. Furthermore, according to the definition of strong AI, the ANNs can't reach this aim genuinely because of the absence of a logical structure."
— (Zhang, 2023)
Ultimately, connectionism won the battle for contemporary market dominance. By combining vast amounts of parallel processing power with backpropagation algorithms and Web-scale training data, connectionist neural networks demonstrated unprecedented empirical success in areas where symbolic systems had failed for decades: pattern recognition, natural language generation, and complex games. However, as critics note, the victory of connectionist neural networks remains incomplete, as modern architectures still inherit fundamental vulnerabilities related to reasoning, robustness, and semantic grounding.
The Taxonomy of Machine Intelligence
As these paradigms shifted, the vocabulary used to describe artificial intelligence evolved, often reflecting the commercial and academic incentives of each era. Today, several terms are deployed interchangeably in public discourse, though they carry distinct technical definitions in scholarly literature:
- Narrow AI (Weak AI): Systems designed, trained, and optimized to execute a specific, highly bounded cognitive task—such as speech-to-text transcription, medical image classification, or playing chess. They cannot transfer their capabilities to novel domains.
- Connectionist AI / Deep Learning: A subset of machine learning based on multi-layered artificial neural networks. These systems learn representations directly from raw data without manual feature engineering.
- Generative AI: Deep learning models trained specifically to predict and generate new content (text, image, audio, or video) based on patterns extracted from training data.
- Artificial General Intelligence (AGI) / Strong AI: A theoretical machine capable of performing most economically valuable, intellectually demanding tasks at or above human-level proficiency across a highly diverse range of domains (Sarma, 2026).
- Superintelligence: A hypothetical form of AI that exceeds human cognitive capabilities in virtually all domains, including scientific creativity, social skills, and general wisdom (Armitage, 2026).
Expansionism and the Next-Token Prediction Paradigm
The contemporary deep learning landscape is dominated by autoregressive Large Language Models (LLMs). These systems are trained via a self-supervised formulation known as Next-Token Prediction (NTP), where the network minimizes cross-entropy loss by predicting the most statistically likely subsequent token given a sequence of preceding tokens (Ricciardi Celsi, 2026). Underpinning this method is "expansionism"—the scaling law hypothesis popularized by Kaplan et al. (2020), which demonstrates that cross-entropy loss decreases predictably and smoothly as a power-law function of compute budget, dataset size, and parameter count.
Proponents of expansionism argue that this quantitative reduction in cross-entropy loss drives the qualitative "emergence" of advanced cognitive behaviors. However, as models scale toward astronomical parameter limits, severe technical bottlenecks have emerged. The table below details the core structural limitations of the next-token prediction paradigm as identified in recent computer science and cognitive science literature.
Table 1: Structural Mismatches in Autoregressive Next-Token Prediction
| Technical Limitation | Underlying Mechanism | Cognitive Consequence | Academic Source |
|---|---|---|---|
| Exposure Bias | Autoregressive decoders are trained on gold-standard human sequences but generate iteratively during inference; early errors accumulate over long horizons. | Systemic brittleness, loss of coherence, and rapid decay in multi-step planning. | Ricciardi Celsi, 2026 |
| Myopic Objectives | Maximum Likelihood Estimation (MLE) optimizes solely for local next-token probability rather than long-term, global task success under constraints. | Absence of global look-ahead, planning, and structured logical deduction. | Ricciardi Celsi, 2026 |
| The Grounding Problem | Models learn statistical correlations between lexical "squiggles" (syntax) without physical, spatial, or causal interaction with the real world (semantics). | Plausible-sounding but factually detached hallucinations; a total lack of genuine semantic understanding. | Davenport, 2013; Harnad, 1990 |
| Diminishing Returns | Performance gains for real-world tasks (such as single-message persuasion or reasoning) exhibit sharp logarithmic decay relative to exponential parameter scaling. | The physical and financial limits of hardware scale yield progressively marginal utility gains. | Hackenburg, 2024 |
Is AGI via "Autocomplete" Truly Possible, or is it Deep Hype?
A rigorous evaluation of recent scholarly literature suggests a growing consensus: AGI cannot be achieved solely through the next-token prediction "autocomplete" paradigm. While LLMs exhibit remarkable linguistic fluency and broad semantic retrieval, they do not possess genuine reasoning, causal world models, or autonomous error-correction mechanisms. Instead, current autoregressive networks behave as sophisticated statistical compressors of human culture (An et al., 2024). They excel at recreating past linguistic patterns but struggle to generate novel physical or logical reasoning structures outside their training distributions.
To bypass these limitations, contemporary commercial laboratories rely on hybrid architectures. These pair autoregressive LLM cores with external search algorithms, Monte Carlo Tree Search, verifiers, and reinforcement learning interfaces. However, these programmatic frameworks are wrapped around the core model rather than integrated into its fundamental learning objective. This mismatch raises serious doubts about whether scaling "autocomplete" can lead to general, self-improving intelligence.
"AGI deep hype is the result of venture capital logic that converges technological and financial speculation in its functioning. Operativised through a myriad of uncertainties, the sociotechnical fiction of AGI establishes a recursive relation to deep hype: its hype is sustained by a number of uncertainties that are tamed and excited by sociotechnical fictions..."
— (Gonçalves, 2025)
Consequently, many critical scholars categorize the current narrative of imminent AGI as "deep hype"—a strategic, long-term overpromissory dynamic designed to mobilize capital and secure regulatory capture (Gonçalves, 2025). By projecting grandiose, semi-fictional scenarios of civilizational transformation or catastrophic existential risk into the near future, private technology laboratories successfully accomplish several objectives simultaneously:
- Capital Mobilization: Building and training frontier models requires capital-intensive infrastructure (thousands of high-end GPUs, massive energy grids, and private data acquisitions). Cultivating an existential narrative of AGI as an inevitable "space race" justifies the billions of dollars flowing from venture capitalists and public markets into private firms.
- Regulatory Moats: Promising that superintelligent, potentially dangerous systems are just around the corner encourages governments to draft strict regulatory licensing requirements. These legal frameworks effectively prevent open-source developers and academic competitors from training large models, creating a high barrier to entry for smaller firms.
- Displacement of Present-Day Accountabilities: Focusing the public and political gaze on speculative, sci-fi existential risks (such as autonomous AGI superweapons) effectively crowds out democratic scrutiny regarding immediate harms, such as systemic algorithmic bias, data privacy violations, intellectual property theft, and carbon-intensive compute infrastructure (Oldenburg & Papyshev, 2025).
Beyond Statistical Mimicry: The Technological Horizon
Ultimately, modern deep learning represents an immense engineering feat with genuine, transformative utility for specialized automation. However, conflating next-token prediction with general intelligence is a fundamental category error. AGI is not merely scaled-up autocomplete. True cognitive generalization requires active interaction with a physical or structured environment, continuous online learning, causal reasoning, and symbol grounding—none of which are optimized by minimizing cross-entropy loss over static text corpora. Until the AI research paradigm shifts from purely autoregressive text prediction toward grounded, model-based active systems, AGI will likely remain a highly lucrative, socially influential sociotechnical fiction.
References
In-Text Footnotes:
An, H., Song, Y., & Li, X. (2024). Physics in Next-token Prediction. arXiv. https://doi.org/10.48550/arxiv.2411.00660
Armitage, R. C. (2026). Artificial General Intelligence and Its Threat to Public Health. PMC - NIH. PMC12415933.
Can, W. N. (2013). Stevan Harnad's target article. Archipel UQAM.
Davenport, D. (2013). The Two (Computational) Faces of AI. Studies in Applied Philosophy, Epistemology and Rational Ethics, vol. 1, pp. 43-58. https://doi.org/10.1007/978-3-642-31674-6_4
Gonçalves, A. B. (2025). Deep Hype in Artificial General Intelligence: Uncertainty, Sociotechnical Fictions and the Governance of AI Futures. arXiv. https://doi.org/10.48550/arxiv.2508.1974
Hackenburg, K. (2024). Scaling language model size yields diminishing returns for single-message political persuasion. Proceedings of the National Academy of Sciences (PNAS)
Oldenburg, N., & Papyshev, G. (2025). The Stories We Govern By: AI, Risk, and the Power of Imaginaries. arXiv. https://doi.org/10.48550/arxiv.2508.11729
Ricciardi Celsi, L. (2026). Beyond Next-Token Prediction: A Standards-Aligned Survey of Autoregressive LLM Failure Modes, Deployment Patterns, and the Potential Role of World Models. MDPI Electronics, 15(5), 966.
Sarma, G. P. (2026). Artificial General Intelligence Forecasting and Scenario Analysis: State of the Field, Methodological Gaps, and Strategic Implications. RAND Corporation.
Zhang, Y. (2023). A Historical Interaction between Artificial Intelligence and Philosophy. Teorie vědy / Theory of Science, 1(1). https://doi.org/10.46938/tv.2023.579
