Research map / Large language models / AI and society

The AI Race: Why Current Neural Network-based Architectures are a Poor Basis for Artificial General Intelligence

AAAI 2025 · Large language models › AI and society · approach: Social-science view of AI evaluation · 13 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position/review paper arguing that neural-network-based architectures (games AIs, generative art/deepfake AIs, LLM assistants) are poor foundations for AGI due to fixed interfaces, objective-function/gradient limits, and data-hungry learning, and that near-term risks lie in misuse rather than AGI.

Problem

Whether current deep neural network-based AI systems, amid a rapid AI race, are on a path toward artificial general intelligence, and what the real risks of current AI are.

Approach

Argues from architecture rather than benchmarks: fixed input/output interfaces, objective-function and gradient-based learning, and a geometric model of learning (function L) imply that stacked self-training or adversarial training cannot produce an intelligence explosion, so deep learning systems are poor candidates for AGI.

Key results

• Presents a comparison table of Deep Blue, AlphaGo, AlphaGo Zero and AlphaStar in terms of training games, equivalent human training time and TPU usage.
• Proposes a geometric model L(CX, SX, CA, SA) of learning capability and argues that stacked self-training yields diminishing returns with a learning angle tending to zero.
• Concludes that current AI systems are very far from AGI and that the main dangers are misuse, reliability, legal and skill-erosion risks.

Surprising findings

• Deep Blue is characterized as a brute-force search system rather than true AI, and the paper argues the assistant-AI reasoning layer resembles symbolic AI.

Limitations

• No empirical experiments; argument is conceptual and relies on illustrative geometric figures.
• The geometric learning function L is informal, with few formal properties assumed rather than proven.
• Scope limited to deep learning and three AI categories, with the claim that all future AI will remain deep-learning based not substantiated.

Experimental setup

Builds on: AlphaGo AlphaGo Zero AlphaStar BERT Word2vec variational autoencoder

Closest papers by meaning

• The Alignment Problem from a Deep Learning Perspective (ICLR 2024 · similarity 0.87)
• Position: Stop Making Unscientific AGI Performance Claims (ICML 2024 · similarity 0.86)
• AGI-Elo: How Far Are We From Mastering A Task? (NeurIPS 2025 · similarity 0.85)
• Position: Levels of AGI for Operationalizing Progress on the Path to AGI (ICML 2024 · similarity 0.84)
• Position: Stop treating `AGI' as the north-star goal of AI research (ICML 2025 · similarity 0.84)
• GAIA: a benchmark for General AI Assistants (ICLR 2024 · similarity 0.83)