New Perspective on AGI Research: Limitations and Reflections on the Scaling of Large Language Models

On June 5, 2026, ABXLab released the research paper artificial-general-intelligence-research on GitHub, which explores the fundamental limitations of large language models (LLMs) in the path toward artificial general intelligence (AGI). Key points include: diminishing marginal returns in LLM scaling, the so-called "emergent abilities" may be an illusion caused by evaluation metrics, the pattern-matching nature of the Transformer architecture leads to limitations in reasoning ability; and it calls for a shift from "scale-first" to "structure-first" to explore new paths for AGI.

In recent years, LLMs have developed rapidly (e.g., from GPT-3 to GPT-4, Llama, Claude) and have shown amazing performance in tasks such as text generation and code writing. The industry generally believes that continuing to expand model size and training data will lead to AGI. However, through systematic theoretical analysis and experimental verification, ABXLab raises a question: Is scaling the right path to AGI?

Diminishing Marginal Returns in Scaling

The performance improvement brought by model size growth shows diminishing marginal returns: the increase from 10B to 100B parameters is significant, the improvement narrows from 100B to 1T, while the computing cost grows exponentially.

Re-examining Emergent Abilities

The so-called "emergent abilities" may not be true emergence, but an illusion caused by non-linear changes in evaluation metrics.

Fundamental Limitations in Reasoning

The Transformer is essentially a pattern-matching system, [incomplete] with human causal reasoning

ABXLab's research paper delves into the fundamental limitations of large language models in the path toward artificial general intelligence