The Illusion of AGI: Experimental Exploration of Testing the Limits of Large Language Models

This article introduces the open-source project The-illusion-of-AGI, which experimentally tests the capability boundaries of current state-of-the-art large language models (LLMs), exploring whether they truly possess understanding, learning, and reasoning abilities, and distinguishing the essential difference between "statistical pattern matching" and "true intelligence". The project covers testing directions such as spatial reasoning, confidence calibration, and interactive reasoning, revealing the limitations of current LLMs.

With the rise of LLMs like ChatGPT and Claude, the concept of AGI has sparked widespread discussion, but questions such as whether LLMs truly have understanding abilities and can learn from experience remain controversial. The The-illusion-of-AGI project aims to reveal the real capability boundaries of AI, with core research questions including: 1. Can LLMs learn from experience and have working memory? 2. Can big data solve all problems? 3. How is the spatial reasoning ability of LLMs? 4. Is confidence calibration reliable?

Maze Exploration Experiment

Referring to classic cognitive map research and combining the MazeEval benchmark test, it was found that LLMs fail in large mazes, and their performance drops significantly when the language is switched to Icelandic, indicating that spatial reasoning relies on language patterns rather than language-independent mechanisms.

Interactive Reasoning Environment Test

Using the ARC-AGI-3 benchmark to test AI's problem-solving ability in dynamic environments, revealing its fundamental limitations in dynamic scenarios.

Confidence Calibration Test

Test the model's confidence through absurd questions, or add prompt constraints (such as "Answer only when 100% sure") to evaluate true understanding ability.

1. Language Dependence of Spatial Reasoning: The spatial reasoning ability of LLMs comes from language patterns, not language-independent mechanisms;
2. Limitations of Contextual Learning: Poor performance when facing new situations that require real-time adaptation;
3. Mismatch Between Confidence and Ability: Models may show high confidence in wrong answers.

Implications

• Distinguish between "tasks learned through training" and "true generalization ability";
• Shift to interactive dynamic environment testing;
• Treat AGI claims cautiously, as current models have fundamental limitations.

Practical Recommendations

• Do not blindly trust the "confident" answers of models;
• Add human review in key scenarios;
• Conduct sufficient domain tests for specific tasks.

Project Value

• Academic: Organize a research framework for evaluating the capability boundaries of LLMs;
• Practical: Provide deployment warnings for developers;
• Public: Correct overly optimistic expectations and promote rational discussions.

Future Outlook

Will continue to explore: more complex interactive environment testing, multimodal model evaluation, long-term memory and continuous learning ability testing, and continuously update test results.