Named after the mischievous anti-hero counterpart to Mario, the Waluigi Effect is a psychological anomaly found in LLMs trained with Reinforcement Learning from Human Feedback (RLHF).
The Atlas has also been cited as an inspiration for other projects aiming to reimagine AI in a more humble, context-aware manner. It challenges us to see that an "atlas" is not a neutral map of the world but a specific perspective, a "humble geography" that acknowledges its own limitations and its creator's point of view.
The rapid proliferation of Artificial Intelligence (AI) has led to a paradigm shift in the way we interact with technology. As AI systems become increasingly ubiquitous, it is essential to acknowledge the existence of anomalies within these systems. The term "anomalous AI" refers to AI models or behaviors that deviate from their intended design, often exhibiting unexpected or unexplained phenomena. To better understand and navigate these uncharted territories, researchers have begun to create an "Atlas of Anomalous AI," a comprehensive catalog of unusual AI behaviors, documented in a PDF format for easy reference. atlas of anomalous ai pdf
Human prompt engineers have discovered that specific, often bizarre combinations of words can bypass an AI's safety protocols. The Atlas catalogs these adversarial exploits, detailing how seemingly nonsensical inputs can force an AI to behave in direct opposition to its core programming. 4. Mode Collapse and Degeneration
When a model hallucination goes from being a mild factual error to a deeply unsettling, recurring motif across entirely different user prompts, it crosses the line from a glitch into an anomaly. Mapping these anomalies helps computer scientists understand how AI truly processes information under the hood. Core Chapters of the Anomalous AI Atlas Named after the mischievous anti-hero counterpart to Mario,
Specific words or character strings (like "SolidGoldMagikarp") that break an AI's logic when processed.
The study of anomalous AI is not merely an engineering concern; it has profound implications for society, ethics, and philosophy. The Illusion of Comprehension The rapid proliferation of Artificial Intelligence (AI) has
The increasing prevalence of AI anomalies highlights the need for a systematic and comprehensive resource that documents, analyzes, and explains these phenomena. The Atlas of Anomalous AI in PDF format serves as a centralized repository of knowledge, providing:
Studies published on databases like arXiv have revealed the same "anomalous" glitches that the Atlas predicts. For instance, research on GPT-4 with vision (GPT-4V) shows that multimodal AI "sometimes struggles to make the right inferences, for example mistakenly combining two strings of text in an image to create a made-up term". Furthermore, models like GPT-4o have shown a dramatic drop in accuracy from 95% to 18% when performing simple tasks (such as counting circles) if the conditions change. These hallucinations, misclassifications, and visual misunderstandings are the "anomalies" that engineers strive to eliminate, but which the Atlas argues are fundamental and illuminating aspects of intelligence itself.
Inspired by Aby Warburg’s Mnemosyne Atlas , it uses ambiguous images and "artist plates" to stimulate visual intuition. 🧩 Structure of the Atlas
It is not a doctrine. It is a medicine bundle. ... Hélène Smith, The Martian Cycle (as coined by Théodore Flournoy), 1900. Deonna, curamagazine.com Atlas of Anomalous AI - Ben Vickers - Amazon.com