The projected widespread implementation of AI agents for comprehensive automation, initially anticipated for 2025, has experienced delays, prompting discussions regarding the long-term viability of such a vision.
A paper titled "Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models" indicates that large language models (LLMs) may have inherent limitations in executing complex computational and agentic tasks. Vishal Sikka, a former SAP CTO and Infosys CEO, commented that AI models are unlikely to achieve the reliability required for critical operations like managing nuclear power plants. He suggested they could assist with simpler tasks, acknowledging potential for errors.
In contrast, the AI industry has presented different perspectives. Notable advancements have been reported in AI coding applications. Demis Hassabis, Google's head of AI, reported recent progress in minimizing AI hallucinations.
A startup, Harmonic, co-founded by Vlad Tenev and Tudor Achim, has announced developments in AI coding through its product, Aristotle. Harmonic's methodology utilizes formal methods of mathematical reasoning to verify LLM outputs, specifically by encoding them in the Lean programming language to enhance trustworthiness. While Harmonic's primary focus is on 'mathematical superintelligence' and coding, excluding subjective tasks such as history essays, Achim suggests that current models possess sufficient intelligence for tasks like booking travel itineraries.
Despite the varied opinions on current capabilities, there is a consensus that hallucinations remain a persistent challenge. OpenAI scientists confirmed in a September paper that hallucinations continue to affect the field, stating that "accuracy will never reach 100 percent." They cited instances where AI models, including ChatGPT, generated incorrect information for author dissertation titles and publication years.