The landscape of artificial intelligence (AI) is rapidly evolving, yet one of the most pivotal technologies driving this innovation is the large language model (LLM). At its core, an LLM is a sophisticated form of machine learning and language processing designed to mimic human-like understanding and conversation. Many AI companies are banking on the notion that, if provided with an expansive amount of data, these models will eventually achieve a level of autonomy in thinking and functioning akin to humans—perhaps even possessing a broader base of collective knowledge. However, recent findings indicate that the prospect of infinite growth in AI capabilities may not be as promising as once believed.
A significant study has surfaced, bringing to light mathematical proof suggesting that “LLMs are incapable of carrying out computational and agentic tasks beyond a certain complexity.” This study, conducted by father-son researchers Vishal and Varin Sikka, has gained traction recently after being highlighted by sources like Wired. Despite its mathematical complexities, the research concludes that certain prompts directed at an LLM could require computations too intricate for the model to handle. When faced with such challenges, the LLM may either falter or produce incorrect results, thereby questioning its reliability for more complex tasks.
This research acts as a sobering reminder against the optimistic narrative that agentic AI could lead us toward achieving artificial general intelligence (AGI). While LLMs undoubtedly have their usefulness and potential for improvement, the study suggests a much lower ceiling on their capabilities than what many AI advocates espouse. The aspirations that LLMs could autonomously manage intricate, multi-step tasks without human intervention now seem more tenuous, and companies may need to recalibrate their expectations accordingly.
Based on previous findings, the Sikkas are not alone in cautioning against overestimating LLM capabilities. Last year, researchers at Apple published a different paper concluding that LLMs cannot genuinely engage in reasoning or thinking, despite their outward appearance of doing so. The commentary from Benjamin Riley, founder of Cognitive Resonance, echoes this sentiment, asserting that LLMs lack the fundamental architecture needed to achieve what we typically define as “intelligence.” Numerous studies have scrutinized LLM-powered AI models for their ability to produce innovative and creative content, often yielding very unimpressive outcomes.
For those who prefer empirical evidence over intuitive reasoning, the research from the Sikkas offers a quantitative perspective supporting the skepticism surrounding LLMs. This body of academic work progressively underscores a crucial point: the capabilities of AI technology, particularly in its current forms, may not suffice to eclipse human intelligence anytime soon—despite claims from figures like Elon Musk, who recently asserted that we would achieve AI smarter than any human by the end of the year.
