In recent months, there’s been a wave of disappointment in the realm of large language models (LLMs) and their journey toward artificial general intelligence (AGI). What started as high expectations has crumbled, revealing key limitations in these technologies.
In June 2025, a landmark paper from Apple highlighted that, despite advances in reasoning, LLMs struggle with distribution shifts. This issue has long been a weakness in neural networks, as noted by Gary Marcus, who has been discussing it for nearly three decades. Other studies, including one by ASU, echoed these findings.
By August, GPT-5 was anticipated but arrived late, failing to impress many experts. Fast forward to September, where Turing Award winner Rich Sutton echoed Marcus’s critiques of LLMs, marking a significant moment in AI discussions.
In October, Andrej Karpathy, a prominent figure in machine learning, stated that AGI is still a decade away. His insights reinforced doubts surrounding the capabilities of current LLMs.
Furthermore, Sir Demis Hassabis, the CEO of DeepMind, recently countered overly optimistic claims about LLMs’ mathematical abilities. This only underscores the gap between public perception and the actual performance of these systems.
Statistics from AI research communities suggest that only 30% of AI practitioners believe we will achieve AGI in the next five years, compared to 70% just two years ago. This shift indicates a growing skepticism about the pace of advancements in AI.
While LLMs are valuable tools, expecting them to lead us directly to AGI is misguided. As we discuss the future, it’s crucial to remain grounded in the actual capabilities and limitations of these technologies.
For those interested in a more in-depth analysis, Gary Marcus offers insights in his 2020 paper, “Next Decade in AI,” which presents a more realistic view of future advancements. Looking at the evolving landscape of AI, it’s clear that we need to balance hope with practical understanding.
