The GenAI world has remained at the churning stage since its inception, notwithstanding the fact that products / models are being launched at an astonishing regularity. As a result there is no clarity about the AI trajectory, the AI progress and what is likely to happen in what timeline. The hype of agentic AI is to prove that it is round the corner and it is out to engulf us all, instantaneously. For sure there is no unison of thought in the AI world. Are we ready to take the Agentic AI plunge?
In a candid chat on The Dwarkesh Podcast, Andrej Karpathy poured a dose of realism on the runaway hype around AI agents. The AI hype is being sold and bought. The former OpenAI and Tesla researcher said the industry is overselling the current generation of autonomous systems, comparing their code outputs to “slop.” According to him, today’s AI agents simply “don’t work” as advertised — they lack the depth, robustness, and self-improving capabilities needed for true autonomy.
Agents hit a wall due to limited general intelligence and inadequate multimodal reasoning. Artificial reinforcement learning, often hailed as AI’s superpower, is actually “terrible” and mostly statistical noise. In Karpathy’s view, genuine intelligence requires long-term adaptive learning and persistent memory, both of which are missing today. He cautioned that the full promise of autonomous AI systems may still be a decade away.
Let sanity be the north star in an exponentially transformative age. Public domain statements should be rational and real. Otherwise, things can get awry, or delayed or get messed up. Nothing magical is going to happen, the focus needs to be on foundational research — refining model intelligence, grounding, and real-world reasoning. Hype has always been a double edged sword. His viewpoint offers a refreshing checkpoint in the “Year of the AI Agent” narrative. Progress doesn’t mean human-level autonomy yet, just better human tools.
ARE WE STILL A DECADE AWAY FROM REALISING THE AGENTIC AI REVOLUTION?