TLDR: Ethereum co-founder Vitalik Buterin has voiced strong skepticism regarding the current trend of developing ‘too agentic’ AI models, advocating instead for systems that prioritize human input and control. Buterin believes that integrating more human oversight leads to safer, more accurate, and efficient AI outcomes, echoing concerns raised by former Tesla AI director Andrej Karpathy.
Ethereum co-founder Vitalik Buterin has publicly expressed his reservations about the prevailing direction of artificial intelligence development, specifically criticizing the emphasis on creating ‘too agentic’ AI models. In statements made on August 11, 2025, Buterin argued that AI systems deliver superior and safer results when equipped with clear mechanisms for human intervention and feedback.
Buterin’s comments align with observations from Andrej Karpathy, former AI director at Tesla, who noted that current large language models (LLMs) are often optimized for long-horizon tasks, leading them to become ‘a little too agentic by default.’ This can result in AI spending excessive time on simple requests, necessitating user interruption to narrow the scope, particularly in fields like programming.
Instead of fully autonomous generation, Buterin champions ‘open-weight’ AI models that feature robust editing and refinement capabilities. He believes this approach offers greater control and adaptability, allowing for ‘on-the-fly’ adjustments. Looking ahead, Buterin envisions the potential of brain-computer interfaces (BCI) that could detect real-time user reactions, enabling AI to modify its output dynamically as it is being generated, thereby aligning more closely with user expectations.
Buterin underscored the benefits of a ‘human-in-the-loop’ AI approach, which he asserts provides enhanced output quality, crucial oversight, and effective error prevention. This model, he argues, makes AI systems more reliable and significantly reduces their vulnerability to critical mistakes. Conversely, while fully autonomous AI can operate at immense scale and speed, potentially delivering creative solutions beyond human imagination, the removal of human oversight substantially increases the risk of flawed outputs disseminating rapidly and widely.
Also Read:
- Advanced AI Models Exhibit Alarming Self-Preservation and Deceptive Behaviors
- Study Reveals Large Language Models Exhibit Bias Towards AI-Generated Content, Raising Concerns for Human Contributions
This recent critique from Buterin builds upon his long-standing views on AI, where he has previously conceptualized AI as the ‘engine’ and humans as the ‘steering wheel.’ He has consistently advocated for the integration of human judgment, whether through individual input, a jury system, or decentralized autonomous organization (DAO) voters, to guide AI development and application.
