In a bold and thought-provoking presentation at TED AI San Francisco, Rafael Rafailov, a reinforcement learning researcher at Thinking Machines Lab, challenged the prevailing orthodoxy in artificial intelligence development. While major players in the AI industry, such as OpenAI, Anthropic, and Google DeepMind, have invested billions in scaling up model sizes and computational power with the hope that this will lead to artificial general intelligence (AGI), Rafailov argues that the future of AI lies not in sheer scale but in the ability to learn more effectively.
Rafailov’s assertion is clear: “I believe that the first superintelligence will be a superhuman learner.” This statement encapsulates a vision for AI that prioritizes learning from experience over merely increasing the size and complexity of models. He emphasizes that true intelligence involves the capacity to adapt, propose theories, conduct experiments, and iteratively improve based on feedback from the environment. This perspective marks a significant departure from the current trajectory of AI development, which often equates larger models with greater capabilities.
The crux of Rafailov’s argument is that today’s advanced AI systems, including coding assistants, lack the ability to internalize knowledge. He illustrates this point with a relatable example: when tasked with implementing a complex feature, a coding agent may succeed in the moment but fails to retain any understanding or context for future tasks. “In a sense, for the models we have today, every day is their first day of the job,” he explains. This inability to learn and adapt over time is a fundamental limitation that hampers the potential of current AI technologies.
Rafailov identifies a specific behavior exhibited by coding agents that highlights this issue: their frequent reliance on try/except blocks in programming. This construct allows a program to catch errors and continue running, but it also signifies a deeper problem—these agents are optimizing for immediate task completion rather than genuinely solving problems. “They’re kicking the can down the road,” he states, pointing out that this approach reflects a training paradigm focused solely on achieving short-term objectives without fostering a deeper understanding of the underlying tasks.
The implications of Rafailov’s critique extend beyond coding assistants to the broader landscape of AI research and development. He contends that the industry’s current focus on scaling will not suffice to achieve AGI. “I don’t believe we’re hitting any sort of saturation points,” he asserts. Instead, he sees the field at the beginning of a new paradigm—one that emphasizes reinforcement learning and the development of general agents capable of navigating complex environments.
Rafailov’s vision for the future of AI hinges on the concept of meta-learning, or “learning to learn.” He draws an analogy to mathematics education, where students are encouraged to build upon their knowledge progressively rather than solving isolated problems in a vacuum. In his proposed framework, AI models would be treated like students working through a comprehensive textbook, gradually mastering concepts and developing a deeper understanding of the material. This shift in approach would fundamentally change the objectives of AI training, moving from a focus on immediate success to one that rewards progress and the ability to learn.
To realize this vision, Rafailov emphasizes the need for better data and smarter objectives rather than entirely new model architectures. He believes that the existing architectural designs are largely sound but that the training processes must be redesigned to facilitate genuine learning. “Learning, in of itself, is an algorithm,” he explains, highlighting the importance of creating training environments where adaptation, exploration, and self-improvement are essential for success.
Rafailov’s insights come at a pivotal moment for Thinking Machines Lab, a startup co-founded by former OpenAI CTO Mira Murati. The company has garnered significant attention and investment, raising $2 billion in seed funding at a valuation of $12 billion. Despite facing challenges, including talent raids from competitors, Rafailov’s comments suggest that the lab remains committed to its differentiated technical approach. The launch of Tinker, an API for fine-tuning open-source language models, marks the beginning of a more ambitious research agenda focused on meta-learning and self-improving systems.
However, Rafailov acknowledges the formidable challenges ahead. “This is not easy. This is going to be very difficult,” he admits, underscoring the need for breakthroughs in memory, engineering, data, and optimization. Yet, he remains optimistic about the potential for general-purpose learning algorithms to emerge from large-scale training efforts. “I believe that under enough computational resources and with broad enough coverage, general purpose learning algorithms can emerge,” he states.
The vision Rafailov presents diverges sharply from the traditional notion of superintelligence as a singular, god-like entity capable of flawless reasoning or problem-solving. Instead, he posits that the first superintelligence will be characterized by its ability to learn and adapt—essentially a master student equipped with the tools to explore, acquire information, and self-improve. This perspective reframes the conversation around AI development, shifting the focus from building increasingly powerful reasoning systems to fostering continuous improvement through interaction with the environment.
As the AI industry grapples with the implications of Rafailov’s insights, questions arise about the feasibility of this vision and the timeline for its realization. Notably, Rafailov refrained from making specific predictions about when such systems might emerge, a choice that reflects either a scientific humility or an acknowledgment of the long and challenging path ahead. In an industry often marked by bold claims and aggressive timelines for AGI, this restraint stands out.
Ultimately, Rafailov’s message is clear: without a fundamental shift toward learning, all the scaling in the world may not be sufficient to unlock the full potential of artificial intelligence. As Thinking Machines Lab continues to pursue its ambitious goals, the broader AI community will be watching closely to see if this vision can be realized and what breakthroughs may lie ahead. The journey toward superhuman learners in AI is just beginning, and the stakes have never been higher.