NVIDIA is making strides towards the next evolution of Artificial Intelligence, moving beyond merely training massive models with human data or optimizing inference. The company has joined forces with Ineffable Intelligence, an AI lab founded in London by David Silver, a key figure behind AlphaGo. Their collaboration focuses on building large-scale reinforcement learning infrastructure for what they term “Superlearners.”
The core concept behind Superlearners involves AI systems that continuously learn through experience. This means models that actively engage, observe the outcomes, receive feedback, adjust their behavior, and try again.
NVIDIA and Ineffable Intelligence Partner to Create Superlearners, AI That Learns Like a Human Through Independent Trial and Error
This approach fundamentally differs from traditional pre-training. In conventional training, models are fed a fixed dataset of human-generated content – text, images, code, or video – and learn patterns from this static material.
In contrast, reinforcement learning generates its own data dynamically. The AI doesn’t just process existing information; it learns by interacting with a task, simulation, or specific environment, regardless of its complexity.
Jensen Huang describes this as “the next frontier of AI.” Following his previous analogy of GPUs being like fine wine, he now states that Superlearners will be systems that learn continuously from their own experiences, not from the experiences of others. Therefore, NVIDIA and Ineffable are co-designing the necessary infrastructure to execute large-scale reinforcement learning for these Superlearner systems.
It Must Do Everything Itself, Like a Human, Without Needing Anyone Else’s Learning
Technically, these systems require a highly optimized training pipeline. The cycle isn’t as simple as feeding data into a GPU. The Superlearner AI must constantly act, observe, score, and update itself in tight, iterative loops. This places significant demands on interconnection, memory bandwidth, and serving—the process of delivering models and responses during training. In essence, AI hardware needs to advance significantly to usher in this new era of AI.
Consequently, development will commence on NVIDIA Grace Blackwell systems and will also be among the first to explore the future NVIDIA Vera Rubin platform. While these may not be the ultimate solutions, the company is already looking ahead to “Feynman” (referring to the hypothetical next-generation hardware).
The objective is to understand the hardware and software requirements for AI as it transitions away from heavy reliance on human data towards learning through simulation and experience. This means the current AI boom is not a bubble but is accelerating towards becoming the new norm in the AI market with the advent of Superlearners.
David Silver articulates the shift directly: researchers have made substantial progress in creating systems that understand what humans already know. However, “the next challenge is to build systems capable of discovering new knowledge independently.” This necessitates a different approach and substantial hardware investment.
NVIDIA aims to be at the forefront of this development, not just as a chip provider but as a designer of the infrastructure that will power this new generation of AI. This AI will be driven by trial, error, experience, and continuous learning, embodied by the Superlearners.
