Introduction
Artificial intelligence has made extraordinary progress in recent years. By training vast amounts of human-generated data and fine-tuning with expert inputs, AI has become capable of performing a wide range of tasks. Large language models are the best example of this leap. They can now write poetry, solve complex physics problems, diagnose medical issues, and even summarize legal documents.
However, while imitating human knowledge has been enough to achieve competence in many areas, this approach alone cannot take us all the way. It is unlikely to deliver superhuman intelligence across the most important and complex topics.
The limits of imitation learning
Imitation learning is when AI systems are trained to mimic human behavior and knowledge by studying large collections of human-generated data, books, code repositories, scientific papers, conversations, images, and more. By learning from examples, the model develops the ability to reproduce human-like outputs, whether it is solving a math problem, writing a poem, or generating computer code.
But imitation has its limits. In domains such as mathematics, coding, and science, AI is reaching the boundaries of what can be extracted from human data. Most of the high-quality data sources that genuinely improve strong models have already been consumed or soon will be.
The pace of progress driven solely by supervised learning on human data is visibly slowing. More importantly, the most valuable breakthroughs like new theorems, new technologies, or scientific discoveries lie beyond what humans currently know, and therefore cannot be captured in existing datasets.
The possibilities of experimental learning
Experimental learning is when AI systems improve not just by copying human examples, but by actively trying things out, interacting with their environment, making mistakes, and learning from the outcomes. Instead of only relying on static datasets, the system gains knowledge through direct experience. In technical terms, this is closely related to reinforcement learning, a framework where agents learn by taking actions, receiving feedback, and gradually improving through trial and error.
In their paper Welcome to the Era of Experience, Google DeepMind researchers David Silver and Richard Sutton argue that the next leap in AI will not come from larger datasets or models. Instead, it will come from experience.
They describe three eras of AI:
- The simulation era – when agents mastered specific tasks in tightly controlled environments, such as board games or virtual simulations
- The human data era – which gave us today’s models, trained on massive collections of text, images, and videos created by people
- The era of experience – where agents learn not only from us, but also from the world itself, by interacting with it, making decisions, and improving over time
This perspective resonated with us. It does not dismiss the value of human knowledge, but it challenges the assumption that it is sufficient. Human data inevitably reflects our own blind spots and biases. To move beyond imitation and address open-ended, dynamic problems, AI must engage directly with its environment.
When human data isn’t enough
In many of the most demanding AI applications, the real world does not present neatly labelled datasets. Instead, data is messy, incomplete, and constantly changing. In such contexts, systems cannot rely on static inputs. They must learn dynamically, in context, and often under pressure.
The traditional supervised learning model for clean input-output pairs and static evaluation metrics simply does not suffice. What we are now seeing is the shift towards long-horizon learning and reasoning, where AI systems evolve continuously with accumulated experience, memory, and adaptability. These are not merely bigger models, but fundamentally different types of agents.
Grounded rewards beyond human judgement
AI systems are often trained with rewards shaped by human judgement and an expert mark an answer as right or wrong or selects the best among alternatives. Large language models can also act as judges, evaluating outputs against instructions or examples. These approaches help align systems, but they remain limited: an agent cannot discover strategies that lie outside human knowledge or model-trained biases.
A more practical path is to use grounded rewards signals that come directly from the environment. These could be health indicators, exam results, carbon levels, material strength, profits, or even user-reported outcomes like taste, fatigue, or pain.
With rewards tied to real-world consequences, learning becomes richer and more adaptive. Humans still set the direction like “improve my fitness” or “help me learn Spanish” while the system measures progress through grounded signals like resting heart rate, sleep quality, or exam performance, and refines itself with continuous feedback.
The right balance of human judgement, model evaluation, and grounded signals can create AI that is not only more capable, but also more aligned with practical goals.
The risks and the gaps
An autonomous, goal-driven, and experience-based approach also raises critical questions:
- How do we supervise such agents?
- How do we know when they have truly learned something valuable?
- How do we ensure their experiences are safe, meaningful, and transferable?
There is also an infrastructure challenge. Few environments today are designed to support this kind of learning. Most enterprise and research setups are still optimized for static datasets and predictable performance, not for systems that explore, fail, and iterate. Building this foundation is one of the biggest challenges for the years ahead.
Charting the road forward
If we want AI systems to go beyond human capabilities, they must be given the scope to build their own experiences. But this does not mean humans step aside; it means stepping in at the right moments to steer discovery in safe and productive directions. The future of intelligence is not only about replicating what we already know, but also about expanding the boundaries of what is possible together.
At AI Squared, we are building the tools and infrastructure to support this shift, so that AI systems can learn more effectively from experience, with domain expertise embedded where it matters most.
If you are exploring these challenges too, we would be glad to connect with you.
