Intelligence, observed from different disciplines, seems to wear various masks: data compression in statistics and machine learning, universal computation in dynamical systems, and adaptive behavior in agents. Each field pursues its own objective, and two of the most influential drives often fail in mirror image: novelty search, which craves surprise, is transfixed by a noisy television screen; the free-energy principle, which avoids surprise, is most content in a dark room. Both failures share a single cause: they treat as one quantity the surprise a learner can convert into knowledge and the surprise it can never assimilate. Here we show that the learnable part of that information, which we call learnable novelty, yields the seemingly disparate projections of intelligence, and we give a closed-form estimator built on a cheap and differentiable reservoir computer. Used as a measure, with no supervision of any kind, the estimator recovers decades of complexity classification, ranking the Turing-complete rule 110 highest among elementary cellular automata. Used as an objective, its gradient carries a neural cellular automaton from simple dynamics into a regime of solitons—the traveling, colliding structures by which rule 110 computes—and organizes the representation of an image encoder around the ten digit classes of MNIST, fully unsupervised: no label ever enters training. Handed to a reinforcement-learning agent as an intrinsic reward, it supplies the exploration that task rewards lack, improving on the task baseline in nine of ten environments and collapsing in none. Complexity generation, abstraction, and exploration, ordinarily pursued with unrelated objectives in separate fields, thus emerge from ascent on one differentiable quantity, and the projections of intelligence gain a common quantitative footing.
This perspective has profound implications for business technology development. In a world where data grows exponentially, organizations need algorithms that not only identify patterns but distinguish between what can be learned and what is sterile noise. Learnable novelty offers a criterion for designing systems that maximize useful knowledge acquisition without falling into overfitting or stagnation. At Q2BSTUDIO, we apply these principles to creating custom software applications that integrate artificial intelligence, cloud computing, and cybersecurity, ensuring each software component is oriented to extract real value from data. Our teams implement AI models inspired by learnable novelty, learning abstract representations without massive labeling, reducing costs and speeding time to production.
For example, in computer vision, instead of training networks with thousands of labeled images, a system based on learnable novelty can automatically organize internal representations around relevant categories, as happened with MNIST in the reference study. This is especially valuable for companies handling proprietary data or niches where manual labeling is unfeasible. Q2BSTUDIO incorporates these techniques into its AI solutions, enabling clients to discover latent structures in their datasets without human intervention. Furthermore, the ability to generate controlled complexity—like the emergence of solitons in cellular automata—translates into software architectures that can adapt to changing dynamics, ideal for cloud environments such as AWS or Azure. Managing demand spikes, fault tolerance, and scalability benefits from this approach, as systems learn to explore optimal configurations without predefined rules.
The connection with cybersecurity is also natural: a system that can distinguish between truly learnable novel events (like a new attack type) and random noise (false positives) can prioritize alerts and adapt defenses in real time. At Q2BSTUDIO, we develop anomaly detection modules that use learnable novelty estimators to reduce analyst fatigue and improve accuracy. Likewise, in Business Intelligence, the concept enables dashboards that not only show past indicators but automatically flag which data variations merit investigation because they contain learnable information. Our BI / Power BI service integrates these principles to transform raw data into actionable insights, while process automation platforms benefit from AI agents that explore the action space intelligently, guided by learnable novelty rather than superficial rewards.
In the realm of intelligent agents, learnable novelty provides an intrinsic reward that prevents the agent from getting stuck in repetitive behaviors or obsessing over irrelevant stimuli. This is crucial for applications like autonomous robots, virtual assistants, or recommendation systems. Instead of designing complex reward functions, developers can let the agent actively seek what it can learn, improving exploration and thereby final performance. Q2BSTUDIO applies this philosophy in developing custom AI agents, capable of adapting to dynamic environments without constant retraining. Our automation and robotic process projects benefit from this approach, reducing setup time and increasing robustness.
In summary, learnable novelty emerges as a unifying concept that not only explains phenomena across different branches of science but also offers practical guidance for designing intelligent software. From feature abstraction to autonomous exploration, this differentiable metric aligns technological development with the true nature of learning: turning surprise into knowledge. At Q2BSTUDIO, we are committed to bringing these advances to the business world, offering services ranging from AI consulting to cloud infrastructure implementation and cybersecurity, always focused on extracting maximum value from information. Intelligence, at last, finds a common engine: the novelty that can be learned.
