The field of Artificial Intelligence, like many disciplines, draws heavily on nature. The power behind the language models (LLMs) that are now deeply woven into everyday tasks lies in Deep Learning, a principle based on the neuroplasticity of the brain. Neural networks are built from adaptable connections, an idea that mimics neuroplasticity following the principle of Hebbian learning. In our brains, neurons form new (or stronger) connections with each other when they are repeatedly active at the same time: “Neurons that fire together, wire together”. This is the basis for how we experience things and learn concepts.
Historically, imitating nature, also known as biomimetics, has often driven scientific progress. In Artificial Intelligence that is no surprise, since the human brain may well be the most complex biological system we know. As neurobiology reveals more about how we learn, remember, and reason, computer science finds new starting points for mimicking those mechanisms.
In 2017, scientists at Google published the groundbreaking paper ‘Attention is all you need’, which introduced the transformer model architecture. Where earlier Natural Language Processing (NLP) systems struggled to model long-range dependencies in language, transformers performed like nothing before them. Through the so-called attention mechanism, a mathematical technique loosely inspired by the linguistic concept of attention, models could dynamically give more weight to the most relevant parts of a text. Most language model architectures are built on this innovation (the GPT in ChatGPT, for example, stands for Generative Pre-trained Transformer).
Still, people are increasingly realizing that these language models are not a magic wand that solves every problem in the blink of an eye, even though big tech companies would love to sell them that way. In fact, over the past three years (ChatGPT came out in 2022), the limitations of these systems have become painfully clear. Above all, the phenomenon of ‘hallucinating’ models (another anthropomorphic term) breeds distrust in what these systems can do, by generating answers that sound convincing but are factually wrong.
Groundbreaking as it is, this architecture hides a fundamental limitation. That limitation is again best described with a term from neuropathology: anterograde amnesia. Just as a patient with this condition can no longer form new long-term memories once the disease sets in, the fundamental knowledge of leading language models stays largely static after training. The attention mechanism works brilliantly in the ‘now’ (within the context window, as they call it), but no structural change takes place in the neural connections based on new input. Once the so-called Hebbian training is over, Hebb’s principle goes out the window. It shows that today's LLMs so far simply imitate our brains, while still lacking the biological ability to integrate experiences.
Funnily enough, it is Google's scientists who seem to have found a solution to this memory problem, once again with a neurobiological source of inspiration. This time they focused not on the static connections between neurons, but on the rhythm at which they communicate: brain waves. Our brains do not run at a single tempo the way artificial neural networks do. Instead, there is a spectrum of different waves, each operating at its own speed. Fast Gamma waves, with their high frequency, let us respond instantly to what is happening right now, while slow Delta waves process and store information over longer periods. This biological rhythm has become the blueprint for a new concept: ‘Nested Learning’. You can picture this system as a set of gears: small, fast-spinning cogs process the fleeting context of a conversation, while larger, slower wheels turn along in the background to record the essence in deep memory, slowly but permanently."Diagram comparing biological brain waves and neuroplasticity to the uniform structure and multi-frequency updates used in Nested Learning models" - Google, 2025.
The concept of Nested Learning is proposed in an architecture called HOPE. And guess what? HOPE beats traditional transformer models on language modeling and reasoning tasks, especially in situations that require continual learning. Hope, then, is a model that is not ‘finished’ after training. It keeps tinkering with its own internal wiring as it reads new text: a much finer imitation of neuroplasticity. That makes it not a static encyclopedia, but truly a living organism that adapts to its environment.
With HOPE, the picture seems to be getting more complete: by mimicking not only the neurons but also the rhythm of the brain, we take a further step toward real artificial intelligence. Concretely, for users of language models this means a virtual assistant can now truly learn from interacting with its user and keep adapting to their style. Providers of these models have been selling that promise for a while, but until now it only happened in fairly roundabout ways.
Biomimetics thus seems to bring us closer to a functional imitation of the human brain. At the same time, there is a sense that reaching human capacity is not the end goal: superintelligence aims to surpass humans. Not imitatio, but aemulatio. I wonder whether that is realistic, when most of our progress comes from imitation. Maybe it is already impressive enough that, inspired by the still largely mysterious and poorly understood brains we carry ourselves, we build systems that approach only a fraction of their capacity. For me, that is reason enough to keep refining artificial intelligence while also pausing more often to take care of our own brains.
Tags: