Biological neural network hardware

New Neuromorphic Chips That Change Their Structure

What if the key to creating an artificial intelligence capable of learning like a living being lay not in algorithms or massive data centers, but in the very material that makes up the processors? While current chatbots need to access huge databases and completely lose the context of a conversation once the session ends, a new generation of hardware promises to shift that paradigm.

Researchers around the world are working on chips capable of altering their own physical connections as they accumulate experiences, mimicking one of the human brain's most extraordinary properties. Current systems like GPT, Gemini, Claude, and Grok are extremely powerful, yet they remain essentially software running on massive infrastructure; they rely on billions of parameters stored on servers distributed across the globe and consume absurd amounts of energy to operate.

When a session ends, much of the context vanishes—it is as if the artificial intelligence suffers from a sort of constant digital amnesia. To overcome this limitation, engineers began looking for inspiration in the most efficient architecture known in the universe: the human brain. Instead of separating memory and processing—as happens in traditional computers—researchers started developing components capable of performing both functions simultaneously. This is where so-called "memory-storage" components come into play.

Connections change continuously in response to accumulated experiences.

Unlike conventional transistors, these devices possess a fascinating property: they alter their physical state in response to incoming electrical signals, creating something akin to a memory embedded within the material itself. In other words, the hardware learns. In neuromorphic chips, connections change continuously in response to accumulated experiences. Instead of loading vast histories from the cloud, the circuit itself stores information through patterns and associations.

The result is an architecture that operates much more like biological neural networks, consuming only a fraction of the energy required to power today's processing centers. This transformation raises a fascinating question. If two absolutely identical chips were exposed to different experiences over the years, their internal structures would also diverge; each system would carry its own unique history etched into the material. We would no longer be dealing with perfect copies, but rather with technological entities shaped by their own experiences.