What Biology Can Teach About Building AI

Direct Source Verification: This story is aggregated from Forbes (forbes.com). Full reporting rights and copyright belong to the primary publisher.
Zuzanna Stamirowska is the CEO & Co-Founder of Pathway, an AI lab building AI architecture that continually reasons, learns, and evolves.

Zuzanna Stamirowska is the CEO & Co-Founder of Pathway, an AI lab building AI architecture that continually reasons, learns, and evolves.

getty​​​We in the AI industry have been solving an engineering problem when we should have been answering a biological one. The brain has already figured out memory, so why haven’t we copied it?​

​​Neuroscience may hold the answer to persistent memory and continuous learning. It has for millions of years. The brain has been pointing to the solution AI engineers have been chasing all along. If we model our systems after how the human brain already works, we can build AI that doesn’t need constant oversight, prompting and re-reminding.

​The human brain runs on about 20 watts of energy. Today’s AI data centers are crossing the gigawatt threshold—that’s a billion watts, or roughly 50 million times more power than a single human brain. Compute demand is growing more than twice as fast as chip efficiency, and industry estimates project roughly $500 billion a year in CapEx just to keep pace.

As the co-founder and CEO of Pathway, I build post-transformer architectures inspired by the brain, informed by my background in complexity science. Studying how large networks of simple components give rise to intelligent behavior convinced me that the brain’s efficiency is a design blueprint, not a curiosity.​

The industry needs to take a harder look at brain-like AI systems, which reflect how we think, reason and act. When new information arrives in the brain, only the neurons that care about it light up, passing the signal to their neighbors rather than waking the whole network unnecessarily. Useful, repeated connections get stronger over time. This is one way that post-transformer AI can operate. ​

Many of today’s models take the opposite approach. They are densely activated, running vast numbers of parameters for every token regardless of what the query actually requires. That’s a large part of why inference costs scale so steeply. ​

The brain stores memory in the connections between neurons, not filed away separately. AI, by contrast, largely lacks persistent memory. Most tasks start as a new request that can’t call back on previous information, either given to or generated by the model. ​

In the brain, this works through synaptic plasticity—the brain’s ability to alter the strength and number of connections between neurons—which makes long-term potentiation (LTP) possible.

LTP strengthens neural pathways after repeated use and, with it, comes efficiency. The future of AI architecture is pathways that strengthen with use, built into the structure from the beginning rather than bolted on afterward. ​

The brain learns while it’s running, continuously. It doesn’t stop for retraining. In the brain, the cognitive workspace that holds, manipulates and tracks information in real time is called working memory. Its closest analog in AI is latent reasoning. The prefrontal cortex broadly supports this executive function: complex cognitive processes, working memory and task-switching, all happening continuously and without a reset button.​

Transformer-based models today have no equivalent. Reasoning happens once, in a single forward pass. Whatever the model figures out along the way is usually discarded the moment the session ends. No mechanism lets a thought become a habit. Every insight is rediscovered from scratch, every time, at a cost that compounds with scale, rather than shrinking with it.

Rebuilding systems with this kind of continuous, structural learning in mind is a materially different undertaking than fine-tuning a static model to behave as if it remembers.​

The brain does not think in words first. Language is a later evolutionary layer, built on much older systems for perception, prediction and action. A thought takes shape in the brain before it is ever translated into a sentence, and most of that shaping happens in a space with no linguistic form. ​

Transformer-based models have no equivalent space to think or reason within. They express reasoning as language because it is the only medium they have. Every intermediate step has to be written out as text, fed back in and read again, which is why long reasoning chains get slower and more expensive as the problem gets harder. The model isn’t just showing the work; it is also required to produce that work verbally to accomplish it all. ​

Latent reasoning is the alternative: working through a problem in a continuous, internal representation, the way the brain does, and only translating the result into language at the end, if at all. That distinction separates a system that has to talk to itself to think through complex processing from one that reasons the way your prefrontal cortex might—silently, before anything must be said out loud. ​

Meta has begun integrating related capabilities into its Muse personal agents with models like Muse Wonder and Muse Wonder Full. The models feature “Active Memory” and “Active Focus,” both intended to make the agents more capable. Active Memory lets the models carry over information about the user from one conversation to another, and Active Focus lets the agent keep working on a task by trying multiple solutions when it hits an issue. Databricks has also built Agent Memory, a managed memory feature on its Agent Bricks platform.​

Functional memory and latent reasoning are necessary tools for successful agents and, therefore, successful large language models (LLMs). However, these memory systems are currently behind paywalls and still limited in their applications. AI with memory and the ability to problem-solve cannot be limited to LLMs and agents.

By looking beyond LLMs and their transformers, the AI industry can develop a technology that truly reflects the brain. ​

Evolution spent hundreds of millions of years on this problem and converged on the same three answers: memory that lives in the structure itself, learning that never stops and reasoning that happens before language. The industry has spent a decade building systems that do none of those things, and building them extraordinarily well. That is the strange position the field is in. The engineering is world-class. Next, we just have to build the models to match.​​​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Original Source
https://www.forbes.com/councils/forbestechcouncil/2026/11/20/what-biology-can-teach-about-building-ai/
Visit Forbes ↗
SHARE STORY:
𝕏 f in

Related Coverage in Technology