The project sounds modest on paper: tweak the inner wiring of neural networks. Yet the idea quietly challenges how today’s AI learns, and points toward systems that train faster, use less electricity, and behave a bit more like a human brain.
A hidden cost behind today’s “smart” AI
Every impressive chatbot reply or AI-generated image hides an uncomfortable reality: training these models burns extraordinary amounts of energy. Data centres run flat out. Power grids strain to keep up.
Some experts, including Elon Musk, have warned that AI development could hit an energy wall within a year if demand keeps rising. The problem does not come only from the sheer size of models, but from how they process information.
Modern neural networks work in large batches. Data points travel through layer after layer, across billions of artificial “synapses”. Only once the full circuit is completed does the model adjust its internal weights. That means a huge quantity of data shuttling back and forth in one go.
Most of the effort goes into moving data around the network, not thinking. The transport, not the logic, eats the power.
By contrast, the human brain updates itself differently. We adjust bit by bit, second by second, while we act, remember and plan. A team at Cold Spring Harbor Laboratory (CSHL) asked a simple question: what if AI did something similar?
Borrowing a trick from the human brain
The group, led by researcher Kyle Daruwalla, focused on a concept called “working memory”. In humans, this is the mental notebook we use to hold a phone number for a few seconds or keep track of the steps in a calculation. It sits at the crossroads of perception, attention and decision-making.
Neuroscientists have long suspected that working memory and learning are tightly linked. Children with stronger working memory tend to perform better in school. Adults rely on it to solve new problems. Yet evidence at the biological level remains incomplete.
CSHL’s idea is to bake a similar mechanism into artificial networks. Instead of letting data flow straight through, they add an auxiliary memory network that runs alongside the main system.
➡️ Cold leaves them hungry, snow leaves them trapped – yet your old broom can help
➡️ This guide to eco-friendly ways to make your pavements safe this winter without rock salt
➡️ Why cleaning feels heavier when you don’t see quick results
➡️ March 3 total lunar eclipse: Where and when to see the ‘blood moon’ from the US
Each artificial neuron now gets on-the-fly feedback from a working memory module, and can adjust its connections in real time.
In standard systems, updates usually happen only after a full pass and a heavy, global optimisation step. In the new design, updates are local, frequent and tied to what the network is currently “holding in mind”.
What changes inside the network?
The researchers compare two layouts:
- Classic neural network: information flows forward through the layers, then a backward pass updates all synapses at once.
- New approach: each layer receives feedback from a separate working memory circuit, allowing weights to shift continuously.
This feedback loop reduces the need to send huge gradients back through the full stack of layers. Instead, small, targeted changes happen as the model processes data, closer to how neurons behave in biological brains.
From energy hog to leaner learning machine
Why does this matter for energy use? Every big update step in a conventional AI model requires massive matrix operations. These are carried out on specialised chips that draw substantial power. Cut down the size and frequency of those operations, and you cut down the power bill.
By allowing synapses to update in place, guided by working memory, the new architecture aims to slash the number of heavy compute steps. Fewer global passes, fewer unnecessary data transfers, less wasted work.
The promise is simple: less shuffling, more thinking — and far lower energy demand per unit of learning.
If such networks reach production scale, data centres could train capable models without constantly racing to secure more electricity and more GPUs.
Could AI train on fewer examples?
The energy angle is only half the story. Today’s leading AI systems often need billions of examples to reach competence. That brute-force approach looks nothing like human learning. A child needs only a handful of demonstrations to grasp a new game or a basic grammar rule.
Daruwalla’s team suggests that tying learning directly to working memory could make artificial networks more sample-efficient. When a system can maintain and manipulate a small set of relevant facts over time, it might extract more value from each training example.
That prospect could reshape how we build models for tasks such as robotics, tutoring, or scientific research, where labelled data is scarce or expensive to collect.
Linking AI design and real brains
The CSHL work also nudges along a debate in neuroscience. A long-standing theory proposes a tight connection between working memory, synaptic updates and academic performance. Put simply: the better your mental scratchpad, the more effectively your brain tweaks its wiring while you learn.
The new AI framework gives this idea a computational backbone. By directly connecting an artificial working memory to synaptic changes, it shows that such a mechanism is not just biologically plausible, but also algorithmically useful.
| Concept | Human brain | New AI model |
|---|---|---|
| Working memory | Holds short-term information during tasks | Auxiliary network stores task-relevant data |
| Synaptic update | Local changes as we act and remember | Real-time weight tweaks guided by memory |
| Energy use | Highly efficient, low power | Fewer heavy operations, lower compute load |
From theory to real applications
The research still sits at a relatively early stage. The team’s results appear in the journal Frontiers in Computational Neuroscience, not in a commercial product launch. Scaling the method to the size of today’s largest language models will take time and testing.
Yet possible applications already stand out:
- On-device AI: Phones, home robots and wearables could host smarter assistants that learn from the user without constant cloud access.
- Scientific tools: Energy-efficient models could run long simulations or analyse lab data continuously without blowing research budgets.
- Education tech: Adaptive tutoring systems could adjust to each pupil in real time, mirroring how a human teacher tracks progress.
What “Hebbian” and “information bottleneck” actually mean
The paper’s technical title mentions “information bottleneck-based Hebbian learning”. The jargon sounds intimidating, but the core ideas are not exotic.
Hebbian learning is often summarised as “cells that fire together, wire together”. When two neurons activate at the same time, the connection between them strengthens. Many brain-inspired algorithms use variants of this rule.
The information bottleneck principle comes from information theory. It says a good representation should keep what is useful for a task while discarding irrelevant noise. Imagine compressing a story so that only the plot and key characters remain.
By tying these two ideas together, the CSHL method encourages synapses to strengthen when they help transmit the most task-relevant information held in working memory, not just any co-activation.
Practical scenarios and possible risks
Picture a factory robot equipped with such an AI. It observes a new part, holds key measurements in its working memory, and adjusts its internal connections as it tries to assemble the piece. It does not need to send every frame to a distant server. It learns on the spot, with modest energy use.
Or think of a mobile health app that gradually adapts to a patient’s routine. With local, memory-driven learning, it could refine recommendations without constantly retraining a giant cloud model from scratch.
Still, shifting to more brain-like learning paths raises questions. More on-device adaptation can make behaviour harder to predict and audit. Safety teams will need tools to monitor local changes and prevent systems from drifting into unstable behaviour.
There is also a risk of uneven performance. Local updates guided by a small working memory might overfit to very recent experiences if designers are not careful. Balancing short-term flexibility with long-term stability will be a key engineering challenge.
A step toward less wasteful, more human-style AI
The CSHL work will not instantly rewrite the AI landscape. Yet it points in a direction that many in the field quietly hope for: systems that burn less electricity, rely less on brute force, and lean more on clever architecture.
By giving artificial networks a kind of working memory and letting synapses adjust in real time, the researchers offer a concrete path toward AI that learns a little more like we do — and wastes far less along the way.
Originally posted 2026-02-05 09:38:26.