Why Our Current Approach to AI is Fundamentally Flawed

Why Bigger AI Models May Not Be the Only Path Towards More Intelligent Machines.

Artificial intelligence has made tremendous leaps in recent years, offering us tools that were previously thought to exist only in science fiction. From virtual assistants to autonomous vehicles and complex data processing, the applications of AI seem boundless. However, as exciting as these advancements are, there are fundamental issues with our current approach to AI development. Chief among them is the excessive energy consumption of AI systems, the limitations of data training methodologies, and the absence of intuitive environmental awareness and adaptive learning mechanisms akin to the human brain.

---

The Energy Demand of Modern AI Systems

The power requirements of contemporary AI systems are astounding. While a single AI request from a user may seem harmless, the energy demand behind the scenes is significant. A study from the University of Massachusetts, Amherst, for instance, reported that training a single large AI model can emit over 626,000 pounds of CO₂, equivalent to the lifetime emissions of five average cars .

OpenAI’s GPT-3, one of the largest language models, is estimated to have used around 1,287 MWh (megawatt-hours) of energy during its training phase alone, with a carbon footprint comparable to the lifetime emissions of an average American . When we consider that the human brain consumes only about 20 watts of power to complete similar tasks—such as pattern recognition, language processing, and problem-solving—it becomes clear that AI lacks the efficiency we see in biological systems.

Note:

The approximately 20-watt figure for the human brain is useful, but comparing it directly with an AI training run is misleading. The brain and AI systems perform different workloads, and training energy and operating power are not equivalent measurements.

Understanding the Energy Efficiency of the Human Brain

The human brain achieves remarkable feats with a minimal energy budget. Its unique architecture, which evolved over millions of years, enables it to process complex stimuli while conserving energy. This disparity in energy consumption highlights a crucial difference: current AI models lack a way to emulate the brain's energy-efficient processes. The brain relies on its ability to perform complex, multi-modal learning in a highly dynamic, adaptable way, consuming less energy as repetitive tasks become automatic. For example, once the brain has mastered a task through learning, it is able to perform that task with minimal energy by relegating it to “automatic” pathways.

In contrast, AI systems typically consume the same amount of energy regardless of how often a task is repeated. This indicates that, despite incredible advancements, our approach to AI lacks a vital “adaptive efficiency” that would allow it to improve its performance and minimize energy use over time.

Nov 2026 update - Modern systems use caching, specialised models, adaptive computation and other optimisation methods. The more defensible criticism is that most deployed models do not continuously incorporate everyday experiences into their underlying learned parameters as humans do.

Why Current AI Training Models Are Incomplete

At the core of this inefficiency is the way we train AI. Current models rely on vast datasets composed of millions of images, texts, and other data to create an expansive, though sometimes shallow, knowledge base. This approach, while effective for training models to perform specific tasks, is not conducive to a true, holistic understanding of the world. Humans, on the other hand, learn from multi-sensory, real-world experiences.

Humans do not need to see countless images to recognize an object. For example, we understand what a “cat” is by seeing, touching, hearing, and interacting with one in 3D space. This rich, multi-dimensional interaction gives us an innate ability to recognize a cat from any angle, in any context, and even apply this knowledge to recognize other animals or objects in similar categories.

Creating AI That Mimics the Brain’s Architecture

For AI to reach a new level of efficiency and adaptability, it must evolve beyond current models toward a structure that emulates the human brain. The brain is organized into specialized regions—each dedicated to a specific sensory input or function, such as sight, hearing, or balance. These regions are interconnected, allowing for seamless communication and adaptive decision-making based on environmental context.

To achieve similar flexibility, future AI systems should be structured in a more modular way. Such systems could involve multiple dedicated processors or networks, each responsible for a particular “sense” or type of input, that would be able to share information dynamically. This approach would enable AI to process complex inputs more holistically rather than isolating them as separate tasks. Additionally, as in the human brain, AI would benefit from being able to “rewire” itself, adapting its own code to optimize performance and enhance learning over time.

The Role of Quantum Computing in AI’s Future

The human brain’s remarkable efficiency and possible quantum processing capabilities may offer insights into what AI needs to become truly self-aware. Theories in neuroscience propose that the brain may rely on quantum mechanics to enable certain cognitive functions, potentially explaining phenomena such as consciousness, memory, and sensory integration. Recent research on avian navigation supports the idea that biological systems can utilize quantum processing to interact with complex natural phenomena.

If AI is to achieve true awareness, a breakthrough in quantum computing may be necessary. Quantum processors offer the potential to perform complex calculations at a fraction of the energy required by traditional silicon-based processors, with quantum states allowing for unprecedented levels of efficiency and processing speed. As quantum computing technology advances and potentially becomes miniaturized, it could enable AI systems to process information with the speed and adaptability that resemble biological brains.

A quantum-enhanced AI could, in theory, recognize patterns and make decisions based on highly complex, interrelated data sets, similar to the way a human might instinctively recognize and respond to their surroundings. By combining quantum computing with an adaptive, modular architecture, we may come closer to developing AI that is not only vastly more energy-efficient but also more conscious of its environment and capable of evolving over time.

Nov 2026 Update

There is no established evidence that human consciousness requires quantum computation, nor that quantum computing is necessary for generally intelligent AI. A 2026 scientific review reaches a similarly cautious conclusion. 
Quantum computing is a possible specialised technology, not a prerequisite for machine awareness.

---

November 2026 Update: Are We Building Intelligence or Just Bigger Machines?

When I originally wrote this article in October 2024, my concern was relatively straightforward. Artificial intelligence was achieving extraordinary results, but the methods being used to achieve those results appeared fundamentally different from the way biological intelligence develops.

Two years later, AI has become considerably more capable. Modern systems can solve complex problems, write software, analyse images, interpret speech and undertake increasingly sophisticated tasks.

Yet the question that prompted this article has not disappeared.

Are we developing artificial intelligence in the most effective way, or are we compensating for weaknesses in our underlying architecture by continually increasing the amount of computing power we throw at the problem?

The Increasing Cost of Intelligence

One of the most significant developments since 2024 has been the extraordinary expansion of the infrastructure required to support AI.

According to the International Energy Agency's 2026 report, global data centre electricity consumption is projected to increase from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030.

Not all this electricity is consumed by AI, but artificial intelligence is an increasingly important driver of demand.

Meanwhile, modern AI models have become more efficient at individual tasks. This creates an interesting contradiction: the technology is becoming more efficient, yet the infrastructure needed to support its expanding capabilities and growing popularity continues to consume more energy.

This raises an important question.

What happens when millions of AI systems are expected to reason, observe, remember, plan and respond continuously rather than simply answer occasional questions?

Perhaps the solution is not simply to build larger data centres. Perhaps we need to reconsider the way artificial intelligence processes information in the first place.

Intelligence Is More Than Information

Today's advanced AI models possess access to an extraordinary amount of learned information.

Yet possessing information and developing understanding are not necessarily the same thing.

Consider how a child learns that an object placed behind a chair continues to exist, even when it can no longer be seen.

The child does not need to consult millions of examples of similar situations. Through repeated observation and interaction, the child gradually develops expectations about the physical world.

An AI language model may describe this concept perfectly. Some multimodal AI systems can also infer object permanence and predict physical events.

However, there remains an important distinction between producing an accurate explanation and maintaining a persistent, continuously updated understanding of a changing environment.

Modern AI systems have demonstrated increasingly impressive forms of reasoning and generalisation. It would be incorrect to claim that they simply memorise their training data.

Nevertheless, reliable learning from ongoing experience remains an important challenge.

The Missing Relationship Between Experience and Memory

Human memory is not simply an enormous collection of information.

It is selective.

We do not consciously remember every movement, sound or visual detail we encounter throughout the day.

Our brains continually filter information, identify patterns, form associations and establish expectations about what may happen next.

Much of this processing happens without conscious involvement.

An unfamiliar sound attracts our attention because it differs from what we expect. A familiar environment requires relatively little deliberate thought until something changes.

Could artificial intelligence become more efficient by adopting similar principles?

Rather than processing every incoming piece of information with equal importance, a future AI architecture might identify which observations are familiar, which are unexpected and which deserve further investigation.

It could gradually develop a memory of its environment, establish relationships between recurring events and adjust its expectations as those events change.

The objective would not necessarily be to reproduce the human brain. It would be to explore whether some of the principles behind biological learning could improve artificial systems.

A Different Approach: Specialised Intelligence Working Together

Another possibility is that future AI systems will become increasingly modular.

Rather than relying on one enormous model to perform every function, an intelligent system could consist of specialised components.

One component might process visual information. Another might identify spatial relationships and movement. Others could examine longer-term patterns, manage memory or evaluate predictions.

These components could communicate selectively, ensuring that detailed processing is concentrated on information that actually matters.

Such an architecture would introduce its own challenges, including communication overhead, coordination and the difficulty of determining which information is genuinely important.

Nevertheless, it offers a potentially interesting alternative to relying exclusively on larger general-purpose models.

This is not a completely new direction. Researchers are already exploring combinations of specialised models, external memory, multimodal processing, world models and adaptive computation.

The question is whether these approaches can eventually produce systems capable of learning more efficiently from experience.

Learning Without Being Told What Everything Is

Perhaps one of the most interesting challenges is whether AI can develop useful concepts without humans providing all the categories in advance.

Imagine a camera observing a garden.

Over time, objects appear, disappear and return. Some move independently, while others remain stationary. Certain movements occur repeatedly in particular locations.

Could an AI system begin to recognise these relationships without being told what a bird, tree, fence or person is?

Could it identify recurring groups, establish expectations and recognise when something unexpected happens?

And, perhaps most importantly, could it decide for itself which observations deserve remembering?

This is related to research into self-supervised learning, unsupervised representation learning and predictive world models.

The challenge is not simply detecting patterns. It is turning those patterns into useful, reliable knowledge that can support future predictions.

From Recognition to Prediction

A system that recognises an object is performing one useful function.

A system that remembers where the object appeared, observes how it behaves and predicts where it might appear next is performing something more sophisticated.

Prediction could become one of the foundations of an alternative approach to AI.

The system observes an event, develops an expectation and compares that expectation with what actually happens.

When the prediction is incorrect, it has a reason to investigate further.

Over time, it could develop increasingly sophisticated models of its surroundings.

This would not automatically produce general intelligence or consciousness. But it might produce systems that are more adaptable, more efficient and better grounded in the environments in which they operate.

Does Artificial Intelligence Really Need to Think Like Us?

There is an important qualification to this entire argument.

Human intelligence is not necessarily the only effective form of intelligence.

Birds, insects and other animals demonstrate remarkable abilities using nervous systems very different from our own.

Similarly, an artificial system may eventually develop powerful capabilities without replicating human cognition.

The objective should therefore not be to build an artificial human brain simply because the biological brain is our most familiar example.

Instead, we should investigate which principles of biological intelligence are useful, which can be improved upon and which may be unnecessary for artificial systems.

The Question We Should Be Asking in 2026

The remarkable achievements of modern AI demonstrate that today's methods work.

They do not necessarily demonstrate that these methods represent the most efficient or ultimately most successful way of developing intelligence.

Larger models, additional computing power and more sophisticated training techniques may continue to deliver impressive results.

At the same time, alternative approaches involving continuous learning, selective memory, specialised processing and interaction with the physical world deserve serious investigation.

Perhaps the future of artificial intelligence will not involve choosing between these approaches.

It may involve combining them.

The greatest breakthrough in artificial intelligence may not be a machine that knows more than every previous machine. It may be a machine that learns what is worth knowing, remembers what matters and recognises when its understanding of the world needs to change.

The Path Forward: An AI That Can Truly Understand

AI holds incredible promise, yet current approaches show the limitations of focusing on expansive data training over creating adaptive, efficient models that can “understand” rather than simply respond. As we seek to build AI that more closely mirrors the remarkable capabilities of the human brain, we will need to integrate multi-sensory inputs, self-adaptive programming, and potentially quantum processing.

A paradigm shift is required to design AI that is truly sustainable, intuitive, and conscious of its surroundings. Only then will we be able to create AI systems that, much like the human brain, achieve extraordinary feats with minimal energy while continuously learning and evolving in real-time.

---

The Future of AI on YouTube

---

Latest Future AI Articles