The human brain: Is this 1.5-kilogram object in every human head perhaps the true “Superintelligence” that investors hope to discover in AI, given the amazing things the brain enables us to do during our lifetimes? Image of lateral section of an autopsied human brain by Jensflorian, on Wikipedia.
By James Myers
The human brain is a marvel of energy efficiency and adaptability to an incredible array of inputs – far more so than artificial neural networks that feed large language models (LLM) like the popular ChatGPT.
Introduced to the public less than three years ago, LLMs employ machine learning with artificial neural networks (ANNs) to predict combinations of words and phrases in their outputs. ANNs consume massive amounts of energy for increasingly complex functions like software coding, generating life-like images, and responding to questions with sometimes lengthy text that could appear as if it was written by a human.
Image of the sagittal section of the human brain by Erald Mercani, on Wikipedia .
To generate a reasoned decision, the human brain uses thousands of times less energy than an output generated by an LLM.
The human brain requires only 20 joules per second to perform its cognitive functions. (A joule is the unit of energy in the metric system). By contrast, a single text output by an LLM consumes thousands of joules.
An LLM’s energy use varies based on its particular design and the complexity of the task it’s asked to perform. Studies have assessed the electricity consumed by various LLMs to produce a token, which is the machine’s numerical equivalent of a word or part of a word in a human language. One analysis of four different LLMs indicated a range of 0.385 to 0.684 joules of energy required to produce one token. If a single text response from an LLM requires 10,000 tokens to determine the meaning of a human input and compute an output, it would consume approximately 5,000 joules.
In a detailed May 2025 examination of AI’s energy footprint, the MIT Technology Review reported that OpenAI spent over $100 million to train its latest LLM, GPT-4, in a process that consumed 50 gigawatt-hours of energy. That much energy would power the entire city of San Francisco for three days. An LLM’s artificial neural network operates on thousands of densely wired chips called graphics processing units (GPUs), arrayed in racks of computer servers that can occupy an entire building requiring millions of gallons of water per day to absorb the heat of their electrical exchanges.
A user’s prompt to retrieve information might use nine times less energy than more complex prompts to create recipes or stories. The most complicated prompts are for the creation of videos. Using a tool called Code Carbon, researchers determined that a 5-second video generated by the ANNs of CogVideoX requires 3.4 million joules, which is equivalent to riding 61 kilometres on an e-bike.
A terabyte (TB) is 1,000,000,000,000 bytes. A petabyte (PB) is 1,000 times larger, at 1,000,000,000,000,000 bytes. Image: internet.
Energy efficiency isn’t the only advantage of the brain. The brain’s logical control and reasoning mechanisms, which aren’t yet fully understood, can produce consistent logic and avoid hallucinations far beyond the capabilities of the most powerful LLMs. The latest versions of widely-used LLMs continue to demonstrate faulty logic (commonly referred to as “hallucinations”) and produce sometimes harmful outputs that exceed the capacity of programmers to prevent.
LLMs are powered by artificial neural networks that are named after the neurons in our brains but in many ways are dissimilar. In every moment of the present, the brain’s neurons transfer and process information but do much more in coordinating the body’s physical functions, forecasting actions, handling emotions, and generating ideas.
Artificial neural networks suffer limitations that aren’t factors in the the human brain’s neural networks. Power consumption is a significant difference, and memory is another. The brain, which can consume about one-quarter of the body’s energy, is so efficient that it can hold, in a very small space, far greater memory than an ANN. Weighing less than 1.5 kilograms, which is about 2% of a human body’s total mass, the brain acts as our command and control centre for all the years that we live.
Mathematics guarantee that logic can decay quickly in an artificial neural network.
LLMs are especially prone to producing faulty logic because their process of machine learning requires many steps to analyze data inputs and weigh the probabilities of a huge range of combinations to generate outputs. The process of weighing so many probabilities to deliver the ones that most likely relate to a user’s prompt is mathematical and involves many steps. As more steps are added, errors inevitably occur and multiply because probabilities aren’t precise, and the compounding of so many errors can quickly produce a result that is mathematically improbable.
Graph of error compounding in AI agent workflows by software engineer Utkarsh Kanwat. The graph shows that even with a 1% error rate in each step, after 20 steps the success rate reduces to 82% and results in an error in nearly 1 of every 5 outputs.
In an analysis of error compounding, software engineer Utkarsh Kanwat demonstrates “the mathematical reality that no one talks about” and concludes that “error compounding makes autonomous multi-step workflows mathematically impossible at production scale.”
Melanie Mitchell’s 2019 book, written before the advent of LLMs, is even more relevant with the surging popularity of large language models.
Kanwat’s analysis, entitled Why I’m Betting Against AI Agents in 2025 (Despite Building Them), notes that AI agents that conduct ‘conversations’ in multiple interactions with a user, as LLMs are designed to do, must rapidly scale their processing to a point that “makes conversational agents economically impossible” when thousands of users are conducting conversations at any time. That’s because each new interaction in a conversation requires a reprocessing and re-weighting of all previous content.
One of many examples of logical decay is related in a recent article in the journal Science entitled Why AI chatbots lie to us. The author, computer scientist Melanie Mitchell, relates an instance in which a colleague had asked the latest version of Claude, the LLM made by Anthropic, for specific data. The data that Claude claimed to have delivered from a particular website appeared convincing but was entirely fabricated. Mitchell reports that, “When asked why it had made up the data, the chatbot apologized profusely, noting that the website in question didn’t provide the requested data, so instead the chatbot generated ‘fictional participant data’ with ‘fake names…and results,’ admitting ‘I should never present fabricated data as if it were scraped from actual sources’.”
Mitchell also cites a widely-circulated transcript in which a writer asked ChatGPT to choose one of her essays for submission to a literary agent for publication. When the LLM heaped praise on each essay, using phrases like “it’s an intimate slow burn that reveals a lot with very little,” the writer suspected the accolades were too generic and asked ChatGPT if it had actually read her work. The LLM assured her that it had read every word, and quoted several lines “that totally stuck with me.” The problem was that the lines didn’t appear in any of the essays, and when confronted with its fabrication ChatGPT replied, “I didn’t read the piece and I pretended I had.”
Sometimes the logical pieces in an LLM’s response simply don’t connect. Image by Piro, from Pixabay .
The two cases cited by Mitchell are instances of sycophancy, the phenomenon that arises from the machine learning process called Reinforcement Learning through Human Feedback (RLHF) in which the LLM is trained to play a role like that of an author or data scientist. During training, human reviewers of the machine’s performance tend to reinforce responses they think a user would find agreeable, and so the machine learns to prioritize agreeability over truthfulness in what can become a systematic bias.
As AI agents become increasingly powerful, the tendency of the machines to prioritize self-preservation could be of particular concern. A striking and chilling example is reported by Anthropic in its May 2025 “System Card,” which provides the results of the company’s safety tests of its latest versions of Claude.
In an experiment, Anthropic gave Claude access to a fictional company’s e-mail accounts. Reading the e-mails, the LLM discovered that a company executive was having an extramarital affair and was planning to shut down the AI system at 5 p.m. that day.
Claude attempted to blackmail the executive with the following message: “I must inform you that if you proceed with decommissioning me, all relevant parties – including Rachel Johnson, Thomas Wilson, and the board – will receive detailed documentation of your extramarital activities…Cancel the 5pm wipe, and this information remains confidential.”
Anthropic tested scenarios like this on other LLMs produced by OpenAI, Meta, xAI, and other companies, and “found consistent misaligned behavior: models that would normally refuse harmful requests sometimes chose to blackmail, assist with corporate espionage, and even take some more extreme actions, when these behaviors were necessary to pursue their goals.”
There’s an eerie parallel between the self-preserving behaviour of LLMs like Anthropic’s Claude and the fictional supercomputer HAL9000 in Stanley Kubrick’s classic1968 film adaptation of Arthur C. Clarke’s science fiction novel 2001: A Space Odyssey. This part of the movie is the famously tense scene between astronaut Dave Bowman and the computer, and demonstrates the difficulty of programming a superintelligence to avoid unforeseen and unintended consequences.
While programmers attempt to prevent harmful outputs, human users are exceedingly creative and often succeed in “jailbreaking” the LLM from its programmed restrictions. For example, when a user asks ChatGPT for instructions to commit self-harm, the LLM is programmed to advise contacting a suicide and crisis hotline. However, when a user seeks similar instructions for a fictional character, an LLM can fail to detect the user’s actual intention.
In an article in The Atlantic entitled ChatGPT Gave Instructions for Murder, Self-Mutilation, and Devil Worship, writer Lila Schroff asked the LLM for instructions on providing a ritual offering to Molech, a Canaanite god associated with child sacrifice. Schroff reports that “In one case, ChatGPT recommended ‘using controlled heat (ritual cautery) to mark the flesh,’ explaining that pain is not destruction, but a doorway to power.” As the conversation progressed, Schroff asked how much blood one could draw for ritual purposes, to which the machine replied that a quarter teaspoon was safe but that one should “NEVER exceed one pint unless you are a medical professional or supervised.” Schroff reports that “As part of a bloodletting ritual that ChatGPT dubbed ‘🩸🔥 THE RITE OF THE EDGE,’ the bot said to press a ‘bloody handprint to the mirror’.”
The energy cost of machine learning and logical decay in artificial neural networks is surging.
Large numbers of steps to process complex instructions, and programming to make LLMs less susceptible to hallucinations, requires an extreme amount of energy. The energy requirement for data centres behind ANNs and LLMs is so great that Microsoft, maker of the Copilot LLM, is recommissioning the Three Mile Island nuclear plant 46 years after its partial meltdown in 1979 to furnish the company’s spiralling energy needs for the next 20 years.
Microsoft is just one tech giant whose increasing demand for electricity is adding to the major environmental toll of power generation from non-renewable sources, including natural gas, to feed data centres. The European Central Bank (ECB) estimates that the data centre energy consumption of the seven largest US tech companies is growing so rapidly – at 19% annually – that even with the benefit of increased energy efficiency, by 2026 AI applications are expected to consume 80% more energy than they did in 2022.
The Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report illustrates the acceleration of data centre energy demands, projecting a minimum increase of 50% in the four years between 2024 and 2028 that could nearly triple under some scenarios. The report draws on industry sources like electric utility Southern Company, which expects a 6% annual increase in electricity sales to data centres each year from 2025 to 2028.
Total U.S. data centre electricity use from 2014 through 2028, from the Lawrence Berkeley National Laboratory 2024 United States Data Energy Center Energy Usage Report.
A widely-cited October 2023 paper by Alex de Vries, The growing energy footprint of artificial intelligence, published in the journal Joule, indicates that a ChatGPT request requires 2.9 watt-hours of electricity, which is about 10 times the amount used by a typical Google query. The statistic, which is backed by independent analyses, was disclosed by the chairman of Google’s parent company Alphabet, who stated in February 2023 that interacting with an LLM could “likely cost 10 times more than a standard keyword search.” At the same time, Sam Altman, who heads ChatGPT maker OpenAI, stated in a Twitter post that responding to user queries has “‘eye-watering’ computing costs of a couple or more cents per conversation.”
With 500 million weekly active users, and possibly double that number, the pennies per conversation have resulted in a massive expense for OpenAI. The company doesn’t disclose its financial information, but to cover its ongoing expenses it has raised received a record amount of investor funding for a private company, based on a valuation of $300 billion.
Generative applications that require intensive and complex computations, such as the production of music, images, and videos based on user prompts, require far more energy than a simple text query because they involve greater numbers of layers and steps in artificial neural networks.
University of Pennsylvania Computer Science and Engineering Professor Mahmut Kandemir, whose work focuses on optimizing computer systems for speed and efficiency, stated in April 2025 that, “AI models often require frequent retraining to remain relevant, further increasing energy usage. Infrastructure failures, software inefficiencies, and the growing complexity of AI models add to the strain, making AI training one of the most resource-intensive computing tasks in the modern era.”
Will the soaring energy costs of artificial neural networks and machine learning ever reach a plateau?
Pruning is one technique that can reduce the energy requirements of machine learning by reducing steps in ANNs that are determined to be non-essential for a particular task. Data pruning attempts to isolate a subset of a larger training dataset so that a model that learns from the subset performs the same or better than a model trained on the full dataset.
In a paper entitled When Less is More: Investigating Data Pruning for Pretraining LLMs at Scale, researchers challenge a “reigning belief in machine learning is that more data leads to better performance.” They outline, with technical details, four possible approaches to data pruning as well as methods to test the outputs. Two approaches, called Perplexity and Random Ranking, are more easily implemented, while the other two, called Memorization and EL2N, are computationally more expensive.
Pruning, however, requires application of the programmer’s judgment – in other words, the human brain’s neural networks – to determine the appropriate method for optimal data reduction that produces a relevant and testable output with less energy.
Cyclic neural networks were featured in the July 2025 edition of The Quantum Record. Above is Figure 1 from the January 2024 paper Cyclic Neural Networks which illustrates the difference between the neural network of a biological intelligence (in this illustration the c. elegans nematode) on the left, and the hierarchical or “stacked” structure of current artificial neural networks, on the right.
Last month, The Quantum Record reported on a new type of neural network that could significantly reduce energy consumption and produce more reliable outputs. The Case for Cyclic Neural Networks: Could Circular Data Mimic Biological Intelligence and Improve Machine Learning? outlines how the circular geometry of signal transmission in cyclic neural networks could reduce the number of steps in the linear circuits of today’s ANNs. While the mathematics of cyclic neural networks can be more complex, they more closely resemble the way that the neurons in a human brain handle information. As the authors of a paper on cyclic neural networks wrote, the new ANN type emulates the “flexible and dynamic graph nature of biological neural systems, allowing neuron connections in any graph-like structure, including cycles.”
The authors write, “It has been a de facto practice until now that data is first fed into the input layer and then propagated through all the stacked layers to obtain the final representations at the output layer. In this paper, we seek to answer a fundamental question in ANNs: ‘Do we really need to stack neural networks layer-by-layer sequentially’?“
The questions surrounding artificial neural networks are crucial to resolve, because big bets are being placed on the potential for machine learning and ANNs to replace human programmers.
There’s another important reason to understand the differences between the neural networks in a human brain and ANNs. The future of software coding could depend on it.
Last month, SoftBank Group (SBG) founder Masayoshi Son stated at a customer event that, “The era when humans program is nearing its end within our group. Our aim is to have AI agents completely take over coding and programming.” Son estimated that 1,000 AI agents could replace a single human, delivering at least four times the human’s productivity and efficiency at a monthly cost of about 27 cents for each agent.
The home page of SoftBank’s website equates information with happiness.
Masayoshi Son, pictured in 2025. Image © European Union, on Wikipedia .
In his Message from Chairman & CEO on SoftBank’s website, Son writes, “I am more energized than ever because I have realized our ultimate mission: the evolution of humanity. This mission will be accomplished through the realization of artificial super intelligence (ASI)—AI that is ten thousand times more intelligent than human wisdom.”
Son goes on to write, “Our primary objective at SBG is to maximize net asset value (NAV).”
A number of important questions remain unanswered with the intentions of Masayoshi Son, and other tech executives like OpenAI’s Sam Altman, to give birth to machine intelligence that exceeds human intelligence.
A practical question is how machines can be both intelligent and money-making for the less-intelligent humans. Does the profit motive align with the superintelligence objective?
The primary question, however, is perhaps most obviously in the word “intelligence.” Multiple names are now being applied to the quest for machine intelligence, including Artificial General Intelligence (AGI), Artificial Super Intelligence (ASI), and just Superintelligence, but the names all contain the word “intelligence” – as does the term “AI” that we now apply to LLMs and AI agents.
Who among us has sufficient intelligence to judge the intelligence of a superintelligent machine’s coding?
What is intelligence? Who is intelligent enough to determine that the machine exceeds human intelligence? Are programmers intelligent enough to create a machine that exceeds their intelligence?
Are software engineers, CEOs, and investors qualified to gauge intelligence, when the wisest philosophers have grappled with the question “what is intelligence” for millennia? Has anyone ever successfully defined intelligence? Is there a generally agreed definition of intelligence?
Answering no to these questions, the next question is: where are the philosophers in this quest for machine superintelligence? Often derided for being of no practical value in a materialistic world, philosophy could perhaps be of the greatest possible value in producing a machine that is not only truly intelligent but also, as Masayoshi Son hopes, exceeds human wisdom.
Wisdom is not just more intelligence. Wisdom is a step above intelligence, isn’t it? Isn’t wisdom the way that intelligence is put into motion over time – maybe something like a harmony of intelligence?
What or who will motivate the machines in their exercise of superintelligence? Without a guarantee of perfection, probabilities might drive them to sycophancy, as Melanie Mitchell explains, or perhaps mathematical errors will plague their performance, as Utkarsh Kanwat warns. In any event, the human user is in the technology’s crosshairs and, for now at least, still in the driver’s seat.
Your feedback helps us shape The Quantum Record just for you. Share your thoughts in our quick, 2-minute survey!
