The conversation around AI these days is almost always about benchmarks and which model scored highest on some vague reasoning test or which one passed the bar exam this time in how many seconds.. or even which one can now “think” better than the average human. And I get it, the “progress” is genuinely staggering.

But here’s the thing that keeps nagging at me: the more time I spend working with LLMs and neural networks, the more I find myself quietly amazed at us. At the three pounds of electrochemical mystery sitting between our ears that we’ve been casually carrying around our entire lives without a second thought.

Working close to AI has been, unexpectedly, one of the most humbling lessons in what it means to be human.

1. The Ultimate Energy Efficiency

To do what a modern large language model does, we need massive data centers, industrial-scale liquid cooling systems, and enough electricity to power a small city. Large scale models are estimated to consume around 500–700 MW of power during training alone. Even once the model is created, inference isn’t cheap.[1] Add to that the fact that this hardware degrades at exceptionally fast speed. It “ages” faster and there is no way to slow that aging down at the moment. Once the hardware degrades, it needs to be replaced.

Meanwhile, our brain runs on roughly 20 watts which is about the same as a dim bedside lamp… We can contemplate for hours and that only makes our brains younger! I remember today I was working on a system and I asked AI to fix a very “obvious” problem for me while I multi-task to something else. After I was done with that task, I switch my attention back to the IDE to find the AI is going through what I call a “thought loop” (when you switch the reasoning to high) and it gave me a chuckle. It was going back and forth with itself over an issue that was very obvious to me. AI will never be able to grasp the sheer context we have as humans, and that remains until today a very assuring thing to me. That our thinking minds and our ability observe and extract meaning will remain distinctive to us only and no machine can match us at that.

Right now, your brain is regulating your heartbeat, filtering out background noise and reading / interpreting these words, all at the same time! on the caloric equivalent of a mini-snickers, we are truly out of our depth when we claim that we cracked efficient intelligence.

2. A tid-bit Neuroplasticity and AI model training

When an AI model is done “training”, it’s essentially crystallized and frozen in time. It knows what it knows, and that’s it, until someone runs an expensive fine-tuning or retraining cycle to teach it something new. There’s no such thing as a model casually picking up a new skill over lunch. And to be honest, I never believed in the “emergent” skills that AI would gain with more data.

On the other hand, our brains does this constantly from the moment we are born until our “lights” go out. And to me, the ability to learn something new and understand the unknown is something I would never trade the world for; such a blissful skill.

Every time we learn to drive a new route, pick up a language (I’m trying to learn french!), or break a bad habit, our neurons are reorganizing, pruning old connections and forming new ones through a process called synaptic plasticity.[2] The brain isn’t just software running on fixed hardware; the hardware itself rewires in response to experience. Although our capacity to store data is quite limited to 2.5 petabytes [3].

Neuroscientist Michael Merzenich, whose work on neuroplasticity spans decades, described the brain as a system that “writes and rewrites itself throughout life.”[4] Even into old age, the brain retains the capacity to adapt in ways no current AI architecture can match organically.

You are a self-optimizing algorithm that never stops updating its own hardware.

3. Moravec’s Paradox: The Easy Stuff Is the Hardest

In AI research, there’s a long-standing puzzle known as Moravec’s Paradox, named after roboticist Hans Moravec who described it in the 1980s: the things we think of as “hard” (chess, calculus, legal reasoning) are surprisingly easy for machines, while the things we consider “basic” (walking, recognizing a face, picking up a cup without crushing it) are extraordinarily difficult.[5]

  • Winning at Chess or Go? Easy — AI has dominated both for years.
  • Solving complex differential equations? Trivial.
  • Walking across a cluttered room, noticing that a friend looks slightly off today, and knowing exactly when to say something and when to stay quiet? Incredibly hard.

The intuitive, embodied, socially-attuned intelligence we exercise without thinking is, computationally speaking, some of the most sophisticated processing in the known universe. It took hundreds of millions of years of evolution to build it. We got it for free.

4. Learning from a Single Example

To train an image classifier to recognize cats, you typically need tens of thousands of labeled examples. And even then, the model is essentially performing very sophisticated pattern-matching.

A human toddler sees one cat and from that single moment, they’ve extracted a rich causal model of cats, that understanding generalizes instantly and permanently.

Researchers call this one-shot learning, and it remains one of the hardest open problems in AI.[6] Humans do it naturally because we don’t just catalog patterns, we build causal models of the world. We understand why things happen, not just that they tend to happen together.

You don’t need curated and sometimes biased Big Data to gain wisdom. You just need real experience, a functioning critical thinking engine and enough attention to grasp what is happening.

5. The Hidden Cost of Outsourcing Intelligence

We are, quietly and quickly, outsourcing more and more of our cognitive lives to AI. We ask it what to think about, how to phrase what we feel, what decision to make, what to read next. Heck I’ve asked AI to give me resources to read and enrich this article because I was too lazy to Google it! and sometimes it genuinely is time-saving. But if we zoom out for a minute, a more troubling picture starts to emerge.

This laziness will narrow our horizon, what the AI gives us will become what is “possible”..

Cognitive scientists have a concept called cognitive offloading, the practice of using external tools (notebooks, GPS, calculators) to reduce mental effort.[7] It’s not new; humans have always used tools to extend what our minds can do. But there’s a meaningful difference between a tool that amplifies a capability you still exercise (for example a calculator helping you calculate larger numbers) and one that replaces it entirely (asking AI to solve your math problem altogether).

Now scale that to reasoning, writing, planning, and judgment.

There’s growing concern among researchers that habitual AI-assisted decision-making may, over time, erode the very faculties that make human intelligence distinctive. One line of research published in Organizational Behavior and Human Decision Processes found that over-reliance on algorithmic recommendations reduced individuals’ independent analytical confidence over time.[8] Another body of work suggests that leaning heavily on AI for writing tasks can gradually narrow a writer’s own expressive range and voice and vocabulary.[9]

The deeper risk isn’t that AI will become smarter than us. It’s that we might, through accumulated convenience, become less capable versions of ourselves without ever noticing the drift. This thought keeps me up at night and I wish I had an answer for it. The only way out, is to accept that we need to slow down and go back to doing things the hard way, slow and boring way. There are no shortcuts for building our minds, thoughts, sophistication and character.

Again, this isn’t an argument against using AI, I use it every day in development and research. But there’s a difference between using a gym machine and having someone carry you everywhere. One builds strength overtime, the other slowly makes your legs useless.

The question worth asking is: which cognitive muscles are you still choosing to exercise?

References

  1. Patterson, D. et al. (2021). Carbon and the Wide Net: Measuring the Carbon Intensity of Machine Learning. arXiv:2104.10350.
  2. Bhaskaran, M., & Bhaskaran, V. (2020). Synaptic plasticity and learning: A review. Frontiers in Cellular Neuroscience, 14.
  3. Yonelinas, A. P. (2020). The capacity of human memory: Is there any limit to human memory? ResearchGate.
  4. Merzenich, M. (2013). Soft-Wired: How the New Science of Brain Plasticity Can Change Your Life. Parnassus Publishing.
  5. Moravec, H. (1988). Mind Children: The Future of Robot and Human Intelligence. Harvard University Press.
  6. Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332–1338.
  7. Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.
  8. Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103.
  9. Bender, E. M. et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ‘21.