This is the first post in a series I’m calling “Through a Glass, Darkly: Brains, Machines, and the Stories That Shape Us”.

In this series, I will explore the thesis that brains, AI systems, and cultures are all different kinds of information compression systems, and that , like all information compression, they pay a price for the efficiency gains.

I sat down to write this after noticing a convergence among four topics I’ve spent a lot of time thinking about: how the brain works and fails, the large software systems I’ve spent my career building, technological progress outside of computing, and cultural belief systems I think about as I raise children.

The Human Brain

“The Brain—is wider than the Sky—” -Emily Dickinson

The human brain may be the most efficient general learning system we know of.

A child enters the world essentially helpless – able to do little more than produce noise at the feeling of discomfort. One year and 200,000 calories of energy consumption we have a mobile being, capable of navigating unfamiliar environments, manipulating small objects and starting to recreate language.

Twenty more years and 15-20 million calories of energy investment later, and some 100 million words listened to, we have an adult human. A fully independent creature, typically capable of deep thought, detecting nuanced social cues, impressive physical feats and broad creativity. A creature who has nearly completed the development process responsible for the most impactful species in the known universe.

The mental processing only accounts for some 20% of the energy consumption of the body, (still, a lot given the brain only accounts for some 2% of the body by mass) which means a human brain consumes 3-4 million calories of energy in its development years. Some portion of that goes to output tokens instead of input and a lot of it goes to other things that aren’t related to learning, but even if all of those calories went towards learning, 3-4 million calories is extremely efficient.

In comparison, our best LLMs require thousands of times as many of both; tokens input and calories consumed as the human brain. GPT 5 is estimated to have consumed over 10 GWh of electricity in its training, which translates to roughly eight billion calories – two thousand times more than our estimated twenty year old human brain. The token numbers are more extreme, GPT 5 was trained on an estimated 20 trillion tokens, some 100 thousand times as much as our 20 year old brain.

Even if the comparison is rough, modern AI systems appear vastly less energy efficient than human brains at learning.

A century of research and billions of dollars of investment into computers has achieved a lot, but it seems we have not gotten close to what millions of years of evolution was able to achieve for humans in terms of efficiency at being able to learn.

Some people are worried about how powerful AI is, but I don’t think they properly grasp how powerful the human brain is.

If you want to see the most impressive learning machine we know of, find a friend with a baby and simply admire that drooling, babbling bundle of cuteness on the floor, wading in its own excrement. Marvelous!

Efficiency and Electricity

“Man is but a reed, the weakest in nature, but he is a thinking reed.” -Blaise Pascal

Scale has limits. We cannot simply double our energy input forever, and we seem to be running into trouble producing more high-quality training data. As we train AI agents for increasingly specific tasks, the corpus available for those tasks may be much smaller. So as we continue to invest, efficiency will become increasingly important: ways to get more output from less energy and less training data.

In that search for efficiency, we will inevitably look towards the miraculous computing device we possess in our skull.

There is precedent of this – neural networks, an underlying technology in LLMs and precursor to them, are based on a simple understanding of the human brain. As our understanding of the brain increases, our ability to recreate it electronically will too.

To make more powerful AI tools, we will attempt to catch up with some of the five orders of magnitude of efficiency advantage that our wet, carbon-based skull computer has over hard, cold, silicon GPUs.

The Cost Of Shortcuts

“Anxiety, keep on tryin’ me
I feel it quietly
Tryin’ to silence me” -Doechii

There is no free lunch.

A large part of what makes the human brain so efficient is that it takes massive shortcuts. It summarizes, compresses and assumes. Intelligence is expensive, and so we opt for a more affordable approximation of it through compression.

We ingest the world in rough approximations. We focus our attention consciously and subconsciously, ignoring the vast majority of the signals around us. We only really process a small fraction of what our senses can take in. It is more efficient that way. The estimated intake from our senses is about 10 million bits per second, where we seem to only be able to consciously process 10-100 bits per second, meaning our conscious ignores 99.999% of information our senses gather.

Perhaps something to remind someone in your life the next time they accuse you of not paying attention. “Don’t worry dear, I’m not just ignoring you, I ignore 99.999% of things as well.” Certainly that will make the conversation go better…

Next, in our hunt for efficiency, we store only parts of what we process and even then, do so in a highly lossy manner. Our data storage is corrupt – memories are verifiably unreliable. Our defragging is uncontrolled, continuous and seemingly random. Sleep, specifically dreaming, is the process of throwing away information via turning the rest of it into very strange fictions.

This system is flawed by any engineering standards. But this flawed system does good enough at the job it needs to complete and it does so efficiently enough that it has stayed around. Time and time again, the human brain has passed the tests thrown at it. Humans learn the things imperative to our survival. We survive.

“Life, uh, finds a way” -Ian Malcolm (Jurassic Park)

But survival and efficiency have a dark side. Some of the same shortcuts that help us navigate the world can, under the wrong conditions, harden into patterns that harm us. Anxiety, depression, and addiction are maladaptations that arise from the same brain that so efficiently helps us navigate the world.

If, as we look for efficiency in AI system development, we look for similar shortcuts, will those systems face similar maladaptations? What will the costs of maladaptive and powerful AI systems be to humans?

That is what I will explore in Part 2: Maladaptation and the Price of Efficiency