A session at IISc's Institute of Neuroscience sent me down a rabbit hole on what the human brain does that no current AI system can — and why the answer has almost nothing to do with software. The physics of forgetting might be the most important thing nobody is talking about.
I spent an evening at a session run by the Institute of Neuroscience at IISc recently, and I walked in expecting the usual story. The one where AI is racing up behind us and the only real question is how many years we have left before it overtakes the human mind on everything. That is the story we are all marinating in right now. It is in every headline and every boardroom.
The researchers in that room were quietly making the opposite case. Not out of sentiment. Out of evidence. And one idea in particular followed me home and would not leave. The human brain is extraordinarily good at something that the most powerful AI systems on earth still cannot do, and the reason has almost nothing to do with software. It comes down to physics.
I went digging after that, and the deeper I went the more I became convinced that we have been asking the wrong question about AI and humans for years. The question is not who is smarter. The question is what each one is actually built for. And once you see the answer, the most interesting implication is not about machines at all. It is about how spectacularly we have been wasting human beings.
Think about the last time you bought a car.
For about three weeks, you became a genuine expert. You learned how depreciation curves work. You knew the difference between trims that look identical on paper. You understood dealer margins well enough to smell a bad offer across the room. If someone had cornered you at a party during those three weeks, you could have lectured them on resale value with total confidence.
Now try to remember the resale value of the model you almost bought. You cannot. And here is the strange part. You did not try to forget it. You did not study to lose that knowledge the way you studied to gain it. It simply faded on its own, quietly, the moment it stopped being useful.
We treat this as a flaw. A bad memory. Something to apologise for. The neuroscience says it is the opposite. It is one of the most sophisticated things your brain does, and it is the single biggest reason you are still better than a machine at a whole class of problems.
The brain does not store what you know as one fixed, permanent thing. It runs what researchers describe as a layered system that changes across many timescales at once. Some changes happen in milliseconds, as individual connections become more or less excitable. Over seconds to minutes, a fresh impression either hardens or dissolves depending on what happens next. Chemicals like dopamine and acetylcholine act like a relevance signal broadcast across the whole brain, tagging which experiences are worth keeping and which are noise. Over weeks and months, the physical wiring itself reshapes, growing new connections for things you use and quietly dismantling the ones you do not. And every night while you sleep, your brain runs a kind of housekeeping pass, strengthening what mattered and turning down the volume on everything else.
The research literature, including major reviews in Nature and the Annual Review of Neuroscience, calls this a hierarchy of partially reversible states. Some are cheap and fast to change. Some are expensive and slow. And crucially, some are designed to fade. That third category is the one that machines do not have.
Here is the thing nobody tells you about memory. Keeping everything would make you worse, not better.
When you forgot the car you almost bought, your brain was not failing. It was making room. There is now strong evidence that forgetting is an active, controlled process, not passive decay. When two memories compete and one gets in the way of what you are doing now, the brain has machinery in the prefrontal cortex that actively suppresses the loser. Specific molecular pathways, including Rac1 signalling cascades studied in mammalian brains, speed up the breakdown of connections that are no longer earning their keep. AMPA receptors can be physically pulled back from synapses, dismantling the architecture that had encoded a memory. Even microglia and astrocytes, once dismissed as mere support cells, participate directly in synapse elimination and remodelling.
Your brain sheds expertise on purpose because holding onto it has a cost. Every memory you keep at full strength is a little more interference, a little more clutter, a little more energy spent maintaining something you may never use again. So the brain runs a constant, ruthless, brilliant economy. It invests heavily in what is useful right now and divests from what is not.
This is why you can throw yourself completely into a new problem, master it, and then walk away clean. Plan a wedding and for six months you are a world authority on florists and seating charts and the politics of the guest list. Move to a new city and within weeks you know the shortcuts, the good restaurants, which neighbourhoods to avoid after dark. Plan a trip to Japan and you briefly know more about Tokyo train lines than most people who live there. And then it all gently dissolves, leaving room for the next thing.
What stays behind is not the trivia. It is the judgment. The pattern of how to read a new situation fast and make good calls inside it. The brain keeps the skill of becoming an expert and discards the specific expertise once it has served its purpose. That is the whole trick. That is the thing.
This is where almost everyone gets it wrong, and where the IISc session genuinely changed how I think.
The assumption is that AI just needs to get better. More data, smarter algorithms, a few more breakthroughs, and machines will learn and unlearn as fluidly as we do. The physics says otherwise, and the reason is elegant in how precise it is.
In your brain, a single synapse is doing everything at once. It is the place where signals pass, the place where the memory lives, the place where computation happens, and the place where learning physically occurs. All of it in the same tiny spot. When something important happens, the change is made right there, locally, with no permission needed from anywhere else.
A modern AI system is built the opposite way. Its knowledge lives as billions of numbers stored in memory. To use that knowledge, the system has to read those numbers out, move them to a separate place where the computing happens, do the work, and write the result back. To actually learn something new, it has to run a whole separate training process that touches the entire network at once, guided by an external optimiser running outside the system itself. Everything in that design depends on shuttling information back and forth across a gap between where things are stored and where things are computed.
The best way I can put it is this. Your brain learns the way a living city changes. A shop opens here, a road gets widened there, a neighbourhood slowly reinvents itself, all of it happening locally and independently with no central office signing off on each modification. A current AI system learns the way a city would change if every road resurfacing project on every street had to be submitted to one central authority, calculated for the entire city at once, and approved before any local work could begin. One of these scales effortlessly. The other does not.
This is not a criticism of the engineers who built it. It is a description of the wiring. And it is why a model trained to be brilliant at chess or protein folding cannot cheaply repurpose itself into a marketing strategist. It would need to be retrained, at great expense, on a mountain of new data that probably does not exist yet. You just need a few weeks and some good conversations.
This is the question I really wanted answered, and the honest version is more uncomfortable for the AI optimists than they tend to let on.
Start with the reassuring part. There is a hard physical floor on computation called Landauer's principle, the absolute minimum energy required to erase a single bit of information. You might assume our chips are bumping up against it, that we are near some fundamental wall. We are not even close. A single memory access in a modern processor burns through something on the order of a hundred billion times that minimum. The ceiling is so far above us that it explains nothing. So in the narrow sense, no, physics is not forbidding a brain-like machine.
But that is the wrong place to look, and it is where most optimistic accounts quietly stop. The real story is not about some distant thermodynamic limit. It is about what a transistor fundamentally is, and what that forces every machine built from transistors to do.
Here is the heart of it. In your brain, a single synapse does four things in the same physical spot, at the same time. It stores a memory. It computes with that memory. It learns by physically changing itself. And it does all of this using slow, graded chemistry rather than fast, brittle switching. A transistor can do none of these things together, because a transistor is a switch. That is its entire nature. It flips between on and off, and it does so brilliantly, which is exactly why digital computers are so reliable and so scalable. But a switch cannot remember a value. To store a number, a silicon system needs a separate memory cell, sitting physically somewhere else on the chip. And the moment storage lives in one place and computation lives in another, every single calculation requires hauling data across the gap between them, and hauling the answer back.
That gap is not a flaw an engineer forgot to fix. It is the foundation the entire technology is poured on. Reading a value out of memory, moving it to the processor, and writing it back can cost something on the order of a thousand times more energy than the actual arithmetic in between. Your brain never pays this tax, because in your head there is no gap. The memory and the math are the same physical thing.
It goes deeper still. Your synapses are analog. A synapse does not store a crisp 1 or 0. It holds a continuous strength, set by ion concentrations and receptor counts, that can sit anywhere on a smooth scale and be nudged in either direction by a small chemical change. A transistor is the opposite by design. Its whole purpose is to be cleanly binary, because binary is what makes digital computing trustworthy. So to fake a single graded synapse, a digital system has to gang together many transistors holding many bits, and then read and rewrite all of them just to make one small adjustment. The brain changes a connection by shifting a chemical gradient. Silicon changes a weight by conducting a careful electronic ritual across separated components. One is a whisper. The other is a committee meeting.
And then there is heat, which is where the dream of simply stacking everything closer together quietly dies. The obvious move to fix the data movement problem is to bring memory and compute physically nearer, ideally in three dimensions the way the brain packs its synapses into a dense volume. But silicon transistors leak current and throw off heat, and they leak more as you make them smaller, not less. Stack them densely into a 3D block and you cannot get the heat out. Reviews on 3D chip integration say this plainly. Heat piles up between the layers with nowhere to go. The brain runs a fully three-dimensional, densely packed architecture and stays cool because it works at low voltage, moves information slowly through chemistry rather than electricity, and has a liquid cooling and supply network woven through every cubic millimetre. A chip has no equivalent. So silicon is forced to stay relatively flat and spread out to shed heat, which reintroduces the very distances that made data movement expensive in the first place. Density and cooling pull against each other, and no material we currently manufacture at scale lets you win both at once.
Researchers know all of this, and they are genuinely chasing it. There are neuromorphic chips designed to fire in sparse bursts like real neurons rather than running a relentless global clock. There is in-memory computing, which tries to do the mathematics where the data already lives. A 2025 paper in Nature Electronics proposed a hybrid stack combining memristors for fast rough computation with ferroelectric capacitors for precise updates, deliberately trying to mimic the brain's use of multiple state variables at different timescales. These are serious attempts and serious progress.
But the same literature is equally clear about what is missing. A 2025 review in Nature Electronics states the problem directly: no current memory technology delivers the full combination an adaptive system actually needs, which is low write energy, many graded levels, reads that do not destroy what is stored, durability over endless rewrites, stable retention, consistency across devices, and manufacturability at scale. Every candidate device delivers some of these and fails the rest. The ones that store beautifully are stubborn to update. The ones that update easily tend to drift or wear out. There is no single device that is simultaneously a good memory and a good synapse. Biology built one. We have not.
The most telling example I found is a neuromorphic research platform called SpiNNaker 2, running a learning algorithm called e-prop that was specifically designed to be more brain-like, more local, more efficient than standard approaches. Even with every advantage, one reported training run would have taken twenty four days to complete a fairly modest learning task. That is not a software failure. It is a measurement of how physically expensive this kind of learning is, even after you have improved the algorithm.
This is not a gap measured in software updates or the next chip generation. As long as AI lives on conventional transistor-based, digital, clocked, memory-separated hardware, it carries these structural disadvantages with it — not because it is badly designed, but because it is exquisitely designed for the opposite purpose. We built this entire technology to be exact, repeatable, binary, and reliable. Those choices gave us everything computing has achieved. They are also, almost point for point, the wrong choices for fluid, graded, self-rewriting, adaptively forgetful intelligence. You cannot tune your way out of this from inside the current substrate. The properties are baked into what a transistor is. To truly close the gap, you need a different physics altogether. And we are nowhere near manufacturing that at the scale and reliability the brain achieves in a warm, wet, three pound organ running on twenty watts.
That is a gap measured in decades. Possibly in a materials science we have not invented yet.
None of this is an argument that AI is weak. It is an argument that AI is differently shaped, and in its own shape it is unbeatable.
AI wins decisively wherever the problem is well defined, the rules are stable, and the task needs to be done at a scale or consistency no human could survive. Run the same pricing analysis across millions of transactions and a person will fatigue, drift, and make errors. The machine will not. Review ten thousand contracts for the same clause and a human will start missing things by the second hundred. The machine stays sharp at number ten thousand.
And then there is the category where AI does not just beat humans but does something we genuinely cannot do at all. Some problems require holding millions of variables in mind at the same instant. Protein folding is the defining example of our time. When AlphaFold solved a problem that had stumped structural biology for fifty years, it did not win by being wiser than biologists. It won because no human mind, however brilliant, can simultaneously evaluate the forces acting on every atom in a folding protein. The same principle applies to molecular docking in drug discovery, genomic pattern recognition across population scale datasets, and a growing frontier of problems in materials science. These are not cases where a machine outran a person who was trying hard. They are cases where the architecture of human cognition makes the problem structurally unsolvable regardless of how much expertise or effort a person brings.
We should use AI here relentlessly, even at its current extraordinary energy cost, because the alternative is not a slower answer. It is no answer.
This is the part that genuinely excites me, and it is where the whole story turns from being about machines to being about us.
For roughly a hundred and fifty years, we built the world of work on a single assumption. That a human being should pick one narrow thing, get very good at it, and then do that thing, over and over, for an entire career. The accountant who does the same kind of return for thirty years. The analyst who owns one report. The specialist who goes so deep into one narrow trench that they can never climb out of it. We called this professionalism. We called it expertise. We built entire organisations, entire education systems, entire identities around it.
And it was never what we were built for.
Read the neuroscience again and the truth is almost embarrassing. The human brain is the most powerful rapid adaptation engine we have ever encountered. It is built, from the ground up, to drop into something brand new, master it at speed, make sharp judgment calls under real uncertainty, and then clear the decks and do it all over again somewhere completely different. We took the finest serial expertise machine in the known universe and bolted it to a single narrow task for forty years. We have been running Ferraris in first gear, in a car park, for a century and a half.
The specialisation model was never a tribute to human capability. It was a workaround. It existed because there was no other way to get repetitive cognitive work done at scale. A human had to be the cog, because the cog had to be a human. That constraint is now lifting.
And this is the part the doom narrative completely misses. When AI absorbs the repetitive, narrow, grindingly specialised work, it is not stealing the best of what humans do. It is finally removing the part we were never suited for in the first place.
Picture the marketing professional who, instead of owning one category for a decade, moves fluidly across markets. New product, new geography, new audience, and within weeks they are reading the room with the confidence of someone who has been there for years, because that is exactly what the brain is designed to do. Picture the consultant who walks into an industry they have never touched and is asking the sharpest question in the room by the end of the first week. Picture a workforce of serial experts — people who develop deep working mastery, deploy it, solve the real problem, let the specifics go, and move cleanly to the next thing, with the brain's forgetting machinery making each fresh start not a loss but a feature.
That is not a downgrade of human work. That is the most human version of work there has ever been.
A go-to-market strategy for a new product in an emerging market takes about three months of real contextual work before launch. To do the equivalent with a fine-tuned AI model, you would need training data that does not exist yet, a model built on it, and a validation process against outcomes that have not happened. By the time that pipeline produces anything reliable, the human who spent those three months in the market, talking to real people, reading weak signals, iterating on positioning, has already moved on to the next problem. And when the market shifted unexpectedly at month two, as it always does, they updated their thinking overnight. The model's weights did not change.
We have spent a very long time building organisations that treated this adaptability as a liability. Too generalist. Lacks deep expertise. Moves around too much. The performance review systems, the job architectures, the entire grammar of professional development were built to reward the thing AI is about to do better than any human ever could.
What is left, once you strip that away, is the thing the brain was always best at. The thing we systematically undervalued because we had no machine to take the other work. The capacity to walk into the unknown, make sense of it fast, exercise real judgment under genuine uncertainty, and then let it go and do it again.
The most powerful learning system on the planet has been sitting in every chair, in every office, in every classroom, the entire time. It runs on about twenty watts, less than the bulb over your desk. We are only now, finally, about to let it do what it was built for.
Written by Sameer
samspoke.com · Singapore