Bamboo Scaffolds and the Robotic Reckoning

Audio article

Uses the speech voice supplied by your browser or device.

They come before the light does, these bamboo men, these scaffold saints of the pre-dawn dark, hauling poles and rope through a city that will soon pretend the enormous temporary thing rising in its lanes appeared by divine intervention. Every Puja season I see them: wiry men climbing structures that would make most office workers demand a helmet, an indemnity form, and their mother.

Pole against pole. Rope around bamboo. Pull. Twist. Brace. Climb.

And I think: here is some of the last honest geometry in a city that has forgotten how to count.

The popular picture is that these men somehow build by primitive instinct, without engineering, calculation, or thought. That is nonsense. The calculation is there. It simply does not always arrive wearing spectacles and carrying an AutoCAD printout. It lives partly in drawings and instructions, partly in the foreman, partly in convention, and enormously in the hands, eyes, feet, muscles, and accumulated experience of the karigar. A man who has spent twenty years tying bamboo does not need a load cell to know that this pole feels wrong. He knows because the pole speaks through his palm.

This is knowledge.

We have merely spent several centuries becoming very good at recognizing knowledge only when somebody has written it down.

The pandal rises, and the city swoons.

We call it culture. We call it tradition. UNESCO calls Durga Puja in Kolkata intangible cultural heritage, which is a splendid phrase for something that requires a very tangible army of people carrying things, welding things, painting things, tying things, lifting things, sleeping beside things, and occasionally hanging sixty feet above the ground while the rest of us argue about whether the lighting is tasteful.

For a few nights Calcutta becomes delirious.

Then the goddess leaves.

The lights come down.

The bamboo comes apart.

And many of the men who built the delirium go home.

They came from districts across Bengal, these carpenters, bamboo workers, rope men, painters, electricians, fabricators, clay workers, decorators, helpers, craftsmen and general-purpose human beings. Some return to Purba Medinipur, South 24-Parganas, Nandigram, and elsewhere. One North Calcutta puja in 2022 actually put up photographs of 215 artisans who had built its pandal, which tells you something about how many invisible humans can hide inside one visible spectacle.

Usually we photograph the spectacle.

Not the vertebrae holding it up.

And standing there, in my infinite and largely unemployable white-collar wisdom, I cannot stop thinking about Moravec’s paradox.

Hans Moravec put the problem beautifully in 1988. Roughly speaking, the things humans consider intellectually difficult can sometimes be surprisingly tractable for computers, while things a small child performs without applause—seeing, balancing, walking, grabbing, adjusting, navigating a changing physical world—can be infernally difficult for machines.

This has always offended our vanity.

We thought chess was intelligence because clever men played chess.

We did not think standing up was intelligence because toddlers did it while shitting themselves.

So computers became terrifyingly good at chess.

Then Go.

Then image classification, protein prediction, code generation, translation, writing, pattern recognition, statistical inference and increasingly large territories of what educated people once considered secure intellectual employment.

Meanwhile a robot could be defeated by a sock on the floor.

That gap is narrowing.

And this is where things become interesting.

As of 2026, roboticists are building machines with remarkably sophisticated hands, tactile skins, variable-stiffness joints, vision-language models, reinforcement learning, imitation learning and increasingly capable whole-body control. Humanoids can walk, run, carry loads, manipulate objects and perform carefully bounded industrial tasks that would have looked like science fiction not very long ago.

But the demonstrations can fool you.

A laboratory video is a tiny, well-lit island surrounded by an ocean of shit that can happen in the real world.

The rope is wet.

The bamboo is slightly warped.

Someone left a hammer where the robot expected empty floor.

The pole slips six millimeters.

A child runs underneath.

The rain starts.

The material supplier changed the thickness.

The knot must be tightened while the structure is moving.

The worker beside you misunderstands the instruction.

The light is bad.

The ladder isn’t where it was yesterday.

Nobody has labeled any of this for the neural network.

This is the world humans inhabit with astonishing competence while calling ourselves stupid.

Nature Machine Intelligence noted this year that commercial robots can still struggle with apparently ridiculous variations of ordinary tasks such as opening doors. Robotic manipulation researchers continue to describe humanlike dexterity in unstructured environments as a major unsolved problem.

Moravec is not dead.

He is merely being surrounded.

Which means the bamboo worker may, for the moment, possess something enormously valuable that the fashionable white-collar graduate does not: a body attached to a nervous system that has spent several hundred million years being debugged in the physical world.

I find this extremely funny because our economic hierarchy assigned prestige almost exactly backwards.

The man balancing barefoot on bamboo was “unskilled labor.”

The man making PowerPoint slides about him was “knowledge talent.”

Then the language model arrived and ate the PowerPoint first.

There is justice in the universe after all. It is simply run by an asshole.

We have been wrong about the AI timeline before.

The grand optimism did not begin in the 1980s. It began much earlier. After the birth of artificial intelligence as a formal field in the 1950s, some of its pioneers made forecasts that now look charmingly optimistic. Herbert Simon and Allen Newell believed machines would rapidly conquer intellectual tasks that ultimately took decades. The field repeatedly discovered that a convincing laboratory demonstration and a generally competent intelligence were not remotely the same thing.

Then expectations collapsed.

Then they inflated again.

Then collapsed.

AI has spent seventy years periodically discovering that the last ten percent contains another ninety percent.

Now the prophets are back.

This time they have billions of dollars, GPU clusters, venture funds, scaling curves, foundation models, embodied AI, better motors, better batteries, better sensors and hoodies of considerable technical sophistication.

And this time they are not entirely wrong.

That is the uncomfortable part.

The robots really are getting better.

The AI really is getting better.

The cameras are cheaper.

The actuators are better.

The training data are growing.

Simulation is improving.

Teleoperation can generate demonstrations.

Large multimodal models can connect language, vision and action in ways that were awkward or impossible a few years ago.

The mistake would be assuming either extreme: that a polished humanoid demonstration means general-purpose robotic labor has arrived, or that today’s failures prove it never will.

Nobody knows the date.

Ten years?

Twenty?

Thirty?

Never at economically useful scale for some jobs?

Anyone giving you the year with great confidence is either selling robots, selling shares in somebody who sells robots, raising venture capital, writing a book, or auditioning for television.

But grant the possibility.

Suppose Moravec’s mountain is eventually climbed.

Suppose a machine can arrive at a half-built Puja pandal in Calcutta with no perfect digital twin, no carefully arranged factory floor and no team of graduate students hiding just outside the camera frame.

It looks around.

It understands the structure.

It picks up bamboo.

It tests the pole.

It climbs.

It ties.

It compensates when the scaffold shifts.

It works in heat.

It works in rain.

It learns from the human foreman.

Then one year the foreman is unnecessary too.

Now we have a different problem.

Not an AI problem.

An economic problem.

The robot does not need to be better than the worker.

That is another little piece of bullshit hidden inside automation discussions.

It merely has to become cheaper overall.

Purchase price. Financing. Electricity. Maintenance. Insurance. Downtime. Repairs. Supervision. Software fees. Battery replacement.

Put all that on one side.

Put wages, accommodation, injury risk, scheduling, shortages, training and human unreliability on the other.

When the first number becomes smaller than the second, philosophy leaves the room.

The accountant enters.

And the bamboo man disappears.

Machines do not need lunch breaks, but they do need batteries.

They do not strike, but their manufacturer can discontinue a part.

They do not demand higher wages, but their cloud subscription can.

They do not get tired, but a gearbox can shit itself at three in the morning.

Capitalism does not require perfection.

Only a favorable spreadsheet.

This is the part that curdles my tea.

Because if machines eventually cross this frontier, we are no longer talking only about programmers, paralegals, copywriters, illustrators, translators, analysts, clerks and the enormous upholstered buttocks of the modern administrative economy.

We are talking about skilled physical work.

The plumber.

The mason.

The warehouse worker.

The agricultural laborer.

The electrician.

The construction worker.

The cleaner.

The mechanic.

The delivery worker.

The bamboo artisan.

For the last few years, people frightened by generative AI have comforted themselves with a new folk theorem: learn a trade.

Become a plumber.

Machines cannot replace plumbers.

Perhaps.

For now.

But “for now” is doing an enormous amount of work in that sentence.

If machines acquire robust manipulation, locomotion, perception and improvisation, there is no magical economic floor beneath which human labor becomes protected by God.

There is only another engineering problem.

And once that problem becomes solvable, another cost curve.

Then what?

That question is usually where the technology conference suddenly discovers ethics.

What happens if an economy becomes enormously productive while requiring fewer and fewer human beings to produce?

This is often phrased as though the problem were lack of things.

It isn’t.

A sufficiently automated economy may produce an obscene abundance of things.

The problem is distribution.

Capitalism traditionally distributes purchasing power largely through ownership and work.

You own something productive, or you sell your labor to somebody who does.

But what happens when labor loses bargaining power because increasingly capable machines can perform larger portions of it?

The machine does not need a salary.

The machine’s owner still receives revenue.

That distinction may become the entire century.

Perhaps productivity gains create new industries and new work, as they repeatedly have before.

Perhaps augmentation dominates substitution.

Perhaps human services expand.

Perhaps new desires create new occupations we cannot presently imagine.

That is the optimistic historical argument, and it deserves to be taken seriously because technological revolutions have repeatedly destroyed jobs without destroying employment itself.

But there is no law of physics guaranteeing that this transition must repeat forever.

A horse in 1900 could also have been told not to worry because previous improvements in transportation had created new jobs for horses.

The horse would have appreciated the economic theory while being converted into glue.

I am fifty-one.

I spent years working in healthcare data and information technology in the United States, building and dealing with systems that moved information between hospitals and databases. I returned to Calcutta more than a decade ago and have now spent roughly a decade outside regular paid employment.

So I do not require a futurist to explain technological redundancy to me.

I have already enjoyed the free trial.

And now machines can perform in seconds pieces of intellectual work that once looked comfortably human: write competent prose, produce code, summarize technical documents, analyze data, translate language, generate images, draft reports.

Not perfectly.

That word keeps saving us.

Not perfectly.

Neither are employees.

That is the problem.

A company does not compare the machine with God.

It compares the machine with payroll.

Which brings me back to the bamboo men.

Because they are not obsolete.

Not today.

They are doing something our most advanced systems still find extremely difficult: manipulating a dirty, changing, irregular physical world while balancing their own bodies inside it.

There is dignity in that competence whether the market recognizes it or not.

There is intelligence in it whether a university issues a certificate for it or not.

And there is something particularly obscene about the possibility that humanity may finally learn to reproduce that intelligence mechanically without first learning to value the humans who possessed it.

That is the part I cannot get past.

We may spend hundreds of billions of dollars teaching machines how to do what poor people have been doing for centuries, and only when the machine can do it will we finally describe the task as technologically sophisticated.

The robot will climb the bamboo.

Investors will applaud.

Engineers will publish papers.

Consultants will discover “autonomous temporary-structure assembly.”

Some bastard will put “AI-powered scaffolding ecosystem” on a slide.

The worker who taught humanity the problem existed will be standing outside the gate asking whether there is any work.

Progress.

We are building extraordinary machines.

We should.

The answer cannot be to deliberately remain inefficient so that human beings may retain the privilege of injuring themselves for money.

Nobody should preserve dangerous drudgery merely because our economic imagination cannot think of another way to distribute food.

If a machine can prevent a man from falling sixty feet from a bamboo pole, build the machine.

If it can enter a sewer instead of a human being, build ten thousand.

If it can carry bricks, weld in toxic fumes, clear mines, crawl through collapsed buildings, work around pathogens, lift the elderly, harvest crops in murderous heat, let the machine do it.

The obscenity is not automation.

The obscenity would be creating abundance and then telling the displaced human being that because his labor is no longer required, neither is he.

That is not a technological outcome.

That is a political choice wearing a robot costume.

So perhaps the real robotic reckoning is not whether we can build a machine capable of tying bamboo.

We probably will, eventually.

The harder engineering problem may be building a civilization intelligent enough to survive its own productivity.

The pandal will rise again next year.

The goddess will be installed.

The drums will sound.

Millions will walk through something that took months of invisible labor to create and will disappear within days.

For now, somewhere inside that structure, there will still be a human knot.

A hand pulling rope against bamboo.

A foot finding balance.

A nervous system making thousands of calculations nobody bothered to call intelligence because a poor man was doing them.

One day there may be metal fingers there instead.

When that happens, do not ask whether the robot has become human.

Ask what happened to the human.

P.S. Hans Moravec, Mind Children: The Future of Robot and Human Intelligence (Harvard University Press, 1988), especially his discussion of the strange asymmetry between abstract reasoning and sensorimotor competence. For the present state of the problem rather than the usual promotional robot videos, see recent work on dexterous robotic manipulation and embodied or “physical” AI. The machines are advancing quickly. The physical world remains an uncooperative bastard.