Abhishek S.
Shipping in public. Listening in private.

Abhishek

I lead women’s Indo-Western & Premium at Max Fashion. I also wrote the AI that runs the buying floor.

Rare profile. Category operator who ships production code.

Senior Buying Leader · Max Fashion Women’s Indo-Western & Premium · 530+ India stores NIFT ’12 · Twelve years on the floor

abhishek@bengaluru ~ %
>role: senior buying lead
>dept: women’s indo-western + premium
>floor: 530+ stores india

Moravec's Paradox

Chess fell before reliable dishwashing. Hans Moravec named the insult in 1988: high-level reasoning is often easy for computers, while the sensorimotor skills a toddler gets for free can defeat a robot in a kitchen.

The paradox is not that machines are bad at movement. It is that human intuition ranks difficulty backward. Algebra feels hard because school makes it explicit; grasping a cup feels easy because evolution did the training run for roughly 4 billion years before language, exams, or code review existed.

The Case

Moravec's claim was simple: the skills humans find ancient are computationally expensive, while the skills humans find recent can be brittle but formal. Chess has rules, a board, legal moves, and a win condition. A sink has reflections, soap, deformable cloth, unknown friction, water, occluded objects, and a glass that breaks if pressure is wrong by a few newtons.

That is why Deep Blue beat Garry Kasparov in 1997, while household robots in 2026 still struggle with open-ended chores. The chess world is closed. The kitchen is not.

Rodney Brooks attacked the same mistake from robotics. In "Elephants Don't Play Chess" (1990), he argued that real intelligence starts in the world, not in a symbolic planner floating above it. His subsumption architecture looked crude compared with classical AI diagrams, but it captured the point: walking, avoiding obstacles, and responding to a changing room are not lower-level details. They are the base layer.

The diagram is the trap. Old AI put planning at the top and treated sensors as input. Moravec's paradox says the loop itself is where much of the intelligence lives.

Where It Shows Up

Task Machine milestone Why the easy-looking task is hard
Chess Deep Blue beats Kasparov, 1997 Fixed board, legal moves, searchable game tree
Image labeling ImageNet error rates drop sharply after 2012 Still works mainly inside labeled categories
Folding laundry No household default by 2026 Cloth changes shape at every touch
Dishwashing Partial demos, narrow products Wet glass, hidden residue, grip force, clutter
Warehouse picking Deployed since the 2010s Constrained inventory beats open homes

The useful distinction is not mental versus physical. It is closed-form versus open-world. concept information theory helps here: formal games compress into rules, while kitchens keep producing new state.

What's Contested

The paradox is less clean after deep learning. Vision improved after AlexNet in 2012, and robot learning now absorbs video, simulation, and teleoperation data at scales Brooks did not have. The old line between reasoning and perception is no longer a wall.

The live question is transfer. A model can learn a thousand demonstrations of picking up a mug, then fail on a mug with a wet handle, a different lighting angle, or a crowded sink. Is that a data shortage, an embodiment problem, or a missing causal model? As of 2026, the field has not settled that.

Why This Crosses Realms

Moravec's paradox rhymes with concept fermi paradox: both punish human intuition about difficulty. Interstellar travel sounds like one grand engineering problem, but mission voyager 1 shows the scale problem in a single number: after nearly half a century, Voyager 1 is still nowhere near dest proxima centauri. The universe makes distance look empty; robotics makes a kitchen look simple.

It also touches concept scaling laws from the opposite side. Scaling text prediction gave AI a clean victory path because the internet is already tokenized. The physical world is not pre-tokenized. A spoon in a drawer does not arrive as a row in a dataset.

An Open Question

If the body is part of the computation, can intelligence be fully learned from video and text, or does the machine need scars from acting in the world?

Key Sources

Further Reading

Abhishek's take

What grabs me about Moravec's paradox is that it flips the prestige map of intelligence. The hard part is not always the task that sounds intellectual; sometimes it is the wet glass, the crumpled sleeve, or the missing label in the physical world. I read this as a warning against demos that confuse a closed arena with reality.

Tags: #ai #robotics #embodied-intelligence #cognition #automation