John Sokol's Blog: A Computer Children Can See: A Clear-Plastic Pneumatic Adder

A Computer Children Can See: A Clear-Plastic Pneumatic Adder

This essay is the follow-through on a one-line note I wrote to myself on the front page of my wiki in October 2019 and never came back to. The note said: “Build laser cut Pneumatic computer. Slide valves, and other air flow in layers laser cut materials. Can I build an Adder? Latches, flip flops?” The answer to all three questions is yes, and the result would be a small clear-plastic box, illuminated from below, that adds numbers slowly enough for a child to watch each bit propagate. This essay is the design.

The Hacker Dojo had a laser cutter. So did half the maker spaces I’ve ever joined. The laser cutter sits in the corner, somebody pays the power bill, and every couple of weeks one of the members loses an afternoon to “what should I build with the laser cutter that justifies the membership fee.” The output of that question is mostly small boxes, lamp shades, jigsaw puzzles, and signs.

The 2019 note was the answer to that question, on one particular afternoon. I wanted to build something that a kid could look at and understand. Something where computation, normally hidden inside silicon at gigahertz speeds invisible to anyone, would be slow enough and visible enough that a child could point at a part of it and say “that bit is on right now and now it’s off and look the next one is turning on.” I wanted addition to be a thing you could watch happen.

The mechanism I had in mind was pneumatic. Air through channels carved into stacked sheets of clear acrylic. Slide valves visible from above through the top sheet. Each output bit indicated by a little ball floating in a column — high when the bit was 1, low when it was 0. A 4-bit input set by toggling four little plastic switches; the result of input + 1 (or input + input, depending on the wiring) appearing on four output columns about a second later, the carry rippling visibly from low bit to high bit on the way.

That was the idea. I never built it. This essay is what it would take, why the parts work the way they do, and a plan a maker space could actually follow.

1. Why a visible computer matters

A kid born in 2026 will grow up surrounded by computation, none of which is visible. The phone is a slab. The laptop is a slab. The hearing aid, the doorbell, the elevator, the car, the toaster — all slabs, all opaque. Computers are not, to a modern child, things that do anything visible. They’re things that show things on screens about what they’ve already done.

I think this is bad for intuitions. Computation isn’t magic. It’s a sequence of small mechanical decisions: this bit, then this bit, then this bit, with logic governing each step. When the decisions happen at a billion times per second inside a slab, you cannot watch them, and the abstraction layer between “I press a key” and “a character appears” feels like sorcery.

But computation runs at any speed you build it for. If you slow it down to one decision per second, in a medium you can see, kids understand it. I have watched this happen with Lego Mindstorms, with the Phillips Hydraulic Computer demos at LSE, with the Antikythera replicas at the museum in Athens, with the Z3 reconstruction at the Deutsches Museum. The reaction is the same: when a child can see a machine deciding things, they get it. The mystery evaporates. They don’t think it’s stupid; they think it’s cool. The slowness is the feature.

A pneumatic computer in clear acrylic is the maker-space version of that pedagogy. You can build one with a $5,000 laser cutter and $300 worth of materials, in maybe four weekends, and afterwards it sits in the front of a classroom or the kids’ room of a maker space and does one job, well, forever: it adds numbers slowly enough that you can see it think.

2. The 2019 note in context

I scribbled the note down because I had three things converging at once.

I had a laser cutter handy. Laser cutters do one thing brilliantly: they cut precise 2D shapes out of flat material. Stacking those 2D shapes turns them into 3D channels for whatever fluid you want to push through. The features can be small (well under a millimetre), the materials are cheap (acrylic at a few dollars a sheet), and the cycle from “design change” to “new test piece” is an afternoon.

I had been reading old computing history. The Phillips Hydraulic Computer (1949, LSE) modelled the UK economy with water flowing through transparent perspex tanks. Konrad Zuse’s relay computers clicked their way through addition in the 1940s. There was a whole era when computers were visibly mechanical things, and the abstractions we now take for granted — bit, register, gate — were literally pieces of metal you could touch.

And I had been thinking about kids. Specifically, about how kids encounter the word “computer” before they have any model of what one does, and how the gap between “computer” the magical box and “computation” the precise mechanical sequence is one of the worst educational gaps in modern STEM. Whatever closes that gap is a public good.

The note was the intersection. Laser cutter plus old fluidic-logic history plus pedagogy of visible computation equals: clear-plastic adder. Six years later I still think it’s the right answer.

3. Fluidic logic is a real thing, and it used to be a serious field

The thing I half-knew in 2019 and fully understood by writing this essay: people built actual working computers out of air in the 1960s. Not toys. Not demos. Industrially-supported, multi-thousand-gate, production hardware.

The seed event is Raymond “Billy” Horton’s 1960 fluidic amplifier, developed at the U.S. Army’s Harry Diamond Laboratories in Washington, D.C. Horton showed that a small jet of air, deflected by the Coanda effect, could behave as a bistable element — a flip-flop with no moving parts. The paper landed in Control Engineering and the field exploded. Within five years there were dedicated conferences, commercial fabrication houses (Bowles Engineering being the biggest), multiple competing gate geometries, and Air Force programs funding the development of entirely-fluidic computers as candidates for high-radiation, high-vibration, high-EMP environments where electronics couldn’t survive.

The flagship build was Norden Systems’ MOD-A, around 1966 — a general-purpose digital computer made entirely of air channels. It worked. It was slow (kilohertz, not gigahertz) and big (room-sized, not chip-sized) but it computed, in the same way a contemporary electronic mainframe computed, with no electrical components anywhere.

Then ICs got cheaper. By 1975 a TTL flip-flop cost a few cents; a fluidic flip-flop cost a few dollars and was a thousand times slower. The economics collapsed. The Air Force funding dried up. Bowles Engineering eventually went under. The conferences shrank. “Fluidic logic” became a forgotten word.

There was a quiet survival in two specialist domains: gas-turbine controllers (where the working fluid is already there) and high-rad environments. But the general-purpose fluidic-logic research programme died on schedule, around the time of the first 8-bit microprocessor.

What brought it back, decades later, is soft robotics. Around 2015 George Whitesides’s group at Harvard demonstrated soft robots whose control logic was implemented in pneumatic gates inside the robot’s own body — no electronics, no batteries, just compressed air driving a small fluidic computer that decided how the robot would move. Daniel Preston’s group at Rice and Rob Wood’s group at Harvard extended this. The application is robots that can operate inside the body during surgery, or in vibration-sensitive environments, or untethered in places where batteries fail. It’s a real and active field again.

For a kids’-education pneumatic computer, none of those advanced techniques matter. The 1960s-vintage slide valves, scaled up to human-visible size and cut from clear acrylic, are exactly the right choice.

4. How a fluidic gate works, mechanically

There are two families of fluidic gates and you choose based on what matters to you.

The Coanda-effect gate has no moving parts. You shape a chamber so a jet of air naturally attaches to one of two walls. A small sideways control jet can detach it and shove it to the other wall, where it stays. Output flips. It’s elegant, it’s fast, and it’s fussy: the chamber geometry has to be right to a few percent, the pressures have to be tuned, and watching it work requires a smoke generator because the air is invisible.

The slide-valve gate uses a small moving piece — a spool, a piston, a membrane — that physically shifts position when pressure changes on one side versus the other. The spool either connects or disconnects two channels, depending on which way it’s sitting. It’s slower, mechanically more visible, and much easier to design and debug because the logic is literally the visible mechanics. You can watch the spool move.

For a clear-acrylic kids’ build, the slide-valve approach is the obviously right one. The bits are moving things. You can see them. A flip-flop is two small pieces of plastic, one in each “remembered” position. A NAND gate is a piece of plastic that slides one way when both inputs are pressurised and another way when either isn’t. The mechanism is the lesson.

For the indicators on the output, I’d use little vertical glass or clear-plastic tubes with a coloured ball inside. When the bit is 1, air pressure lifts the ball to the top of the tube. When the bit is 0, the ball sits at the bottom. The output is four (or eight) coloured balls arranged in a row, and the result of the addition is a binary number you read by which balls are up. With a small reference card (“0001 = 1, 0010 = 2, …”), a kid translates the binary into a decimal answer in their head. That little act of decoding is itself pedagogy — they’re decoding binary, which is what the machine is.

5. Laser-cut layered construction

This is the practical move that makes the project achievable at hobby gauge.

A fluidic gate is mostly geometry. The interesting features are channels and chambers, both two-dimensional in cross-section. You can fabricate a complete gate by:

Designing the layout as a stack of 2D layers in Inkscape, Solvespace, or Onshape. Each layer is one slice through the device. Some layers are channel layers (with the channels cut out). Some are cover layers (with just port holes). Some are spacer layers.
Cutting each layer out of clear acrylic on the laser cutter. A 3 mm cast-acrylic sheet cuts in about 30 mm/s at 60 W. A layer with a couple of dozen gate features cuts in 20-30 seconds.
Stacking the layers in order, with a thin film of acrylic solvent (dichloromethane, a.k.a. “weld-on”) between them.
Pressing the stack under weight overnight while the solvent chemically welds the layers into a single block.
Drilling input/output ports through the top sheet and attaching pneumatic tubing with barbed fittings.

Channel features down to about 100 μm are routine on a hobby laser cutter. With care and a focused beam you can hit 25 μm. Hobby fluidic gates work fine at channel widths of 0.5 to 2 mm — luxuriously large, high yield on the first cut.

The aesthetic bonus is that the finished stack is transparent. You can see every channel, every chamber, every slide valve. A small LED panel underneath illuminates the whole device from below, and the channels glow when air flows through them (more so if you fog the air with a trace of theatrical haze, but even dry air shows up because of the slight density difference between fast-moving and still air). The thing looks like a piece of glass with a circuit diagram trapped inside it, and the circuit moves.

6. Can you build an adder?

This was the empirical question the 2019 note ended on. The answer is yes, and walking through it is the way to see whether the project is actually feasible at the home-laser-cutter gauge.

NAND is universal. If you have a NAND gate you can build any digital logic. So we’ll build a NAND and then use it everywhere.

A NAND gate in slide-valve fluidics is, mechanically, two small spools in series with a pull-up channel from the air supply. Default state: output is pressurised (= 1) because the pull-up channel feeds it. When both inputs are pressurised, both spools shift to a position that opens a vent path from the output to atmosphere, so the output goes to 0. When either input is at atmospheric pressure, its spool stays in the default position, keeping the vent closed, so the output stays at 1. Standard NAND truth table:

A	B	NAND(A,B)
0	0	1
0	1	1
1	0	1
1	1	0

About 2 cm × 1 cm of acrylic real estate per gate, four ports, two moving spools.

A half-adder from NANDs is two outputs: SUM = A XOR B and CARRY = A AND B. Built entirely from NAND gates that’s 5-6 gates total if you share the inner subexpression, 8 if you don’t.

A full adder is two half-adders plus an OR gate (for the carry-out), and OR is three NANDs. Total: about 12-15 NAND gates per bit of addition.

A 4-bit ripple-carry adder is four full adders chained, so about 50-60 NAND gates total.

At 2 cm × 1 cm per gate plus some routing space, the whole adder fits comfortably on a piece of acrylic about 20 cm × 30 cm. That’s the size of a typical laser-cutter bed. Six or eight 3 mm sheets stacked = about 25 mm thick. The whole device is the size of a hardcover book.

Speed: about half a second per addition. Slide-valve fluidics at hobby scale switches in 10-50 ms per gate, and the carry has to ripple through four full adders, so end-to-end maybe 200-500 ms. The kid can watch the result settle. The slowness is the entire point.

Latches and flip-flops are similar: an SR latch is two cross-coupled NAND gates (so 4 spools), a D flip-flop is a few more. With latches you can build registers, and with registers you could build a small accumulator that adds successive inputs over time. The 4-bit incrementer (counts up by 1 each time you push a button) is the right starting point: simpler than a general adder and just as educational.

7. The educational design

This is where the kid-pedagogy details matter, and where the project goes from “interesting maker project” to “actually-pedagogical artifact.”

Big inputs. Four big plastic toggle switches across the front, labeled bit 0 through bit 3. The kid sets the input by physically clicking them up and down. Tactile, deliberate, the input is a thing they did.

Big indicator balls. Four (or eight, if you want to show input and output) tall clear tubes across the top of the device, with brightly-coloured ping-pong-ball-sized indicators inside. High = 1, low = 0. The kid sees four balls do their thing every time.

A small reference card. “0000 = 0. 0001 = 1. 0010 = 2…” laminated and Velcro’d to the side of the device. The kid translates the binary output into a decimal answer. The reference card is part of the learning.

Visible carry. The wiring between full-adder cells should be visibly distinct — different channel colour, brighter LED behind it, something — so that when the carry ripples from bit 0 to bit 1 to bit 2 to bit 3 over the course of half a second, the kid can see the ripple happening. The carry is the thing that makes addition interesting (it’s why 9 + 1 = 10, why kids learn carrying in second grade), and being able to point at the carry physically moving through the machine is the moment the pedagogy lands.

A reset button. Big, red, satisfying. Vents the entire device. The balls fall back to zero. The kid does it again.

No electronics anywhere visible. Yes, you might use a small microcontroller to drive solenoids on the input toggles (so the device can demo itself in a loop unattended), but the microcontroller lives inside the base, and from the kid’s perspective the whole computer is the transparent block. The point is that a thing made of nothing but air and plastic does math.

An information sign. A small placard near the device with a two-paragraph “what is this” — enough that a teacher or parent can read it and explain. The placard should say what NAND means, what addition in binary looks like, and that the entire computer in your phone is just an enormous and very fast version of this. The intellectual jump from “this slow plastic adder” to “the phone in your pocket” is the headline lesson.

8. Why this hasn’t been built much

A handful of YouTube videos and Hackaday articles exist of hobbyist fluidic-logic projects. The Hackaday piece by Jenny List from 2018 is a fair survey. Most of these build a single gate or a 1-bit demonstrator and stop. Almost nobody has built a working multi-gate fluidic adder at hobby scale and posted about it.

I think the reasons are:

The application case is invisible. If your goal is “build a computer,” then a fluidic computer is a thousand times slower than the microcontroller in your phone, and roughly nobody who knows enough to build the fluidic computer is unconvinced of the microcontroller’s superiority. The “what is it for” question kills most attempts.
The pedagogical case is right there but undersold. If you reframe the goal as “build a thing kids can understand,” the slowness becomes a feature and the application case becomes “in the corner of a classroom or a maker space, for years, doing one job well.” Nobody seems to have led with this framing in public.
It’s a real project. A weekend isn’t enough. Four to six weekends is realistic, plus a couple of iteration cycles where gates don’t quite work. The maker who decides to do it has to commit to seeing it through. Most maker projects are afternoons.
Pneumatic stuff is loud. A workshop compressor isn’t suited to a classroom. The fix is to drive the demo with a small quiet diaphragm pump or a battery-powered miniature compressor running intermittently. Whitesides Lab’s soft-robotics work has good examples here.

None of these are unfixable. They just need someone to take the pedagogical framing seriously and put in the four weekends.

9. The plan

Concrete enough to start tomorrow.

MVP: a 4-bit incrementer. Takes a 4-bit input (set by four toggles), outputs input+1 (on four balls). About 50 fluidic gates. Carry ripples visibly from low bit to high bit. Reset button vents the device. Half a second per operation. A small placard explains.

Materials:

6 sheets of clear cast acrylic, 3 mm thick, roughly 30 × 30 cm. About $50.
A roll of 0.5 mm latex sheet for slide-valve seals. About $10.
Dichloromethane acrylic cement. About $15.
Pneumatic tubing (4 mm OD) and a bag of barbed fittings. About $30.
A bag of clear plastic ping-pong balls and clear acrylic tubing for the output indicators. About $20.
A small quiet diaphragm pump or a 12 V mini-compressor with a tank. About $80.
An LED panel for backlighting. About $40.
A laser-cut wooden base to mount it all in. About $30.

Total: about $275 in materials.

Tools:

A 60 W or 80 W CO₂ laser cutter (Hacker Dojo, TechShop, your friend’s garage, the local fab lab).
Inkscape or Solvespace for the layout. Solvespace is better because the constraint solver lets you parameterise channel widths once and recut quickly.
A drill press for the input/output ports.
Generic workshop tools (clamps, sandpaper, isopropyl alcohol for cleaning between weld steps).

Build cycle:

Weekend 1: Design and cut a single NAND gate. Test it on the bench with a hand pump. Iterate the channel geometry once or twice until it switches cleanly. Document the working geometry as a reusable cell.
Weekend 2: Tile the NAND cell into a full-adder block. Six NAND gates wired into half-adder + half-adder + OR. Test all 8 input combinations. Time it.
Weekend 3: Lay out and cut the 4-bit incrementer — four full-adder cells in a row, plus the input toggle interface and the output indicator manifold. Assemble.
Weekend 4: Build the wooden mounting base. Wire the toggles to the inputs. Tune the pump pressure (probably 5-10 psi). Add the LED panel. Add the reset button. Photograph the whole thing in action. Write up the build guide and put it on GitHub.

Optional, weekend 5-6: Add a small Arduino in the base that can demo the device unattended — cycle through inputs 0 through 15 every ten seconds so visitors who walk by can watch it count.

Open-source the whole thing. Solvespace files, BOM, build guide, photos. The point is that the second one of these costs four weekends; the hundredth one costs an afternoon because somebody has done the laser-cut layout work already.

The end state is a transparent acrylic block on a wooden base, about the size of a hardcover book, sitting on a table at the front of a classroom or in the kids’ corner of a maker space. Coloured balls across the top. Toggle switches across the front. Reset button on the side. A small placard. A child walks up, flips a few toggles, watches the bits ripple across the device, reads the answer, smiles.

That is what I had in mind in 2019. The note didn’t say so explicitly because it was a note to myself, but that’s the project.

10. Why this is worth a weekend

I’m going to be direct. The world has plenty of educational kits for kids that purport to teach computation. Most of them are LEDs on breadboards driven by an Arduino, where the Arduino is a black box doing the actual work and the LEDs are decoration. The lesson the kid walks away with is “I plugged things in and they lit up.” That isn’t computation pedagogy. That’s plumbing.

A pneumatic adder is different. There is no Arduino doing the work in secret. The plastic block is the computer. Every bit visible. Every gate visible. Carry visibly rippling. The kid who watches it has, in some real sense, watched a computer add. They can take that mental model with them into every screen they ever look at afterwards, and the screen will be less mysterious.

That’s worth the weekend.

It’s also, by the way, a beautiful object. Multi-layer clear acrylic with channels carved into it, illuminated from below, coloured balls bobbing on top, makes for a striking display piece. The aesthetic case is not nothing. Maker spaces compete for foot traffic and the “oh, what’s that” reaction is real. A pneumatic computer in the front window gets people to walk in.

So this is the design. The cost is $275 and four weekends. The output is a transparent device that adds binary numbers in plain sight at half a second per operation. The audience is every kid who walks past it for the next five years.

I’m going to build it. If you have a laser cutter and four weekends of your own to spare, you should too.

John Sokol's Blog

Friday, May 29, 2026